Re: [PATCH V13 07/10] mmc: block: blk-mq: Add support for direct completion

[Adrian Hunter <adrian.hunter@intel.com>]

On 09/11/17 14:34, Linus Walleij wrote:
> On Thu, Nov 9, 2017 at 8:27 AM, Adrian Hunter <adrian.hunter@intel.com> wrote:
>> On 08/11/17 11:28, Linus Walleij wrote:
>>> On Fri, Nov 3, 2017 at 2:20 PM, Adrian Hunter <adrian.hunter@intel.com> wrote:
>>>
>>>> For blk-mq, add support for completing requests directly in the ->done
>>>> callback. That means that error handling and urgent background operations
>>>> must be handled by recovery_work in that case.
>>>>
>>>> Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
>>>
>>> I tried enabling this on my MMC host (mmci) but I got weird
>>> DMA error messages when I did.
>>>
>>> I guess this has not been tested on a non-DMA-coherent
>>> system?
>>
>> I don't see what DMA-coherence has to do with anything.
>>
>> Possibilities:
>>         - DMA unmapping doesn't work in an atomic context
>>         - requests' DMA operations have to be synchronized with each other
> 
> So since MMCI need the post_req() hook called with
> an error code to properly tear down any DMA operations,
> I was worried that maybe your error path is not doing this
> (passing an error code or calling in the right order).
> 
> I had a bunch of fallouts in my own patch set relating
> to that.
> 
>>> I think I might be seeing this because the .pre and .post
>>> callbacks need to be strictly sequenced, and this is
>>> maybe not taken into account here?
>>
>> I looked at mmci but that did not seem to be the case.
>>
>>> Isn't there as risk
>>> that the .post callback of the next request is called before
>>> the .post callback of the previous request has returned
>>> for example?
>>
>> Of course, the requests are treated as independent.  If the separate DMA
>> operations require synchronization, that is for the host driver to fix.
> 
> They are treated as independent by the block layer but
> it is the subsystems duty to serialize them for the hardware,
> 
> MMCI strictly requires that pre/post hooks per request
> happen in the right order, so if you have prepared a second
> request after submitting the first, and the first fails, you have
> to back out by unpreparing the second one before unpreparing
> the first. It is also the only host driver requireing to be passed
> an error code in the last parameter to the post hook in
> order to work properly.
> 
> I think your patch set handles that nicely though, because I
> haven't seen any errors, it's just when we do this direct
> completion I see problems.

If a request gets an error, then we always do the post_req before starting
another request, so the driver can assume that the first request finished
successfully if it is asked to do post_req on the second request.  So the
driver does have enough information to keep the DMA unmapping in order if
necessary.