[U-Boot] [PATCH] mmc: Avoid HS400 mode when accessing boot partitions

Sat Jun 15 15:15:37 UTC 2019

On 6/14/19 5:27 PM, Jean-Jacques Hiblot wrote:
> Marek, Faiz,
> 
> On 11/06/2019 17:59, Faiz Abbas wrote:
>> Hi Marek,
>>
>> On 11/06/19 3:34 PM, Marek Vasut wrote:
>>> On 6/11/19 10:12 AM, Faiz Abbas wrote:
>>>> Peng, Marek,
>>>>
>>>> On 11/06/19 6:47 AM, Peng Fan wrote:
>>>>>> partitions
>>>>>>
>>>>>> On 6/10/19 7:59 AM, Peng Fan wrote:
>>>>>>>> Subject: Re: [U-Boot] [PATCH] mmc: Avoid HS400 mode when accessing
>>>>>>>> boot partitions
>>>>>>>>
>>>>>>>> Hi Marek, Peng,
>>>>>>>>
>>>>>>>> On 03/06/19 12:04 PM, Peng Fan wrote:
>>>>>>>>>> Subject: [PATCH] mmc: Avoid HS400 mode when accessing boot
>>>>>>>>>> partitions
>>>>>>>>>>
>>>>>>>>>> According to JEDEC JESD84-B51.pdf section 6.3.3 Boot operation ,
>>>>>>>>>> HS200 & HS400 mode is not supported during boot operation. The
>>>>>>>>>> U-Boot code currently only applies this restriction to HS200
>>>>>>>>>> mode,
>>>>>>>>>> extend this to
>>>>>>>>>> HS400 mode as well.
>>>>>>>> The spec in section 6.3.3 (according to my understanding) is
>>>>>>>> talking
>>>>>>>> about "boot operation" which is a way of getting data from the the
>>>>>>>> eMMC without going through the Device identification mode (Section
>>>>>>>> 6.4.4) i.e. without sending any commands. All the host has to do is
>>>>>>>> hold the command line low in Pre-Idle mode to automatically receive
>>>>>>>> data at the preconfigured frequency and bus width.
>>>>>>>>
>>>>>>>> When U-boot is accessing the partition, it has already gone through
>>>>>>>> the Device identification mode and is in data transfer mode
>>>>>>>> (i.e. it
>>>>>>>> needs to send commands for read/write to happen). In this case, we
>>>>>>>> need to switch the partition in Extended CSD to access the boot
>>>>>>>> partition (Section 6.2.5). The spec doesn't say anything about
>>>>>>>> HS200 and
>>>>>> HS400 not being supported here.
>>>>>>> Yes, the spec does not mention this. It only mentions HS200/400 not
>>>>>>> supported during boot operation.
>>>>>>>
>>>>>>>> Also, I don't see linux kernel switching down speed when trying to
>>>>>>>> access a boot partition (unless its being very sneaky about it). So
>>>>>>>> if you are seeing issues with accessing boot partitions at
>>>>>>>> HS200/HS400 then you should probably look at how linux code is
>>>>>>>> working
>>>>>> instead.
>>>>>>> There might be bug in U-Boot code.
>>>>>> So are we gonna leave this inconsistency in for current release or
>>>>>> what's it
>>>>>> gonna be ? Like I said, we're in rc3, it's fine to do bigger
>>>>>> changes in next
>>>>>> release, but we should at least fix this in current release.
>>>>> I'll pick up your patch in this release.
>>>>>
>>>> The issue that Marek is facing is not a regression, is it? Are we
>>>> really
>>>> going to merge something that we know to be wrong just so we are
>>>> consistently wrong?
>>> First of all, you established this is "wrong" without any real backing
>>> except for your interpretation of the specification. I would still like
>>> to hear from Jean the real reason why he added this filtering in the
>>> first place.
>> I think Peng agrees with my interpretation. The backing for it being
>> "right" is also JJ's and your interpretation of spec. The additional
>> justification that I am trying to give is that there is no code to
>> fallback in kernel and I have observed it working in kernel with no
>> issues. I needed your observations (with any HS200/HS400 supporting
>> platform) in kernel for additional data points.
>>
>>> That said, we're in rc4 , the release is just around the corner. I would
>>> like to avoid big changes in the MMC subsystem , or any subsystem for
>>> that matter. That's for next release , and if you have a patch for next,
>>> please post it, I am happy to test it on the hardware I have available.
