When using DRI3+Present with PRIME render offload, sometimes there is
a mismatch between the stride of the to-be-presented Pixmap and the
frontbuffer. The current code would reject a pageflip present in this
case if atomic modesetting is not enabled, ie. always, as atomic
modesetting is disabled by default due to brokeness in the current
modesetting-ddx.
Fullscreen presents without page flipping however trigger the copy
path as fallback, which causes not only unreliable presentation timing
and degraded performance, but also massive tearing artifacts due to
rendering to the framebuffer without any hardware sync to vblank.
Tearing is extra awful on modesetting-ddx because glamor afaics seems
to use drawing of a textured triangle strip for the copy implementation,
not a dedicated blitter engine. The rasterization pattern creates extra
awful tearing artifacts.
We can do better: According to a tip from Michel Daenzer (thanks!),
at least atomic modesetting capable kms drivers should be able to
reliably change scanout stride during a pageflip, even if atomic
modesetting is not actually enabled for the modesetting client.
This commit adds detection logic to find out if the underlying kms
driver is atomic_modeset_capable, and if so, it no longer rejects
page flip presents on mismatched stride between new Pixmap and
frontbuffer.
We (ab)use a call to drmSetClientCap(ms->fd, DRM_CLIENT_CAP_ATOMIC, 0);
for this purpose. The call itself has no practical effect, as it
requests disabling atomic mode, although atomic mode is disabled by
default. However, the return value of drmSetClientCap() tells us if the
underlying kms driver is atomic modesetting capable: An atomic driver
will return 0 for success. A legacy non-atomic driver will return a
non-zero error code, either -EINVAL for early atomic Linux versions
4.0 - 4.19 (or for non-atomic Linux 3.x and earlier), or -EOPNOTSUPP
for Linux 4.20 and later.
Testing on a MacBookPro 2017 with Intel Kabylake display server gpu +
AMD Polaris11 as prime renderoffload gpu, X-Server master + Mesa 21.0.3
show improvement from unbearable tearing to perfect, despite a stride
mismatch between display gpu and Pixmap of 11776 Bytes vs. 11520
Bytes. That this is correct behaviour was also confirmed by comparing the
behaviour and .check_flip implementation of the patched modesetting-ddx
against the current intel-ddx SNA Present implementation.
Please consider merging this patch before the server-1.21 branch point.
This patch could also be cherry-picked into the server 1.20 branch to
fix the same limitation.
Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com>
In some scenarios, the Wayland compositor might have more knowledge
than the X11 server and may be able to perform pointer emulation for
touch events better. Add a command-line switch to allow compositors
to turn Xwayland pointer emulation off.
Signed-off-by: Simon Ser <contact@emersion.fr>
This reverts commit 9b89994110.
Turns out that defaulting glamor_egl->dmabuf_capable = TRUE
breaks kms page-flipping on various Mesa+Linux/DRM-KMS+hardware
combos, resulting in broken presentation timing, degraded performance
and awful tearing. E.g., my testing shows that X-Server master +
Mesa 21.2 + Linux 5.3 on Intel Kabylake has broken pageflipping.
Similar behaviour was observed in the past on RaspberryPi 4/400
with VideoCore-6 by myself and others, and iirc by myself on some
AMD gpu's, although my memories of the latter are a bit dim.
Cfe. https://gitlab.freedesktop.org/mesa/mesa/-/issues/3601 and
possibly https://gitlab.freedesktop.org/xorg/xserver/-/merge_requests/254
for related problems.
The reason for pageflip failure on the modesetting-ddx under
DRI3/Present seems to be the following sequence:
1. Atomic modesetting for the modesetting-ddx is broken and therefore
both disabled by default in the modesetting-ddx itself and also
force-disabled by the Linux kernel since quite a while. If the kernel
detects drmSetClientCap(fd, DRM_CLIENT_CAP_ATOMIC, 1); from the
X-Server, it will reject the request, as a countermeasure to all the
past and current brokeness.
2. Without DRM_CLIENT_CAP_ATOMIC we don't get the implied universal
planes support (DRM_CLIENT_CAP_UNIVERSAL_PLANES).
