Create the intermediate shadow framebuffer for use exclusively when a
shadowfb is required.
Keep the previous offscreen framebuffer is as an intermediate
framebuffer for transformations only.
This way, we can apply transformations between in-memory framebuffers
prior to blit the result to screen, and achieve acceptable performance
even with software rendering on discrete GPU.
https://gitlab.gnome.org/GNOME/mutter/merge_requests/877
It was not the lack of forcing the shadow fb that caused slowness, but
rather due to the method the shadow fb content was copied onto the
scanout fb. With 'clutter: Use cogl_blit_framebuffer() for shadow FB'
we'll use a path that shouldn't be slow when copying onto the scanout
fb.
Also 437f6b3d59 accidentally enabled
shadow fb when using hw accelerated contexts, due to the cap being set
to 1 in majority of drivers. While the kernel documentation for the
related field says "hint to userspace to prefer shadow-fb rendering",
the name of the hint when exposed to userspace is
DRM_CAP_DUMB_PREFER_SHADOW, thus should only be taken into consideration
for dumb buffers, not rendering in general.
This reverts commit 437f6b3d59.
https://gitlab.gnome.org/GNOME/mutter/merge_requests/818
The commit 'renderer/native: Use shadow fb on software GL if preferred'
attempted to force using a shadow fb when using llvmpipe in order to
speed up blending, but instead only did so when llvmpipe AND the drm
device explicityl asked for it.
Now instead always force it for llvmpipe and other software rendering
backends, and otherwise just query the drm device (i.e.
DRM_CAP_DUMB_PREFER_SHADOW).
https://gitlab.gnome.org/GNOME/mutter/merge_requests/807
COPY_MODE_PRIMARY has two paths, automatically chosen. For debugging purposes,
e.g. why is my DisplayLink screen slowing down the whole desktop, it will be
useful to know which copy path is taken. Debug prints are added to both when
the primary GPU copy succeeds the first time and when it fails the first time.
This is not the full truth, because theoretically the success/failure could
change every frame, but we don't want to spam the logs (even in debug mode)
every frame. In practise, it should be rare for the success or failure to ever
change. Hence, saying what happened on the first time is enough. This does
indicate if it ever changes even once, too, so we know if that unexpected thing
happens.
The debug prints are per secondary GPU since there could be several.
https://gitlab.gnome.org/GNOME/mutter/merge_requests/615
When the preferred path META_SHARED_FRAMEBUFFER_COPY_MODE_SECONDARY_GPU cannot
be used, as is the case for e.g. DisplayLink devices which do not actually have
a GPU, try to use the primary GPU for the copying before falling back to
read-pixels which is a CPU copy.
When the primary GPU copy works, it should be a significant performance win
over the CPU copy by avoiding stalling libmutter for the duration.
This also renames META_SHARED_FRAMEBUFFER_COPY_MODE_* because the new names are
more accurate. While the secondary GPU copy is always a GPU copy, the primary
copy might be either a CPU or a GPU copy.
https://gitlab.gnome.org/GNOME/mutter/merge_requests/615
Simplify the call site a bit and make the native renderer know it should
queue mode reset itself when views have been rebuilt. This is done
partly due to more things needing to be dealt with after views have been
rebuilt.
https://gitlab.gnome.org/GNOME/mutter/merge_requests/630
When building without EGL device support, the following compiler warning
is seen:
```
src/backends/native/meta-renderer-native.c:2637:20: warning: unused
variable ‘cogl_renderer_egl’ [-Wunused-variable]
```
Fix the warning by placing the relevant variable declarations within the
`#ifdef HAVE_EGL_DEVICE/#endif` statement.
https://gitlab.gnome.org/GNOME/mutter/merge_requests/656
This commit introduces, and makes use of, a transactional API used for
setting up KMS state, later to be applied, potentially atomically. From
an API point of view, so is always the case, but in the current
implementation, it still uses legacy drmMode* API to apply the state
non-atomically.
The API consists of various buliding blocks:
* MetaKmsUpdate - a set of configuration changes, the higher level
handle for handing over configuration to the impl backend. It's used to
set mode, assign framebuffers to planes, queue page flips and set
connector properties.
