Re: [PATCH V8 00/14] mmc: Add Command Queue support

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

On 11/10/17 16:58, Ulf Hansson wrote:
> On 11 October 2017 at 14:58, Adrian Hunter <adrian.hunter@intel.com> wrote:
>> On 11/10/17 15:13, Ulf Hansson wrote:
>>> On 10 October 2017 at 15:31, Adrian Hunter <adrian.hunter@intel.com> wrote:
>>>> On 10/10/17 16:08, Ulf Hansson wrote:
>>>>> [...]
>>>>>
>>>>>>>>>
>>>>>>>>> I have also run some test on my ux500 board and enabling the blkmq
>>>>>>>>> path via the new MMC Kconfig option. My idea was to run some iozone
>>>>>>>>> comparisons between the legacy path and the new blkmq path, but I just
>>>>>>>>> couldn't get to that point because of the following errors.
>>>>>>>>>
>>>>>>>>> I am using a Kingston 4GB SDHC card, which is detected and mounted
>>>>>>>>> nicely. However, when I decide to do some writes to the card I get the
>>>>>>>>> following errors.
>>>>>>>>>
>>>>>>>>> root@ME:/mnt/sdcard dd if=/dev/zero of=testfile bs=8192 count=5000 conv=fsync
>>>>>>>>> [  463.714294] mmci-pl18x 80126000.sdi0_per1: error during DMA transfer!
>>>>>>>>> [  464.722656] mmci-pl18x 80126000.sdi0_per1: error during DMA transfer!
>>>>>>>>> [  466.081481] mmci-pl18x 80126000.sdi0_per1: error during DMA transfer!
>>>>>>>>> [  467.111236] mmci-pl18x 80126000.sdi0_per1: error during DMA transfer!
>>>>>>>>> [  468.669647] mmci-pl18x 80126000.sdi0_per1: error during DMA transfer!
>>>>>>>>> [  469.685699] mmci-pl18x 80126000.sdi0_per1: error during DMA transfer!
>>>>>>>>> [  471.043334] mmci-pl18x 80126000.sdi0_per1: error during DMA transfer!
>>>>>>>>> [  472.052337] mmci-pl18x 80126000.sdi0_per1: error during DMA transfer!
>>>>>>>>> [  473.342651] mmci-pl18x 80126000.sdi0_per1: error during DMA transfer!
>>>>>>>>> [  474.323760] mmci-pl18x 80126000.sdi0_per1: error during DMA transfer!
>>>>>>>>> [  475.544769] mmci-pl18x 80126000.sdi0_per1: error during DMA transfer!
>>>>>>>>> [  476.539031] mmci-pl18x 80126000.sdi0_per1: error during DMA transfer!
>>>>>>>>> [  477.748474] mmci-pl18x 80126000.sdi0_per1: error during DMA transfer!
>>>>>>>>> [  478.724182] mmci-pl18x 80126000.sdi0_per1: error during DMA transfer!
>>>>>>>>>
>>>>>>>>> I haven't yet got the point of investigating this any further, and
>>>>>>>>> unfortunate I have a busy schedule with traveling next week. I will do
>>>>>>>>> my best to look into this as soon as I can.
>>>>>>>>>
>>>>>>>>> Perhaps you have some ideas?
>>>>>>>>
>>>>>>>> The behaviour depends on whether you have MMC_CAP_WAIT_WHILE_BUSY. Try
>>>>>>>> changing that and see if it makes a difference.
>>>>>>>
>>>>>>> Yes, it does! I disabled MMC_CAP_WAIT_WHILE_BUSY (and its
>>>>>>> corresponding code in mmci.c) and the errors goes away.
>>>>>>>
>>>>>>> When I use MMC_CAP_WAIT_WHILE_BUSY I get these problems:
>>>>>>>
>>>>>>> [  223.820983] mmci-pl18x 80126000.sdi0_per1: error during DMA transfer!
>>>>>>> [  224.815795] mmci-pl18x 80126000.sdi0_per1: error during DMA transfer!
>>>>>>> [  226.034881] mmci-pl18x 80126000.sdi0_per1: error during DMA transfer!
>>>>>>> [  227.112884] mmci-pl18x 80126000.sdi0_per1: error during DMA transfer!
>>>>>>> [  227.220275] mmc0: Card stuck in wrong state! mmcblk0 mmc_blk_card_stuck
>>>>>>> [  228.686798] mmci-pl18x 80126000.sdi0_per1: error during DMA transfer!
>>>>>>> [  229.892150] mmci-pl18x 80126000.sdi0_per1: error during DMA transfer!
>>>>>>> [  231.031890] mmci-pl18x 80126000.sdi0_per1: error during DMA transfer!
>>>>>>> [  232.239013] mmci-pl18x 80126000.sdi0_per1: error during DMA transfer!
>>>>>>> 5000+0 records in
>>>>>>> 5000+0 records out
>>>>>>> root@ME:/mnt/sdcard
>>>>>>>
>>>>>>> I looked at the new blkmq code from patch v10 13/15. It seems like the
