Use cogl_framebuffer_read_pixels_into_bitmap () instead of
glReadPixels () for the CPU copy path in multi-GPU support.
The cogl function employs several tricks to make the read-pixels as fast
as possible and does the y-flip as necessary. This should make the copy
more performant over all kinds of hardware.
This is expected to be used on virtual outputs (e.g. DisplayLink USB
docks and monitors) foremost, where the dumb buffer memory is just
regular system memory. If the dumb buffer memory is somehow slow, like
residing in discrete VRAM or having an unexpected caching mode, it may
be possible for the cogl function perform worse because it might do the
y-flip in-place in the dumb buffer. Hopefully that does not happen in
any practical scenario.
Calling meta_renderer_native_gles3_read_pixels () here was conceptually
wrong to begin with because it was done with the Cogl GL context of the
primary GPU, not on the GL ES 3 context of a secondary GPU. However,
due eglBindAPI being a no-op in Mesa and the glReadPixels () arguments
being compatible, it worked.
This patch adds a pixel format conversion table between DRM and Cogl
formats. It contains more formats than absolutely necessary and the
texture components field which is currently unused for completeness. See
Mutter issue #323. Making the table more complete documents better how
the pixel formats actually map so that posterity should be less likely
to be confused. This table could be shared with
shm_buffer_get_cogl_pixel_format () as well, but not with
meta_wayland_dma_buf_buffer_attach ().
On HP ProBook 4520s laptop (Mesa DRI Intel(R) Ironlake Mobile, Mesa
18.0.5), without this patch copy_shared_framebuffer_cpu () for a
DisplayLink output takes 5 seconds with a 1080p frame. Obviously that
makes Mutter and gnome-shell completely unusable. With this patch, that
function takes 13-18 ms which makes it usable if not fluent.
On Intel i7-4790 (Mesa DRI Intel(R) Haswell Desktop) machine, this patch
makes no significant difference (the copy takes 4-5 ms).
The format will be needed in a following commit in the CPU copy path
which stops hardcoding another format and starts using the format the
dumb FB was created with.
Change the callers of init_dumb_fb () to use DRM format codes. DRM and
GBM format codes are identical, but since this is about dumb buffers,
DRM formats fit better.
The header /usr/include/gbm.h installed by Mesa says:
* The FourCC format codes are taken from the drm_fourcc.h definition, and
* re-namespaced. New GBM formats must not be added, unless they are
* identical ports from drm_fourcc.
That refers to the GBM_FORMAT_* codes.
The order and way include macros were structured was chaotic, with no
real common thread between files. Try to tidy up the mess with some
common scheme, to make things look less messy.
There was a race in setting next_fb_id when a secondary GPU was using
the CPU copy path. Losing this race caused the attempt to
drmModePageFlip () to FB ID 0 which is invalid and always fails. Failing
to flip causes Mutter to fall back to drmModeSetCrtc () permanently.
In meta_onscreen_native_swap_buffers_with_damage ():
- update_secondary_gpu_state_pre_swap_buffers ()
- copy_shared_framebuffer_cpu () but only on the CPU copy path
- secondary_gpu_state->gbm.next_fb_id is set
- wait_for_pending_flips ()
- Waits for any remaining page flip events and executes and destroys
the related page flip closures.
- on_crtc_flipped ()
- meta_onscreen_native_swap_drm_fb ()
- swap_secondary_drm_fb ()
- secondary_gpu_state->gbm.next_fb_id = 0;
- meta_onscreen_native_flip_crtcs ()
- meta_onscreen_native_flip_crtc ()
- meta_gpu_kms_flip_crtc () gets called with fb_id = 0
This race was observed lost when running 'mutter --wayland' on a machine
with two outputs on Intel and one output on DisplayLink USB dock, and
wiggling around a weston-terminal window between the Intel and
DisplayLink outputs. It took from a second to a minute to trigger. For
testing with DisplayLink outputs Mutter also needed a patch to take the
DisplayLink output into use, as it would have otherwise been ignored
being a platform device rather than a PCI device.
