Previously the first card with non-empty connectors, encoders
and crtcs would be selected, however KMSDRM_VideoInit could still reject
it if the connector was not connected. This allow finding the first card
(in a multi GPU setup) that is actually connected to a display.
Manuel Alfayate Corchete
I noticed pt2-clone had problems with it's optional hardware mouse on the KMSDRM backend: cursor had a transparent block around it.
So I was investigating and it seems that a GBM cursor needs it's pixels to be alpha-premultiplied instead of straight-alpha.
A
lso, I was previously relying on "manual testing" for the cursor size, but it's far better to use whatever the DRM driver recommends via drmGetCap(): any working driver should make a size recommendation via drmGetCap(), so that's what we use now. I took this decision because I found out that the AMDGPU driver reported working cursor sizes that would appear garbled on screen, and only the recommended cursor size works.
The two are only ever called together, and combining them makes it possible
to eliminate redundant symbol loading and redundant attempts to connect
to a display server.
Manuel Alfayate Corchete
This small patch fixes the KMSDRM_CreateSurfaces() call in KMSDRM_CreateWindow(), that was segfaulting deeper into SDL internals because the windata->viddata pointer wasn't set before the KMSDRM_CreateSurfaces() call.
So that's what this small patch does.
Now, L?VE2D works perfectly well on the Raspberry Pi 3, instead of just segfaulting.
Manuel Alfayate Corchete
This patch is needed so programs that do this work as expected:
1) Start in a different video mode than the mode used by the system and then...
2) Try to go fullscreen with the mode originally used by the system via SetWindowFullScreen() with the SDL_WINDOW_FULLSCREEN_DESKTOP flag.
An example would be pt2-clone in https://github.com/8bitbubsy/pt2-clone.
This program does this:
Starts with:
video.window = SDL_CreateWindow("", SDL_WINDOWPOS_CENTERED,
SDL_WINDOWPOS_CENTERED, screenW, screenH, windowFlags);
and then, *IF* the user has configured it in fullscreen mode in its .ini, it tries to go fullscreen with the desktop mode:
SDL_SetWindowFullscreen(video.window, SDL_WINDOW_FULLSCREEN_DESKTOP);
This sequence of operations is currently failing because SDL_SetDisplayModeForDisplay() in SDL_video.c fails because display->desktop_mode is not being initialized with its correct value: SetDisplayMode() in SDL_kmsdrmvideo.c will not be able to set the mode because it detects the mode to have a driverdata of 0x0 ("if (!modedata)") and rightfully returns an error.
So, the included patch fixes this small problem, and programs that first change the video mode and then try to go fullscreen with the system video mode will now work.
The patch simply fixes an small omission, but its really needed now that dynamic video mode changing was implemented on the KMSDRM backend.
Manuel Alfayate Corchete
On the KMSDRM backend, there is no such thing as a desktop, yet some programs could (and DO) use SDL_GetGlobalMouseState().
So I think its good idea that, in KMSDRM, it returns the same mouse coordinates anyway as SDL_GetMouseState() would return. There is nothing else it could return, as far as I can understand, since there is no desktop anyway.
This small patch does precisely that.
Manuel Alfayate Corchete
The KMSDRM backend was doing things wrong because of some small (but important) misconceptions on how KMS/DRM works: to implement a largely broken non-vsync refresh mechanism, the SwapWindow() function was issuing new pageflips before previous ones had completed, thus causing EBUSY returns, buffer mismanagement, etc... resulting in general breakage on vsync disabling from apps, that would not allow vsync to work again without KMSDRM video re-initialization.
To further clarify, on most DRM drivers async pageflips are NOT working nowadays, so all issued pageflips will complete on next VBLANK, NOT ASAP (calling drmModePageFlip() with the DRM_MODE_PAGE_FLIP_ASYNC flag will return error).
The old code was assuming that can just issue a synchronous (=on VBLANK) pageflip and then pass a 0 timeout to the pull() function so we do not wait for the pageflip event, thinking that this will lead to correct non-vsynced screen updates from the program: That is plain wrong.
Each pageflip has to be waite before issuing a new one, ALWAYS. And if we do not support ASYNC pageflips on the DRM driver level, then we are forced to wait for the next VBLANK. There is no way around it.
I have also added many comments on the KMSDRM code. This is needed for future reference for me or others who may need to look at this code: KMS/DRM terminology regarding what SYNC and ASYNC mean in pageflip terms, and where to do certain things and why, is not trivial. It is not desirable or possible to invest time on researching the same concepts every time there is need to dive into this code. So please leave all these comments in the patch.
