Thanks — this is genuinely useful, and your config-C correction closes a
loop on our side as well.

**Config C is now ruled out on both hardware sets.** Your 2026-09-15
occurrence is essentially a mirror of our occurrence #6 (config:
`UseKernelSuspendNotifiers=0`, `PreserveVideoMemoryAllocations=1`,
session freeze re-enabled): deadlock at **suspend entry**, freeze
verifiably active, the same ~60 s `user.slice` freeze timeout, and
`setfont` queued behind the same `fbcon_register_existing_fbs` kworker.
Combined with your two September events, every configuration lever
either of us has tried is now exhausted on both machines — which
retroactively validates the caution you attached to your original
config-C numbers.

**On the kernel -34 → -43 counter-datapoint — our wedge set cuts the
other way:**

| kernel | our occurrences |
|---|---|
| 7.0.0-42 | 2 |
| 7.0.0-43 | 4 (including both 2026-11-11 events) |

All six on 615.71.09 open. When -43 first shipped here we explicitly
treated it as a natural experiment (same driver code, new kernel): the
race recurred on the first -43 overnight resume, with a full hung-task
trace, and then three more times. So whatever changed between -31 and
-34 does not clear the race on this platform — which supports your own
closing hypothesis that the difference between our datasets is more on
the hardware/platform side (Blackwell GB205 + Intel Z890 here vs Turing
TU104 + AMD X470 there) than in the software stack. Your caveats are
also fair: at a 1-in-20 background rate, 73 clean cycles is suggestive
rather than conclusive.

One extra unconfound from our side: we wedge on -43 in **every**
configuration tried — freeze on and off, notifiers on and off,
preservation forced on — so on this machine, kernel version and
configuration shift the race's timing and entry point around but never
remove it.

**TLB precondition — agreed from both sides now.** Our occurrences also
include both shapes: TLB retry storms preceding the wedge, and clean
logs with zero TLB failures. The invariant in every single one is the
kworker/console_lock/nvkms-rwsem cycle itself, consistent with the rwsem
being held across the FBSR window rather than any TLB event being the
trigger.

**On testing 620.70.06 — the packaging situation changed this week.**
`nvidia-driver-620-open 620.70.06-0ubuntu0.26.04.1` has landed in
`resolute-proposed` on the Ubuntu 26.04 archive (pin priority 100, so it
installs only on explicit request). Note the proposed build is DKMS-
based (`nvidia-dkms-620-open`), unlike the precompiled 615 modules — it
needs kernel headers, and if Secure Boot is enabled the module must be
MOK-enrolled. Since you offered to test 620 on Turing:

    sudo apt install -t resolute-proposed nvidia-driver-620-open

One caution before committing a working machine to it: the 620.70.06
feedback thread on the NVIDIA forums is accumulating **resume-correlated
display faults** — flip-event timeouts on two heads on an RTX 5080 (post
#57), and today an RTX 2070 SUPER user (post #66) tallying across 7
boots that flip-event faults track suspend/resume count and *not* GPU
load. None of those are the fbcon/rwsem signature, but they suggest
620's resume path has its own unresolved issues. A clean 620 run on your
Turing box would be a genuinely valuable datapoint either way.

For our part, auto-suspend is disabled on this machine until the race is
understood — the cost of a wedged resume here is a dead unattended
machine — so we cannot rapidly generate fresh cycles. Full SysRq-T
captures from all six occurrences are available if any of them would
help, and we'll report back if anything moves on 615.71.09 / -43.

-- 
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2166914

Title:
  [regression 7.0.0-30 -> 7.0.0-31] Suspend deadlocks in
  pm_restore_console: fbcon takeover holds console_lock while blocked on
  nvidia-modeset rwsem

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/linux/+bug/2166914/+subscriptions


-- 
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs

Reply via email to