Fix this race by first waiting for pending flips and only then
proceeding with the swap operations. This should be safe, because the
pending flips could have completed already before entering
meta_onscreen_native_swap_buffers_with_damage ().
meta_renderer_native_gles3_read_pixels() was assuming that the target
buffer stride == width * 4. This is not generally true. When a DRM
driver allocates a dumb buffer, it is free to choose a stride so that
the buffer can actually work on the hardware.
Record the driver chosen stride in MetaDumbBuffer, and use it in the CPU
copy path. This should fix any possible stride issues in
meta_renderer_native_gles3_read_pixels().
Track the allocated dumb buffer size in MetaDumbBuffer. Assert that the
size is as expected in copy_shared_framebuffer_cpu().
This is just to ensure that Cogl and the real size match. The size from
Cogl was used in the copy, so getting that wrong might have written
beyond the allocation.
This is a safety measure and has not been observed to happen yet.
If drmModeAddFB2() does not work, the fallback to drmModeAddFB() can
only handle a single specific format. Make sure the requested format is
that one format, and fail the operation otherwise.
This should at least makes the failure mode obvious on such old systems
where the kernel does not support AddFB2, rather than producing wrong
colors.
The "backends: Move MetaOutput::crtc field into private struct"
accidentally changed the view transform calculation code to assume that
"MetaCrtc::transform" corresponds to the transform of the CRTC; so is
not the case yet; one must calculate the transform from the logical
monitor, and check whether it is supported by the CRTC using
meta_monitor_manager_is_transform_handled(). This commit restores the
old behaviour that doesn't use MetaCrtc::transform when calculating the
view transform.
Fixes: https://gitlab.gnome.org/GNOME/mutter/issues/216
Commit c0d9b08ef9 replaced the old GBM API calls
with the multi-plane GBM API. However, the call to gbm_bo_get_handle_for_plane
fails for some DRI drivers (in particular i915). Due to missing error checks,
the subsequent call to drmModeAddFB[2] fails and the screen output locks up.
This commit adds the missing error checks and falls back to the old GBM API
(non-planar) if necessary.
v5: test success of gbm_bo_get_handle_for_plane instead of errno
This commit adopts solution proposed by Daniel van Vugt to check the return
value of gbm_bo_get_handle_for_plane on plane 0 and fall back to old
non-planar method if the call fails. This removes the errno check (for
ENOSYS) that could abort if mesa ever sets a different value.
Related to: https://gitlab.gnome.org/GNOME/mutter/issues/127
Commit 712ec30cd9 added the logic to only
choose EGL configs that match the GBM_FORMAT_XRGB8888 pixel format.
However, there won't be any EGL config satisfying such criteria for
non-GBM backends, such as EGLDevice.
This change will let us choose the first EGL config for the EGLDevice
backend, while still forcing GBM_FORMAT_XRGB8888 configs for the GBM
one.
Related to: https://gitlab.gnome.org/GNOME/mutter/issues/2
drmModeAddFB2 allows userspace to specify a real format enum on
non-ancient kernels, as an improvement over the legacy drmModeAddFB
which derives format from a fixed depth/bpp mapping.
As an optimisation, Weston used to decide at the first failure of
drmModeAddFB2 that the ioctl was unavailable: as non-existent DRM
ioctls return -EINVAL rather than -ENOSYS or similar, bad parameters are
not distinguishable from the ioctl not being present.
Mutter has also implemented the same optimisation for dumb framebuffers,
which potentially papers over errors for the gain of avoiding one ioctl
which will rapidly fail on ancient kernels. Remove the optimisation and
always use AddFB2 where possible.
Closes: #14
When using the EGLStream backend, the MetaRendererNative passed a
GClosure to KMS when using EGLStreams, but KMS flip callback event
handler in meta-gpu-kms.c expected a closure wrapped in a closure
container, meaning it'd instead crash when using EGLStreams. Make the
flip handler get what it expects also when using EGLStreams by wrapping
the flip closure in the container before handing it over to EGL.
https://bugzilla.gnome.org/show_bug.cgi?id=790316
Before we just set it to "none", but this was not enough since various
calls will depend on not just the context being active, but the main
rendering surface.
