Creating an EGLImage is rather expensive and is taking the bulk of the
time the secondary GPU copy path is using for each frame. By caching
these per GBM BO we avoid this expensive recreation, which seems to
significantly improve FPS throughput in these scenarios, e.g. an
AMD or Intel iGPU with an NVIDIA dGPU.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/4027>
Not with VRR though, because the deadline timer fires at most every 33ms
then.
Previously, the deadline timer was used only for cursor-only updates.
Using it for other updates means they pick up the latest cursor position
available at the deadline, resulting in the lowest possible input→output
latency for cursor movement.
TTBOMK this unlocks the full potential of the KMS thread given the
current atomic KMS API.
v2:
* Don't call meta_kms_update_merge_from with twice the same update
pointer.
v3:
* Don't arm deadline timer if crtc_frame->pending_page_flip is true.
v4:
* Tweak want_deadline_timer indentation per check-code-style CI job.
v5:
* Also check crtc_frame->await_flush for want_deadline_timer.
v6:
* Tweak coding style to keep lines shorter. (Jonas Ådahl)
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3964>
Previously it was just dropped on the floor with no feedback, which
could result in the main thread waiting for something which never
happens.
Fixes: 0d9fd1ead7a7 ("kms/impl-device: Destroy submitted update in disarm_all_deadline_timers")
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3964>
To swap_buffer_result_feedback from page_flip_feedback_discarded. The
former is where META_KMS_ERROR_DISCARDED from disarm_all_frame_sources
gets handled here.
Fixes: af250506fbe8 ("kms/impl-device: Queue result when discarding submitted update")
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3964>
We closed all the DMA buffer fds after creating the DMA buffer backed
CoglFramebuffer, which meant that the fds we passed to CoglDmaBufHandle
were already closed. This broke screen casting with DMA buffers.
Fixes: 8509b74532 ("Make DMA buffer allocation format and modifier aware")
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/4177>
This makes it possible to use udev even for backends which are not
native. Specifically we want to start controling backlights in mutter
which has to happen for both native and X11.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/4130>
The generic backend now provides the launcher which is the last native
specific dependency. This makes it possible to move the udev component
to the generic backend.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/4130>
For now we only pause/resume the rendering in the base class but in the
next few commits, we'll move udev from native to generic and need to
pause/resume it as well.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/4130>
By connecting to changes of the new property in backend native itself.
This removes the dependency on the native backend and allows us to move
it to the generic backend.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/4130>
We create the cursor view data when we update the cursor, but allow the
texture to be realized earlier than that. The texture change listener
invalidates the hw cursor validity state, but since it's initialized to
invalid, the hardware cursor will still be correctly handled.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3859>
Fixes leak:
==9634== 60 (16 direct, 44 indirect) bytes in 1 blocks are definitely lost in loss record 8,198 of 13,049
==9634== at 0x48489F3: calloc (vg_replace_malloc.c:1340)
==9634== by 0x5DDCD29: drmModeGetPlaneResources (xf86drmMode.c:1243)
==9634== by 0x4AD94C9: init_planes (meta-kms-impl-device.c:1010)
==9634== by 0x4ADC7B5: meta_kms_impl_device_init_mode_setting (meta-kms-impl-device.c:2350)
==9634== by 0x4AD27BC: meta_kms_impl_device_atomic_initable_init (meta-kms-impl-device-atomic.c:1416)
==9634== by 0x4DD3B32: g_initable_new_valist (ginitable.c:249)
==9634== by 0x4DD3C19: g_initable_new (ginitable.c:163)
==9634== by 0x4ACEAE8: meta_create_kms_impl_device (meta-kms-device.c:662)
==9634== by 0x4ACECD1: create_impl_device_in_impl (meta-kms-device.c:722)
==9634== by 0x4B01837: meta_thread_impl_dispatch (meta-thread-impl.c:542)
==9634== by 0x4B00D06: impl_source_dispatch (meta-thread-impl.c:175)
==9634== by 0x4C5C81E: g_main_dispatch (gmain.c:3357)
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/4149>
The same MetaCrtcMode objects can be reached via the
MetaOutputInfo::modes array & MetaMonitorMode::crtc_modes arrays and via
the MetaGpuPrivate::modes list, so all of them need to hold their own
references.
