[PATCH V2 3/3] mmc: dw_mmc: Dont cut off vqmmc and vmmc

[Doug Anderson]

Ulf,

On Mon, Aug 25, 2014 at 1:13 AM, Ulf Hansson <ulf.hansson at linaro.org> wrote:
> On 22 August 2014 20:27, Sonny Rao <sonnyrao at chromium.org> wrote:
>> On Fri, Aug 22, 2014 at 8:31 AM, Ulf Hansson <ulf.hansson at linaro.org> wrote:
>>> On 22 August 2014 15:47, Yuvaraj Kumar C D <yuvaraj.cd at gmail.com> wrote:
>>>> Exynos 5250 and 5420 based boards uses built-in CD# line for card
>>>> detection.But unfortunately CD# line is on the same voltage rails
>>>> as of I/O voltage rails. When we cut off vqmmc,the consequent card
>>>> detection will break in these boards.
>>>
>>> I am not sure I follow here.
>>>
>>> Is the card detect mechanism handled internally by the dw_mmc controller?
>>
>> Yes
>
> Just out of curiosity.
>
> Do you know how the power to the actual dw_mmc controller is handled?
> I expect it to be SoC specific and I am guessing power domain
> regulators may be involved!?

You can likely read the dw_mmc registers when vqmmc is off.  Is that
what you're asking?  Certainly if vqmmc is not powered then the lines
themselves will be useless, won't they?  The "vqmmc" supply goes to
the "VDDQ_MMC2" pin on 5420.  In my 5420 user manual, I see that
"clk", "cmd", "cd", "datN", "wp" and "biuvr" pins are all in this same
voltage (VDDQ_MMC2) domain.  Can you really read a pin without
powering that part of the SoC?


>>> I thought HW engineers long time ago realized that this should be done
>>> separately on a GPIO line to be able to save power while waiting for a
>>> card to be inserted. But that's not case then?
>>
>> At least in my limited experience, this seems to be common among SoC
>> vendors who are using dw_mmc, as we've seen this elsewhere as well and
>> after seeing it here we know that we need to ignore the CD pin that's
>> routed to dw_mmc and use a separately powered GPIO on the board, but
>> still there are probably many SoCs/boards which are doing it this way.
>>
>>>>
>>>> These hosts (obviously) need to keep vqmmc (and thus vmmc) on all the
>>>> time, even when the mmc core tells them to power off. However, one
>>>> problem is that these cards won't properly handle mmc_power_cycle().
>>>> That's needed to handle error cases when trying to switch voltages
>>>> (see 0797e5f mmc:core: Fixup signal voltage switch).
>>>>
>>>> This patch adds a new MMC_POWER_OFF_HARD mode when it's doing a power
>>>> cycle.  This mode differs from the normal MMC_POWER_OFF mode in that
>>>> the mmc core will promise to power the slot back on before it expects
>>>> the host to detect card insertion or removal.
>>
>> This patch is based off of one that Doug wrote (sent privately to
>> Yuvaraj) which just modifies the MMC core, and should be split into
>> two patches.
>> One that modifies the mmc core and one that implements this in dw_mmc.
>
> I looked at the mmc core parts, it seems like the wrong approach.
>
> I think you shall be able use MMC_CAP_NEEDS_POLL, to handle this
> broken card detect mechanism. We even have a DT binding for that,
> "broken-cd".

I don't think this is possible, but let me explain why I think so and
you can correct me.

The voltage domain of the "card detect" pin on the SoC is vqmmc,
right?  That means that you won't be able to read the pin without
turning on vqmmc.  Even if you could read the pin without turning on
vqmmc, the pullup on this line is connected to vqmmc too.  ...so if
vqmmc is off then there's no pulup and you can't use card detect.

Are you suggesting that we should flip the voltage of vqmmc (and thus
vmmc to prevent damaging the card) during polling?  That seems ugly.


One other thing to mention: we didn't find any power savings by
actually turning off vmmc and vqmmc when there was no card inserted.
There's no current running through the lines when there is no card
inserted and apparently everything is efficient enough that there was
no problem.

-Doug