Re: [PATCH 1/3] mtd: nand: omap: Revert to using software ECC by default

[Roger Quadros <rogerq@ti.com>]

On 08/07/2014 01:55 AM, Grazvydas Ignotas wrote:
> On Wed, Aug 6, 2014 at 11:02 AM, Roger Quadros <rogerq@ti.com> wrote:
>> Hi Gražvydas,
>>
>> On 08/05/2014 07:15 PM, Grazvydas Ignotas wrote:
>>> On Tue, Aug 5, 2014 at 1:11 PM, Roger Quadros <rogerq@ti.com> wrote:
>>>> For v3.12 and prior, 1-bit Hamming code ECC via software was the
>>>> default choice. Commit c66d039197e4 in v3.13 changed the behaviour
>>>> to use 1-bit Hamming code via Hardware using a different ECC layout
>>>> i.e. (ROM code layout) than what is used by software ECC.
>>>>
>>>> This ECC layout change causes NAND filesystems created in v3.12
>>>> and prior to be unusable in v3.13 and later. So revert back to
>>>> using software ECC by default if an ECC scheme is not explicitely
>>>> specified.
>>>>
>>>> This defect can be observed on the following boards during legacy boot
>>>>
>>>> -omap3beagle
>>>> -omap3touchbook
>>>> -overo
>>>> -am3517crane
>>>> -devkit8000
>>>> -ldp
>>>> -3430sdp
>>>
>>> omap3pandora is also using sw ecc, with ubifs. Some time ago I tried
>>> booting mainline (I think it was 3.14) with rootfs on NAND, and while
>>> it did boot and reached a shell, there were lots of ubifs errors, fs
>>> got corrupted and I lost all my data. I used to be able to boot
>>> mainline this way fine sometime ~3.8 release. It's interesting that
>>> 3.14 was able to read the data, even with wrong ecc setup.
>>
>> This is due to another bug introduced in 3.7 by commit 65b97cf6b8deca3ad7a3e00e8316bb89617190fb.
>> Because of that bug (i.e. inverted CS_MASK in omap_calculate_ecc), omap_calculate_ecc() always fails with -EINVAL and calculated ECC bytes are always 0. I'll be sending a patch to fix that as well. But that will only affect the cases where OMAP_ECC_HAM1_CODE_HW is used which happened for pandora from 3.13 onwards.
>>
>>>
>>> Do you think it's safe again to boot ubifs created on 3.2 after
>>> applying this series?
>>>
>>
>> Yes. If you boot pandora using legacy boot (non DT method), it passes 0 for .ecc_opt in pandora_nand_data. This used to mean OMAP_ECC_HAMMING_CODE_DEFAULT which is software ecc. i.e. NAND_ECC_SOFT with default ECC layout. Until the above mentioned commits changed the meaning. We now call that option OMAP_ECC_HAM1_CODE_SW.
>>
>> Please let me know if it works for you. Thanks.
> 
> Yes it does, thank you.
> Tested-by: Grazvydas Ignotas <notasas@gmail.com>
> 
> Found something new in dmesg though:
> [    1.542755] nand: device found, Manufacturer ID: 0x2c, Chip ID: 0xbc
> [    1.549621] nand: Micron MT29F4G16ABBDA3W
> [    1.553894] nand: 512MiB, SLC, page size: 2048, OOB size: 64
> [    1.560058] nand: WARNING: omap2-nand.0: the ECC used on your
> system is too weak compared to the one required by the NAND chip
> 
> Do you think it's best to migrate to different ECC scheme? It would be
> better to avoid that so that users can freely change kernels and the
> bootloader wouldn't have to be changed..
> 
I'm not sure why these boards were using Software ECC scheme in the first place.
So moving to a better ECC scheme should be considered with a warning that backward
compatibility will be broken.

There is a limitation with the OMAP3 ROM code loader. So if you want uniform ECC scheme
for MLO, u-boot and kernel partitions then we are limited to Hamming code for SLC NAND with
512B, 2KB and 4KB pages.

For MLC NAND, the ROM code uses a proprietary layout using checksum and BCH and I'm not very sure
if this is compatible with the newer OMAP platforms and AM33xx platforms.

For details see OMAP35x TRM. (spruf98y.pdf)
http://www.ti.com/lit/ug/spruf98y/spruf98y.pdf
sections
25.4.7.4.2 SLC NAND Read Sector Procedure
25.4.7.4.3 MLC NAND Read Sector Procedure

cheers,
-roger