Segfault in fullscreen on Wayland with NVIDIA Vulkan WSI (hybrid GPU)

[pablopda]

mpv version and platform

• mpv v0.40.0
• libplacebo v7.349.0
• FFmpeg 7.1.1
• Ubuntu 25.10 (kernel 6.17.0-14-generic)
• GNOME Shell 49.0 / Mutter 49.0 (Wayland)
• NVIDIA driver 590.48.01 (proprietary, nvidia-dkms-590)
• Vulkan loader 1.4.321.0

Hardware

• Dell XPS 17 9700 (hybrid GPU, no MUX)
• iGPU: Intel CometLake-H GT2 [UHD Graphics] (PCI 00:02.0)
• dGPU: NVIDIA TU106M [GeForce RTX 2060 Max-Q] (PCI 01:00.0)
• Internal eDP panel (1920x1200 @ 60Hz), physically wired to Intel iGPU
• NVIDIA set as primary compositor GPU via udev rule (mutter-device-preferred-primary)

Description

mpv segfaults in the "vo" thread when entering Wayland fullscreen with Vulkan rendering on NVIDIA. The crash is inside libnvidia-glcore.so.590.48.01 during swapchain recreation triggered by xdg_toplevel.set_fullscreen.

Windowed mode works perfectly. Only true Wayland fullscreen crashes.

Reproducer

# Crashes — native Wayland Vulkan fullscreen:
mpv --fs --vo=gpu --gpu-api=vulkan \
    --vulkan-device="NVIDIA GeForce RTX 2060 with Max-Q Design" \
    any_video_file.mkv

# Works — windowed:
mpv --no-fs --vo=gpu --gpu-api=vulkan \
    --vulkan-device="NVIDIA GeForce RTX 2060 with Max-Q Design" \
    any_video_file.mkv

# Works — XWayland Vulkan context:
mpv --fs --vo=gpu --gpu-api=vulkan --gpu-context=x11vk \
    --vulkan-device="NVIDIA GeForce RTX 2060 with Max-Q Design" \
    any_video_file.mkv

# Works — maximized borderless (avoids xdg fullscreen):
mpv --no-fs --window-maximized=yes --no-border --vo=gpu --gpu-api=vulkan \
    --vulkan-device="NVIDIA GeForce RTX 2060 with Max-Q Design" \
    any_video_file.mkv

Key observations

• Not format-specific: crashes with any video file
• Not hwdec-specific: crashes with --hwdec=no (pure software decode)
• Not renderer-specific: crashes with both vo=gpu and vo=gpu-next
• Wayland WSI-specific: gpu-context=waylandvk crashes; gpu-context=x11vk works
• Fullscreen-specific: only xdg_toplevel.set_fullscreen triggers it; maximized does not
• OpenGL mode (--gpu-api=opengl) doesn't segfault but Mutter fails with: failed to import supplied dmabufs: Could not bind the given EGLImage to a CoglTexture2D

Stack trace

Thread 31 "vo" received signal SIGSEGV, Segmentation fault.
#0  0x00007fff19301132 in ??? () at /lib/x86_64-linux-gnu/libnvidia-glcore.so.590.48.01
#1  0x00007fff19301ea8 in ??? () at /lib/x86_64-linux-gnu/libnvidia-glcore.so.590.48.01
#2  0x00007fff19302a16 in ??? () at /lib/x86_64-linux-gnu/libnvidia-glcore.so.590.48.01
#3  0x00007ffff2f20635 in ??? () at /lib/x86_64-linux-gnu/libvulkan.so.1
#4  0x00007ffff39421d8 in ??? () at /lib/x86_64-linux-gnu/libplacebo.so.349
#5  0x00007ffff3943473 in ??? () at /lib/x86_64-linux-gnu/libplacebo.so.349
#6  0x0000555555759d44 in ??? ()  [mpv vo thread]
#7  0x00005555556dc81c in ??? ()  [mpv vo thread]
#8  0x00005555556e6030 in ??? ()  [mpv vo thread]
#9  0x00007ffff2ca3d64 in start_thread () at ./nptl/pthread_create.c:448

Registers at crash:
  rsi = 0xffffffffffffff  (corrupt size/offset value)

Analysis

The crash path is:

1. mpv enters fullscreen → Wayland compositor sends xdg_toplevel.configure with new dimensions
2. mpv's vo thread handles resize → libplacebo calls pl_swapchain_resize
3. libplacebo calls vkCreateSwapchainKHR to recreate swapchain
4. Inside libnvidia-glcore.so, the swapchain creation/teardown dereferences an invalid pointer, causing SIGSEGV

This appears to be the same underlying NVIDIA driver bug reported in:

• NVIDIA Forums: Vulkan fullscreen swapchain creation failure causes segfault on surface destruction
• Possibly related to libplacebo #200 (race in swapchain recreation)

Related mpv issues

• Fullscreen video always freezing in Wayland session, sound still playing (Fedora 40, Gnome 46, Nvidia GTX 1060) #14385 — Fullscreen video freezing in Wayland (NVIDIA)
• While in fullscreen and hovered by the mouse cursor the video hangs with a bunch of swapchain errors on the terminal #17239 — Fullscreen + swapchain errors (NVIDIA 580/590)

Workaround

Use XWayland Vulkan context:

# ~/.config/mpv/mpv.conf
gpu-context=x11vk

[CounterPillow]

thanks Claude but why report this nvidia driver bug with an outdated mpv version to mpv?

[pablopda]

Confirmed the crash also reproduces with mpv 0.41.0 (installed via Homebrew):

mpv v0.41.0
libplacebo v7.360.1
FFmpeg 8.0.1

Same segfault in libnvidia-glcore.so.590.48.01 with --gpu-context=waylandvk fullscreen. XWayland (--gpu-context=x11vk) still works fine.

