Re: [PATCH 2/2] mmc: rtsx: add support for async request

[micky <micky_ching@realsil.com.cn>]

On 06/18/2014 07:03 PM, Ulf Hansson wrote:
> On 18 June 2014 12:08, micky <micky_ching@realsil.com.cn> wrote:
>> On 06/18/2014 03:39 PM, Ulf Hansson wrote:
>>> On 18 June 2014 03:17, micky <micky_ching@realsil.com.cn> wrote:
>>>> On 06/17/2014 03:45 PM, Ulf Hansson wrote:
>>>>> On 17 June 2014 03:04, micky <micky_ching@realsil.com.cn> wrote:
>>>>>> On 06/16/2014 08:40 PM, Ulf Hansson wrote:
>>>>>>> On 16 June 2014 11:09, micky <micky_ching@realsil.com.cn> wrote:
>>>>>>>> On 06/16/2014 04:42 PM, Ulf Hansson wrote:
>>>>>>>>>> @@ -36,7 +37,10 @@ struct realtek_pci_sdmmc {
>>>>>>>>>>>             struct rtsx_pcr         *pcr;
>>>>>>>>>>>             struct mmc_host         *mmc;
>>>>>>>>>>>             struct mmc_request      *mrq;
>>>>>>>>>>> +       struct workqueue_struct *workq;
>>>>>>>>>>> +#define SDMMC_WORKQ_NAME       "rtsx_pci_sdmmc_workq"
>>>>>>>>>>>
>>>>>>>>>>> +       struct work_struct      work;
>>>>>>>>> I am trying to understand why you need a work/workqueue to implement
>>>>>>>>> this feature. Is that really the case?
>>>>>>>>>
>>>>>>>>> Could you elaborate on the reasons?
>>>>>>>> Hi Uffe,
>>>>>>>>
>>>>>>>> we need return as fast as possible in mmc_host_ops
>>>>>>>> request(ops->request)
>>>>>>>> callback,
>>>>>>>> so the mmc core can continue handle next request.
>>>>>>>> when next request everything is ready, it will wait previous done(if
>>>>>>>> not
>>>>>>>> done),
>>>>>>>> then call ops->request().
>>>>>>>>
>>>>>>>> we can't use atomic context, because we use mutex_lock() to protect
>>>>>>> ops->request should never executed in atomic context. Is that your
>>>>>>> concern?
>>>>>> Yes.
>>>>> Okay. Unless I missed your point, I don't think you need the
>>>>> work/workqueue.
>>>> any other method?
>>>>
>>>>> Because, ops->request isn't ever executed in atomic context. That's
>>>>> due to the mmc core, which handles the async mechanism, are waiting
>>>>> for a completion variable in process context, before it invokes the
>>>>> ops->request() callback.
>>>>>
>>>>> That completion variable will be kicked, from your host driver, when
>>>>> you invoke mmc_request_done(), .
>>>> Sorry, I don't understand here, how kicked?
>>> mmc_request_done()
>>>       ->mrq->done()
>>>           ->mmc_wait_done()
>>>               ->complete(&mrq->completion);
>>>
>>>> I think the flow is:
>>>> - not wait for first req
>>>> - init mrq->done
>>>> - ops->request()                         ---         A.rtsx: start queue
>>>> work.
>>>> - continue fetch next req
>>>> - prepare next req ok,
>>>> - wait previous done.                --- B.(mmc_request_done() may be
>>>> called
>>>> at any time from A to B)
>>>> - init mrq->done
>>>> - ops->request()                         ---         C.rtsx: start queue
>>>> next work.
>>>> ...
>>>> and seems no problem.
>>> Right, I don't think there are any _problem_ by using the workqueue as
>>> you have implemented, but I am questioning if it's correct. Simply
>>> because I don't think there are any reasons to why you need a
>>> workqueue, it doesn't solve any problem for you - it just adds
>>> overhead.
>> Hi Uffe,
>>
>> we have two driver under mfd, the rtsx-mmc and rtsx-ms,
>> we use mutex lock(pcr_mutex) to protect resource,
>> when we handle mmc request, we need hold the mutex until we finish the
>> request,
>> so it will not interruptted by rtsx-ms request.
> Ahh, I see. Now, _that_ explains why you want the workqueue. :-) Thanks!
>
>> If we not use workq, once the request hold the mutex, we have to wait until
>> the request finish,
>> then release mutex, so the mmc core is blocking at here.
>> To implement nonblocking request, we have to use workq.
> One minor suggestion below, please consider this as an optimization
> which goes outside the context of this patch.
>
> There are cases when I think you should be able to skip the overhead
> from scheduling the work from ->request(). Those cases can be
> described as when the mutex are available which can be tested by using
> mutex_trylock().
Thanks for your suggestion.

we need schedule the work every time mmc core call ops->request(),
if we want to handle request, we need hold mutex and do the work.
so mutex_trylock() will not help decrease overhead.
if we not schedule the work, the ops->request will do nothing.

Best Regards.
micky
> Kind regards
> Uffe
> .
>