Fixes https://gitlab.gnome.org/GNOME/mutter/issues/21
Make it re-enable:able by a hidden "experimental feature". To enable, add
"kms-modifiers" to the org.gnome.mutter.experimental-features GSettings entry.
If we attempt GBM surface allocation with a set of modifiers but the
allocation fails, fall back to non-modifier allocations. This fixes
startup on Pineview-based Atom systems, where KMS provides us a set of
modifiers but the GBM implementation doesn't support modifier use.
Closes: https://gitlab.gnome.org/GNOME/mutter/issues/84
Rendering the next frame (which mostly happens as part of the flush done
in swap buffers) is a task that the GPU can complete independently of
the CPU having to wait for previous page flips. So reverse their order
to get the GPU started earlier, with the aim of greater GPU-CPU
parallelism.
We just arbitrarily chose the first EGL config matching the passed
attributes, but we then assumed we always got GBM_FORMAT_XRGB8888. That
was not a correct assumption. Instead, make sure we always pick the
format we expect.
Closes: https://gitlab.gnome.org/GNOME/mutter/issues/2
We were retrieving the supported KMS modifiers for all GPUs even
though what we really need to intersect between these sets of
modifiers:
1) KMS supported modifiers for primary GPU if the GPU is used for
scanout;
2) EGL supported modifiers for secondary GPUs (different than the
primary GPU used for rendering);
3) GBM supported modifiers when creating the surface (already
taken care of by gbm_surface_create_with_modifiers());
https://gitlab.gnome.org/GNOME/mutter/issues/18
Now that we have the list of supported modifiers from the monitor
manager (via the CRTCs to the primary planes), we can use this to inform
EGL it can use those modifiers to allocate the GBM surface with. Doing
so allows us to use tiling and compression for our scanout surfaces.
This requires the Mesa commit in:
Mesa 10.3 (08264e5dad4df448e7718e782ad9077902089a07) or
Mesa 10.2.7 (55d28925e6109a4afd61f109e845a8a51bd17652).
Otherwise Mesa closes the fd behind our back and re-importing will fail.
See FDO bug #76188 for details.
https://bugzilla.gnome.org/show_bug.cgi?id=785779
Newer versions of GBM support buffer modifiers, including multi-plane
buffers. Use this new API to explicitly pull the information from GBM,
and feed it to drmModeAddFB2WithModifiers.
https://bugzilla.gnome.org/show_bug.cgi?id=785779
Some x86 clamshell design devices use portrait tablet LCD panels while
they should use a landscape panel, resoluting in a 90 degree rotated
picture.
Newer kernels detect this and rotate the fb console in software to
compensate. These kernels also export their knowledge of the LCD panel
orientation vs the casing in a "panel orientation" drm_connector property.
This commit adds support to mutter for reading the "panel orientation"
and transparently (from a mutter consumer's pov) fixing this by applying
a (hidden) rotation transform to compensate for the panel orientation.
Related: https://bugs.freedesktop.org/show_bug.cgi?id=94894https://bugzilla.gnome.org/show_bug.cgi?id=782294
Proprietary drivers such as ARM Mali export EGL_KHR_platform_gbm instead
of EGL_MESA_platform_gbm. As such, GBM platform check should be done for
both MESA and non-MESA drivers.
https://bugzilla.gnome.org/show_bug.cgi?id=780668
The DRM properties container must be destroyed with
drmModeFreeObjectProperties, and the connectors must be freed on every
caller. Also make it sure that gbm_device structs are destroyed with the
MetaRendererNativeGpuData that owns them.
https://bugzilla.gnome.org/show_bug.cgi?id=789984
A hybrid GPU system is a system where more than one GPU is connected to
connectors. A common configuration is having a integrated GPU (iGPU)
connected to a laptop panel, and a dedicated GPU (dGPU) connected to
one or more external connector (such as HDMI).
This commit adds support for rendering the compositor stage using the
iGPU, then copying the framebuffer content onto a secondary framebuffer
that will be page flipped on the CRTC of the dGPU.