* MetaKmsPlaneAssignment - the assignment of a framebuffer to a plane.
Currently used to map a framebuffer to the primary plane of a CRTC. In
the legacy KMS implementation, the plane assignment is used to derive
the framebuffer used for mode setting and page flipping.
This also means various high level changes:
State, excluding configuring the cursor plane and creating/destroying
DRM framebuffer handles, are applied in the end of a clutter frame, in
one go. From an API point of view, this is done atomically, but as
mentioned, only the non-atomic implementation exists so far.
From MetaRendererNative's point of view, a page flip now initially
always succeeds; the handling of EBUSY errors are done asynchronously in
the MetaKmsImpl backend (still by retrying at refresh rate, but
postponing flip callbacks instead of manipulating the frame clock).
Handling of falling back to mode setting instead of page flipping is
notified after the fact by a more precise page flip feedback API.
EGLStream based page flipping relies on the impl backend not being
atomic, as the page flipping is done in the EGLStream backend (e.g.
nvidia driver). It uses a 'custom' page flip queueing method, keeping
the EGLStream logic inside meta-renderer-native.c.
Page flip handling is moved to meta-kms-impl-device.c from
meta-gpu-kms.c. It goes via an extra idle callback before reaching
meta-renderer-native.c to make sure callbacks are invoked outside of the
impl context.
While dummy power save page flipping is kept in meta-renderer-native.c, the
EBUSY handling is moved to meta-kms-impl-simple.c. Instead of freezing the
frame clock, actual page flip callbacks are postponed until all EBUSY retries
have either succeeded or failed due to some other error than EBUSY. This
effectively inhibits new frames to be drawn, meaning we won't stall waiting on
the file descriptor for pending page flips.
https://gitlab.gnome.org/GNOME/mutter/issues/548https://gitlab.gnome.org/GNOME/mutter/merge_requests/525
Fix the following compiler warning:
../src/backends/native/meta-renderer-native.c: In function ‘meta_renderer_native_create_view’:
/usr/include/glib-2.0/glib/gmacros.h:523:17: warning: ‘formats’ may be used uninitialized in this function [-Wmaybe-uninitialized]
523 | { if (_ptr) (cleanup) ((ParentName *) _ptr); } \
| ^
../src/backends/native/meta-renderer-native.c:773:22: note: ‘formats’ was declared here
773 | g_autoptr (GArray) formats;
| ^~~~~~~
https://gitlab.gnome.org/GNOME/mutter/merge_requests/632
Make sure to destroy the EGL surface after releasing held buffers,
otherwise we'll get the following valgrind warnings:
==24016== Invalid read of size 8
==24016== at 0x1739943F: release_buffer (platform_drm.c:73)
==24016== by 0x49AC355: meta_drm_buffer_gbm_finalize (meta-drm-buffer-gbm.c:213)
==24016== by 0x4B75B61: g_object_unref (gobject.c:3346)
==24016== by 0x49B4B41: free_current_bo (meta-renderer-native.c:991)
==24016== by 0x49B816F: meta_renderer_native_release_onscreen (meta-renderer-native.c:2971)
==24016== by 0x5209441: _cogl_onscreen_free (cogl-onscreen.c:167)
==24016== by 0x5208D81: _cogl_object_onscreen_indirect_free (cogl-onscreen.c:51)
==24016== by 0x51C8066: _cogl_object_default_unref (cogl-object.c:103)
==24016== by 0x5207989: _cogl_framebuffer_unref (cogl-framebuffer.c:1814)
==24016== by 0x51C80B1: cogl_object_unref (cogl-object.c:115)
==24016== by 0x53673C7: clutter_stage_view_dispose (clutter-stage-view.c:304)
==24016== by 0x4B75AF2: g_object_unref (gobject.c:3309)
==24016== Address 0x18e742a8 is 536 bytes inside a block of size 784 free'd
