[CRASH] OpenGL acceleration except gtk: bad interaction between NVIDIA usermode opengl libraries and QEMU seccomp -sandbox on,spawn=deny, crashes immediately on startup with Bad system call

Host environment

• Operating system: Arch Linux

• OS/kernel version: Linux hostname 6.7.0-arch3-1 #1 SMP PREEMPT_DYNAMIC Sat, 13 Jan 2024 14:37:14 +0000 x86_64 GNU/Linux

• Architecture: x86_64

• QEMU flavor: system-x86_64? all qemu-system-* seem to have this issue

• QEMU version: 8.2.0

• QEMU command line:

  qemu-system-x86_64 -spice gl=on -sandbox on,spawn=deny
  qemu-system-x86_64 -display egl-headless,rendernode=/dev/dri/renderD128 -sandbox on,spawn=deny

Emulated/Virtualized environment

• Operating system: crashes before loaded
• OS/kernel version: N/A
• Architecture: all? aarch64, alpha, i386, riscv64 are affected

Description of problem

When running any of the above command lines, QEMU crashes with Bad system call (core dumped). Not exclusive to spice; it seems this is caused by QEMU forking during OpenGL initialization after seccomp takes effect.

Steps to reproduce

1. Run the above commandline
2. Notice a Bad system call (core dumped)

Additional information

This crash only happens if spawn=deny is set, resourcecontrol/obsolete/elevateprivileges don't cause crashes.

The crash happens around the same time as an audit event is generated in dmesg: audit: type=1326 audit(1705775880.776:14): auid=MYUSERID uid=MYUID gid=MYGID ses=REDACTED pid=REDACTED comm="qemu-system-x86" exe="/usr/bin/qemu-system-x86_64" sig=31 arch=c000003e syscall=56 compat=0 ip=REDACTED code=REDACTED

ausyscall c000003e 56 tells me it's clone which (iirc) is the syscall used by glibc to implement fork() (I might be wrong about glibc part)

Suggested solution: move seccomp activation until just before guest code starts executing? make frontends (ie -display gtk/sdl/whatever, including -spice) initialize before seccomp?

Workaround: chmod -x /bin/nvidia-modprobe if not using the NVIDIA gpu or use this wrapper script (untested, not enterprise-ready, I am not responsible if unexpected things happen):

• rename /bin/qemu-system-x86_64 to qemu-system-x86_64.real
• put this in /bin/qemu-system-x86_64 and chmod +x it

#!/usr/bin/env sh
chmod -x /bin/nvidia-modprobe
qemu-system-x86_64.real $@ & disown
sleep 10 # excessive but maybe safer?
chmod +x /bin/nvidia-modprobe

Also, you can use -display gtk,gl=on instead, or (unknown security implications) remove spawn=deny from -sandbox args

original bug report was libvirt/libvirt#585 (closed) but I realized this was more of a qemu issue than a libvirt one