Message-ID: <c216f131-6f83-c9c9-9d17-8d44ec06972d@nvidia.com>
Date:   Wed, 4 Mar 2020 09:26:06 -0800
From:   Sowjanya Komatineni <skomatineni@...dia.com>
To:     Ulf Hansson <ulf.hansson@...aro.org>
CC:     Jon Hunter <jonathanh@...dia.com>,
        Bitan Biswas <bbiswas@...dia.com>,
        Adrian Hunter <adrian.hunter@...el.com>,
        Naresh Kamboju <naresh.kamboju@...aro.org>,
        Jens Axboe <axboe@...nel.dk>,
        Alexei Starovoitov <ast@...nel.org>,
        linux-block <linux-block@...r.kernel.org>,
        <lkft-triage@...ts.linaro.org>,
        open list <linux-kernel@...r.kernel.org>,
        "linux-mmc@...r.kernel.org" <linux-mmc@...r.kernel.org>,
        Arnd Bergmann <arnd@...db.de>,
        John Stultz <john.stultz@...aro.org>,
        Faiz Abbas <faiz_abbas@...com>,
        Thierry Reding <treding@...dia.com>,
        Anders Roxell <anders.roxell@...aro.org>,
        Kishon <kishon@...com>
Subject: Re: LKFT: arm x15: mmc1: cache flush error -110


On 3/4/20 9:21 AM, Sowjanya Komatineni wrote:
>
> On 3/4/20 8:56 AM, Sowjanya Komatineni wrote:
>>
>> On 3/4/20 2:18 AM, Ulf Hansson wrote:
>>> External email: Use caution opening links or attachments
>>>
>>>
>>> [...]
>>>
>>>> So, from my side, me and Anders Roxell, have been collaborating on
>>>> testing the behaviour on a TI Beagleboard x15 (remotely with limited
>>>> debug options), which is using the sdhci-omap variant. I am trying to
>>>> get hold of an Nvidia jetson-TX2, but not found one yet. These are the
>>>> conclusions from the observed behaviour on the Beagleboard for the
>>>> CMD6 cache flush command.
>>>>
>>>> First, the reported host->max_busy_timeout is 2581 (ms) for the
>>>> sdhci-omap driver in this configuration.
>>>>
>>>> 1. As we all know by now, the cache flush command (CMD6) fails with
>>>> -110 currently. This is when MMC_CACHE_FLUSH_TIMEOUT_MS is set to 30 *
>>>> 1000 (30s), which means __mmc_switch() drops the MMC_RSP_BUSY flag
>>>> from the command.
>>>>
>>>> 2. Changing the MMC_CACHE_FLUSH_TIMEOUT_MS to 2000 (2s), means that
>>>> the MMC_RSP_BUSY flag becomes set by __mmc_switch, because of the
>>>> timeout_ms parameter is less than max_busy_timeout (2000 < 2581).
>>>> Then everything works fine.
>>>>
>>>> 3. Updating the code to again use 30s as the
>>>> MMC_CACHE_FLUSH_TIMEOUT_MS, but instead forcing the MMC_RSP_BUSY to be
>>>> set, even when the timeout_ms becomes greater than max_busy_timeout.
>>>> This also works fine.
>>>>
>>>> Clearly this indicates a problem that I think needs to be addressed in
>>>> the sdhci driver. However, of course I can revert the three discussed
>>>> patches to fix the problem, but that would only hide the issues and I
>>>> am sure we would then get back to this issue, sooner or later.
>>>>
>>>> To fix the problem in the sdhci driver, I would appreciate if someone
>>>> from TI and Nvidia can step in to help, as I don't have the HW on my
>>>> desk.
>>>>
>>>> Comments or other ideas of how to move forward?
>>> [...]
>>>
>>>> Hi Ulf,
>>>>
>>>> I could repro during suspend on Jetson TX1/TX2 as when it does mmc 
>>>> flush cache.
>>> Okay, great.
>>>
>>>>
>>>> Timeout I see is for switch status CMD13 after sending CMD6 as 
>>>> device side CMD6 is still inflight while host sends CMD13 as we are 
>>>> using R1 response type with timeout_ms changes to 30s.
>>>>
>>>>
>>>>
>>>> Earlier we used timeout_ms of 0 for CMD6 flush cache, and with it 
>>>> uses R1B response type and host will wait for busy state followed 
>>>> by response from device for CMD6 and then data lines go High.
>>>>
>>>>
>>>>
>>>> Now with timeout_ms changed to 30s, we use R1 response and SW waits 
>>>> for busy by checking for DAT0 line to go High.
>>> If I understand correctly, because of the timeout now set to 30s,
>>> MMC_RSP_BUSY becomes disabled in __mmc_switch() for your case in
>>> sdhci-tegra as well?
>> Yes
>>>
>>> In other words, mmc_poll_for_busy() is being called, which in your
>>> case means the ->card_busy() host ops (set to sdhci_card_busy() in
>>> your case) will be invoked to wait for the card to stop signal busy on
>>> DAT0.
>>>
>>> This indicates to me, that the ->card_busy() ops returns zero to
>>> inform that the card is *not* busy, even if the card actually signals
>>> busy? Is that correct?
>> Yes
>>>
>>>>
>>>>
>>>> With R1B type, host design after sending command at end of 
>>>> completion after end bit waits for 2 cycles for data line to go low 
>>>> (busy state from device) and waits for response cycles after which 
>>>> data lines will go back high and then we issue switch status CMD13.
>>>>
>>>>
>>>>
>>>> With R1 type, host after sending command and at end of completion 
>>>> after end bit, DATA lines will go high immediately as its R1 type 
>>>> and switch status CMD13 gets issued but by this time it looks like 
>>>> CMD6 on device side is still in flight for sending status and data.
>>> So, yes, using R1 instead of R1B triggers a different behaviour, but
>>> according to the eMMC spec it's perfectly allowed to issue a CMD13
>>> even if the card signals busy on DAT0. The CMD13 is not using the DATA
>>> lines, so this should work.
>>>
>>> If I understand correctly, your driver (and controller?) has issues
>>> with coping with this scenario. Is it something that can be fixed?
>>>
>>>>
>>>> 30s timeout is the wait time for data0 line to go high and 
>>>> mmc_busy_status will return success right away with R1 response 
>>>> type and SW sends switch status CMD13 but during that time on 
>>>> device side looks like still processing CMD6 as we are not waiting 
>>>> for enough time when we use R1 response type.
>>> Right, as stated above, isn't sdhci_card_busy() working for your case?
>>> Can we fix it?
>>
>> sdhci_card_busy() returned 0 indicating its not busy.
>>
>> Based on our host design, When CMD6 is issued with R1 type, we 
>> program it as NO_RESPONSE and with this command complete interrupt 
>> happens right at end bit of command and there will be no transfer 
>> complete interrupt.
> *[Correction] Based on our host design, When CMD6 is issued with R1 
> type as we program it as NO_RESPONSE and with this command complete 
> interrupt happens right at end bit of command and there will be no 
> transfer complete interrupt.

