[PATCH v4 2/2] drm/amdgpu: add reset register dump trace on GPU reset

[Andrey Grodzovsky]

On 2022-02-16 05:46, Somalapuram, Amaranath wrote:
>
> On 2/15/2022 10:09 PM, Andrey Grodzovsky wrote:
>>
>> On 2022-02-15 05:12, Somalapuram Amaranath wrote:
>>> Dump the list of register values to trace event on GPU reset.
>>>
>>> Signed-off-by: Somalapuram Amaranath <Amaranath.Somalapuram at amd.com>
>>> ---
>>>   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 17 ++++++++++++++++-
>>>   drivers/gpu/drm/amd/amdgpu/amdgpu_trace.h  | 16 ++++++++++++++++
>>>   2 files changed, 32 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c 
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>>> index 1e651b959141..ff21262c6fea 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>>> @@ -4534,6 +4534,19 @@ int amdgpu_device_pre_asic_reset(struct 
>>> amdgpu_device *adev,
>>>       return r;
>>>   }
>>>   +static int amdgpu_reset_reg_dumps(struct amdgpu_device *adev)
>>> +{
>>> +    uint32_t reg_value;
>>> +    int i;
>>> +
>>> +    for (i = 0; i < adev->n_regs; i++) {
>>> +        reg_value = RREG32(adev->reset_dump_reg_list[i]);
>>> + trace_amdgpu_reset_reg_dumps(adev->reset_dump_reg_list[i], 
>>> reg_value);
>>> +    }
>>> +
>>> +    return 0;
>>> +}
>>> +
>>>   int amdgpu_do_asic_reset(struct list_head *device_list_handle,
>>>                struct amdgpu_reset_context *reset_context)
>>>   {
>>> @@ -4567,8 +4580,10 @@ int amdgpu_do_asic_reset(struct list_head 
>>> *device_list_handle,
>>>                   tmp_adev->gmc.xgmi.pending_reset = false;
>>>                   if (!queue_work(system_unbound_wq, 
>>> &tmp_adev->xgmi_reset_work))
>>>                       r = -EALREADY;
>>> -            } else
>>> +            } else {
>>> +                amdgpu_reset_reg_dumps(tmp_adev);
>>>                   r = amdgpu_asic_reset(tmp_adev);
>>> +            }
>>
>>
>> Is there any particular reason you only dump registers in single ASIC 
>> case and not for XGMI ?
>>
>> Andrey
>>
> Not really, should I move it to the top of function?
>
> Regards,
>
> S.Amarnath


Yes, no reason to avoid dumping this info for XGMI case.

Consider also that maybe you don't want to print this when
the reset is not due to any error state - like manual GPU reset
from debugfs - but to know this you will have to wire a flag from
high above in the call stack to here and so probably this not worth
all this changes.

Andrey


>
>>
>>>                 if (r) {
>>>                   dev_err(tmp_adev->dev, "ASIC reset failed with 
>>> error, %d for drm dev, %s",
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_trace.h 
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_trace.h
>>> index d855cb53c7e0..b9637925e85c 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_trace.h
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_trace.h
>>> @@ -537,6 +537,22 @@ TRACE_EVENT(amdgpu_ib_pipe_sync,
>>>                 __entry->seqno)
>>>   );
>>>   +TRACE_EVENT(amdgpu_reset_reg_dumps,
>>> +        TP_PROTO(uint32_t address, uint32_t value),
>>> +        TP_ARGS(address, value),
>>> +        TP_STRUCT__entry(
>>> +                 __field(uint32_t, address)
>>> +                 __field(uint32_t, value)
>>> +                 ),
>>> +        TP_fast_assign(
>>> +               __entry->address = address;
>>> +               __entry->value = value;
>>> +               ),
>>> +        TP_printk("amdgpu register dump 0x%x: 0x%x",
>>> +              __entry->address,
>>> +              __entry->value)
>>> +);
>>> +
>>>   #undef AMDGPU_JOB_GET_TIMELINE_NAME
>>>   #endif