==24016== at 0x4839A0C: free (vg_replace_malloc.c:540)
==24016== by 0x17399764: dri2_drm_destroy_surface (platform_drm.c:231)
==24016== by 0x1738550A: eglDestroySurface (eglapi.c:1145)
==24016== by 0x5440286: eglDestroySurface (in /home/jonas/Dev/gnome/install/lib/libEGL.so.1.1.0)
==24016== by 0x49613A5: meta_egl_destroy_surface (meta-egl.c:432)
==24016== by 0x49B80F9: meta_renderer_native_release_onscreen (meta-renderer-native.c:2954)
==24016== by 0x5209441: _cogl_onscreen_free (cogl-onscreen.c:167)
==24016== by 0x5208D81: _cogl_object_onscreen_indirect_free (cogl-onscreen.c:51)
==24016== by 0x51C8066: _cogl_object_default_unref (cogl-object.c:103)
==24016== by 0x5207989: _cogl_framebuffer_unref (cogl-framebuffer.c:1814)
==24016== by 0x51C80B1: cogl_object_unref (cogl-object.c:115)
==24016== by 0x53673C7: clutter_stage_view_dispose (clutter-stage-view.c:304)
==24016== Block was alloc'd at
==24016== at 0x483AB1A: calloc (vg_replace_malloc.c:762)
==24016== by 0x173997AE: dri2_drm_create_window_surface (platform_drm.c:145)
==24016== by 0x17388906: _eglCreateWindowSurfaceCommon (eglapi.c:929)
==24016== by 0x5440197: eglCreateWindowSurface (in /home/jonas/Dev/gnome/install/lib/libEGL.so.1.1.0)
==24016== by 0x49612FF: meta_egl_create_window_surface (meta-egl.c:396)
==24016== by 0x49B752E: meta_renderer_native_create_surface_gbm (meta-renderer-native.c:2538)
==24016== by 0x49B7E6C: meta_onscreen_native_allocate (meta-renderer-native.c:2870)
==24016== by 0x49B8BCF: meta_renderer_native_create_view (meta-renderer-native.c:3387)
==24016== by 0x48D274B: meta_renderer_create_view (meta-renderer.c:78)
==24016== by 0x48D27DE: meta_renderer_rebuild_views (meta-renderer.c:111)
==24016== by 0x49BB4FB: meta_stage_native_rebuild_views (meta-stage-native.c:142)
==24016== by 0x49A733C: meta_backend_native_update_screen_size (meta-backend-native.c:517)
https://gitlab.gnome.org/GNOME/mutter/merge_requests/622
When making a new surface/context pair current, mesa may want to flush
the old context. Make sure we don't try to flush any freed memory by
unmaking a surface/context pair current before freeing it.
Not doing this results in the following valgrind warnings:
==15986== Invalid read of size 8
==15986== at 0x69A6D80: dri_flush_front_buffer (gbm_dri.c:92)
==15986== by 0x1750D458: intel_flush_front (brw_context.c:251)
==15986== by 0x1750D4BB: intel_glFlush (brw_context.c:296)
==15986== by 0x1739D8DD: dri2_make_current (egl_dri2.c:1461)
==15986== by 0x17393A3A: eglMakeCurrent (eglapi.c:869)
==15986== by 0x54381FB: InternalMakeCurrentVendor (in /home/jonas/Dev/gnome/install/lib/libEGL.so.1.1.0)
==15986== by 0x5438515: eglMakeCurrent (in /home/jonas/Dev/gnome/install/lib/libEGL.so.1.1.0)
==15986== by 0x522A782: _cogl_winsys_egl_make_current (cogl-winsys-egl.c:303)
==15986== by 0x49B64C8: meta_renderer_native_create_view (meta-renderer-native.c:3076)
==15986== by 0x48D26E7: meta_renderer_create_view (meta-renderer.c:78)
==15986== by 0x48D277A: meta_renderer_rebuild_views (meta-renderer.c:111)
==15986== by 0x49BF46E: meta_stage_native_rebuild_views (meta-stage-native.c:142)
==15986== Address 0x1b076600 is 0 bytes inside a block of size 48 free'd
==15986== at 0x4839A0C: free (vg_replace_malloc.c:540)
==15986== by 0x49B59F3: meta_renderer_native_release_onscreen (meta-renderer-native.c:2651)
==15986== by 0x5211441: _cogl_onscreen_free (cogl-onscreen.c:167)
==15986== by 0x5210D81: _cogl_object_onscreen_indirect_free (cogl-onscreen.c:51)
==15986== by 0x51D0066: _cogl_object_default_unref (cogl-object.c:103)
