[Bug 219492] amdgpu: failed to write reg 28b4 wait reg 28c6 and amdgpu: failed to write reg 1a6f4 wait reg 1a706

[bugzilla-daemon@kernel.org]

https://bugzilla.kernel.org/show_bug.cgi?id=219492

Denis Pisarev (pisarevden+bugzilla@gmail.com) changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |pisarevden+bugzilla@gmail.c
                   |                            |om

--- Comment #5 from Denis Pisarev (pisarevden+bugzilla@gmail.com) ---
"failed to write reg 28b4 wait reg 28c6" persists for hours after every S4
(hibernate) resume, Cezanne, kernel 7.1.8

Adding another reproduction on a different setup (desktop use, no ROCm),
with two observations that go beyond the original report:

1. The failures are NOT limited to early resume. They continue
   indefinitely (80-140 per hour for 9+ hours) until reboot.
2. On this machine there are no "TLB flush failed for PASID" lines at
   all - only the two reg write/wait pairs.

HARDWARE / ENVIRONMENT
----------------------
CPU:     AMD Ryzen 7 PRO 5850U
GPU:     AMD Cezanne iGPU [1002:1638] (rev d1), amdgpu, DMUB 0x0101002B
Kernel:  7.1.8-1-MANJARO
Desktop: KDE Plasma / KWin Wayland, no ROCm or compute workloads
Trigger: resume from hibernation (S4), 100% reproducible over multiple
         cycles across two days; does not occur on a fresh boot

LOG EXCERPT (resume at Aug 19 10:52:39)
---------------------------------------
Aug 19 10:52:39 horsehead kernel: Freezing remaining freezable tasks completed
(elapsed 0.001 seconds)
Aug 19 10:52:39 horsehead kernel: amdgpu 0000:07:00.0: failed to write reg 28b4
wait reg 28c6
Aug 19 10:52:39 horsehead kernel: amdgpu 0000:07:00.0: failed to write reg
1a6f4 wait reg 1a706
Aug 19 10:52:39 horsehead kernel: amdgpu 0000:07:00.0: failed to write reg 28b4
wait reg 28c6
Aug 19 10:52:39 horsehead kernel: amdgpu 0000:07:00.0: failed to write reg
1a6f4 wait reg 1a706
Aug 19 10:52:39 horsehead systemd-sleep[1062034]: System returned from sleep
operation 'hibernate'.

The first failures appear in the same second as the thaw, interleaved
with the hibernation-exit messages. On an earlier cycle the resume also
produced:

Aug 18 10:04:42 horsehead kernel: amdgpu 0000:07:00.0: MODE2 reset
Aug 18 10:04:42 horsehead kernel: usb usb5: root hub lost power or was reset
(... usb6, usb7, usb1 ...)

ONGOING RATE AFTER RESUME (Aug 19, single cycle)
------------------------------------------------
hour 10 (resume 10:52): 138 errors
hour 11:               140 errors
hour 12:                78 errors (still ongoing at time of writing)
Previous cycle (Aug 18, resumed 10:04): 1414 errors over the day,
still occurring 9+ hours later. No "TLB flush failed for PASID"
messages appear on this machine, only the 28b4/28c6 and 1a6f4/1a706
pairs, always together.

USER-VISIBLE IMPACT
-------------------
The retry path makes the whole desktop sluggish (stalls in the
compositor and apps). CPU and memory are idle; the disk is not involved
(NVMe <2% busy, fsync latency normal), yet PSI "some" io pressure reads
~80% because tasks spend their time sleeping in the retry loop.

WORKAROUND
----------
Reboot clears it until the next S4 resume. Using suspend-to-RAM avoids
triggering it (S4 is the only trigger observed).

NOTE
---------------
the resume_gpu_stable proposal from January
(https://lore.kernel.org/amd-gfx/20260106125929.25214-6-sunlightlinux@gmail.com/)
was rejected because it disabled the KIQ path permanently, but the
persistent-hours-long failure pattern seen here suggests KIQ does not
recover on its own after S4 resume on gmc_v9 + Cezanne - at least not
under a plain Wayland desktop workload.

-- 
You may reply to this email to add a comment.

You are receiving this mail because:
You are watching the assignee of the bug.