I agree the root cause is in the NVIDIA driver — the crash is inside libnvidia-glcore.so during swapchain recreation. Filed on the NVIDIA forums as well. Reporting here for visibility to other mpv users on hybrid NVIDIA+Intel laptops hitting this on Wayland, since gpu-context=x11vk is an effective mpv-side workaround.

[llyyr]

could you link your nvidia forum post to the issue? thanks. nothing here is actionable on mpv side

[kasper93]

Does --vo=gpu-next work? Also try with libplacebo 7.360.1

[pablopda]

Good catch — tested with mpv 0.41.0 + libplacebo 7.360.1:

• --vo=gpu-next --gpu-context=waylandvk --fs → Works (plays 10+ seconds, no crash)
• --vo=gpu --gpu-context=waylandvk --fs → Segfaults (same crash in libnvidia-glcore.so)

So this is specific to the legacy vo=gpu renderer. gpu-next handles the fullscreen swapchain transition correctly.

[pablopda]

@llyyr NVIDIA forum post: https://forums.developer.nvidia.com/t/bug-590-48-01-segfault-in-libnvidia-glcore-so-during-vulkan-wayland-fullscreen-swapchain-transition-hybrid-gpu/363509

To clarify the findings: with mpv 0.41.0 + libplacebo 7.360.1, --vo=gpu-next does not crash in Wayland fullscreen. Only the legacy --vo=gpu renderer triggers the segfault. So gpu-next handles the swapchain recreation differently and avoids the NVIDIA driver bug.

The crash was originally found on mpv 0.40.0 where we initially tested with both vo=gpu and vo=gpu-next — both crashed there. On 0.41.0 with libplacebo 7.360.1, only vo=gpu crashes, suggesting the libplacebo update may have fixed the code path used by gpu-next.

[llyyr]

Can you also see what happens on mpv 0.40.0 with libplacebo 7.360.1? It'd be nice if you could bisect and see what fixes the issue in either mpv or libplacebo

[pablopda]

Unfortunately I can't easily test mpv 0.40.0 + libplacebo 7.360.1 without building from source — Ubuntu 25.10 ships mpv 0.40.0 linked against libplacebo 7.349.0, and my current install is mpv 0.41.0 (Homebrew) with libplacebo 7.360.1.

What I can confirm:

• mpv 0.40.0 + libplacebo 7.349.0 (Ubuntu package): vo=gpu-next crashed (same segfault)
• mpv 0.41.0 + libplacebo 7.360.1 (Homebrew): vo=gpu-next works, only vo=gpu crashes

So the fix is somewhere between libplacebo 7.349.0→7.360.1 or mpv 0.40.0→0.41.0 (or both). Given that gpu-next relies heavily on libplacebo for its Vulkan swapchain handling, and libplacebo #200 addressed a race condition in swapchain recreation, the libplacebo update is the more likely fix. But I haven't bisected to confirm.

[pablopda]

@llyyr I was able to test mpv 0.40.0 + libplacebo 7.360.1 (built mpv 0.40.0 from Homebrew's extracted formula, linking against the current libplacebo 7.360.1):

mpv	libplacebo	vo=gpu-next --fs	vo=gpu --fs
0.40.0	7.349.0 (Ubuntu)	Segfault	Segfault
0.40.0	7.360.1 (Homebrew)	Segfault	Segfault
0.41.0	7.360.1 (Homebrew)	Works	Segfault

The fix for gpu-next is in mpv 0.41.0, not in libplacebo. mpv 0.40.0 crashes with gpu-next even with the latest libplacebo.

The vo=gpu crash persists across all combinations — that one is likely an NVIDIA driver issue that neither mpv nor libplacebo can work around.

[CounterPillow]

This is not a bisect.

[pablopda]

@CounterPillow thanks, you are been really helpful.

[pablopda]

Correction: After more thorough testing, --vo=gpu-next also crashes consistently (5/5 runs segfault). My earlier result showing it working was not reproducible — likely a race condition in the driver that occasionally doesn't trigger.

Updated results:

Config	Result
--vo=gpu --gpu-context=waylandvk --fs	Segfault (always)
--vo=gpu-next --gpu-context=waylandvk --fs	Segfault (always)
--vo=gpu --gpu-context=x11vk --fs	Works
--vo=gpu-next --gpu-context=x11vk --fs	Works
--vo=gpu-next --gpu-context=waylandvk --no-fs	Works
--vo=gpu-next maximized borderless	Works

Both renderers crash — the bug is entirely in the NVIDIA Vulkan Wayland WSI (libnvidia-glcore.so.590.48.01), triggered by xdg_toplevel.set_fullscreen. A bisect of mpv won't find anything actionable.

The only workaround is --gpu-context=x11vk or avoiding true fullscreen.

Sorry for the earlier incorrect data. NVIDIA forum post: https://forums.developer.nvidia.com/t/363509

[llyyr]

After more thorough testing, --vo=gpu-next also crashes consistently

With libplacebo 7.360.1 or 7.349.0?

[pablopda]

Both. Tested --vo=gpu-next --gpu-context=waylandvk --fs:

mpv	libplacebo	Result
0.40.0	7.349.0 (Ubuntu apt)	Segfault
0.40.0	7.360.1 (built via Homebrew)	Segfault
0.41.0	7.360.1 (Homebrew)	Segfault

All crash consistently. The segfault is inside libnvidia-glcore.so, not in mpv or libplacebo code.

[llyyr]

Can you try with direct-scanout disabled on Mutter? I don't know the env var to do that, but MUTTER_DEBUG_DISABLE_HW_CURSORS=1 should also disable scanout.

Set that env variable before launching Mutter, then launch mpv and see if you get it to crash on fullscreen