In turn, those references need to dropped in meta_output_info_unref and
meta_monitor_mode_free. (meta_gpu_finalize is already dropping the
MetaGpuPrivate::modes list references)
Fixes leak:
==4092== 123 (32 direct, 91 indirect) bytes in 1 blocks are definitely lost in loss record 11,396 of 13,056
==4092== at 0x4F81D57: g_type_create_instance (gtype.c:1929)
==4092== by 0x4F64ABF: g_object_new_internal.part.0 (gobject.c:2606)
==4092== by 0x4F676BA: g_object_new_internal (gobject.c:2923)
==4092== by 0x4F676BA: g_object_new_valist (gobject.c:2945)
==4092== by 0x4F67A0C: g_object_new (gobject.c:2418)
==4092== by 0x4AB578B: meta_crtc_mode_virtual_new (meta-crtc-mode-virtual.c:51)
==4092== by 0x4B081BB: meta_virtual_monitor_native_new (meta-virtual-monitor-native.c:83)
==4092== by 0x4AC58CC: meta_monitor_manager_native_create_virtual_monitor (meta-monitor-manager-native.c:603)
==4092== by 0x496803B: meta_monitor_manager_create_virtual_monitor (meta-monitor-manager.c:621)
==4092== by 0x4A4A204: create_virtual_monitor (meta-screen-cast-virtual-stream-src.c:625)
==4092== by 0x4A4A39D: ensure_virtual_monitor (meta-screen-cast-virtual-stream-src.c:663)
==4092== by 0x4A4A461: meta_screen_cast_virtual_stream_src_notify_params_updated (meta-screen-cast-virtual-stream-src.c:684)
==4092== by 0x4A568C5: on_stream_param_changed (meta-screen-cast-stream-src.c:1565)
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/4149>
This makes DMA buffer allocation in all layers take an array of possible
modifiers to allocate using, or zero modifiers if implicit modifiers
should be used.
The format hard coding previously used is moved to the screen cast code,
or in case of the (unused by default) shadow buffer buffers use the same
format as the the CoglOnscreen.
This also means the CoglDmaBufHandle and MetaDrmBuffer got taught how to
distinguish between planes. It's mostly unused in practice, so rather
untested.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3900>
It's only implemented by the GBM render device backend, and uses EGL,
thus does not distinguish scanout capable modifiers.
A filter enum is added to Cogl, since it'll be used via the Cogl API,
but the actual Cogl API isn't added yet.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/3900>
On big endian machine, the native 32bits format is
DRM_FORMAT_BGRX8888, so add this format to the list.
Tested on a s390x VM, using virtio-gpu driver.
Signed-off-by: Jocelyn Falempe <jfalempe@redhat.com>
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/4088>
In add_drm_device, calling g_hash_table_insert with NULL for the hash
table (which happens for any device added after mutter finished starting
up) would presumably just drop the key/value pointers on the floor,
leaking the string & MetaRenderDevice object referenced by them. Fixing
this should help for the referenced issue.
In meta_backend_native_take_render_device, g_hash_table_steal_extended
would presumably return FALSE (and log a warning), so the behaviour would
be "correct" by chance.
Issue: https://gitlab.gnome.org/GNOME/mutter/-/issues/3785
Fixes: 58c3734d78e9 ("backend/native: Prepare render devices earlier than KMS")
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/4136>
In contrast to realtime scheduling, this doesn't risk us getting
SIGKILL:ed when the kernel is doing busy looping in
drmModeAtomicCommit() for some reason, but will according to testing,
right now, give us more or less the same benefit when it comes to
dispatch lateness and commit lateness.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/4124>
Either prefer "normal" or "realtime", via an enum, instead of a boolean.
Also make it configurable with an env var
`MUTTER_DEBUG_KMS_SCHEDULING_PRIORITY`, which can be set to either
`normal` or `realtime`.
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/4124>
Some laptops don't have the integrated GPU come earlier in the PCI topology
which means that when mutter enumerates the GPUs it may select the dGPU
as the primary.
In a laptop context, this is unlikely to work because the eDP panel can
only be actively connected to a single GPU at a time.
Closes: https://gitlab.gnome.org/GNOME/mutter/-/issues/3746
Suggested-by: Jonas Ådahl
Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
Part-of: <https://gitlab.gnome.org/GNOME/mutter/-/merge_requests/4114>