==15986== by 0x520F989: _cogl_framebuffer_unref (cogl-framebuffer.c:1814)
==15986== by 0x51D00B1: cogl_object_unref (cogl-object.c:115)
==15986== by 0x536F3C7: clutter_stage_view_dispose (clutter-stage-view.c:304)
==15986== by 0x4B7DAF2: g_object_unref (gobject.c:3309)
==15986== by 0x4A9596C: g_list_foreach (glist.c:1013)
==15986== by 0x4A9599A: g_list_free_full (glist.c:223)
==15986== by 0x48D2737: meta_renderer_rebuild_views (meta-renderer.c:100)
==15986== Block was alloc'd at
==15986== at 0x483AB1A: calloc (vg_replace_malloc.c:762)
==15986== by 0x69A76B2: gbm_dri_surface_create (gbm_dri.c:1252)
==15986== by 0x69A6BFE: gbm_surface_create (gbm.c:600)
==15986== by 0x49B4E29: meta_renderer_native_create_surface_gbm (meta-renderer-native.c:2221)
==15986== by 0x49B57DB: meta_onscreen_native_allocate (meta-renderer-native.c:2569)
==15986== by 0x49B6423: meta_renderer_native_create_view (meta-renderer-native.c:3062)
==15986== by 0x48D26E7: meta_renderer_create_view (meta-renderer.c:78)
==15986== by 0x48D277A: meta_renderer_rebuild_views (meta-renderer.c:111)
==15986== by 0x49BF46E: meta_stage_native_rebuild_views (meta-stage-native.c:142)
==15986== by 0x49A75B5: meta_backend_native_update_screen_size (meta-backend-native.c:520)
==15986== by 0x48B01BB: meta_backend_sync_screen_size (meta-backend.c:224)
==15986== by 0x48B09B7: meta_backend_real_post_init (meta-backend.c:501)
https://gitlab.gnome.org/GNOME/mutter/merge_requests/622
Currently the EGLDevice code gets the display and calls eglInitialize.
As a follow-up it checks the required EGL extensions - technically it
could check the EGL device extensions earlier.
In either case, eglTerminate is missing. Thus the connection to the
display was still bound.
This was highlighted with Mesa commit d6edccee8da ("egl: add
EGL_platform_device support") + amdgpu.
In that case, since the eglTerminate is missing, we end up reusing the
underlying amdgpu_device due to some caching in libdrm_amdgpu. The
latter in itself being a good solution since it allows buffer sharing
across primary and render node of the same device.
Note: we should really get this in branches all the way back to 3.30.
https://gitlab.gnome.org/GNOME/mutter/merge_requests/619
Fixes: 934184e23 ("MetaRendererNative: Add EGLDevice based rendering support")
Cc: Jonas Ådahl <jadahl@gmail.com>
Signed-off-by: Emil Velikov <emil.velikov@collabora.com>
Extract the next buffer -logic into a new function. This allows to
simplify copy_shared_framebuffer_cpu () making it more readable.
This change is a pure refactoring, no functional changes.
https://gitlab.gnome.org/GNOME/mutter/merge_requests/593
There is no reason why we should have an internal type enum when we have
all the infrastructure to just use multiple GObject types. Also there
was no code sharing between the old "types", the only common API was
getting the framebuffer ID, so lets make that a vfunc.
https://gitlab.gnome.org/GNOME/mutter/merge_requests/584
Since "renderer/native: make EGL initialization failure not fatal" it is
possible, under specific failure conditions, to end up with a primary GPU whose
EGL initialization failed. That cannot work.
https://gitlab.gnome.org/GNOME/mutter/merge_requests/521
The failure to initialize EGL does not necessarily mean the KMS device cannot
be used. The device could still be used as a "secondary GPU" with the CPU copy
mode.
If meta_renderer_native_create_renderer_gpu_data () fails,
meta_renderer_native_get_gpu_data () will return NULL, which may cause crashes.