3. Without DRM_CLIENT_CAP_UNIVERSAL_PLANES, drmModeGetPlaneResources()
will only return overlay planes, but not primary- or cursor planes.
4. As modesetting-ddx drmmode_crtc_create_planes() function can only
operate on primary planes, but can't get any from drmModeGetPlaneResources(),
the drmmode_crtc_create_planes() mostly turns into a no-op, never
executes populate_format_modifiers() and therefore the Linux kernels
DRM-KMS driver is not ever queried for the list of scanout/pageflip
capable DRM format modifiers. Iow. the drmmode_crtc->formats[i].modifiers
list stays empty with zero drmmode_crtc->formats[i].num_modifiers.
5. The list from step 4 provides the format+modifiers for intersection
which would get returned by the X-Servers DRI3 backend as response to
a xcb_dri3_get_supported_modifiers_window_modifiers() request. Given
an empty list was returned in step 4, this will lead to return of an
empty modifiers list by xcb_dri3_get_supported_modifiers_window_modifiers().
6. Both Mesa's DRI3/Present OpenGL backbuffer allocation logic and iirc
Mesa/Vulkan/WSI/X11's swapchain image allocation logic use the list
from xcb_dri3_get_supported_modifiers_window_modifiers() for format+
modifier selection for scanout/pageflip capable buffers. Cfe. Mesa's
dri3_alloc_render_buffer() function.
Due to the empty list, the Mesa code falls back to the format+modifiers
reported by xcb_dri3_get_supported_modifiers_screen_modifiers()
instead. This list contains all modifiers reported by GLAMOR as
result of glamor_get_formats() and glamor_get_modifiers(), which
in turn are query results from Mesa eglQueryDmaBufFormatsEXT()
and eglQueryDmaBufModifiersEXT(). Iow. all format+modifiers which
are supported for rendering are considered for the OpenGL backbuffers
and Vulkan swapchain buffers.
7. Depending on kms driver + gpu combo and Mesa version, such buffers
are often not direct-scanout / pageflip capable, and so pageflipping
can't be used for DRI3/Present of fullscreen windows. Whenever the
system has to fallback to copies instead of pageflips, the results
are broken presentation timing, degraded performance and quite
horrible tearing, as the current DRI3/Present implementation does not
perform any hardware synchronization of copy presents to the start
of vblank or similar.
By defaulting glamor_egl->dmabuf_capable = FALSE instead, as the server
1.20 branch does, we avoid this failure:
1. glamor_get_modifiers() turns into a no-op and returns false, not
reporting any supported dmabuf modifiers to the servers DRI3 code,
ie. the servers cache_formats_and_modifiers() function can't retrieve
and cache any format+modifiers. Therefore the servers DRI3 code now
also reports an empty format+modifiers list when Mesa does a
xcb_dri3_get_supported_modifiers_screen_modifiers() query.
2. Mesa's buffer allocation code therefore falls back to using the old
DRI image extensions createImage() function to allocate buffers
with use flags __DRI_IMAGE_USE_SCANOUT | __DRI_IMAGE_USE_BACKBUFFER
and our OpenGL backbuffers / Vulkan swapchain images get allocated
in a direct-scanout / pageflip capable format. Pageflipping works,
timing and performance is good, presentation is tear-free.
Please consider merging this for branching the X-Server 1.21 branch.
Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com>
misc.h has complex logic (checking MAXSHORT is undefined etc.)
controlling if it includes assert.h or not.
Including windows.h from w32api 9.0.0 now trips over that, causing
assert.h to not be included, causing various errors, e.g.
In file included from ../include/cursor.h:53,
from ../include/dix.h:54,
from ../os/osdep.h:139,
from ../hw/xwin/winauth.c:40:
../include/privates.h: In function ‘dixGetPrivateAddr’:
../include/privates.h:121:5: error: implicit declaration of function ‘assert’ [-Werror=implicit-function-declaration]
Fix this by IWYU in privates.h
A misplaced error check can cause this failure scenario, and does
so reliably as tested on Ubuntu 21.04 with KDE Plasma 5 desktop
within the first few seconds of login session startup, rendering
VRR under modesetting-ddx unusable:
1. Some X11 client application changes some window property.
2. ms_change_property() is called as part of the property change
handling call chain (client->requestVector[X_ChangeProperty]).
