Message-ID: <CAPDyKFq5st=eBW98A1VUjnGwNbocay6gf-7trcUXbUNbUThSBA@mail.gmail.com>
Date:   Tue, 28 Mar 2017 11:58:07 +0200
From:   Ulf Hansson <ulf.hansson@...aro.org>
To:     Adrian Hunter <adrian.hunter@...el.com>
Cc:     Chaotian Jing <chaotian.jing@...iatek.com>,
        Matthias Brugger <matthias.bgg@...il.com>,
        Jaehoon Chung <jh80.chung@...sung.com>,
        Shawn Lin <shawn.lin@...k-chips.com>,
        Masahiro Yamada <yamada.masahiro@...ionext.com>,
        "linux-mmc@...r.kernel.org" <linux-mmc@...r.kernel.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        "linux-arm-kernel@...ts.infradead.org" 
        <linux-arm-kernel@...ts.infradead.org>,
        linux-mediatek@...ts.infradead.org,
        srv_heupstream <srv_heupstream@...iatek.com>
Subject: Re: [PATCH] mmc: core: Do not hold re-tuning during CMD6 commands

On 28 March 2017 at 11:01, Adrian Hunter <adrian.hunter@...el.com> wrote:
> On 28/03/17 11:30, Ulf Hansson wrote:
>> [...]
>>
>>>>
>>>> If there is a problem in __mmc_switch(), let's try to fix it there first.
>>>>
>>> Anyway, it is a bug of retry 3 times at max but without check current
>>> card status and ensure it's in transfer state before next retry.
>>
>> Correct. Do you want to send a patch that fixes this? Otherwise I can do it...
>>
>>>>>>> I think the purpose of "re-tune" is trying to cover particular case(eg.
>>>>>>> voltage fluctuate or EMI or some glitch of host/device which caused CRC
>>>>>>> error)
>>>>>>
>>>>>> No, re-tuning is to compensate for drift caused primarily by temperature change.
>>>>>>
>>>>> Yes, by JEDEC spec, temperature change cause timing drift of EMMC
>>>>> device, but, as you mentioned, maybe I have a hardware problem of host,
>>>>> but needs Software to cover it. so that we are doing our best to do
>>>>> re-tune if got CRC error. if could recover it, then  it's better than
>>>>> system hung.
>>>>
>>>> Exactly in what cases do you get CRC errors for CMD6. We need a full
>>>> cmd log to understand and to help.
>>>>
>>>>>>> error) , but in such cases, too many cases are disable re-tune function
>>>>>>> by mmc_retune_hold(), for example, in this case, if a response CRC error
>>>>>>> got then we never have chance to recover it. then cause system cannot
>>>>>>> access emmc or suspend/resume fail.
>>>>>>
>>>>>> Maybe you have a hardware problem.
>>>>
>>>> There is no way I am going to accept patches touching this part of the
>>>> mmc core, without providing real evidence for how it solves a problem.
>>>> To me, it seems like you are applying a workaround for another issue.
>>>>
>>>> Again, try to provide us with some more data and logs, then perhaps we
>>>> can help narrow down the issues.
>>>>
>>>> Kind regards
>>>> Uffe
>>>
>>> Below is the fail log of suspend fail.
>>> the normal command tune result should be 0xffffff9ff, but some time, we
>>> get the tune result of 0xffffffff, then we choose the 10 as the best
>>> tune parameter, which is not stable.
>>> I know that we should focus on why we get the result of 0xffffffff, this
>>> may be result of device/host timing shifting while tuning. but what I
>>> want to do is that when get a response CRC error, we can do re-tune to
>>> recovery it, but not only return the -84 and cause suspend fail
>>> eventually. if all hardware are perfect, then we don't need the re-tune
>>> mechanism.
>>
>> Thanks for elaborating!
>>
>> Can you please also tell exactly which of the CMD6 commands in the
>> suspend sequence that is triggering this problem? Cache flush? Power
>> off notification?
>>
>>>
>>> as Adrian's comment, if temperature change at here caused CMD6 response
>>> CRC error, then how to recovery it ?
>>
>> So in your case, allowing re-tuning a little longer in __mmc_switch()
>> solves your problem. Clearly there are cases when we need to prevent
>> re-tuning when sending CMD6, however maybe not in all cases as we do
>> today.
>>
>> For example it seems reasonable to not hold retuning before sending
>> CMD6 for cache flush, but instead it should be sufficient to hold it
>> before polling for busy in __mmc_switch().
>>
>> Adrian, what's your thoughts on this?
>
> mmc_retune_hold() and mmc_retune_release() are designed to go around a group
> of commands, but re-tuning can still be done before the first command. i.e.
>
>         mmc_retune_hold
>         <re-tune can happen here>
>         cmd A
>         <re-tune not allowed here>
>         cmd B
>         <re-tune not allowed here>
>         cmd C
>         mmc_retune_release
>
> That is the same in the retry case
>
>         mmc_retune_hold
>         <re-tune can happen here>
>         cmd A
>         <re-tune not allowed here>
>         retry cmd A
>         <re-tune not allowed here>
>         cmd B
>         <re-tune not allowed here>
>         cmd C
>         mmc_retune_release

Ohh, right! Thanks for the detailed description.

>
> The retry mechanism provided by mmc_wait_for_cmd() and friends really only
> makes sense for simple commands.  In other cases, like this, we need to
> consider what state the card is in.  For __mmc_switch we need to consider
> whether the card is busy or whether a timing change been made.

I definitely agree. We should remove retries for CMD6 and perhaps also
for some other cases.

When we have changed the above in __mmc_switch(), the change Chaotian
suggest gets a different impact, as it would potentially allow a
re-tuning to happen before the next CMD1to poll for busy or to check
the switch status. This isn't okay.

This all sounds to me that Chaotian's issue may not all be related to
tuning, but to the CMD6 switch sequence itself. However I may be wrong
- of course. :-)

[...]

Kind regards
Uffe

Powered by blists - more mailing lists