This patch removes most of the failures, but does not fix the NULL dereferences
that will still happen if creating gpu data fails.
This patch reorders create_renderer_gpu_data_gbm () so that it fails hard only
if GBM device cannot be created, and otherwise always returns an initialized
gpu data structure. Users of the gpu data structure are responsible for
checking egl_display validity.
The GBM device creation failure is a hard failure because presumably GBM is
necessary for cursors.
Fixes: https://gitlab.gnome.org/GNOME/mutter/issues/542https://gitlab.gnome.org/GNOME/mutter/merge_requests/521
We're currently always waiting for unfinished page flips before flipping
again. This is awkward when we are in an asynchronous retry-page-flip
loop, as we can synchronously wait for any KMS page flip event.
To avoid ending up with such situations, just freeze the frame clock
while we're retrying, then thaw it when we succeded.
https://gitlab.gnome.org/GNOME/mutter/merge_requests/506
We rely on the frame clock to compress input events, thus if the frame
clock stops, input events are not dispatched. At the same time, there
is no reason to redraw at a full frame rate, as nothing will be
presented anyway, so slow down to 10Hz (compared to the most common
60Hz). Note that we'll only actually reach 10Hz if there is an active
animation being displayed, which won't happen e.g. if there is a screen
shield in the way.
https://gitlab.gnome.org/GNOME/mutter/merge_requests/506
When we're in a page-flip retry loop due to the FIFO being full
(drmModePageFlip() failing with EBUSY), we should not continue to try
when starting to power save, as that means we're blocking new frames,
which itself blocks input events due to them being compressed using the
frame clock.
We'd also hit an assert assuming we only try to page flip when not power
saving.
Thus, fake we flipped if we ended up reaching a power saving state while
retrying.
Fixes: https://gitlab.gnome.org/GNOME/mutter/issues/509https://gitlab.gnome.org/GNOME/mutter/merge_requests/506
It tried to add a (implicitly casted) float to a uint64_t, and due to
floating point precision issues resulted in timestamps intended to be
in the future to actually be in the past. Fix this by first casting the
delay to an uint64_t, then add it to the time stamp.
https://gitlab.gnome.org/GNOME/mutter/merge_requests/506
We might fail to page flip a new buffer, often after resuming, due to
the FIFO being full. Prior to this commit, we handled this by switching
over to plain mode setting instead of page flipping. This is bad because
we won't be synchronized to the refresh rate anymore, but just the
clock.
Instead, deal with this by trying again until the FIFO is no longer
full. Do this on a v-sync based interval, until it works.
This also changes the error handling code for drivers not supporting
page flipping to rely on them returning -EINVAL. The handling is moved
from pretending a page flip working to explicit mode setting in
meta-renderer-native.c.
Fixes: https://gitlab.gnome.org/GNOME/mutter/issues/460
A renderer view will, under the native backend, since long ago always
have a logical monitor associated with it, so remove the code handling
the legacy non-stage view case.
https://gitlab.gnome.org/GNOME/mutter/issues/460
If the extension is missing, the GPU copy path would not work. The code sets
the error, but forgets to return a failure. Fix this.
While adding the necessary return FALSE, also destroy the EGL context we just
created. Code refactoring shares the destroying code.
Found by reading code.
https://gitlab.gnome.org/GNOME/mutter/merge_requests/416
If the GPU copy path would use a software renderer, fall back to the CPU
copy path. The CPU copy path is possibly faster and avoids screen
corruption issues that were observed on an Intel Haswell desktop. The
corruption was likely due to texturing from an unfinished rendering or
memory caching issues.
https://gitlab.gnome.org/GNOME/mutter/merge_requests/325
Print the pixel format chosen for an output on a secondary GPU for
debugging. Knowing the format can aid in debugging e.g. red/blue channel
swaps and CPU copy performance issues.
This adds a DRM format printing helper in meta-crtc-kms.h. This header
is included in most native backend files making it widely available,
while DRM formats are specific to the native backend. It could be shared
with Wayland bits, DRM format codes are used there too.