It removes itself temporarily from the call chain - or so it
thinks, hooking up saved_change_property instead.
3. ret = saved_change_property(client) is called and fails
temporarily for some non-critical reason.
4. The misplaced error check returns early (error abort), without
first restoring ms_change_property() as initial X_ChangeProperty
handler in the call chain again.
-> Now ms_change_property() has removed itself permanently from the
property handler call chain for the remainder of the X session
and VRR property changes on windows are no longer handled, ie.
VRR no longer gets enabled/disabled in response to window VRR
property changes.
Place the error check at the proper place, just as it is correctly
done by amdgpu-ddx, and in modesetting-ddx ms_delete_property()
function.
Verified to fix VRR handling with an AMD gpu under KDE desktop
session.
Please consider merging before branching the server 1.21 branch.
Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com>
I get this:
In function ‘TryCopyStr’,
inlined from ‘CopyISOLockArgs’ at ../xkb/xkbtext.c:875:9:
../xkb/xkbtext.c:720:13: warning: ‘tbuf’ may be used uninitialized [-Wmaybe-uninitialized]
720 | strcat(to, from);
| ^~~~~~~~~~~~~~~~
../xkb/xkbtext.c: In function ‘CopyISOLockArgs’:
<built-in>: note: by argument 1 of type ‘const char *’ to ‘__builtin_strlen’ declared here
../xkb/xkbtext.c:871:10: note: ‘tbuf’ declared here
871 | char tbuf[64];
| ^~~~
Just initialize tbuf so it definitely works.
[45/388] Compiling C object Xi/liblibxserver_xi.a.p/xichangehierarchy.c.o
../Xi/xichangehierarchy.c:61:32: warning: argument 1 of type ‘int[256]’ with mismatched bound [-Warray-parameter=]
61 | XISendDeviceHierarchyEvent(int flags[MAXDEVICES])
| ~~~~^~~~~~~~~~~~~~~~~
In file included from ../Xi/xichangehierarchy.c:54:
../Xi/xichangehierarchy.h:42:37: note: previously declared as ‘int[]’
42 | void XISendDeviceHierarchyEvent(int flags[]);
| ~~~~^~~~~~~
The dix-config.h file is not installed, but dix.h is. The include makes the
compilation of external drivers fail (for example the libinput driver).
The Xserver compilation also works without the include, so just remove it.
Signed-off-by: Roman Gilg <subdiff@gmail.com>
The xf86CVTMode() was implemented in a standalone source file because it
was being used for both the xfree86 API and the standalone cvt utility.
Now that the cvt utility is removed (as part of libxcvt) we can move the
small xf86CVTMode() function with the rest of the xf86Modes sources.
Closes: https://gitlab.freedesktop.org/xorg/xserver/-/issues/1142
Signed-off-by: Olivier Fourdan <ofourdan@redhat.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
The cvt utility is now replaced by the standalone version found in
libxcvt, no need to build the one in xfree86 anymore.
Closes: https://gitlab.freedesktop.org/xorg/xserver/-/issues/1142
Signed-off-by: Olivier Fourdan <ofourdan@redhat.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Replace the local implementation of the VESA CVT standard timing
modelines generator with the one from libxct to avoid code duplication.
Closes: https://gitlab.freedesktop.org/xorg/xserver/-/issues/1142
Signed-off-by: Olivier Fourdan <ofourdan@redhat.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
Xwayland is using a copy of the CVT generator found in Xorg.
Rather than duplicating the code within the xserver tree, use the
libxcvt implementation instead.
Closes: https://gitlab.freedesktop.org/xorg/xserver/-/issues/1142
Signed-off-by: Olivier Fourdan <ofourdan@redhat.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
libxcvt is a library providing a standalone version of the X server
implementation of the VESA CVT standard timing modelines generator.