This can work in two different ways: GPU accelerated using Open GL ES
3, or CPU unaccelerated.
When supported, GPU accelerated copying works by exporting the iGPU
onscreen framebuffer as a DMA-BUF, importing it as a texture on a
separate dGPU EGL context, then using glBlitFramebuffer(), blitting it
onto a framebuffer on the dGPU that can then be page flipped on the dGPU
CRTC.
When GPU acceleration is not available, copying works by creating two
dumb buffers, and each frame glReadPixels() from the iGPU EGL render
context directly into the dumb buffer. The dumb buffer is then page
flipped on the dGPU CRTC.
https://bugzilla.gnome.org/show_bug.cgi?id=785381
When creating a renderer with a custom winsys (which is always how
mutter uses cogl) make it possible to pass a user data with the winsys.
Still unused.
https://bugzilla.gnome.org/show_bug.cgi?id=785381
Make dumb buffer creation/destruction reusable by introducing a
MetaDumbBuffer type (private to meta-renderer-native.c). This will
later be used for software based fallback paths for copying render GPU
buffers onto secondary GPUs.
https://bugzilla.gnome.org/show_bug.cgi?id=785381
Get rid of some technical dept by removing the support in the native
backend for drawing the the whole stage to one large framebuffer.
Previously the only way to disable stage views was to set the
MUTTER_STAGE_VIEWS environment variable to 0; doing that now will cause
the native backend to fail to initialize.
https://bugzilla.gnome.org/show_bug.cgi?id=785381
When drmHandleEvent() returns an error and errno is set to EAGAIN,
instead of ending up in a busy loop, poll() the fd until there is
anything to read.
https://bugzilla.gnome.org/show_bug.cgi?id=785381
The prefix, if any, of a variable name often contains information about
the namespace (such as clutter_backend is the ClutterBackend, while
backend is a MetaBackend). Clean up some more inconsistencies in
meta-renderer-native.c where various variable names were egl_ prefixed
but in fact was Cogl types.
https://bugzilla.gnome.org/show_bug.cgi?id=785381
In order to eventually support multilpe GPUs with their own connectors,
split out related meta data management (i.e. outputs, CRTCs and CRTC
modes) into a new MetaGpu GObject.
The Xrandr backend always assumes there is always only a single "GPU" as
the GPU is abstracted by the X server; only the native backend (aside
from the test backend) will eventually see more than one GPU.
The Xrandr backend still moves some management to MetaGpuXrandr, in
order to behave more similarly to the KMS counterparts.
https://bugzilla.gnome.org/show_bug.cgi?id=785381
Move finding, opening and managment of the KMS file descriptor to
MetaMonitorManagerKms. This means that the monitor manager creation can
now fail, both if more than one GPU with connectors is discovered, or
if finding or opening the primary GPU fails.
https://bugzilla.gnome.org/show_bug.cgi?id=785381
Turn MetaCrtc into a GObject and move it to a separate file. This
changes the storage format, resulting in changing the API for accessing
MetaCrtcs from using an array, to using a GList.
https://bugzilla.gnome.org/show_bug.cgi?id=785381
Turn MetaOutput into a GObject and move it to a separate file. This
changes the storage format, resulting in changing the API for accessing
MetaOutputs from using an array, to using a GList.
https://bugzilla.gnome.org/show_bug.cgi?id=785381
The reverted commit seems to cause
https://bugzilla.gnome.org/show_bug.cgi?id=787240 for some reason. Lets
be safe and revert it for now, as the code freeze is just around the
corner.
This partly (it doesn't reintroduce a whitespace issue) reverts commit
dbc63430d8.
Due to rounding issues, we can't assume a floating point calculation
will end up on an integer, even if we got the factor from the reverse
calculation. Thus, to avoid casting away values like N.999... to N,
when they should really be N+1, round the resulting floating point
calculation before casting it to int.
This fixes an issue where using the scale ~1.739 on a 1920x1080 mode
resulted in error when setting the mode, as the calculated size of the
framebuffer was only 1919x1080.
https://bugzilla.gnome.org/show_bug.cgi?id=786918