Build is broken without EGL since version 2.0.12 and
https://hg.libsdl.org/SDL/rev/72cc4740dec2:
/home/giuliobenetti/autobuild/run/instance-1/output-1/build/sdl2-2.0.12/src/video/kmsdrm/SDL_kmsdrmvideo.c: In function 'KMSDRM_CreateSurfaces':
/home/giuliobenetti/autobuild/run/instance-1/output-1/build/sdl2-2.0.12/src/video/kmsdrm/SDL_kmsdrmvideo.c:394:5: error: unknown type name 'EGLContext'
EGLContext egl_context;
^
Fixes:
- http://autobuild.buildroot.org/results/fafd20a01591032662f9ca025fcea3478239cf3c
Signed-off-by: Fabrice Fontaine <fontaine.fabrice@gmail.com>
Anthony Pesch
I was just communicating with one of the Retropie developers regarding this.
This change removed the forced window focus change on creation (3534cb3793) as part of the change no longer assumes there's only a single window being created. This was perhaps an over-aggressive removal.
Due to that change, joystick events are only received if SDL_SetKeyboardFocus is called explicitly, or if the app has specified SDL_HINT_JOYSTICK_ALLOW_BACKGROUND_EVENTS.
I think that part of my change should be reverted to continue setting mouse / keyboard focus to the window being created. If SDL_WINDOW_INPUT_FOCUS is to be used as an input flag the code could be conditional, but that would still leave existing software broken.
Charles Huber
This patch fixes the segfault on my Pi, though the valid display index range reported by the CHECK_DISPLAY_INDEX() macro in src/video/SDL_video.c is a little weird:
$ SDL_VIDEO_EGL_DRIVER=libEGL.so SDL_VIDEO_GL_DRIVER=libGLESv2.so ./a.out
SDL_Init(): displayIndex must be in the range 0 - -1
Malte Kie?ling
At the moment i get following warnings from kmsdrm:
* in SDL_kmsdrmvideo.c KMSDRM_DestroySurfaces is return type int, but thats never returned or checked against
* in SDL_kmsdrmvideo.c KMSDRM_DestroySurfaces the variable viddata is not used
* in SDL_kmsdrmopengles.c KMSDRM_GLES_LoadLibrary a cast to NativeDisplayType is missing
I attached a patch for them :)
Malte Kie?ling
Since https://hg.libsdl.org/SDL/rev/f908bd722523 / bug 4966 i cannot build SDL anymore. The error i get is, essentially, caused by -Werror=declaration-after-statement in SDL_kmsdrmvideo.c and SDL_kmsdrmopengles.c
Anthony Pesch
* Remove triple buffering support. As far as I can tell, this goes against the libdrm API; the EGL implementations themselves control the buffering. Removing it isn't absolutely necessary as it seemingly works on the Pi at least, but I noticed this while doing my work and explained my reasoning in the commit.
* Replace the crtc_ready logic which allocates an extra bo to perform the initial CRTC configuration (which is required before calling drmModePageFlip) with a call to drmModeSetCrtc after the front and back buffers are allocated, avoiding this allocation.
* Standardized the SDL_*Data variable names and null checks to improve readability. Given that there were duplicate fields in each SDL_*Data structure, having generic names such as "data" at times was very confusing.
* Removed unused fields from the SDL_*Data structures and moves all display related fields out of SDL_VideoData and into SDL_DisplayData. Not required since the code only supports a single display right now, but this was helpful in reading and understanding the code initially.
* Implement KMSDRM_GetDisplayModes / KMSDRM_SetDisplayMode to provide dynamic modeset support.
These changes have been tested on a Raspberry Pi 4 and a Dell XPS laptop with an HD 520.
As an update, I went back over the triple buffer changes and left them in. I didn't entirely get the code originally, I had just seen it calling KMSDRM_gbm_surface_lock_front_buffer twice for a single swap and had removed it because I was paranoid of bugs stemming from it while working on the modeset changes.
I've made a few small changes to the logic that had thrown me off originally and rebased the changes:
* The condition wrapping the call to release buffer was incorrect.
* The first call to KMSDRM_gbm_surface_lock_front_buffer has been removed. I don't understand why it existed.
* Added additional comments describing what was going on in the code (as it does fix the buffer release pattern of the original code before it).
For some obscure reason, the order in which the libdrm/libgbm libraries
are loaded matters.
Without this fix, the first call to check_modesetting() will work and
load then unload all symbols properly, but the second call to this
function will lock up as soon as dlopen() is called on libdrm.
Swapping the order in which the libdrm and libgbm libraries are loaded
is enough to fix (or work around?) this issue.
Fixes#4891:
https://bugzilla.libsdl.org/show_bug.cgi?id=4891
Signed-off-by: Paul Cercueil <paul@crapouillou.net>
If KMSDRM_drmModeSetCrtc is called when the swap interval is
set to 0, the driver behaves as though vertical sync is engaged by
limiting framerate to the refresh rate, but performance is much worse
than with vertical sync enabled.
Resolve this issue by ensuring that the Crtc is only set up once,
and KMSDRM_drmModePageFlip is called, albeit without any followup
queueing or waiting for flips.
