Re: [PATCH] mmc: mmc: Fix HS setting in mmc_hs400_to_hs200()

[Adrian Hunter]

On 4/02/19 12:54 PM, Ulf Hansson wrote:
> On Mon, 4 Feb 2019 at 10:58, Adrian Hunter <adrian.hunter@xxxxxxxxx> wrote:
>>
>> On 1/02/19 10:10 AM, Ulf Hansson wrote:
>>> On Fri, 1 Feb 2019 at 02:38, Chaotian Jing <chaotian.jing@xxxxxxxxxxxx> wrote:
>>>>
>>>> On Thu, 2019-01-31 at 16:58 +0100, Ulf Hansson wrote:
>>>>> On Thu, 31 Jan 2019 at 08:53, Chaotian Jing <chaotian.jing@xxxxxxxxxxxx> wrote:
>>>>>>
>>>>>> mmc_hs400_to_hs200() begins with the card and host in HS400 mode.
>>>>>> Therefore, any commands sent to the card should use HS400 timing.
>>>>>> It is incorrect to reduce frequency to 50Mhz before sending the switch
>>>>>> command, in this case, only reduce clock frequency to 50Mhz but without
>>>>>> host timming change, host is still in hs400 mode but clock changed from
>>>>>> 200Mhz to 50Mhz, which makes the tuning result unsuitable and cause
>>>>>> the switch command gets response CRC error.
>>>>>
>>>>> According the eMMC spec there is no violation by decreasing the clock
>>>>> frequency like this. We can use whatever value <=200MHz.
>>>>>
>>>>> However, perhaps in practice this becomes an issue, due to the tuning
>>>>> for HS400 has been done on the "current" frequency.
>>>>>
>>>>> As as start, I think you need to clarify this in the changelog.
>>>>>
>>>> Yes, reduce clock frequency to 50Mhz is no Spec violation, but it may
>>>> cause __mmc_switch() gets response CRC error, decreasing the clock but
>>>> without HOST mode change, on the host side, host driver do not know
>>>> what's operation the core layer want to do and can only set current bus
>>>> clock to 50Mhz, without tuning parameter change, it has a chance lead to
>>>> response CRC error. even lower clock frequency, but with the wrong
>>>> tuning parameter setting(the setting is of hs400 tuning @200Mhz).
>>>
>>> Right, makes sense.
>>>
>>>>>>
>>>>>> this patch refers to mmc_select_hs400(), make the reduce clock frequency
>>>>>> after card timing change.
>>>>>>
>>>>>> Signed-off-by: Chaotian Jing <chaotian.jing@xxxxxxxxxxxx>
>>>>>> ---
>>>>>>  drivers/mmc/core/mmc.c | 8 ++++----
>>>>>>  1 file changed, 4 insertions(+), 4 deletions(-)
>>>>>>
>>>>>> diff --git a/drivers/mmc/core/mmc.c b/drivers/mmc/core/mmc.c
>>>>>> index da892a5..21b811e 100644
>>>>>> --- a/drivers/mmc/core/mmc.c
>>>>>> +++ b/drivers/mmc/core/mmc.c
>>>>>> @@ -1239,10 +1239,6 @@ int mmc_hs400_to_hs200(struct mmc_card *card)
>>>>>>         int err;
>>>>>>         u8 val;
>>>>>>
>>>>>> -       /* Reduce frequency to HS */
>>>>>> -       max_dtr = card->ext_csd.hs_max_dtr;
>>>>>> -       mmc_set_clock(host, max_dtr);
>>>>>> -
>>>>>
>>>>> As far as I can tell, the reason to why we change the clock frequency
>>>>> *before* the call to __mmc_switch() below, is probably to try to be on
>>>>> the safe side and conform to the spec.
>>>>>
>>>> Agree, it Must be more safe with lower clock frequency, but the
>>>> precondition is to make the host side recognize current timing is not
>>>> HS400 mode. it has no method to find a safe setting to ensure no
>>>> response CRC error when reduce clock from 200Mhz to 50Mhz.
>>>>> However, I think you have a point, as the call to __mmc_switch(),
>>>>> passes the "send_status" parameter as false, no other command than the
>>>>> CMD6 is sent to the card.
>>>>>
>>>> yes, the send status command was sent only after __mmc_switch() done.
>>>>>>         /* Switch HS400 to HS DDR */
>>>>>>         val = EXT_CSD_TIMING_HS;