>>>>>>> MMC_CAP_WAIT_WHILE_BUSY is used to determine whether the async request
>>>>>>> mechanism should be used or not. Perhaps I didn't looked close enough,
>>>>>>> but maybe you could elaborate on why this seems to be the case!?
>>>>>>
>>>>>> MMC_CAP_WAIT_WHILE_BUSY is necessary because it means that a data transfer
>>>>>> request has finished when the host controller calls mmc_request_done(). i.e.
>>>>>> polling the card is not necessary.
>>>>>
>>>>> Well, that is a rather big change on its own. Earlier we polled with
>>>>> CMD13 to verify that the card has moved back to the transfer state, in
>>>>> case it was a write. And that was no matter of MMC_CAP_WAIT_WHILE_BUSY
>>>>> was set or not. Right!?
>>>>
>>>> Yes
>>>>
>>>>>
>>>>> I am not sure it's a good idea to bypass that validation, it seems
>>>>> fragile to rely only on the busy detection on DAT line for writes.
>>>>
>>>> Can you cite something from the specifications that backs that up, because I
>>>> couldn't find anything to suggest that CMD13 polling was expected.
>>>
>>> No I can't, but I don't see why that matters.
>>>
>>> My point is, if we want to go down that road by avoiding the CMD13
>>> polling, that needs to be a separate change, which we can test and
>>> confirm on its own.
>>>
>>>>
>>>>>
>>>>>>
>>>>>> Have you tried V9 or V10.  There was a fix in V9 related to calling
>>>>>> ->post_req() which could mess up DMA.
>>>>>
>>>>> I have used V10.
>>>>>
>>>>>>
>>>>>> The other thing that could go wrong with DMA is if it cannot accept
>>>>>> ->post_req() being called from mmc_request_done().
>>>>>
>>>>> I don't think mmci has a problem with that, however why do you want to
>>>>> do this? Wouldn't that defeat some of the benefits with the async
>>>>> request mechanism?
>>>>
>>>> Perhaps - but it would need to be tested.  If there are more requests
>>>> waiting, one optimization could be to defer ->post_req() until after the
>>>> next request is started.
>>>
>>> This is already proven, because this how the existing mmc async
>>> request mechanism works.
>>>
>>> In ->post_req() callbacks, host drivers may do dma_unmap_sg(), which
>>> is something that could be costly and therefore it's better to start a
>>> new request before, such these things can go on in parallel.
>>
>> OK I will make a patch that takes care of both issues.  That will also mean
>> the request is not completed in the ->done() callback because ->post_req()
>> must precede block layer completion.
> 
> Right.
> 
> Actually completing the request in the ->done callback, may still be
> possible, because in principle it only needs to inform the other
> prepared request that it may start, before it continues to post
> process/completes the current one.

It already does that.

> 
> However, by looking at for example how mmci.c works, it actually holds
> its spinlock while it calls mmc_request_done(). The same spinlock is
> taken in the ->request() function, but not in the ->post_req()
> function. In other words, completing the request in the ->done()
> callback, would make mmci to keep the spinlock held throughout the
> post processing cycle, which then prevents the next request from being
> started.

It is in interrupt context also. So if the block layer work is on the same
CPU it will anyway wait.

> 
> So my conclusion is, let's start a as you suggested, by not completing
> the request in ->done() as to maintain existing behavior. Then we can
> address optimizations on top, which very likely will involve doing
> changes to host drivers as well.

Ideally it should be possible to start the next already prepared request
from the ->done() callback of the current request.  Perhaps add a new host API.