This reverts commit 9c23076b9e.
Some scenarios have come to light where this failed to ensure the docker
image exists:
* If the master branch of a forked repository is used for an MR which
doesn't modify .gitlab-ci.yml, the docker-image job may not run.
* If the docker-image job of the first pipeline in a forked repository
is cancelled or fails for any reason, and .gitlab-ci.yml isn't
modified for the next pipeline run.
drm_queue_handler just puts the event on the signalled list; without
calling drm_queue_handle_deferred, actual processing of the event may be
delayed indefinitely, e.g. until another event arrives from the kernel.
This could result in DRI2 clients hanging during DPMS off.
Fixes: 739181c8d3 "Add amdgpu_drm_handle_event wrapper for
drmHandleEvent"
Reviewed-by: Aaron Liu <aaron.liu@amd.com>
Tested-by: Aaron Liu <aaron.liu@amd.com>
If they don't, flipping will result in corrupted display.
Test case:
* Run Xorg at 1920x1080 with no window manager
* glxgears -geometry 2048x1080
The Present extension code in xserver 1.21 will check for this.
Tested-by: Jax Lin <jax.lin@amd.com>
If the window is currently flipping.
This might make a difference when the property gets disabled: Variable
refresh will now be disabled immediately in that case, instead of only
when the window can no longer use page flipping at all.
Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Instead of scanning for PropertyNotify events. Reasons:
* Works even if no client listens to PropertyNotify events for the
window.
* No overhead on delivery of unrelated events, and no overhead at all
if Option "VariableRefresh" is disabled.
v2:
* Use shorter variable name amdgpu_vrr_atom.
* Call MakeAtom regardless of info->instance_id, for robustness vs VRR
being enabled in some but not all AMDGPU screens.
Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Preparation for the following change.
v2:
* Add comments explaining what the wrappers are wrapping.
* Use global amdgpu_property_vectors_wrapped to keep track of whether
the vectors need to be (un)wrapped, for robustness against VRR being
enabled in some but not all AMDGPU screens.
Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Windows aren't associated with GPU screens, and amdgpu_present_flip is
never called for them, so VRR can never actually be enabled for them.
Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Otherwise there's normally no need to run it. It will also run when a
new branch is created, which ensures that the docker image always exists
(e.g. in a newly forked repository).
Inspired by https://gitlab.freedesktop.org/mesa/mesa/merge_requests/143
drm_wait_pending_flip stopped waiting if drm_handle_event returned 0,
but that might have processed only some unrelated DRM events. As long as
the flip is pending, we have to keep waiting for its completion event.
Noticed while working on the previous fix.
Acked-by: Alex Deucher <alexander.deucher@amd.com>
drmHandleEvent can be interrupted by a signal in read(), in which case
it doesn't process any events but returns -1, which
drm_handle_event propagated to its callers. This could cause the
following failure cascade:
1. drm_wait_pending_flip stopped waiting for a pending flip.
2. Its caller cleared drmmode_crtc->flip_pending before the flip
completed.
3. Another flip was attempted but got an unexpected EBUSY error because
the previous flip was still pending.
4. TearFree was disabled due to the error.
The solution is to call drmHandleEvent if it was interrupted by a
signal. We can do that in drm_handle_event, because when that is called,
either it is known that there are events ready to be processed, or the
caller has to wait for events to arrive anyway.
v2:
* Use ErrorF instead of xf86DrvMsg with hard-coded screen 0.
Bugzilla: https://bugs.freedesktop.org/109364
Reviewed-by: Alex Deucher <alexander.deucher@amd.com> # v1
And only clear it if it matches the framebuffer of the completed flip
being processed.
Fixes
(WW) AMDGPU(0): flip queue failed: Device or resource busy
(WW) AMDGPU(0): Page flip failed: Device or resource busy
(EE) AMDGPU(0): present flip failed
due to clobbering drmmode_crtc->flip_pending.
Reproducer: Enable TearFree, run warzone2100 fullscreen, toggle
Vertical sync on/off under Video Options. Discovered while investigating
https://bugs.freedesktop.org/109364 .
