Re: [PATCH RFC 09/10] arm64: dts: qcom: hamoa: Enable LLCC/DDR/DDR_QOS DVFS

From: Birk Skyum

Date: Tue Nov 10 2026 - 22:08:00 EST


Hi Pragnesh, Sibi,

I tried this on a Lenovo Yoga Slim 7x (X1E80100, 83ED, firmware
NHCN62WW) running 7.3 with patches 1-8 of this series and the
protocol@80 node from this patch added to the board's DTB.

Memlat probes, and the ddr, llcc and ddr-qos devfreq devices come up
under the remote governor. CPU read bandwidth goes from about 51 GB/s
to 117 GB/s with six threads, and suspend and resume work.

Under GPU load it isn't stable here, though. I ran Speedometer 3.1 in
Chrome twice with memlat active:

- The first time the run finished, but after a warm reboot into a
kernel whose DTB lacks protocol@80, the GPU never came up:

adreno_gmu 3d6a000.gmu: [drm:a6xx_hfi_wait_for_msg_interrupt [msm]] *ERROR* Message HFI_H2F_MSG_GX_BW_PERF_VOTE id 34 timed out waiting for response
adreno_gmu 3d6a000.gmu: [drm:a6xx_gmu_set_oob [msm]] *ERROR* Timeout waiting for GMU OOB set GPU_SET: 0x4

A cold boot of the same kernel was fine.

- The second time the machine reset partway through the run, with
nothing in pstore.

Four runs without the GPU load (a CPU bandwidth test, with and without
a suspend), each followed by the same warm reboot, were clean, as were
about ten boots without memlat.

Separately, the memlat driver sends MEMLAT_STOP_TIMER only from its
.remove callback, and the SCMI bus has no shutdown hook, so on a
reboot the monitors keep running on the CPUCP into whatever boots
next. That would fit the first failure if the GMU's GX bandwidth votes
and the CPUCP's DDR votes get in each other's way. Would it make sense
to stop the timers from a reboot notifier as well?

I'm happy to test a follow-up on this machine.

Thanks,
Birk