mpv crash with v4l2 camera output #17215

Activity

1. ndren

   added

   os:linux

   on Jan 1, 2026
2. ndren

   added the

   Bug

   issue type on Jan 1, 2026
3. kasper93

   added

   down-upstream:ffmpeg

   on Aug 9, 2026
4. kasper93 commented on Aug 9, 2026

   

   kasper93

   on Aug 9, 2026

   Member

   More actions

   For reference: https://code.ffmpeg.org/FFmpeg/FFmpeg/issues/21358, https://code.ffmpeg.org/FFmpeg/FFmpeg/issues/21621

   Not a mpv bug, need upstream fix to not orphan AVPackets.
5. jeffscottward commented on Sep 22, 2026

   

   jeffscottward

   on Sep 22, 2026

   More actions

   Possibly related V4L2 shutdown crash on ARM64/Asahi

   Historical diagnostics from September 21, 2026 may show another manifestation of the V4L2 packet-lifetime issue linked above. The immediate crash site differs, so I have not established that it is the same bug.

   A webcam overlay crashed during controller-requested shutdown (SIGTERM) in three runs: twice with gpu-next/Asahi Vulkan, and once with gpu-next after adding --gpu-api=opengl. All three coredumps report SIGSEGV / SEGV_MAPERR. The OpenGL run logged:

   VO: [gpu-next] 640x720 nv12
   [ffmpeg/demuxer] video4linux2,v4l2: Some buffers are still owned by the caller on close.
   Exiting... (Quit)
   

   Historical launch command, with only the IPC socket pathname normalized:

   mpv av://v4l2:/dev/video0 \
     --profile=low-latency --untimed --no-cache \
     --demuxer-lavf-o=video_size=1280x720,framerate=30 \
     '--vf=lavfi=[crop=ih*8/9:ih]' \
     --title=WebcamOverlay --wayland-app-id=WebcamOverlay-medium \
     --no-border --no-audio --no-osc --osd-level=0 \
     --input-ipc-server="$XDG_RUNTIME_DIR/webcam.sock"

   For the OpenGL comparison, --gpu-api=opengl was added. No --no-config reproduction was performed.

   Both Vulkan stacks start with:

   libvulkan_asahi.so + 0x302d6c
   libvulkan_asahi.so + 0x12bbc8
   libplacebo.so.360 + 0xa2834
   pl_tex_upload   (libplacebo.so.360 + 0x712b0)
   pl_upload_plane (libplacebo.so.360 + 0x83454)
   mpv + 0x15f158
   

   The OpenGL stack starts with:

   memcpy (libc.so.6 + 0x9ddd8)
   libgallium-26.2.3-arch1.1.so + 0x10dcdc4
   libplacebo.so.360 + 0x8e474
   libplacebo.so.360 + 0x74298
   pl_tex_upload   (libplacebo.so.360 + 0x712b0)
   pl_upload_plane (libplacebo.so.360 + 0x83454)
   mpv + 0x15f158
   

   Both backend paths continue through libplacebo's frame queue and pl_queue_update. Unnamed frames were unresolved in systemd's traces; stack location alone does not establish a libplacebo/Mesa defect.

   Environment: Arch Linux ARM/Asahi, aarch64. Package history and the current matching versions show mpv 1:0.41.0-6, FFmpeg 2:9.0.1-4, libplacebo 7.360.1-6, mesa and vulkan-asahi 1:26.2.3-1. No intervening updates to those packages are recorded after the incident. mpv --version reports v0.41.0, libplacebo v7.360.1 and FFmpeg n9.0.1.

   Changing the overlay renderer to --vo=gpu --gpu-api=opengl avoided new coredumps in two subsequent recording/stop cycles, including a 17.856-second recording that decoded successfully. This is a limited workaround observation, not a demonstrated root-cause fix.

   These are recovered historical diagnostics, not a fresh latest/master reproduction. A full --gpu-debug crash log and fully symbolized backtrace are unavailable in the recovered evidence. I am adding the ARM/Asahi, shutdown and cross-backend observations as possibly related evidence for the existing FFmpeg investigation, not asserting a confirmed duplicate or requesting a separate GPU-driver bug.

   Filed by GPT-6 Astra via Oh My Pi.
6. kasper93 commented on Sep 22, 2026

   

   kasper93

   on Sep 22, 2026

   Member

   More actions

   The root cause is already documented and reported upstream, see my comment above. No need to reiterate it again. Feel free to propose fixes upstream for this issue.