The helper makes the pixel format much more readable than a "%x".
https://gitlab.gnome.org/GNOME/mutter/merge_requests/341
When setting up an output on a secondary GPU with the CPU copy mode,
allocate the dumb buffers with a DRM format that is advertised supported
instead of hardcoding a format.
Particularly, DisplayLink devices do not quite yet support the hardcoded
DRM_FORMAT_XBGR8888. The proprietary driver stack actually ignores the
format assuming it is DRM_FORMAT_XRGB8888 which results the display
having red and blue channels swapped. This patch fixes the color swap
right now, while taking advantage if the driver adds support for XBGR
later.
The preferred_formats ordering is somewhat arbitrary. Here it is written
from glReadPixels point of view, based on my benchmarks on Intel Haswell
Desktop machine. This ordering prefers the format that was hardcoded
before.
https://gitlab.gnome.org/GNOME/mutter/merge_requests/341
As with the commits earlier, this also adds const qualifiers where
expected. However, the const variables are casted to non-const variants
so they can be passed to glib functions that take non-const variants but
expect const-like input.
Mutter prefers platform devices over anything else as the primary GPU.
This will not work too well, when a platform device does not actually
have a rendering GPU but is a display-only device. An example of this
are DisplayLink devices with the proprietary driver stack, which exposes
a DRM KMS platform device but without any rendering driver.
Mutter cannot rely on EGL init failing on such devices either, because
nowadays Mesa supports software renderers on GBM, so the initialization
may well succeed.
The hardware rendering capability is recognized by matching the GL
renderer string to the known Mesa software renderers. At this time,
there is no better alternative to detecting this.
The secondary GPU data is abused for the GL renderer, as the Cogl
context may not have been created yet. Also, the Cogl context would
only be created on the primary GPU, but at this point the primary GPU
has not been chosen yet. Hence, GPU copy path GL context is used as a
proxy and predictor of what the Cogl context might be if it was created.
Mind, that even the GL flavour are not the same between Cogl and
secondary contexts, so this is stretch but it should be just enough.
The logic to choose the primary GPU is changed to always prefer hardware
rendering devices while also maintaining the old order of preferring
platform over boot_vga devices.
Co-authored by: Emilio Pozuelo Monfort <emilio.pozuelo@collabora.co.uk>
https://gitlab.gnome.org/GNOME/mutter/merge_requests/271
Moves the primary GPU choosing to after all secondary gpu data has been
created.
This makes it possible for a future patch to start looking at secondary
gpu data in choose_primary_gpu () to determine if it is using a hardware
driver or a software renderer.
Co-authored by: Pekka Paalanen <pekka.paalanen@collabora.com>
https://gitlab.gnome.org/GNOME/mutter/merge_requests/271
Initialize the secondary GPU data for all GPUs, even the primary one. By
not looking at the primary_gpu_kms member, a future patch is allowed to
postpone choosing the primary GPU.
A future patch will use the secondary GPU data to decide which GPU will
become the primary GPU.
Co-authored by: Pekka Paalanen <pekka.paalanen@collabora.com>
https://gitlab.gnome.org/GNOME/mutter/merge_requests/271
create_renderer_gpu_data_egl_device () relied on the primary GPU being
already chosen for the "EGLDevice currently only works with single GPU
systems" error message. A future patch will choose the primary GPU after
this, not before, so this check needs to be rewritten before the
initialization order is changed.
The new check is implemented exactly as the error message says: there
must be exactly one GPU, otherwise fail.
https://gitlab.gnome.org/GNOME/mutter/merge_requests/271
Make the choosing and identity of the primary GPU an internal detail to
the native renderer. MonitorManagerKms did not need it for anything.
The primary GPU logic remains unchanged.
This allows follow-up patches to change how the renderer chooses the
primary GPU. It will be easier for the renderer to use private
information for choosing.
https://gitlab.gnome.org/GNOME/mutter/merge_requests/271
If a KMS device has the DRM_CAP_DUMB_PREFER_SHADOW and a software based
GL driver is used, always use a shadow fb. This will speed up read backs
in the llvmpipe OpenGL implementation, making blend operations faster.
Fixes: https://gitlab.gnome.org/GNOME/mutter/issues/106
