Chasing down a complete host freeze

Proxuserad

New Member
Nov 16, 2026
3
0
1
Hello everyone,

I'm trying to troubleshoot an issue with my setup. I'm on the latest 9.2.26 and up to date on all updates as of today, on Kernel Linux 7.0.14-28. I have an ASRock Intel Arc A380 that I believe was fine until I ran into a system freeze issue. My setup is as follows: I have a Supermicro X10SRL-F motherboard running an Ubuntu Server no-GUI VM, also running on Linux kernel 7.0+ and ubuntu 24.04, with a Docker farm that hosts a lot of containers; two of them use the passthrough GPU for transcoding: the Jellyfin container and Immich. I had no issues, or so I thought, until I started to transcode my whole Immich library. The system would just hang; the whole host is hanging, not just the VM, which led me down this rabbit hole of troubleshooting. I've tried so many things, so I decided to pull the trigger and buy a new GPU (exactly the same model to isolate the testing). GPU came in, and it's basically the same problem, I'm now debating if it could be the board but I really doubt it. The symptoms are the system only hangs if I transcode heavily and/or simultaneously, which seems to be some type of starvation issue or hand off issue as immich does cpu decoding and the gpu does the transcoding, i tried on the original gpu to do decoding as well but it would freeze. I'm not sure if the kernel's might be pointint to an issue, I guess i would have to go to an older kernel and try it, but at this point I'm running out of ideas.
 
Since the whole host dies, try to capture logs off the box first: serial console or netconsole, plus pstore/ramoops if available. Then limit Immich and Jellyfin to one concurrent transcode and test. Also test CPU-only transcoding with the Arc removed from the VM; if that is stable, focus on passthrough/reset behavior, ASPM, and above-4G/IOMMU settings. Swapping in the same GPU model does not fully rule out host-side PCIe or power issues under heavy simultaneous load.
 
Ok huge update, I think I'm narrowing down the issue. It seems to only happen with QSV in immich specifically. Made a lot of changes but I think it has to do with how immich does QSV (see below). I patched my bios with ReBAR using the ReBarUEFI Mod (https://github.com/xCuri0/ReBarUEFI). I now have 8G showing for the GPU. Had to manually change all my setting in BIOS back ASPM disabled, Above 4G encoding enabled etc (same as before no changes here). Updated kernels to the latest possible (doubt much difference here just went up one revision). I then run a test with local files on the host and it froze again, but then I recall not having an issue with Jellyfin, and I decided to do a test. I transcoded two sessions in jellyfin of several heavy 4k movies (4-6) simultaneously without an issue for 1 hour, I actually had to keep changing movies as it would finish the transcode during those sessions. Then, I did a second 45 minute session a different day with the same result. (Jellyfin on QSV mode). I then changed immich to VAAPI and transcoded more than 1000+ videos with zero problems, then I decided to swap my old GPU card back and same result no issues.

I now believe there is something with QSV setting in immich. As the local host tests were done with the same exact command and ffmpeg version immich uses with the QSV command (I you add the flag for verbose in immich, you can find the exact command it uses for transcoding in immich). I'm going to do more tests, vut so far thankfully, it seems its something with immich QSV. I also found an open tracked issue in github with what seems to be the same issue: https://github.com/immich-app/immich/issues/30083