[Bug 221694] [regression v7.0->v7.1] AMD Strix Point (Ryzen AI 9 HX 370) intermittently hangs on resume from s2idle

[bugzilla-daemon@kernel.org]

https://bugzilla.kernel.org/show_bug.cgi?id=221694

Hongri Cui (hongri.cui@gmail.com) changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|---                         |CODE_FIX

--- Comment #4 from Hongri Cui (hongri.cui@gmail.com) ---
CONFIRMED FIXED in 7.1.6. Thanks Mario for pointing at ISM, and Peter for the
7.1.y backport.

The fix landed in stable 7.1.6 as commit 95776812e6b8 ("drm/amd/display: Fix
ISM
dc_lock deadlock during suspend", [ Upstream commit 3714fe242592 ]). It is not
in
7.1.4 or 7.1.5 -- I checked those ChangeLogs when they were released and stayed
on
a pre-ISM kernel until 7.1.6 appeared.

Test: dropped my 6.18 workaround pin, booted 7.1.6, and ran 30 rtcwake s2idle
cycles (30 s asleep, 15 s awake between cycles):

  resumed 30/30 | no-s0i3 0 | counted-fail 0
  last_hw_sleep us  min=29514099  max=30090332  mean=29935818  (n=30)

Every cycle reached deep s0i3 with essentially the full sleep duration in
hardware. The whole run stayed inside one boot -- no hard reset -- and pstore
picked up no new entries.

For contrast, the same harness on unfixed kernels:

  7.1.1  wedged on cycle 5/30
  7.1.3  wedged on cycle 28/30   (cleared 27 clean cycles first)
  6.18.x clean 30/30             (predates ISM 754003486c3c)
  7.1.6  clean 30/30             (ISM present + fix present)

The 7.1.3 result is why I ran a long loop rather than a handful of cycles: the
deadlock needs an ISM worker in flight at suspend time, so it can hide behind
many consecutive good cycles. 30 cycles is not a proof of absence, but 7.1.6
matching the pre-ISM baseline exactly -- on a kernel that does have ISM -- is
consistent with the race being closed rather than just rarer.

This also retires the amd_pmc line of investigation from comment 1: with
amd_pmc
unbound the box still wedged, which was the opposite of bug 221664 and pointed
away from the PMC/SMU handshake. That now makes sense -- it was the display
driver all along, and the clean PMC mailbox traffic in that dyndbg trace
(attachment 310381) was a red herring.

Hardware, for the record: AMD Ryzen AI 9 HX 370 (Strix Point) + Radeon 890M,
s2idle-only (/sys/power/mem_sleep = [s2idle]), Micro Computer (HK) Tech Limited
"AI Series", board F8BAC, AMI BIOS 1.02 (2025-03-03), unchanged throughout --
so
no firmware update is involved in this result. Running 7.1.6 unpinned since.

#regzbot fixed-by: 3714fe242592

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

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