[U-Boot] [PATCH v4 2/3] mmc: dw_mmc: Support bypass mode with the get_mmc_clk() method

[Simon Glass]

Hi Marek,

On 12 August 2015 at 08:40, Marek Vasut <marex at denx.de> wrote:
> On Wednesday, August 12, 2015 at 03:55:59 PM, Simon Glass wrote:
>> Hi Marek,
>>
>> On 12 August 2015 at 07:53, Marek Vasut <marex at denx.de> wrote:
>> > On Wednesday, August 12, 2015 at 03:51:07 PM, Simon Glass wrote:
>> >> Hi Marek,
>> >>
>> >> On 12 August 2015 at 07:48, Marek Vasut <marex at denx.de> wrote:
>> >> > On Wednesday, August 12, 2015 at 03:04:15 PM, Simon Glass wrote:
>> >> >> Hi Marek,
>> >> >
>> >> > Hi!
>> >> >
>> >> > [...]
>> >> >
>> >> >> >> >> > Why are you passing the @freq into get_mmc_clk() ? Shouldn't
>> >> >> >> >> > you call some clock framework function to determine the
>> >> >> >> >> > input frequency of the DWMMC block from within the
>> >> >> >> >> > get_mmc_clk() implementation instead ? What do you think
>> >> >> >> >> > please ?
>> >> >> >> >>
>> >> >> >> >> Well, yes. If such a clock frame work existed I would call it
>> >> >> >> >> :-) We do have a uclass now so we are getting there.
>> >> >> >> >
>> >> >> >> > Excellent, so do you really need this kind of patch ? :) Why
>> >> >> >> > don't you make just some kind of function -- get_dwmmc_clock()
>> >> >> >> > -- and call it instead ?
>> >> >> >>
>> >> >> >> This is sort-of what is happening. It is calling a function in the
>> >> >> >> host controller - i.e. the SoC's MMC controller. It is one step
>> >> >> >> closer to knowing the input clock to the dwmmc input clock. Note
>> >> >> >> that it is not the clock of the MMC bus itself, but the input
>> >> >> >> clock to the dwmmc logic block.
>> >> >> >
>> >> >> > I don't think I quite understand what you mean here. We're talking
>> >> >> > about obtaining the frequency of the clock which go into the DWMMC
>> >> >> > IP block, right ?
>> >> >> >
>> >> >> > So, if you implement a function, say -- dwmmc_get_upstream_clock()
>> >> >> > -- and call it from within the implementation of the
>> >> >> > .get_mmc_clk(), which is specific for that particular chip of
>> >> >> > yours*, you don't need this patch. Or am I really missing
>> >> >> > something fundamental ?
>> >> >> >
>> >> >> > *the .get_mmc_clk() is specific to a chip, see for example
>> >> >> > exynos_dw_mmc.c
>> >> >>
>> >> >> The purpose of the existing code (before my change) is to find out
>> >> >> the input frequency of the IP block. By knowing this, the dw_mmc
>> >> >> driver can work out what divisor it needs to achieve a particular
>> >> >> MMC bus clock.
>> >> >>
>> >> >> The implementation of get_mmc_clk() (which will be in the
>> >> >> SoC-specific MMC driver) is indeed the place where the clock is
>> >> >> figured out. My only change is to add a parameter which is the
>> >> >> desired bus clock. This parameter can be ignored, but for
>> >> >> implementations which can select the source clock such that it
>> >> >> matches this bus clock, then they can do this and dw_mmc can just
>> >> >> use bypass mode.
>> >> >
>> >> > I see now, this wasn't really clear from the patch description.
>> >> > Shouldn't you introduce another callback for this purpose then, like
>> >> > .set_mmc_clk() instead ?
>> >>
>> >> We could do, but I don't like introducing another interface for one
>> >> client. Also I think the right solution is to move it to use the
>> >> generic clock infrastructure, when it exists (well we have it, but
>> >> nothing uses it yet).
>> >
>> > OK, but making a .get_mmc_clk() function actually configure something
>> > is a behavior I wouldn't expect from a getter function. It's a bit odd
>> > and illogical in my opinion.
>>
>> Yes fair enough, it is odd. I did start an MMC uclass so perhaps that
>> will lead to a better solution. It's unfortunately that dw_mmc need
>> its own callback infrastructure.
>
> I hope we can iron that out shortly. The good thing is that you now have
> a board with the DWMMC and SoCFPGA also has one, so we have at least two
> pairs of eyes on it.
>
> Also, what do you prefer to do about this patch ? Shall we go with the
> .set_mmc_clock() callback and be done with it or do you want to stick
> with the current approach ? I'm inclined to the former as it's less
> confusing in my opinion.

Let's revisit it when I get back to the rockchip series.

Regards,
Simon