It was observed on sc7180 (A618 gpu) that GPU votes for GX rail and CNOC BCM nodes were not removed after GPU suspend. This was because we skipped sending 'prepare-slumber' request to gmu during suspend sequence in some cases. So, make sure we always call prepare-slumber hfi during suspend. Also, calling prepare-slumber without a prior oob-gpu handshake messes up gmu firmware's internal state. So, do that when required.
Fixes: 4b565ca5a2cb ("drm/msm: Add A6XX device support") Cc: stable@vger.kernel.org Signed-off-by: Akhil P Oommen quic_akhilpo@quicinc.com --- Changes in v2: - Minor update to commit text and CC'ed Stable - Link to v1: https://lore.kernel.org/r/20250226-adreno-sys-suspend-fix-v1-1-054261bba114@... --- drivers/gpu/drm/msm/adreno/a6xx_gmu.c | 72 +++++++++++++++++++---------------- 1 file changed, 39 insertions(+), 33 deletions(-)
diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c index 699b0dd34b18f0ec811e975779ba95991d485098..38c94915d4c9d6d33354502651a77c1f9e4648df 100644 --- a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c +++ b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c @@ -1169,49 +1169,50 @@ static void a6xx_gmu_shutdown(struct a6xx_gmu *gmu) struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; u32 val; + int ret;
/* - * The GMU may still be in slumber unless the GPU started so check and - * skip putting it back into slumber if so + * GMU firmware's internal power state gets messed up if we send "prepare_slumber" hfi when + * oob_gpu handshake wasn't done after the last wake up. So do a dummy handshake here when + * required */ - val = gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE); + if (adreno_gpu->base.needs_hw_init) { + if (a6xx_gmu_set_oob(&a6xx_gpu->gmu, GMU_OOB_GPU_SET)) + goto force_off;
- if (val != 0xf) { - int ret = a6xx_gmu_wait_for_idle(gmu); + a6xx_gmu_clear_oob(&a6xx_gpu->gmu, GMU_OOB_GPU_SET); + }
- /* If the GMU isn't responding assume it is hung */ - if (ret) { - a6xx_gmu_force_off(gmu); - return; - } + ret = a6xx_gmu_wait_for_idle(gmu);
- a6xx_bus_clear_pending_transactions(adreno_gpu, a6xx_gpu->hung); + /* If the GMU isn't responding assume it is hung */ + if (ret) + goto force_off;
- /* tell the GMU we want to slumber */ - ret = a6xx_gmu_notify_slumber(gmu); - if (ret) { - a6xx_gmu_force_off(gmu); - return; - } + a6xx_bus_clear_pending_transactions(adreno_gpu, a6xx_gpu->hung);
- ret = gmu_poll_timeout(gmu, - REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS, val, - !(val & A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB), - 100, 10000); + /* tell the GMU we want to slumber */ + ret = a6xx_gmu_notify_slumber(gmu); + if (ret) + goto force_off;
- /* - * Let the user know we failed to slumber but don't worry too - * much because we are powering down anyway - */ + ret = gmu_poll_timeout(gmu, + REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS, val, + !(val & A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB), + 100, 10000);
- if (ret) - DRM_DEV_ERROR(gmu->dev, - "Unable to slumber GMU: status = 0%x/0%x\n", - gmu_read(gmu, - REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS), - gmu_read(gmu, - REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS2)); - } + /* + * Let the user know we failed to slumber but don't worry too + * much because we are powering down anyway + */ + + if (ret) + DRM_DEV_ERROR(gmu->dev, + "Unable to slumber GMU: status = 0%x/0%x\n", + gmu_read(gmu, + REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS), + gmu_read(gmu, + REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS2));
/* Turn off HFI */ a6xx_hfi_stop(gmu); @@ -1221,6 +1222,11 @@ static void a6xx_gmu_shutdown(struct a6xx_gmu *gmu)
/* Tell RPMh to power off the GPU */ a6xx_rpmh_stop(gmu); + + return; + +force_off: + a6xx_gmu_force_off(gmu); }
--- base-commit: 72d0af4accd965dc32f504440d74d0a4d18bf781 change-id: 20250110-adreno-sys-suspend-fix-c5bc7beea0c4
Best regards,