Sorry to correct wordings, I meant sdhci driver programs response type 
as NO_RESPONSE for CMD6.

When CMD6 is issued with R1 type and as NO_RESPONSE, Based on our host 
design  command complete interrupt happens right at end bit of command 
and there will be no transfer complete interrupt.


>>
>> When CMD6 is issued with R1B type, we program is as R1B RESP_SHORT 
>> and with this command complete is end bit of device resp and transfer 
>> complete interrupt will be when DAT0 LOW -> HIGH.
>>
>> Regardless of R1/R1B, device side CMD6 will always have busy state on 
>> D0 and response on CMD lines.
>>
>> There will be 2 clock cycles period after sending CMD6 for device to 
>> send busy state on data0.
>>
>> In case of R1 type, after sending command DAT will stay high and 
>> looks like we are polling for busy early before busy state has 
>> started and sending CMD13 while device is busy and sending response 
>> on CMD line is causing timeout.
>>
>> Probably with this specific case of CMD6 with R1 type, to wait for 
>> card busy we should poll for DAT0 to go Low first and then to go High??
>>
>>>
>>>>
>>>>
>>>>
>>>> Actually we always use R1B with CMD6 as per spec.
>>> I fully agree that R1B is preferable, but it's not against the spec to
>>> send CMD13 to poll for busy.
>>>
>>> Moreover, we need to cope with the scenario when the host has
>>> specified a maximum timeout that isn't sufficiently long enough for
>>> the requested operation. Do you have another proposal for how to
>>> manage this, but disabling MMC_RSP_BUSY?
>>>
>>> Let's assume you driver would get a R1B for the CMD6 (we force it),
>>> then what timeout would the driver be using if we would set
>>> cmd.busy_timeout to 30ms?
>>>
>>> Kind regards
>>> Uffe

Powered by blists - more mailing lists