Re: [PATCH V2 2/2] mmc: sdhci-msm: Rectify DLL programming sequence for SDCC

[Ram Prakash Gupta <quic_rampraka@quicinc.com>]

On 8/9/2025 3:10 PM, Dmitry Baryshkov wrote:
> On Thu, Aug 07, 2025 at 01:28:28AM +0530, Ram Prakash Gupta wrote:
>> On 7/31/2025 7:49 PM, Dmitry Baryshkov wrote:
>>> On 31/07/2025 14:46, Ram Prakash Gupta wrote:
>>>> On 7/30/2025 11:26 PM, Dmitry Baryshkov wrote:
>>>>> On Wed, Jul 23, 2025 at 03:43:37PM +0530, Ram Prakash Gupta wrote:
>>>>>> On 1/22/2025 3:20 PM, Dmitry Baryshkov wrote:
>>>>>>> On Wed, Jan 22, 2025 at 02:57:59PM +0530, Sachin Gupta wrote:
>>>>>>>> On 1/7/2025 8:38 PM, Dmitry Baryshkov wrote:
>>>>>>>>> On Tue, Jan 07, 2025 at 11:13:32AM +0530, Sachin Gupta wrote:
>>>>>>>>>> On 12/27/2024 12:23 AM, Dmitry Baryshkov wrote:
>>>>>>>>>>> On Thu, Dec 26, 2024 at 11:22:40AM +0530, Sachin Gupta wrote:
>>>>>>>>>>>> On 12/19/2024 11:24 AM, Dmitry Baryshkov wrote:
>>>>>>>>>>>>> On Wed, Dec 18, 2024 at 02:40:57PM +0530, Sachin Gupta wrote:
>>>>>>>>>>>>>> +
>>>>>>>>>>>>>> +static unsigned int sdhci_msm_get_clk_rate(struct sdhci_host *host, u32 req_clk)
>>>>>>>>>>>>>> +{
>>>>>>>>>>>>>> +    struct sdhci_pltfm_host *pltfm_host = sdhci_priv(host);
>>>>>>>>>>>>>> +    struct sdhci_msm_host *msm_host = sdhci_pltfm_priv(pltfm_host);
>>>>>>>>>>>>>> +    struct clk *core_clk = msm_host->bulk_clks[0].clk;
>>>>>>>>>>>>>> +    unsigned int sup_clk;
>>>>>>>>>>>>>> +
>>>>>>>>>>>>>> +    if (req_clk < sdhci_msm_get_min_clock(host))
>>>>>>>>>>>>>> +        return sdhci_msm_get_min_clock(host);
>>>>>>>>>>>>>> +
>>>>>>>>>>>>>> +    sup_clk = clk_round_rate(core_clk, clk_get_rate(core_clk));
>>>>>>>>>>>>>> +
>>>>>>>>>>>>>> +    if (host->clock != msm_host->clk_rate)
>>>>>>>>>>>>>> +        sup_clk = sup_clk / 2;
>>>>>>>>>>>>>> +
>>>>>>>>>>>>>> +    return sup_clk;
>>>>>>>>>>>>> Why?
>>>>>>>>>>>> Sorry, I did not understand your question. Can you please explain in detail.
>>>>>>>>>>> Please explain the maths. You get the rate from the clock, then you
>>>>>>>>>>> round it, but it is the rate that has just been returned, so there
>>>>>>>>>>> should be no need to round it. And after that there a division by two
>>>>>>>>>>> for some reason. So I've asked for an explanation for that code.
>>>>>>>>>>>
>>>>>>>>>> clk_round_rate is used in case of over clocking issue we can round it to the
>>>>>>>>>> usable frequency.
>>>>>>>>> If it is a frequency _returned_ by the clock driver, why do you need to
>>>>>>>>> round it? It sounds like that freq should be usable anyway.
>>>>>>>>>
>>>>>>>> I agree, rounding will be taken care by clock driver. Will remove in my next
>>>>>>>> patch.
>>>>>>>>
>>>>>>>>>> Divide by 2 is used as for HS400 the tuning happens in
>>>>>>>>>> HS200 mode only so to update the frequency to 192 Mhz.
>>>>>>>>> Again, is it really 192 MHz? Or 19.2 MHz?
>>>>>>>>> Also if it is for HS400, then shouldn't /2 be limited to that mode?
>>>>>>>>>
>>>>>>>> Yes, It is 192 MHz.
>>>>>>> Good, thanks for the confirmation.
>>>>>>>
>>>>>>>> As part of eMMC Init, driver will try to init with the best mode supported
>>>>>>>> by controller and device. In this case it is HS400 mode, But as part of
>>>>>>>> HS400 mode, we perform Tuning in HS200 mode only where we need to configure
>>>>>>>> half of the clock.
>>>>>>> This isn't an answer to the question. Let me rephrase it for you: if the
>>>>>>> /2 is only used for HS400, why should it be attempted in all other
>>>>>>> modes? Please limit the /2 just to HS400.
>>>>>> Hi Dmitry,
>>>>>>
>>>>>> like updated earlier by Sachin, HS400 tuning happens in HS200 mode, so if
>>>>>> we try to use "ios->timing == MMC_TIMING_MMC_HS400" that wont help, as at
>>>>>> this stage timing can be MMC_TIMING_MMC_HS200/MMC_TIMING_MMC_HS400 for
>>>>>> hs200 tuning and hs400 selection. In this case we must divide clk by 2
>>>>>> to get 192MHz and we find this as host->clock wont be equal to
>>>>>> msm_host->clk_rate.
>>>>> What are host->clock and msm_host->clk_rate at this point?
>>>>> What is the host->flags value? See sdhci_msm_hc_select_mode().
>>>> There are 2 paths which are traced to this function when card initializes
>>>> in HS400 mode, please consider below call stack in 2 paths
>>>>
>>>> sdhci_msm_configure_dll
>>>> sdhci_msm_dll_config
>>>> sdhci_msm_execute_tuning
>>>> mmc_execute_tuning
>>>> mmc_init_card
>>>> _mmc_resume
>>>> mmc_runtime_resume
>>>>
>>>> with values of host->clock as 200000000 & msm_host-clk_rate as 400000000
>>> Please check the rates explicitly in the code rather than just checking that they are not equal.
>> in function msm_get_clock_mult_for_bus_mode(), clk multiplier returns 2, with HS400
>> DDR52 and DDR50 modes which is called from sdhci_msm_set_clock() and
>> sdhci_msm_execute_tuning. And in sdhci_msm_execute_tuning(), we are calling
>> sdhci_msm_dll_config() when SDHCI_HS400_TUNING is set and this flag is cleared
>> immediately after return. And sdhci_msm_dll_config() is called after that.
>>
>> Now when the card is supporting HS400 mode, then from mmc_hs200_tuning(),
>> sdhci_prepare_hs400_tuning is getting called, and there SDHCI_HS400_TUNING
>> flag is set, and clock set is multiplying the clk rate by 2 in below call stack
>>
>> msm_set_clock_rate_for_bus_mode
>> sdhci_msm_execute_tuning
>> mmc_execute_tuning
>> mmc_init_card
>>
>> so this clk rate is doubling only with HS400 mode selection and while setting up
>> dll in HS400 dll configuration path sup_clk need to divide by 2 as msm_host->clk_rate
>> is twice the host->clock as mentioned above.
> I don't see how it's relevant. I'm asking you to check for the rate
> values explicitly in the driver code rather than just checking that
> rateA != rateB. You might error out if rateA != rateB and they are not
> 192 MHz / 384 MHz

