[RFC] Change ECC algorithm from userspace
Florian Fainelli
ffainelli at freebox.fr
Mon Oct 31 05:02:35 EDT 2011
Hi,
On Friday 28 October 2011 12:30:51 Matthieu CASTET wrote:
> Hi,
>
> Javier Martinez Canillas a écrit :
> > Hello,
> >
> > I want to be able to use 1-bit ECC for the first partition where I
> > save the loader binary and has to be accessed by the ROM boot but use
> > a 4-bit ECC for my rootfs partition.
> >
> > Does anyone have this same issue?
>
> We use raw programming and compute the ecc in software.
We are doing something similar here as well. Our bootloader also requires the
data to be layed out differently (data + ecc interleaved inside a page + oob).
>
> > What is the best approach to store data in a NAND device using
> > different ECC techniques?
> >
> > I've think of two approaches:
> >
> > 1- Adding an ioctl to mtdchar (something like ECCSETBITS) to change
> > the ECC technique used.
>
> But this won't work if there is concurrent acess to mtd. One program may
> want 1 bit ecc other want 4 bits ecc.
>
> > 2- Use a platform data field to notify the omap2 nand driver that the
> > ROM boot only supports 1-bit ECC. So it can use a 1-bit ECC to write
> > and read the first 4 sectors but a 4-bit ECC for the rest.
>
> This may be better.
Would not it better to add infrastructure for allowing per-partition ECC
scheme? This should allow the kernel to also be able to properly handle the
bootloader partitions (bad-block scanning ...).
>
> Matthieu
>
> PS : note that some OMAP ROM support a better protection than Hamming (but
> the details are not public AFAIK)
>
> From OMAP34xx Multimedia Device, Silicon Revision 3.1.x, public version :
>
> Pages can contain errors caused by memory alteration. To correct these
> errors, the ROM code uses ECC,
> based on Hamming codes for SLC NAND and BCH (Bose, Ray-Chaudhuri,
> Hocquenghem) code for
> multilevel cell (MLC) devices. The computed ECC is compared to ECC stored in
> the spare area of the
> corresponding page. If there are uncorrectable errors, the ROM code returns
> with FAIL.
--
Florian
More information about the linux-mtd
mailing list