[PATCH 0/3] mmc: Wait for card_busy before starting sdio requests

[Hans de Goede]

Hi,

On 25-09-15 18:14, Doug Anderson wrote:
> Hi,
>
> I think Jaehoon has already responded to much of this, but...
>
> On Fri, Sep 25, 2015 at 12:53 AM, Hans de Goede <hdegoede at redhat.com> wrote:
>>> 1. Only one of the two callers of dw_mci_wait_while_busy() is handled
>>> by your patch.  mci_send_cmd() is used internally in dw_mmc to throw
>>> something in the CMD register without going through the normal MMC
>>> path.  This is used exclusively to update the clock registers in
>>> dw_mmc.  I'm pretty sure this needs the wait, too.  It's always seemed
>>> weird / awkward to me that you need to use the CMD register to update
>>> clock settings in dw_mmc, but c'est la vie.
>>
>>
>> I would not expect the card to signal busy when trying to change clocks
>> though, so I do not think this will really be a problem.
>
> I just can't quite remember what problem I was seeing.  Let's see...
> I'll comment out that and see if I can find any errors.
>
> OK, I commented it out and ran a bit of stress testing for the 4
> devices / 4 cards sitting at my desk.  I saw no issues.  I also ran
> some of the MMC tests that have caused me problems in the past and saw
> no problems.  I also looked back at
> <https://chromium-review.googlesource.com/#/c/244850/> where this
> landed in our tree and I see that my comment is:
>
>> Actually, while this works I may need to extend it to also be used for mci_send_cmd().
>
> That indicates no problems on my end without the check before updating
> clocks.  ...but, oh, I see why this is.  It was even posted upstream!
> :)
>
> https://patchwork.kernel.org/patch/5858221/
>
> I said:
>
>> Sorry for the quick spin.  After I posted this I re-read all the
>> messages and it looks like Addy has an actual SD card (not SDIO) that
>> is also asserting busy.  He's seeing it assert busy around the clock
>> update.
>
>   ...so I guess the answer here is that I personally haven't seen any
> problems, but adding this check in mci_send_cmd() shouldn't hurt,
> should be safe, and might avoid some problems.  Note that it's
> possible that Addy was seeing some other bug somewhere that simply
> resulted in the "busy" line being asserted, but technically the
> Designware databook recommends waiting for "not busy" before updating
> clocks IIRC.

OK, so maybe we need to add a busy-check in the core before updating
the clocks, as you say below this stuff is likely not designware
specific ...

>>> 2. If I remember correctly, we ran into other instances where non-SDIO
>>> cards needed the busy check.  It wasn't terribly common, but I think I
>>> ran into this when stress testing, but only on a few cards.
>>
>>
>> Hmm, that would be a problem yes.
>
> So perhaps it would be good to update your patch to check for all data commands?

Agreed, since you seem to know this stuff much better then I do can you do
a follow-up patch expanding my patch to do the busy check / wait for
all data commands?

>>> The patch referenced here only seems to check for SDIO commands.  As I
>>> understand it, to be correct, it should check for all data commands
>>> (other than stop or voltage change commands).
>>
>>
>> But that is not what the patch does, it actually waits for all commands,
>> including non data commands. An earlier attempt of mine to fix the sdio
>> wifi issues with the sunxi driver copied your approach, and I actually
>> got reports of regressions with using normal micro-sd memory cards
>> from several people testing that patch.
>
> You're saying that my patch waits for all commands including non-data
> commands.  I'm pretty sure that's not true.

Ok I believe you, I was merely going by the commitdiff which adds
the checks unconditionally, but it may very well be that it is added
to a data only path.

> It relies on a whole
> bunch of other logic in dw_mmc that figures out that we have a data
> command (and that logic also looks for stop commands).  Specifically
> my patches keys off SDMMC_CMD_PRV_DAT_WAIT.  Looking how that gets set
> in dw_mci_prepare_command():
> * We don't set it for the various "stop" commands
> * Else we set it for all commands with cmd->data, except "send status"

Ack.

>> And if you're right that we should wait for all data commands, then
>> I wonder if this is a designware thing (I believe the allwinner
>> mmc controller is designware derived) or a generic mmc / sdio thing ?
>
> It seems hard to believe it would be Designware specific.  If I
> understand correctly, "busy" is signaled by the card holding the data
> lines low.  ...so if a normal SD card was really asserting busy then
> you'd better not send a command.

Agreed, so as said above, you seem to be more knowledgeable on this
then me, can you do a follow-up patch extending the check I added to
not just apply to mmc-io commands ?

>>> The Designware Databook
>>> makes no reference to only needing the wait for SDIO commands.
>>
>>
>> Yet your commit message references problems with sdio wifi cards, and
>> on sunxi we've only been seeing this problem with sdio wifi cards / sdio
>> commands.
>
> Yup.  Though I did some amount in parallel, I definitely focused on
> SDIO problems first before focusing on UHS SD cards.  ...so my primary
> focus was on SDIO here but I tried to code it in a generic way so it
> would be useful for all data commands (since it seemed like it could
> technically affect them).

Ok.

Regards,

Hans