>>>>>>         err = __mmc_switch(card, EXT_CSD_CMD_SET_NORMAL, EXT_CSD_HS_TIMING,
>>>>>> @@ -1253,6 +1249,10 @@ int mmc_hs400_to_hs200(struct mmc_card *card)
>>>>>>
>>>>>>         mmc_set_timing(host, MMC_TIMING_MMC_DDR52);
>>>>>>
>>>>>> +       /* Reduce frequency to HS */
>>>>>> +       max_dtr = card->ext_csd.hs_max_dtr;
>>>>>> +       mmc_set_clock(host, max_dtr);
>>>>>> +
>>>>>
>>>>> Perhaps it's even more correct to change the clock frequency before
>>>>> the call to mmc_set_timing(host, MMC_TIMING_MMC_DDR52). Otherwise you
>>>>> will be using the DDR52 timing in the controller, but with a too high
>>>>> frequency.
>>>>>
>>>> for Our host, it has no impact to change the clock before or after
>>>> change timing, as the mmc_set_timing() is only for host side, not
>>>> related to MMC card side and no commands sent do card before the
>>>> timing/clock change completed.
>>>
>>> Alright. After a second thought, it actually looks more consistent
>>> with mmc_select_hs400() to do it after, as what you propose in
>>> $subject patch.
>>>
>>> So, let's keep it as is.
>>>
>>>>>>         err = mmc_switch_status(card);
>>>>>>         if (err)
>>>>>>                 goto out_err;
>>>>>> --
>>>>>> 1.8.1.1.dirty
>>>>>>
>>>>>
>>>>> Finally, it sounds like you are trying to fix a real problem, can you
>>>>> please provide some more information what is happening when the
>>>>> problem occurs at your side?
>>>>>
>>>> Yes, I got a problem with new kernel version. with
>>>> commit:57da0c042f4af52614f4bd1a148155a299ae5cd8, this commit makes
>>>> re-tuning every time when access RPMB partition.
>>>
>>> Okay, could you please add this as fixes tag for the next version of the patch.
>>>
>>>>
>>>> in fact, our host tuning result of hs400 is very stable and almost never
>>>> get response CRC error with clock frequency at 200Mhz. but cannot ensure
>>>> this tuning result also suitable when running at HS400 mode @50Mhz. as I
>>>> mentioned before, the host side does not know the reason of reduce clock
>>>> frequency to 50Mhz at HS400 mode, so what's the host side can do is only
>>>> reduce the bus clock to 50Mhz, even it can just only set the tuning
>>>> setting to default when clock frequency lower than 50Mhz, but both card
>>>> & host side are still at HS400 mode, still cannot ensure this setting is
>>>> suitable.
>>>
>>> Right, thanks for clarifying.
>>>
>>> So I am expecting a new version with a fixes tag and some
>>> clarification of the changelog, then I am ready to apply this to give
>>> it some test.
>>
>> The switch from HS400 mode is done for tuning at times when CRC errors are a
>> possibility e.g. after a CRC error during transfer.  So if the frequency is
>> not to be reduced, then some mitigation is needed for the possibility that
>> the CMD6 response itself will have a CRC error.
> 
> That's a good point!
> 
> However, how can we know that a CMD6 command is successfully
> completed, if there is CRC errors detected during the transmission? I
> guess we can't!?

Yes, in that case, the only option is to assume the CMD6 was successful,
like in

  commit ef3d232245ab7a1bf361c52449e612e4c8b7c5ab
  Author: Adrian Hunter <adrian.hunter@xxxxxxxxx>
  Date:   Fri Dec 2 13:16:35 2016 +0200

      mmc: mmc: Relax checking for switch errors after HS200 switch

If we are going to do that, then we could stick with lowering the frequency
first.

Also I wonder if the mediatek driver could change to fixed sampling in
->set_ios() when the frequency drops for HS400 mode?