ECC errors on padded images

[Thomas Gleixner]

On Sat, 2004-12-04 at 15:33 +0200, Jarkko Lavinen wrote:
> I recently encountered ECC errors with padded JFFS2 images (mkfs.jffs2
> -p) on Nand flash.
> 
> The padded pages were written into flash with a driver which uses
> noninverted ECC code. Pages with all bytes either 00 or FF will
> produce 00s as ECC bytes.
> 
> When JFFS2 scans such padded image, it fails to check the zero ECC
> bytes of seemingly unused pages and will use them. Then the already
> once written page is written again and later an ECC error is produced
> upon read -- because ECC bytes remain zero.
> 
> We have now three options:
> 
> 1. We couuld blame JFFS2, because it fails to check ECC bytes of
>    seemingly unused pages.
> 
>    But should JFFS2 care if somebody produces an image that JFFS2
>    would never produce by itself?
> 
>    JFFS2 itself won't pad erase blocks and won't write unused blank
>    pages.  Even if writing a long node would result writing an all FF
>    page, JFFS2 would know in next mount from node header the page is
>    being used and not available for writing.
> 
> 
> 2. We could blame Mkfs.jffs2, because it is the one who produced
>    padded erase block. Without padding there would be no error.
> 
> 
> 3. We could blame the Ecc logic and insist that writing a blank, all
>    FF page will result ECC code of FFs and it would allow writing
>    first a blank page and then real data page.
> 
>    Conceptually an unused, erased page is not the same as used,
>    written blank page.

4. We could blame you, that you did not read the mail archive and found
the thread discussing this problem and for not using the -p=YOURPAGESIZE
option :)

tglx