Question about checking rate_spi in pwrap_init_reg_clock
Axel Lin
axel.lin at ingics.com
Fri May 8 02:33:15 PDT 2015
CC Leilk Liu
2015-04-21 23:42 GMT+08:00 Matthias Brugger <matthias.bgg at gmail.com>:
> 2015-04-21 14:52 GMT+02:00 Sascha Hauer <s.hauer at pengutronix.de>:
>> On Tue, Apr 21, 2015 at 01:43:19PM +0200, Matthias Brugger wrote:
>>> 2015-04-21 12:31 GMT+02:00 Sascha Hauer <s.hauer at pengutronix.de>:
>>> > On Tue, Apr 21, 2015 at 09:58:07AM +0800, Axel Lin wrote:
>>> >> hi,
>>> >> The implementation in pwrap_init_reg_clock seems has off-by-one bug.
>>> >> If rate_spi is 26000000, current code set ck_mhz to 18 rather than 26.
>>> >>
>>> >> I guess it needs below fix, but I'm not 100% sure as I don't have the datasheet.
>>> >> Can someone confirm if this is a bug or not?
>>> >
>>> > Yes, seems to be a bug. Thanks for noting. Will you send a formal patch
>>> > or should I do it?
>>>
>>> Did you have a look on both datasheets/reference kernels, mt8135 and mt8173?
>>> Reading the datasheet and reference kernel code for mt6589 this SPI
>>> waveform configuration looks like magic.
>>>
>>> But reading the clock frequency which can have clk_spi in mt8135, they
>>> are 0 MHz, 26 MHz and bigger then 26 MHz, which leads me to the
>>> conclusion that the code is correct. Otherwise the check for bigger
>>> then 18 MHz is useless and should be deleted, if we have a similar
>>> clock distribution in mt8173.
>>
>> The CSHEXT and CSLEXT registers extend the time the chipselect stays
>> high after a SPI transfer or low before a new transfer by the given
>> number of spi clk cycles.
>> Looking at the code maybe the real question is not whether the border is
>> at 25999999Hz or at 26000000Hz, but why the transfers are throttled with
>> slower clock frequencies. I would assume the other way round makes
>> sense.
>
> After re-reading the data sheet and the reference kernel
> implementation, it seems as if three clocks exist. clk_wrap and
> clk_spi and "pmic reg_clk". And some how depending on the three clocks
> you have to calculate the values for CSHEXT, CSLEXT etc.
>
> I think we need some clarification here from Mediatek to really
> understand what is going on.
Hi Leilk,
I found you are working on mtk's spi driver.
Any comment about this discussion?
Regards,
Axel
More information about the Linux-mediatek
mailing list