Re: [PATCH v2 08/17] drm/msm/adreno: Coredump on GPU/GMU init failures

[Akhil P Oommen <akhilpo@oss.qualcomm.com>]

On 3/31/2026 1:21 AM, Rob Clark wrote:
> On Thu, Mar 26, 2026 at 5:15 PM Akhil P Oommen <akhilpo@oss.qualcomm.com> wrote:
>>
>> Capture coredump on GPU or GMU errors during initialization to help in
>> debugging the issues. To be consistent with the locks while calling
>> msm_gpu_crashstate_capture(), call pm_runtime_get(gpu) always with
>> msm_gpu->lock.
>>
>> Signed-off-by: Akhil P Oommen <akhilpo@oss.qualcomm.com>
>> ---
>>  drivers/gpu/drm/msm/adreno/a6xx_gmu.c      | 1 +
>>  drivers/gpu/drm/msm/adreno/adreno_device.c | 5 +++--
>>  drivers/gpu/drm/msm/adreno/adreno_gpu.c    | 5 ++++-
>>  drivers/gpu/drm/msm/msm_gpu.c              | 5 +++--
>>  drivers/gpu/drm/msm/msm_gpu.h              | 2 ++
>>  5 files changed, 13 insertions(+), 5 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
>> index 1b44b9e21ad8..916c5d99c4d1 100644
>> --- a/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
>> +++ b/drivers/gpu/drm/msm/adreno/a6xx_gmu.c
>> @@ -1335,6 +1335,7 @@ int a6xx_gmu_resume(struct a6xx_gpu *a6xx_gpu)
>>
>>  disable_irq:
>>         disable_irq(gmu->gmu_irq);
>> +       msm_gpu_crashstate_capture(gpu, NULL, NULL, NULL, NULL);
>>         a6xx_rpmh_stop(gmu);
>>  disable_clk:
>>         clk_bulk_disable_unprepare(gmu->nr_clocks, gmu->clocks);
>> diff --git a/drivers/gpu/drm/msm/adreno/adreno_device.c b/drivers/gpu/drm/msm/adreno/adreno_device.c
>> index 4edfe80c5be7..ca5f96e16870 100644
>> --- a/drivers/gpu/drm/msm/adreno/adreno_device.c
>> +++ b/drivers/gpu/drm/msm/adreno/adreno_device.c
>> @@ -105,6 +105,8 @@ struct msm_gpu *adreno_load_gpu(struct drm_device *dev)
>>          */
>>         pm_runtime_enable(&pdev->dev);
>>
>> +       guard(mutex)(&gpu->lock);
> 
> so, I am a fan of guard(), but I realized this means we are holding
> gpu->lock across the debugfs_init() (only a5xx).. which I suspect will
> anger lockdep, although I don't have a good way to test that atm.

Shall we drop this patch for now? I will send a respin later.

-Akhil.

> 
> BR,
> -R
> 
>> +
>>         ret = pm_runtime_get_sync(&pdev->dev);
>>         if (ret < 0) {
>>                 pm_runtime_put_noidle(&pdev->dev);
>> @@ -112,10 +114,9 @@ struct msm_gpu *adreno_load_gpu(struct drm_device *dev)
>>                 goto err_disable_rpm;
>>         }
>>
>> -       mutex_lock(&gpu->lock);
>>         ret = msm_gpu_hw_init(gpu);
>> -       mutex_unlock(&gpu->lock);
>>         if (ret) {
>> +               msm_gpu_crashstate_capture(gpu, NULL, NULL, NULL, NULL);
>>                 DRM_DEV_ERROR(dev->dev, "gpu hw init failed: %d\n", ret);
>>                 goto err_put_rpm;
>>         }
>> diff --git a/drivers/gpu/drm/msm/adreno/adreno_gpu.c b/drivers/gpu/drm/msm/adreno/adreno_gpu.c
>> index 785e99fb5bd5..8475802fdde2 100644
>> --- a/drivers/gpu/drm/msm/adreno/adreno_gpu.c
>> +++ b/drivers/gpu/drm/msm/adreno/adreno_gpu.c
>> @@ -391,10 +391,13 @@ int adreno_get_param(struct msm_gpu *gpu, struct msm_context *ctx,
>>                 return 0;
>>         case MSM_PARAM_TIMESTAMP:
>>                 if (adreno_gpu->funcs->get_timestamp) {
>> +                       mutex_lock(&gpu->lock);
>>                         pm_runtime_get_sync(&gpu->pdev->dev);
>> +
>>                         *value = adreno_gpu->funcs->get_timestamp(gpu);
>> -                       pm_runtime_put_autosuspend(&gpu->pdev->dev);
>>
>> +                       pm_runtime_put_autosuspend(&gpu->pdev->dev);
>> +                       mutex_unlock(&gpu->lock);
>>                         return 0;
>>                 }
>>                 return -EINVAL;
>> diff --git a/drivers/gpu/drm/msm/msm_gpu.c b/drivers/gpu/drm/msm/msm_gpu.c
>> index 995549d0bbbc..472db2c916f9 100644
>> --- a/drivers/gpu/drm/msm/msm_gpu.c
>> +++ b/drivers/gpu/drm/msm/msm_gpu.c
>> @@ -361,7 +361,7 @@ static void crashstate_get_vm_logs(struct msm_gpu_state *state, struct msm_gem_v
>>         mutex_unlock(&vm->mmu_lock);
>>  }
>>
>> -static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
>> +void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
>>                 struct msm_gem_submit *submit, struct msm_gpu_fault_info *fault_info,
>>                 char *comm, char *cmd)
>>  {
>> @@ -886,7 +886,8 @@ void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
>>
>>         pm_runtime_get_sync(&gpu->pdev->dev);
>>
>> -       msm_gpu_hw_init(gpu);
>> +       if (msm_gpu_hw_init(gpu))
>> +               msm_gpu_crashstate_capture(gpu, NULL, NULL, NULL, NULL);
>>
>>         submit->seqno = submit->hw_fence->seqno;
>>
>> diff --git a/drivers/gpu/drm/msm/msm_gpu.h b/drivers/gpu/drm/msm/msm_gpu.h
>> index 666cf499b7ec..eb5b3a7b81f9 100644
>> --- a/drivers/gpu/drm/msm/msm_gpu.h
>> +++ b/drivers/gpu/drm/msm/msm_gpu.h
>> @@ -747,6 +747,8 @@ static inline void msm_gpu_crashstate_put(struct msm_gpu *gpu)
>>  }
>>
>>  void msm_gpu_fault_crashstate_capture(struct msm_gpu *gpu, struct msm_gpu_fault_info *fault_info);
>> +void msm_gpu_crashstate_capture(struct msm_gpu *gpu, struct msm_gem_submit *submit,
>> +               struct msm_gpu_fault_info *fault_info, char *comm, char *cmd);
>>
>>  /*
>>   * Simple macro to semi-cleanly add the MAP_PRIV flag for targets that can
>>
>> --
>> 2.51.0
>>