criu with VirtualGL

[phil294]

General GPU dumping is not supported because it is highly driver-dependent (#384). Could it be possible to do so using virtualized OpenGL? https://virtualgl.org/ (OS)

The error I am getting is Error (criu/proc_parse.c:549): Can't handle non-regular mapping on 19509's map 7fef8da4c000. Is this intended? Is there a way to get it working?

Reproduce:

• Steps explained at https://criu.org/VNC (awesome!!)
• Xvnc also needs the option -depth 24 passed
• Install VirtualGL, start program with vglrun. For demonstration purposes: vglrun glxheads
• Dump as specified (--tcp-established) -> error

[avagin]

I don't know how vglrun works, but it doesn't allow us to avoid /drm mappings, criu doesn't support them right now.

[avagin@laptop linux-task-diag]$ ps -C glxgears
  PID TTY          TIME CMD
  544 pts/7    00:01:08 glxgears
[avagin@laptop linux-task-diag]$ ls -l /proc/544/map_files/
...
lrw-------. 1 avagin avagin 64 Jan 17 09:22 7fc41b732000-7fc41b73a000 -> /drm mm object (deleted)
lrw-------. 1 avagin avagin 64 Jan 17 09:22 7fc41b762000-7fc41b7ba000 -> /SYSV00000000 (deleted)
lrw-------. 1 avagin avagin 64 Jan 17 09:22 7fc41b7ba000-7fc41b7ca000 -> /drm mm object (deleted)
lrw-------. 1 avagin avagin 64 Jan 17 09:22 7fc41b7ca000-7fc41b8ca000 -> /drm mm object (deleted)
lrw-------. 1 avagin avagin 64 Jan 17 09:22 7fc41b8fe000-7fc41b8ff000 -> /drm mm object (deleted)
lrw-------. 1 avagin avagin 64 Jan 17 09:22 7fc41b8ff000-7fc41b900000 -> /drm mm object (deleted)
lrw-------. 1 avagin avagin 64 Jan 17 09:22 7fc41b900000-7fc41b901000 -> /memfd:xshmfence (deleted)
lrw-------. 1 avagin avagin 64 Jan 17 09:22 7fc41b901000-7fc41b902000 -> /memfd:xshmfence (deleted)
lrw-------. 1 avagin avagin 64 Jan 17 09:22 7fc41b903000-7fc41b904000 -> /drm mm object (deleted)
lrw-------. 1 avagin avagin 64 Jan 17 09:22 7fc41b904000-7fc41b90c000 -> /drm mm object (deleted)
lrw-------. 1 avagin avagin 64 Jan 17 09:22 7fc41b90c000-7fc41b92d000 -> /SYSV00000000 (deleted)

[phil294]

@avagin ok thanks. I am not seeing any /drm mappings, but some /dev/nvidiactl and /dev/nvidia0. The issue is probably the same?
Any idea on how to workaround this? Criu+VNC is doable, OpenGL/VirtualGL is not.

I assume that the possibility of checkpointing these kind of mappings would result in various new usecases.

(related: https://arxiv.org/abs/1312.6650 someone once made a plugin for DMTCP that was apparently able to achieve this. couldnt find any code base)

[github-actions]

A friendly reminder that this issue had no activity for 30 days.