[SRU][S][PATCH 0/1] AMD system hangs on resume when a Thunderbolt dock is plugged in during s2idle

[AceLan Kao]

From: "Chia-Lin Kao (AceLan)" <acelan.kao at canonical.com>

BugLink: https://bugs.launchpad.net/bugs/2168944

[Impact]
On Dell systems with an AMD APU, plugging in a Thunderbolt dock while the system is in s2idle wakes it up early. Then the system can't resume. The power LED blinks an EC power-rail failure code (3 amber, 5 white). The only way out is a forced power off.

The EC reports SLP_S3#, PLTRST# and RUNPWROK as abnormal. The board-level cause is SoC leakage on the 0.75V rail. That trips over-voltage protection (OVP) and the rail powers off.

The amd_pmc driver waits 2.5 seconds before a new HW sleep cycle to avoid this. The wait is skipped when the dock wakes the system very briefly.

Failure rate: 100% on affected hardware. Seen on 6.17 OEM kernels. It does not show up with the amd_pmc driver blacklisted.

[Fix]
The driver used the metrics table (s0i3_last_entry_status) to detect an intermediate wakeup. The table does not update for every fast wakeup, so the 2.5 second wait was sometimes skipped.

The patch drops the metrics table check. It uses the driver's own is_first_check_after_suspend state instead.

Upstream commit, in linux-next (not in a released kernel tag yet):
9d45679a02f5e platform/x86/amd: Stop counting intermediate cycles with metrics table

Fixes: 4dbd11796f3a8 ("platform/x86/amd: pmc: Clear metrics table at start of cycle"), v6.16.

It needs 037f0b03c663a ("platform/x86/amd/pmc: Don't log during intermediate wakeups", v7.2), which added is_first_check_after_suspend. That commit is already in this tree as e1a1c7606406b.

Patch link:
https://patch.msgid.link/20260914141734.2703545-1-mario.limonciello@amd.com

[Test Plan]
On a Dell system with an AMD APU and a Thunderbolt dock:
1. Boot to the desktop.
2. Suspend the system:
   $ sudo systemctl suspend
3. Plug in the dock while suspended.
4. Unplug the dock, then suspend again.
5. Plug in the dock again.

Without the patch: the system fails to resume and the LED shows 3A5W. A forced power off is needed.
With the patch: the system wakes up on its own after the dock is plugged in, on both cycles. No hang, no LED error.

[Where problems could occur]
This could break s2idle suspend and resume on AMD systems that use the amd_pmc driver.

The wait now depends on is_first_check_after_suspend. If that flag is wrong, the wait could be skipped and the hang comes back. Or the wait could run when it isn't needed. Then suspend takes 2.5 seconds longer after a brief wakeup from USB, Thunderbolt or another device.

The change only removes the metrics table read from the suspend path. It adds no new hardware calls.

Mario Limonciello (1):
  platform/x86/amd: Stop counting intermediate cycles with metrics table

 drivers/platform/x86/amd/pmc/pmc.c | 15 +--------------
 1 file changed, 1 insertion(+), 14 deletions(-)

-- 
2.53.0