Acked-by: Alex Deucher <alexander.deucher@amd.com>
We were using a relative target of 0, meaning "complete the flip ASAP".
This could result in the flip sometimes, but not always completing in
the same vertical blank period where the corresponding drawing occurred,
potentially causing judder artifacts with applications updating their
window contents synchronized to the display refresh. A good way to test
this is the vsynctester.com site in a windowed browser, where the judder
results in the large "VSYNC" text intermittently appearing red or cyan
instead of the expected gray.
To avoid this, use a relative target MSC of 1, meaning that if a
vertical blank period is in progress, the flip will only complete in the
next one.
Reported by Julian Tempel and Brandon Wright in
https://bugs.freedesktop.org/106175 .
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
If the compositing manager uses direct rendering (as is usually the case
these days), the storage of a pixmap allocated by glamor_create_pixmap
needs to be reallocated for sharing it with the compositing manager.
Instead, allocate pixmap storage which can be shared directly.
Acked-by: Alex Deucher <alexander.deucher@amd.com>
There's no point in listening for hotplug events if RandR is disabled,
as there's no other mechanism for them to be propagated. We were already
mostly ignoring them in that case.
Inspired by
1a489142c8
(via https://bugs.freedesktop.org/109230#c11).
(Ported from radeon commit 38db1bbcfc019c92884c7819a6630c70e543f6b2)
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
present_unflip can get called between present_check_flip and
present_flip, in which case the latter would pass a NULL WindowPtr to
the former, resulting in a crash.
present_flip should never get called for a window which has already been
destroyed, so there's no need to clear info->flip_window.
Bugzilla: https://bugs.freedesktop.org/109067
Fixes: 2d18b37159 "Check last flip window instead of screen root
before flipping"
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
amdgpu_do_pageflip() indexed the flipdata->fb[] array
indexing over config->num_crtc, but the flip completion
routines, e.g., drmmode_flip_handler(), index that array
via the crtc hw id from drmmode_get_crtc_id(crtc).
This is mismatched and causes indexing into the wrong
array slot at flip completion -> Server crash.
Always use drmmode_get_crtc_id(crtc) for indexing into
the array to fix this.
Tested on a dual-X-Screen setup with one video output
assigned to each X-Screen, page-flipping an OpenGL app
on either of both X-Screens. This used to crash when
flipping on X-Screen 1, now it doesn't anymore.
Fixes: 9b6782c821 "Store FB for each CRTC in drmmode_flipdata_rec"
(Ported from radeon commit 0058fd2ebf4c900b12f129984e98886a7ac84b2f)
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com>
Switch to the other buffer when xf86_config->cursor changes. Avoids
these issues possible when re-using the same buffer:
* The HW may intermittently display a mix of the old and new cursor
images.
* If the hotspot changes, the HW may intermittently display the new
cursor image at the location corresponding to the old image's hotspot.
Bugzilla: https://bugs.freedesktop.org/108832
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
This should be functionally equivalent to what drmModeSetCursor(2) do
behind the scenes, but allows for new tricks in following changes.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
GBM doesn't really buy us anything for the cursor BOs. This simplifies
the code and following changes.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Specifically, after both the page flip and vblank ioctls failed, but
then the vblank ioctl started working again. This can happen
intermittently e.g. when hotplugging a DP display. Previously, TearFree
would stay disabled in that case until a modeset was triggered somehow.
Bugzilla: https://bugs.freedesktop.org/103791
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
drmmode_crtc_scanout_update does the equivalent of a scanout update,
so no need to do it again. This might also avoid issues if there's a
pending scanout update at this point.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
This removes the dependency on an externally generated docker image, and
should make it easier to update the docker image or make other changes
related to it.
v2:
* If the image doesn't exist, try pulling it from the main repo's
registry.
* Use debian:testing-slim as the base, might result in a slightly
smaller image.
v3:
* Prevent installation of packages which are only recommended, for an
even smaller image.
* Add recommendation to remove new image from source repository in
favour of the main repository's.
