[16:38, 14/10/2026] syslog-oct14-wedge.txt Posting because the kernel told me to 🙂 — Please file a _new_ bug report at https://gitlab.freedesktop.org/drm/xe/kernel/issues/new — and because early Battlemage server users seem thin on the ground. Two documented incidents, same signature, with a reliable way to tell this apart from container misconfiguration. System - Unraid 7.3.0, kernel 6.18.29-Unraid - Intel Arc Pro B50 (Battlemage G21, 8086:e212, 16GB), Kernel driver in use: xe - AMD Ryzen 9 5950X, 128GB RAM - GPU passthrough to Emby Docker container (emby/embyserver:beta, /dev/dri passed, QSV/VAAPI hardware transcoding) - Firmware after last boot: GuC xe/bmg_guc_70.bin v70.65.0, HuC xe/bmg_huc.bin v8.2.10, DMC i915/bmg_dmc.bin v2.6 Incident A —(syslog-oct14-wedge.txt), wedge ~1 min after recreating the Emby container Recreated the emby/embyserver:beta container with hardware transcoding enabled. Within ~1 minute: Oct 14 12:27:13 kernel: xe 0000:2d:00.0: [drm] Tile0: GT0: G2H fence (6935) not found! Oct 14 12:27:14 kernel: xe 0000:2d:00.0: [drm] *ERROR* Tile0: GT0: Timed out wait for G2H, fence 7002, action 3003, done no Oct 14 12:27:14 kernel: xe 0000:2d:00.0: [drm] *ERROR* Tile0: GT0: GuC PC query task state failed: -ETIME Oct 14 12:27:14 kernel: xe 0000:2d:00.0: [drm] Xe device coredump has been created Oct 14 12:27:14 kernel: xe 0000:2d:00.0: [drm] Tile0: GT0: trying reset from guc_exec_queue_timedout_job [xe] Oct 14 12:27:15 kernel: xe 0000:2d:00.0: [drm] *ERROR* TLB invalidation fence timeout, seqno=320657 recv=320655 Oct 14 12:27:16 kernel: xe 0000:2d:00.0: [drm] *ERROR* Tile0: GT0: failed to enable GuC opt-in features: -ETIME Oct 14 12:27:16 kernel: xe 0000:2d:00.0: [drm] *ERROR* Tile0: GT0: reset failed (-ETIME) Oct 14 12:27:16 kernel: xe 0000:2d:00.0: [drm] *ERROR* CRITICAL: Xe has declared device 0000:2d:00.0 as wedged. Oct 14 12:27:16 kernel: IOCTLs and executions are blocked. Only a rebind may clear the failure Oct 14 12:27:16 kernel: Please file a _new_ bug report at https://gitlab.freedesktop.org/drm/xe/kernel/issues/new Oct 14 12:27:16 kernel: xe 0000:2d:00.0: [drm] device wedged, needs recovery Followed by 9x Timedout job: seqno=... guc_id=... flags=0x4 in no process [-1] / VM job timed out on non-killed execqueue, each with: WARNING: CPU: 15 PID: 4014753 at drivers/gpu/drm/xe/xe_guc_submit.c:1388 guc_exec_queue_timedout_job+0x79c/0x980 [xe] RIP: 0010:guc_exec_queue_timedout_job+0x79c/0x980 [xe] Repeated reset attempts all failed with -ETIME. The devcoredump was auto-deleted ~1h later (Xe device coredump has been deleted, 13:28:03) before I could capture it — lesson learned. Host rebooted 15:47. The trap: it looks like a container misconfig, not a GPU wedge After the wedge, from inside the container: - vainfo → VA-API error 18 (invalid parameter) / vaGetDriverNames failed - DRM_IOCTL_VERSION → Operation canceled - QSV encoder → segfault - Meanwhile /sys/class/drm/card0/device/power/runtime_status still reports active If you debug from the container side only, this reads exactly like a broken container/media-driver setup. It isn't — the device is wedged. Discriminator Compare timestamps of guc_exec_queue_timedout_job lines against the container's start time. If the timeouts begin right at/after a container (re)start with hw acceleration, it's this wedge — not your container config. Confirmed twice now. Incident B — 2026-09-23, same signature (log storm) [16:38, 14/10/2026]: Same card/driver. 113+ Tile0: GT0: trying reset from guc_exec_queue_timedout_job [xe] lines between 14:34:20 and 14:40:26 (then continuing), spamming /var/log until the tmpfs filled. Full Unraid diagnostics attached (bmc-diagnostics-20260923-1441.zip, logs/syslog.txt). Recovery Host reboot is the only thing that has cleared it (three occurrences total: Jun — PMT/PL2 message spam; Sep 23; Oct 14). After the Oct 14 reboot: vainfo clean in the container, full profile list, and I benchmarked 12 concurrent 1080p HEVC QSV encode sessions all sustaining >1x realtime — so the card itself is completely healthy when the driver doesn't wedge. Questions 1. Known issue for BMG (Battlemage) on xe 6.18? Anything I can provide alongside the devcoredump next time? 2. How do I preserve the devcoredump before the auto-delete — is 10 min the window, or did the ~1h deletion only happen because the queue kept resetting? 3. Any workaround shorter than a full reboot (GPU reset via sysfs didn't exist in the moment — the wedge blocked it)? syslog-oct14-wedge.txt bmc-diagnostics-20260923-1441.zip