Re: [PATCH V2 2/3] dmaselftest: add memcpy selftest support functions

[Sinan Kaya <okaya@codeaurora.org>]

On 11/3/2015 2:44 AM, Dan Williams wrote:
> On Mon, Nov 2, 2015 at 10:30 PM, Vinod Koul <vinod.koul@intel.com> wrote:
>> On Mon, Nov 02, 2015 at 11:18:37PM -0500, Sinan Kaya wrote:
>>>
>>>
>>> On 11/2/2015 11:15 PM, Vinod Koul wrote:
>>>> On Mon, Nov 02, 2015 at 01:07:38AM -0500, Sinan Kaya wrote:
>>>>> This patch adds supporting utility functions
>>>>> for selftest. The intention is to share the self
>>>>> test code between different drivers.
>>>>>
>>>>> Supported test cases include:
>>>>> 1. dma_map_single
>>>>> 2. streaming DMA
>>>>> 3. coherent DMA
>>>>> 4. scatter-gather DMA
>>>>
>>>> This seems quite similar to dmatest, any reason why you cannot use/enhance
>>>> that?
>>>>
>>> Dmatest is a standalone kernel module intended for stress testing
>>> DMA engines from userspace with N number of threads and M size
>>> combinations etc.
>>>
>>> This one; on the other hand, is selftest to verify hardware is
>>> working as expected during power up.
>>>
>>> Almost all DMA engine drivers come with some sort of selftest code
>>> called from probe. I followed the same design pattern.
>>
>> which ones ?
>>
>>>


>>> I think the goal is to remove the duplicate self test code in all
>>> drivers over time.
>>
>> and what prevents us from having common selftest plus dmatest code. Most of
>> the code here to do selftest is _same_ dmaengine routine code used in
>> dmatest
>>
>> We can have common code which is used for dmatest as well as selftest. I do
>> not want to see same code duplicated..
>
> Originally ioatdma and iop-adma had local self tests before Haavard
> created dmatest.  I agree having the drivers also do a test each boot
> is redundant, but then again dmatest is not automatic and I saw the
> local self test catch an interrupt setup regression.

I see the following files still have some sort of self test code in them.

ioat\dma.c
ioat\dma_v2.c
ioat\dma_v3.c
iop-adma.c
mv_xor.c


>
> Maybe you could arrange for drivers to do a quick autorun through
> dmatest on load if dmatest is enabled, but otherwise load without
> testing?  Just my 2 cents from a dmaengine spectator.
>

I'm on the same boat. Almost all kernel configurations that I have seen 
do not have dmatest enabled. Dmatest is considered debug only and is not 
included into the kernel binaries.

I have no problem for moving the code from one location to the other but 
I still want to be able to run self-test code in mission mode before 
enabling DMA as today.

I'm open to suggestions.

-- 
Sinan Kaya
Qualcomm Technologies, Inc. on behalf of Qualcomm Innovation Center, Inc.
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a 
Linux Foundation Collaborative Project