[v3d][KMS/DRM][Raspberry Pi 5] drmModePageFlip() fails when DRM_MODE_PAGE_FLIP_ASYNC flag is set.

[bluestang2006]

Describe the bug

Linking a previously corrupted gfx issue when using OpenGL & SDL2 - https://gitlab.freedesktop.org/mesa/mesa/-/issues/10242

Calling drmModePageFlip() with the DRM_MODE_PAGE_FLIP_ASYNC flag fails on the Raspberry Pi 5.
This is a major problem since it breaks SDL2 on KMS/DRM on the platform.
We detect the DRM_CAP_ASYNC_PAGE_FLIP cap here:
https://github.com/libsdl-org/SDL/blob/058213366bccf7f62d4c4c6061477d7a69ea336d/src/video/kmsdrm/SDL_kmsdrmvideo.c#L957
...where it's detected as available on both Pi 4 and Pi5...
And then we set the flag here:
https://github.com/libsdl-org/SDL/blob/058213366bccf7f62d4c4c6061477d7a69ea336d/src/video/kmsdrm/SDL_kmsdrmopengles.c#L176
And call drmModePageFlip here with the DRM_CAP_ASYNC_PAGE_FLIP flag here:
https://github.com/libsdl-org/SDL/blob/058213366bccf7f62d4c4c6061477d7a69ea336d/src/video/kmsdrm/SDL_kmsdrmopengles.c#L179
[@apinheiro](https://github.com/apinheiro) Can you please take a look? There are many projects using SDL2 on the Pi4/5.

Steps to reproduce the behaviour

This can easily be duplicated with Quake3e using the opengl1 renderer. Just compile the source code, switch over to the TTY and run the game with the ogl renderer, which is the default one I believe.

NOTE: This is only happening when vsync is off. It does not seem to happen when vsync is on...

Device (s)

Other

System

System:
  Host: raspberrypi Kernel: 6.1.0-rpi7-rpi-2712 arch: aarch64 bits: 64
    compiler: gcc v: 12.2.0 Desktop: LXDE v: 0.10.1 wm: Openbox dm: LightDM
    Distro: Debian GNU/Linux 12 (bookworm)
CPU:
  Info: quad core model: N/A variant: cortex-a76 bits: 64 type: MCP
    arch: ARMv8 rev: 1
  Speed (MHz): avg: 2400 min/max: 1000/2400 cores: 1: 2400 2: 2400 3: 2400
    4: 2400 bogomips: 432
  Features: Use -f option to see features
Graphics:
  Device-1: bcm2712-hdmi0 driver: vc4_hdmi v: N/A bus-ID: N/A
    chip-ID: brcm:107c701400
  Device-2: bcm2712-hdmi1 driver: vc4_hdmi v: N/A bus-ID: N/A
    chip-ID: brcm:107c706400
  Display: x11 server: X.Org v: 1.21.1.7 with: Xwayland v: 22.1.9
    compositor: xcompmgr driver: X: loaded: modesetting unloaded: fbdev dri: vc4
    gpu: vc4-drm,vc4_crtc,vc4_dpi,vc4_dsi,vc4_firmware_kms,vc4_hdmi,vc4_hvs,vc4_txp,vc4_v3d,vc4_vec
    display-ID: :0 screens: 1
  Screen-1: 0 s-res: 1920x1080 s-dpi: 96
  Monitor-1: HDMI-A-1 mapped: HDMI-1 model: Gigabyte M27Q res: 1920x1080
    dpi: 82 diag: 684mm (26.9")
  API: OpenGL v: 3.1 Mesa 23.2.1-1~bpo12+rpt2 renderer: V3D 7.1
    direct-render: Yes

Logs

No response

Additional context

The solution was posted by one of the Igalia devs:

This issue was introduced in [5fd6ee](https://github.com/raspberrypi/linux/commit/5fd6ee7fd084838e09d4e463ae53cd9aaa7fce70) and if you revert it, you will be able to fix the problem. 
For example, in the 6.7 kernel, this commit was already reverted [9ef6db](https://github.com/raspberrypi/linux/commit/9ef6dbb755b6445888418aa7212fe085543fbbc5).
I'll try to let RPi developers know about this info. Thanks for sharing the issue!

[6by9]

We're already discussing it directly with Igalia.

6.7 had reverted the iommu patches due to a kernel API change that hasn't been resolved as yet. Doing so means all image buffers have to come from the CMA heap which will then fragment, so that is a regression.

#5827 is the correct fix - the driver wasn't updating the display list correctly for Pi5, so ended up writing an incorrect address as the image pointer.

[bluestang2006]

We're already discussing it directly with Igalia.

6.7 had reverted the iommu patches due to a kernel API change that hasn't been resolved as yet. Doing so means all image buffers have to come from the CMA heap which will then fragment, so that is a regression.

#5827 is the correct fix - the driver wasn't updating the display list correctly for Pi5, so ended up writing an incorrect address as the image pointer.

thank you! I assume this will be back-ported into 6.1? Feel free to close the issue as necessary.

[6by9]

Fix and backport have been merged.

[6by9]

Closing as the fix has been merged.