[Bug 221694] New: [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

            Bug ID: 221694
           Summary: [regression v7.0->v7.1] AMD Strix Point (Ryzen AI 9 HX
                    370) intermittently hangs on resume from s2idle
           Product: Drivers
           Version: 2.5
    Kernel Version: 7.1.1
          Hardware: All
                OS: Linux
            Status: NEW
          Severity: high
          Priority: P3
         Component: Platform_x86
          Assignee: drivers_platform_x86@kernel-bugs.osdl.org
          Reporter: hongri.cui@gmail.com
                CC: mario.limonciello@amd.com, shyam-sundar.s-k@amd.com
        Regression: Yes

Created attachment 310380
  --> https://bugzilla.kernel.org/attachment.cgi?id=310380&action=edit
amd-s2idle 0.2.18 report (7.1.1) — all prereqs pass, 7 deep-s0i3 cycles then
silent wedge

#regzbot introduced: v7.0..v7.1
#regzbot title: amd s2idle: Strix Point HX 370 silently hangs on resume from
s2idle

Summary
-------
On an AMD Ryzen AI 9 HX 370 (Strix Point) laptop whose only sleep state is
s2idle, resume from suspend-to-idle intermittently never completes (~10-40%
of cycles, no correlation to sleep duration). The machine wedges silently and
needs a hard power-off. 7.0.x is reliable; the failure appeared with 7.1 and
still reproduces on 7.1.1, so this is a 7.0 -> 7.1 regression, not one bad
point release. No crash: nothing in pstore; the journal simply ends at
"PM: suspend entry (s2idle)" because userspace is frozen and the box never
resumes to flush the ring buffer.

Hardware
--------
- AMD Ryzen AI 9 HX 370 w/ Radeon 890M (Strix Point, family 1ah)
- GPU: AMD Strix [Radeon 880M/890M] 1002:150e rev c1 (gfx1150)
- Wi-Fi: MediaTek MT7925 (14c3:7925, mt7925e)
- PMC: amd_pmc AMDI000A:00, Program 11, SMU firmware 93.4.0
- Board: Minisforum whitebox F8BAC, AMI BIOS 1.02 (2025-03-03, latest
available)
- /sys/power/mem_sleep = [s2idle] only (no S3 fallback)

Kernel versions
---------------
Good: 7.0.13  |  Bad: 7.1.0 and 7.1.1  (vanilla mainline-stable, NixOS,
unpatched)

Reproducer
----------
Root rtcwake loop driving real s2idle, with pm_debug_messages on:
    for i in $(seq 1 30); do
        echo "cycle $i PRE-SUSPEND" | tee /dev/kmsg; sync
        rtcwake -m mem -s 30
        sleep 15
    done
30 cycles reliably trips at least one wedge on 7.1.x; 7.0.13 completes 30/30.
AMD's own amd-s2idle (amd-debug-tools 0.2.18) reproduces it identically.

Observed
--------
A/B, identical 30-cycle stress:
- 7.0.13: resumed 30/30, full deep s0i3 every cycle (last_hw_sleep mean
~29.5s).
- 7.1.1: clean for several cycles (full residency) then a silent wedge; the
last
  line persisted for the whole boot is the pre-suspend breadcrumb. Last good
  cycle showed normal "PM: suspend-to-idle" -> "amd_pmc: SMU idlemask s0i3:
  0x7fff1afd" -> "PM: suspend exit".

Ruled out
---------
- Not a crash: no oops/panic in efi_pstore across the reset.
- Not a misconfiguration: amd-s2idle prerequisite checks ALL pass (System
  configured for s2idle, ACPI FADT supports Low-power S0 idle, LPS0 _DSM
  enabled, amd_pmc loaded, amdgpu bound, IOMMU OK, PC6/CC6 enabled). See
attached.
- Not a wake-storm / shallow sleep: successful cycles reach deep s0i3 with full
  residency; zero amd_pmc "didn't reach deepest state" warnings, no
AER/MCE/lockup.
- Below the Wi-Fi driver: unloading mt7925e across suspend did not stop the
wedge
  (and caused a separate page_pool_release_retry stall, reverted). MT7925 is
not
  armed as a wake source.
- Not the ISP4 i2c-designware-amdisp resume race: no camera / no amd_isp4 / no
  isp_4_1_0.bin, and that race was fixed before 7.1.

Signature points at a silent PMC/SMU s0i3 entry<->exit handshake.

Not a duplicate of bug 221664 (same iGPU family): that box never reaches s0ix
(last_hw_sleep=0) due to a broken DSDT \_SB.PEP._DSM and fails on 7.0 already.
This box reaches full s0i3 (~76% residency), has a healthy LPS0 _DSM (no
AE_NOT_FOUND), and is a clean 7.0→7.1 regression. cf. 221451 (GNOME/EC race)
and 220812 (drm/amd VPE commit 31ab31433c9b, fixed pre-7.0) — neither applies
here.

Can provide
-----------
git bisect v7.0..v7.1 (slow due to intermittency but doable), additional
amd-s2idle runs, serial/netconsole + no_console_suspend capture of the actual
hang point, and full per-cycle logs from both kernels. Please advise which is
most useful first.

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

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