Closes: https://gitlab.freedesktop.org/xorg/xserver/-/issues/1142
Signed-off-by: Olivier Fourdan <ofourdan@redhat.com>
Reviewed-by: Matt Turner <mattst88@gmail.com>
It is copied using memcpy() and not modified so we can add const. This
fixes a -Wincompatible-pointer-types-discards-qualifiers compiler warning
that was failing a -Werror XVnc build for me.
Signed-off-by: Alex Richardson <Alexander.Richardson@cl.cam.ac.uk>
If there is an explicit configuration, assign the RandR provider
of the GPUDevice to the screen it was specified for.
If there is no configuration (default case) the screen number is
still 0 so it doesn't change behaviour.
The result is e.g:
# DISPLAY=:0.2 xrandr --listproviders
Providers: number : 2
Provider 0: id: 0xd2 cap: 0x2, Sink Output crtcs: 1 outputs: 1 associated providers: 0 name:modesetting
Provider 1: id: 0xfd cap: 0xb, Source Output, Sink Output, Sink Offload crtcs: 2 outputs: 2 associated providers: 0 name:Intel
Signed-off-by: Zoltán Böszörményi <zboszor@gmail.com>
Currently, when main hardware screen is powered-off,
X server initializes fake screen's timer with
1 second update interval.
Streaming software like Nomachine or Vnc, as well as
desktop input automation suffers from it, since it
will forever be stuck on 1 fps until the display is
turned back on.
This commit adds command line option -fakescreenfps <int>
that allows the user to change the default fake screen
timer.
Signed-off-by: Baranin Alexander <ismailsiege@gmail.com>
Since crtc can belong to secondary output that may not have present
extension enabled we should fallback to first enabled crtc or fake crtc.
Fix for issue xorg/xserver#1195
We are handling two cases here: the active flip or the pending flip.
For the pending flip (event->pending == TRUE), we called
xwl_present_release_pixmap.
For the active flip (event->pending == FALSE), we called
xwl_present_release_event. However, xwl_present_flip_notify_vblank
already unhooked event->vblank.event_queue. So this was effectively the
same as calling xwl_present_release_pixmap.
Acked-by: Olivier Fourdan <ofourdan@redhat.com>
Use present_vblank_rec::event_queue instead.
The changes in xwl_present_execute shouldn't really be needed, since
we should never hit queue_vblank in present_execute_wait. But let's be
safe rather than sorry, plus this simplifies the code.
Acked-by: Olivier Fourdan <ofourdan@redhat.com>
Can just call xwl_present_execute directly.
This allows dropping the window member from struct xwl_present_window as
well.
Acked-by: Olivier Fourdan <ofourdan@redhat.com>
We clear the vblank->pixmap field, so next time xwl_present_execute
falls through to present_execute_post.
Acked-by: Olivier Fourdan <ofourdan@redhat.com>
This allows for various simplifications.
Use the pointer to the struct memory as the event ID. In contrast to
the SCMD code for Xorg (where pending DRM events cannot be cancelled),
this is safe here, because we can destroy pending Wayland callbacks. So
we can't get a callback with a stale pointer to freed memory.
Remove xwl_present_window::release_list in favour of
present_vblank_rec::window_list.
Remove xwl_present_event::xwl_present_window in favour of
present_vblank_rec::window.
xwl_present_free_event is never called for a NULL pointer anymore, no
need to check.
v2:
* Restore DestroyWindow wrapping order to make sure
present_destroy_window doesn't call xwl_present_abort_vblank.
Acked-by: Olivier Fourdan <ofourdan@redhat.com>
We can call xwl_present_free_event unconditionally from
xwl_present_abort_vblank, since the sync_callback is already destroyed
in xwl_present_cleanup.
Acked-by: Olivier Fourdan <ofourdan@redhat.com>
Mario Kleiner
Simon Ser
Jon Turney
Povilas Kanapickas
Adam Jackson
Demi Marie Obenour
Roman Gilg
Olivier Fourdan
Alex Richardson
Zoltán Böszörményi
Boris-Barboris
Łukasz Spintzyk
Michel Dänzer