Re: [PATCH v6 00/17] memory: omap-gpmc: mtd: nand: Support GPMC NAND on non-OMAP platforms

From: Roger Quadros
Date: Tue Apr 19 2016 - 10:50:57 EST


On 19/04/16 17:26, Roger Quadros wrote:
> On 19/04/16 16:22, Boris Brezillon wrote:
>> On Mon, 18 Apr 2016 17:39:01 +0300
>> Roger Quadros <rogerq@xxxxxx> wrote:
>>
>>> On 18/04/16 17:10, Boris Brezillon wrote:
>>>> On Mon, 18 Apr 2016 16:48:26 +0300
>>>> Roger Quadros <rogerq@xxxxxx> wrote:
>>>>
>>>>> Boris,
>>>>>
>>>>> On 18/04/16 16:13, Boris Brezillon wrote:
>>>>>> Hi Roger,
>>>>>>
>>>>>> On Mon, 18 Apr 2016 15:52:58 +0300
>>>>>> Roger Quadros <rogerq@xxxxxx> wrote:
>>>>>>
>>>>>>> On 18/04/16 15:31, Roger Quadros wrote:
>>>>>>>> On 16/04/16 11:57, Boris Brezillon wrote:
>>>>>>>>> On Fri, 15 Apr 2016 09:19:51 -0700
>>>>>>>>> Tony Lindgren <tony@xxxxxxxxxxx> wrote:
>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>> Or should I just pull this immutable branch in my current nand/next and
>>>>>>>>>>> let you pull the same immutable branch in omap-soc. I mean, would this
>>>>>>>>>>> prevent conflicts when our branches are merged into linux-next, no
>>>>>>>>>>> matter the order.
>>>>>>>>>>
>>>>>>>>>> Ideally just one or more branches with just minimal changes in
>>>>>>>>>> them against -rc1. But you may have other dependencies in
>>>>>>>>>> your NAND tree so that may no longer be doable :) Usually if
>>>>>>>>>> I merge something that may need to get merged into other
>>>>>>>>>> branches, I just apply them into a separate branch against -rc1
>>>>>>>>>> to start with, then merge that branch in.
>>>>>>>>>
>>>>>>>>> Okay, in this case, that's pretty much what I did from the beginning,
>>>>>>>>> except the immutable branch was provided by Roger (based on 4.6-rc1).
>>>>>>>>> Thanks for this detailed explanation, I'll try to remember that when
>>>>>>>>> I'll need to provide an immutable branch for another subsystem.
>>>>>>>>>
>>>>>>>>> Roger, my request remains, could you check/test my conflict resolution
>>>>>>>>> (branch nand/next-with-gpmc-rework)?
>>>>>>>>
>>>>>>>> I couldn't test that branch yet as nand/next is broken on omap platforms
>>>>>>>> (at least on dra7-evm).
>>>>>>>>
>>>>>>>> The commit where it breaks is:
>>>>>>>> a662ef4 mtd: nand: omap2: use mtd_ooblayout_xxx() helpers where appropriate
>>>>>>>>
>>>>>>>> I'm trying to figure out what went wrong there. Failure log below.
>>>>>>>
>>>>>>> OK. I was able to fix it when at commit a662ef4 with the below patch.
>>>>>>
>>>>>> Thanks for debugging that.
>>>>>>
>>>>>>>
>>>>>>> Looks like we need to read exactly the ECC bytes through the ECC engine and not
>>>>>>> the entire OOB region.
>>>>>>
>>>>>> Hm, it looks like there's a bug somewhere else, because I don't see any
>>>>>> reason why the controller wouldn't be able to read the full OOB region.
>>>>>
>>>>> The controller can read the full OOB region but we only want it to read just
>>>>> the ECC bytes because that is the way the ELM ECC engine works.
>>>>
>>>> Ok, I think I got it: the ECC correction is pipelined with data read,
>>>> and the controller expect to have ECC bytes right after the in-band
>>>> data, is that correct?
>>>
>>> That is correct.
>>>
>>>>
>>>>>
>>>>>>>
>>>>>>> diff --git a/drivers/mtd/nand/omap2.c b/drivers/mtd/nand/omap2.c
>>>>>>> index e622a1b..46b61d2 100644
>>>>>>> --- a/drivers/mtd/nand/omap2.c
>>>>>>> +++ b/drivers/mtd/nand/omap2.c
>>>>>>> @@ -1547,8 +1547,8 @@ static int omap_read_page_bch(struct mtd_info *mtd, struct nand_chip *chip,
>>>>>>> chip->read_buf(mtd, buf, mtd->writesize);
>>>>>>>
>>>>>>> /* Read oob bytes */
>>>>>>> - chip->cmdfunc(mtd, NAND_CMD_RNDOUT, mtd->writesize, -1);
>>>>>>> - chip->read_buf(mtd, chip->oob_poi, mtd->oobsize);
>>>>>>> + chip->cmdfunc(mtd, NAND_CMD_RNDOUT, mtd->writesize + chip->ecc.layout->eccpos[0], -1);
>>>>>>
>>>>>> The whole point of this series is to get rid of chip->ecc.layout, so
>>>>>> we'd rather use the mtd_ooblayout_find_eccregion() instead of
>>>>>> chip->ecc.layout->eccpos[0].
>>>>>
>>>>> We just need the position of the first ECC byte offset.
>>>>> Is that the most optimal way to get it?
>>>>
>>>> For the BCH case, it seems that ECC bytes always start at offset
>>>> BADBLOCK_MARKER_LENGTH, so you can just pass
>>>> mtd->writesize + BADBLOCK_MARKER_LENGTH.
>>>>
>>>> Let me know if this works, and I'll squash those changes into the
>>>> faulty commit (I know this implies a rebase + push -f, but IMO that's
>>>> better than breaking bisectability).
>>>>
>>>>
>>>
>>> So, the below patch works as well. Please feel free to fold it with your patch.
>>>
>>> --
>>> cheers,
>>> -roger
>>>
>>> diff --git a/drivers/mtd/nand/omap2.c b/drivers/mtd/nand/omap2.c
>>> index e622a1b..eb85d6b 100644
>>> --- a/drivers/mtd/nand/omap2.c
>>> +++ b/drivers/mtd/nand/omap2.c
>>> @@ -1547,8 +1547,8 @@ static int omap_read_page_bch(struct mtd_info *mtd, struct nand_chip *chip,
>>> chip->read_buf(mtd, buf, mtd->writesize);
>>>
>>> /* Read oob bytes */
>>> - chip->cmdfunc(mtd, NAND_CMD_RNDOUT, mtd->writesize, -1);
>>> - chip->read_buf(mtd, chip->oob_poi, mtd->oobsize);
>>> + chip->cmdfunc(mtd, NAND_CMD_RNDOUT, mtd->writesize + BADBLOCK_MARKER_LENGTH, -1);
>>> + chip->read_buf(mtd, chip->oob_poi, chip->ecc.total);
>>
>> Are you sure this patch works? Cause it seems to me that it should be
>
> For my limited test case yes. My test case involves reading an existing ubifs partition
> and creating a new one and then reading it back using an older kernel.
>
>>
>> chip->read_buf(mtd, chip->oob_poi + BADBLOCK_MARKER_LENGTH,
>> chip->ecc.total);
>>
>
> You are right. Else we'd have wrong OOB data during reads.

Strange, but nanddump is showing the correct OOB data with or without this
change. I'm puzzled.

--
cheers,
-roger