Re: [PATCH 4/7] mtd: rawnand: sunxi: Fix ECC strength maximization

[Samuel Holland]

Hi Miquèl,

On 1/2/23 10:45, Miquel Raynal wrote:
>>>> This is already accounted for in the subtraction for OOB, since the BBM
>>>> overlaps the first OOB dword. With this change, the driver picks the
>>>> same ECC strength as the vendor driver.
>>>>
>>>> Fixes: 4796d8655915 ("mtd: nand: sunxi: Support ECC maximization")
>>>> Signed-off-by: Samuel Holland <samuel@xxxxxxxxxxxx>
>>>> ---
>>>>
>>>>  drivers/mtd/nand/raw/sunxi_nand.c | 3 +--
>>>>  1 file changed, 1 insertion(+), 2 deletions(-)
>>>>
>>>> diff --git a/drivers/mtd/nand/raw/sunxi_nand.c b/drivers/mtd/nand/raw/sunxi_nand.c
>>>> index 1bddeb1be66f..1ecf2cee343b 100644
>>>> --- a/drivers/mtd/nand/raw/sunxi_nand.c
>>>> +++ b/drivers/mtd/nand/raw/sunxi_nand.c
>>>> @@ -1643,8 +1643,7 @@ static int sunxi_nand_hw_ecc_ctrl_init(struct nand_chip *nand,
>>>>  		ecc->size = 1024;
>>>>  		nsectors = mtd->writesize / ecc->size;
>>>>  
>>>> -		/* Reserve 2 bytes for the BBM */
>>>> -		bytes = (mtd->oobsize - 2) / nsectors;
>>>> +		bytes = mtd->oobsize / nsectors;  
>>>
>>> I'm sorry but I don't think we can make this work. This change would
>>> break all existing users...  
>>
>> OK, it is not too much of an issue because I can manually specify the
>> ECC parameters in the devicetree. Do you think it makes sense to fix
>> this when adding new hardware variants/compatible strings?
> 
> Actually, looking at the code again, I don't get how the above diff
> could be valid. The "maximize strength" logic (in which this diff is)
> looks for the biggest region to store ECC bytes. These bytes cannot
> be stored on the BBM, which "mtd->oobsize - 2" tries to avoid, so we
> cannot get rid of this.

Right, we cannot overlap the BBM, but the BBM is accounted for in the
line below:

  /* 4 non-ECC bytes are added before each ECC bytes section */
  bytes -= 4;

Normally those 4 bytes are all free OOB, but for the first ECC step,
those are split into 2 free bytes and 2 BBM bytes:

  /*
   * The first 2 bytes are used for BB markers, hence we
   * only have 2 bytes available in the first user data
   * section.
   */
  if (!section && ecc->engine_type == NAND_ECC_ENGINE_TYPE_ON_HOST) {
          oobregion->offset = 2;
          oobregion->length = 2;

          return 0;
  }

So if we subtract 4 bytes for the each free OOB area, including the
first one, and also subtract 2 bytes for the BBM, we are double-counting
the BBM. I should have made my commit message clearer. But I am going to
drop this patch anyway.

Regards,
Samuel