ok I see, but I got one another alternate to this, I can try use similar checks used in
function msm_get_clock_mult_for_bus_mode(), ios.timing == MMC_TIMING_MMC_HS400 
host->flags & SDHCI_HS400_TUNING, but for this I ll have to move check
host->flags &= ~SDHCI_HS400_TUNING in function sdhci_msm_execute_tuning () at the bottom
of this function, this will make code more readable and consistent to checks at other
places but I am testing it right now and will update.

If this doesn't work then I ll explicitly use the rate value in check.

Thanks,
Ram

>
>>
>>>> and host->flags as 0x90c6.
>>>>
>>>> and
>>>>
>>>> sdhci_msm_configure_dll
>>>> sdhci_msm_dll_config
>>>> sdhci_msm_set_uhs_signaling
>>>> sdhci_set_ios
>>>> mmc_set_clock
>>>> mmc_set_bus_speed
>>>> mmc_select_hs400
>>>> mmc_init_card
>>>> _mmc_resume
>>>> mmc_runtime_resume
>>>>
>>>> with values of host->clock as 200000000 & msm_host-clk_rate as 400000000
>>>> and host->flags as 0x90c6 which are same as 1st.
>>>>
>>>> Now if card is initialized in HS200 mode only below is the call stack
>>>>
>>>> sdhci_msm_configure_dll
>>>> sdhci_msm_dll_config
>>>> sdhci_msm_execute_tuning
>>>> mmc_execute_tuning
>>>> mmc_init_card
>>>> _mmc_resume
>>>> mmc_runtime_resume
>>>>
>>>> with values of host->clock as 200000000 & msm_host-clk_rate as 200000000
>>>> and host->flags as 0x90c6.
>>>>
>>>> now if you see the host->flags value, its same across the modes, and if
>>>> I am getting it right from the pointed out function
>>>> sdhci_msm_hc_select_mode(), your suggestion seems to be using the check
>>>> host->flags & SDHCI_HS400_TUNING which is bit[13], but in above dumped
>>>> host->flags SDHCI_HS400_TUNING bit is not set where we are using the /2.
>>>>
>>>> and the reason is, this bit is getting cleared in sdhci_msm_execute_tuning()
>>>> before sdhci_msm_dll_config() call.
>>>>
>>>> so this /2, is eventually called only for HS400 mode.
>>>>
>>>> Thanks,
>>>> Ram
>>>>
>>>>>> Now if we go for only HS200 mode supported card, there
>>>>>> the supported clock value would be 192Mhz itself and we need to pass
>>>>>> clk freq as 192MHz itself, hence division by 2 wont be needed, that is
>>>>>> achieved there as host->clock would be equal to msm_host->clk_rate. Hence
>>>>>> no other check is needed here.
>>>>> Please think about the cause, not about the symptom. Clocks being
>>>>> unequal is a result of some other checks being performed by the driver.
>>>>> Please use those checks too.
>>>>>
>>>>>> sorry for it took time to update as I was gathering all this data.
>>>>> 6 months? Well, that's a nice time to "gather all this data".
>>>> Took it up from sachin last month but still its a long gap.
>>>> Thanks for helping revive.
>>>>
>>>>>> since Sachin have already pushed patchset #3, and if this explanation
>>>>>> helps, let me know if we can continue on patchset #3.
>>>>>>
>>>>>> Thanks,
>>>>>> Ram
>>>>>>
>>>