>> I am not saying we try to fix it before this release. All I am saying is
>> that we don't mask real errors (none of which are regressions) with this
>> "fix" that we are not even sure of.
>>
>>> Also note that this patch does not have any impact on general use case,
>>> the regular bulk of the eMMC can be accessed at HS200/HS400, it's just
>>> the boot partitions which are accessed in HS52 or lower.
>> Exceedingly, the general usecase is to put boot images in boot partition
>> and root filesystem in the user data area. In that case, the boot area
>> is all that will be accessed in SPL at HS52 even if
>> CONFIG_SPL_MMC_HS200/HS400 is enabled.
>>
>>> However, right now, the behavior is not consistent between HS200 and
>>> HS400 modes, and I would very much like to have it consistent in the
>>> upcoming release.
>> I don't think consistency is a big enough reason to make this change. If
>> my interpretation is correct, you would be masking real issues for
>> everyone with this change and any platforms which enable HS400/HS200
>> will be blissfully unaware that they are not accessing data at the
>> required speed. If things are failing for others, we can get their
>> datapoints in kernel as well.
>>
>> Having said that, if the maintainer still wants to pull this fix as is,
>> I would at least change the commit message to reflect our uncertainty
>> here so people are not mislead by this patch.
>>
>>>> Marek, I understand you do not want to debug this right now but this
>>>> patch will 1) Mislead people into thinking that we know what we are
>>>> doing (two patches went in with pretty confident patch descriptions)
>>>> and
>>>> 2) Mask issues for people who could take the time to help debug this.
>>> Wrong, I want to debug this, I just don't want to do big changes in the
>>> upcoming release this late in the release cycle. But if you propose a
>>> patch for next, I am happy to test it on the hardware I have available.
>>> Can you send such a patch ?
>>>
>> Agreed on the no big changes this release. Hopefully we can also agree
>> on not having *this* change in the release either. I do not have a fix
>> yet but plan to look into this next week.
> 
> 
> Have you tried to use the boot partitions with HS200 lately ?
> 
> I'm running a test on a DRA76 and haven't seen any issue. I wonder why
> it didn't work properly when I tested it back then.
> 
> I also rand the same test with Linux and checked that the clock was also
> at 192 MHz
> 
> 
> test context:
> 
> The boot partition (8MB) is accessed in HS200 mode (real freq is
> measured at 192 MHz with a scope)
> 
> The data is fresh random data
> 
> The test command is:
> 
> setenv test_boot_part 'random 0x81000000 0x800000; mmc write 81000000 0
> 0x4000; mmc read 82000000 0 0x4000; cmp.b 81000000 82000000 0x800000' ;
> while run test_boot_part ; do echo -------------; done
> 
> I'll post the patch for the 'random' command.

Do you think you can add a test.py test for this ? Then I can
continuously run this test on the boards I have available. It should
also propagate onto nvidia and xilinx boards, which I think do HS modes too.

> If we could get this tested OK on most of the platforms that support
> HS200, I suggest that we remove this limitation.

Yes, but not this close to the release. It might work on TI board(s),
but it could very well break on other boards which we cannot test. I had
stability issues with those HS200/HS400 modes, so I am seriously
concerned about removing this limitation this close to the release.

I would much rather see a release where the eMMCs work stable, albeit
slower, possibly with a downstream patch in some BSP , then a release
where the eMMCs randomly fail when loading data from them .

I am happy to reword the commit message if that helps ?

-- 
Best regards,
Marek Vasut