Thanks - this is very helpful, and I can confirm your Datapoint 1 from the
original reporting hardware. I also need to correct my earlier comment, which
presented config C more favourably than my own later data supports.

CORRECTION: config C does not prevent the deadlock here either

When I wrote up config C I had six clean cycles over 24h and said I could not
claim it prevented the deadlock, only that the GPU fault had not occurred. That
caution turned out to be warranted: it deadlocked twice afterwards, on 
2026-09-15
and 2026-09-18, with the freeze override still in place.

The 2026-09-15 occurrence closely matches your occurrence #6 - freeze verifiably
active, the same 60s user.slice timeout, and the deadlock forming at suspend
entry rather than on resume:

  18:06:01.987  systemd-sleep: Failed to freeze unit 'user.slice': Connection 
timed out
  18:06:01.998  PM: suspend entry (deep)
  18:06:22.034  Freezing user space processes failed after 20.002 seconds (1 
tasks refusing to freeze, wq_busy=0)
  18:06:22.034  fbcon: Taking over console
  18:10:19.450  INFO: task kworker/1:0:11661 blocked for more than 122 seconds.
  18:10:19.453  INFO: task systemd-sleep:58419 blocked on a semaphore likely 
last held by task kworker/1:0:11661
  18:10:19.454  INFO: task setfont:58433 blocked on a semaphore likely last 
held by task kworker/1:0:11661

Last journal entry is 18:12:22 ("Future hung task reports are suppressed"), no
PM: suspend exit - hard power-off required. setfont queued behind the same
kworker, as in your report.

Same kworker path as the original report:

  Workqueue: events fbcon_register_existing_fbs
   rwsem_down_read_slowpath+0x261/0x5a0
   down_read+0x48/0xd0
   nvkms_ioctl_from_kapi+0xdc/0xf0 [nvidia_modeset]
   GetDynamicDisplayInfo+0x9c/0x190 [nvidia_modeset]
   ? __nv_drm_connector_detect_internal+0x1a2/0x350 [nvidia_drm]
   ? nv_drm_connector_detect+0xe/0x20 [nvidia_drm]

Environment for that occurrence: RTX 2070 SUPER (TU104), Ryzen 7 2700X, Gigabyte
X470 AORUS ULTRA GAMING, Ubuntu 26.04.1, GNOME Wayland, S3 (deep), NVIDIA
610.57.04 open, kernel 7.0.0-31-generic.

SUPPORTS DATAPOINT 2: no TLB failure preceded it

Worth noting for your "left held, no visible writer" reading: that entire boot
logged zero kgmmuInvalidateTlb_GM107 failures. The deadlock formed with no TLB
timeout anywhere beforehand.

So on this hardware the TLB invalidation failure is not a precondition. It
appears in some occurrences (my 2026-09-18 event had 8) and is absent in others,
while the kworker/console_lock/nvkms-rwsem cycle is present in all of them. That
is consistent with the rwsem being left held around the FBSR window rather than 
a
TLB timeout being the trigger.

COUNTER-DATAPOINT: this stopped here on kernel 7.0.0-34 and later

Since moving past 7.0.0-31 I have had no recurrence:

  10.09.-18.09.  kernel 7.0.0-31            39 suspends    2 failures
  since 24.09.   kernel 7.0.0-34 -> -43     73 suspends    0 failures

Across those 73 cycles: zero TLB invalidation failures, zero kernel freeze
failures, zero hung task reports, zero nvidia-drm errors on resume. That 
includes
NVIDIA 615.71.09 open since 25.10. and kernel 7.0.0-43-generic since 07.11. -
the same driver and kernel you are seeing six deadlocks on.

Caveat on how much that is worth: by mid-September my failure rate had already
fallen to roughly 1 in 20 suspends, so 73 clean cycles would occur by chance
about 2% of the time. Suggestive, not conclusive, and config C is still active
here, so I cannot cleanly separate kernel version from configuration.

Still, since we now differ mainly in GPU generation and platform (Turing on AMD
X470 vs Blackwell on Intel Z890) while sharing driver, kernel, distro and
compositor, the difference may be more on the hardware or platform side than in
the software stack.

ON TESTING 620.70.06

Not easily from here - the Ubuntu 26.04 archive currently tops out at
nvidia-driver-615-open (615.71.09-0ubuntu0.26.04.2); there is no 620 package. It
would need NVIDIA's .run installer, which I would rather not do on a working
machine. If anyone has 620 packaged for 26.04, I am happy to test it on Turing
and report back.

Happy to run specific tests or provide fuller logs for any of the above.

-- 
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