[Jetson AGX Thor] devfreq_wq deadlock with rmapiLockAcquire causes GPU tasks to hang in D state ( JetPack 7.2)

nvidia-support-192.168.3.209-20260910-172345.tar.gz (3.4 MB)

NVIDIA Jetson kernel issue support bundle

Collected from jetson agx thor (hostname: lzc-pod-rwZVvx) on 2026-09-10, approximately 17:24–17:27 UTC+08:00.

Platform

  • NVIDIA Jetson AGX Thor

  • Ubuntu 24.04.4 LTS

  • Kernel: 6.8.12-1021-tegra, #1 SMP PREEMPT Tue Jul 21 15:58:47 CST 2026

  • NVIDIA L4T: R39, revision 2.0 (GCID 45755727)

Observed issue

The reports a kernel problem. Exact reproduction steps and the visible application symptoms were not provided.

Collected evidence shows repeated hung-task warnings involving kworker/u28:0, nv_queue, nvidia-modeset and other processes. The current boot includes warnings at 16:54:14 through 17:00:16 (UTC+08:00), with blocked durations up to 483 seconds. The preceding boot also contains hung-task warnings.

Live process stacks show nvidia-smi in uninterruptible D state, waiting in os_acquire_rwlock_read → portSyncRwLockAcquireRead → rmapiLockAcquire inside the NVIDIA kernel module. Several nvidia-smi processes were already blocked before collection. This identifies an observed wait location, not a confirmed root cause.

The kernel logs also contain repeated nvethernet xpcs_write_safety failures. Their relationship to the GPU/kernel blocking is unknown.

I noticed that it says here that similar issues have already been fixed in JetPack7.2, but I am still experiencing similar phenomena in JetPack7.2:
https://forums.developer.nvidia.com/t/jetson-thor-r38-4-jp7-1-devfreq-rmapilockacquire-deadlock-wedges-all-gpu-tasks-in-d-state/372833/6#main-container#1

Files to inspect first

  • nvidia-bug-report-safe.log.gz: official NVIDIA GPU report, successfully completed with --safe-mode (exit 0).

  • nvidia-bug-report-jetson.log: official Jetson report with platform details, retained system journal, dmesg and kernel configuration. Collection was interrupted after it invoked the standard GPU report, which blocked in nvidia-smi --version. It therefore lacks its final completion marker.

  • kernel/highlights.txt: selected warning/error lines with surrounding kernel context, derived from the per-boot logs; originals remain included.

  • system/blocked-task-stacks.txt: live process list and /proc//stack + status for D-state processes.

  • kernel/boot-6d6a9d630639478583170ec6bc724422.log.gz: current boot kernel log.

  • kernel/boot-756ee3b97d034aa8ad059b117ec6b7c3.log.gz: preceding boot kernel log.

  • system/boots.txt: retained boot IDs and time ranges.

  • system/journal-boot-0.log.gz and system/journal-boot–1.log.gz: full system journals for the current and preceding boot.

  • kernel/dmesg-*.txt: current kernel ring buffer, captured before and after collection.

  • system/packages.txt, kernel-module-details.txt, identity.txt: package, driver and OS versions.

  • persistent/: available rotated kernel/boot logs and kernel/boot configuration.

  • collection-status.txt, collection-notes.txt and *console.txt: collection status and limitations.

  • SHA256SUMS: checksums of all other files in this directory.

Collection limitations

  • Standard nvidia-smi and nvpmodel read-only queries blocked in kernel waits (os_acquire_rwlock_read and min_freq_show respectively). Their collection process groups were stopped; exit 137 is recorded. D-state tasks may remain until the underlying kernel wait resolves.

  • The standard GPU report started by the Jetson script also blocked. Its small output is preserved as nvidia-bug-report.partial.log.gz; use the completed safe-mode report for GPU diagnostics.

  • The boot list has 9 boot IDs. The oldest boot (2026-09-02, ID 0dc45e4d84a840b28157a7f1d2b78c72) has no retained kernel entries; its kernel export contains “-- No entries --”.

  • An initial parser also treated the boot-list header as an ID, generating an empty boot-BOOT.log.gz. That empty derived file was removed locally; its failed attempt remains in collection-status.txt for transparency.

  • /sys/fs/pstore and /var/lib/systemd/pstore were empty: no persistent kernel crash dump was available.

  • /var/crash contained an application crash report for ChatGPT. Only the crash-file listing was collected; that application memory dump is not included.

  • Some display/PCI queries reported unavailable data; see the official collector console logs.

  • Remote-to-local archive SHA-256 matched: 59c803be85471ed3e16a707190cf7dfbbae0692884b18b2dd083f18063354053. The final bundle additionally contains this README, derived highlights and checksums.

  • No reboot, kernel/driver changes or service restart was performed as part of collection.

Hi,

Please try the below patch to see if it helps:

Thanks.

Thank you for your reply, but I have found a problem: after loading the nvgpu.ko with this patch added, jtop and the Jetson Power GUI are unable to read the GPU usage rate. How should this situation be handled?

Hi,

Please apply all the steps in the kernel customization document:

Thanks