Acked-by: Alex Deucher <alexander.deucher@amd.com> # v2
This patch adds support for setting the CRTC variable refresh property
for suitable windows flipping via the Present extension.
The "VariableRefresh" Option is added to AMDGPU in this patch. This
option defaults to false, and must be set to "true" in an X conf
file for variable refresh support in the driver.
In order for a window to be suitable for variable refresh it must have
the _VARIABLE_REFRESH property with a 32-bit CARDINAL value of 1.
Then the window must pass the checks required to be suitable for
Present extension flips - it must cover the entire X screen and no
other window may already be flipping.
With these conditions met every CRTC for the X screen will have their
variable refresh property set to true.
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
A significant amount of time can pass between the X call into
check_flip for a window and when amdgpu_present_flip actually occurs.
To ensure that flipping is still possible there was an additional check
performed on screen->root in amdgpu_present_flip - but what should
be checked instead is the window itself. This only really worked before
because X ensures that the window has the same dimensions as the screen
to allow for present extension flipping.
This patch tracks the flip window between calls to check_flip and flip
and uses that window instead of screen->root.
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
The un-premultiplied R/G/B values would overflow the gamma LUT, so just
pass through the data unchanged, and leave it up to the HW how to
interpret such weird premultiplied alpha pixels.
Bugzilla: https://bugs.freedesktop.org/108355
It was still possible for nested xorg_list_for_each_entry_safe loops
to occur over the drm_vblank_signalled list, which could mess up that
list. Moving deferred events to a separate list allows processing the
drm_vblank_signalled list without xorg_list_for_each_entry_safe.
v2:
* Refactor drm_handle_vblank_signalled helper function, less code
duplication => better readability (Alex Deucher)
Bugzilla: https://bugs.freedesktop.org/108600
Acked-by: Alex Deucher <alexander.deucher@amd.com>
When an async flip is performed, and TearFree is enabled on the CRTC
used for timing, we schedule a vblank event for completing the page
flip. The DRM event queuing code treated this event like a vblank event,
but it needs to be treated like a page flip event.
Acked-by: Alex Deucher <alexander.deucher@amd.com>
At this point, we've already established that e->handler is NULL, no
need to check again in drm_queue_handle_one. This also makes it clearer
what's happening.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
The stricter detection broke the cursor in some games. Apparently those
use cursor data with premultiplied alpha, but with some pixels having
r/g/b values larger than the alpha value (which corresponds to original
r/g/b values > 1.0), triggering the workaround.
Relax the detection to match what's in the X server since 1.18.4, but
keep the workaround for older versions.
Bugzilla: https://bugs.freedesktop.org/108650
Acked-by: Alex Deucher <alexander.deucher@amd.com>
X server >= 1.18 already had code for this, but it only caught cases
where some pixels have 0 for alpha and non-0 for a non-alpha component.
Turns out some apps (e.g. the Civilization VI game) use
non-premultiplied cursor data which doesn't have such pixels, but can
still result in visual artifacts.
This uses the method suggested by Kamil in
https://bugs.freedesktop.org/92309#c19: check for pixels where any
colour component value is larger than the alpha value, which isn't
possible with premultiplied alpha.
There can still be non-premultiplied data which won't be caught by this,
but that should result in slightly incorrect colours and/or blending at
the worst, not wildly incorrect colours such as shown in the bug report
below.
v2:
* Disable the check with current xserver Git master, which already does
the same check now.
Bugzilla: https://bugs.freedesktop.org/108355
Suggested-by: Kamil Páral <kamil.paral@gmail.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Corresponding to up to six CRTCs being available in the hardware.
v2:
* Move instance overflow check from PreInit to the probe hooks, in
order to further minimize wasted effort.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com> # v1
Builds the driver against all supported versions of xserver, with both
gcc and clang for xserver >= 1.18 (older versions cause warnings with
clang). Compiler warnings are treated as errors.
The xserver 1.15 build uses standalone glamor, the xserver 1.13 & 1.14
builds use --disable-glamor.
With the latest xserver version, make install and make distcheck are
tested as well.
