[PATCH 2/4] Fine split S3C arch dependencies from generic code

Juergen Beisert jbe at pengutronix.de
Fri May 4 07:13:42 EDT 2012


Alexey Galakhov wrote:
> On 04.05.2012 15:58, Juergen Beisert wrote:
> > Alexey Galakhov wrote:
> >> On 03.05.2012 23:41, Juergen Beisert wrote:
> >>>> create mode 100644 arch/arm/mach-samsung/include/mach/s3c-nand.h
> >>>> delete mode 100644 arch/arm/mach-samsung/include/mach/s3c24xx-nand.h
> >>>
> >>> That is a really bad idea. I just renamed these files at January 2012
> >>> to reflect their CPU (refer b29b8f43d56b62e406349a5cf1ed56f17454c1f7).
> >>> Why do you revert this change again? The NAND controller in the S3C24XX
> >>> CPU is unique to this CPU. The NAND controller in the S3C6410 differs
> >>> from it, and I guess the same is true in the S5 CPU. So, the newer CPUs
> >>> need their own NAND drivers.
> >>>
> >>> What is the sense of renaming these files?
> >>
> >> In fact, exactly the opposite is true. The NAND controller in S5PV210 is
> >> almost exactly the same as in S3C24xx. The only difference is the
> >> numbering of registers. Also S5P suports 1-bit and 4-bit HW ECC while
> >> S3C has only 1-bit.
> >
> > Does the S5P not support 8 bit ECC? That would be a step backwards and
> > would make it useless for recent MLC NAND devices.
>
> It has two ECC modes, SLC and MLC.
>
> The SLC mode is exactly the same as S3C2440. The MLC mode has "4-8-16-32
> bit support". Looks like this is achieved by using the same 4-bit
> hardware calculator.
>
> >> The algorithm is the same, even s3c2440_nand_read_buf() works correctly.
> >
> > These routines are useless for NANDs of type MLC which are the standard
> > today.
>
> According to the datasheet, the only difference between SLC and MLC
> handling from the software point of view is ECC calculation. The block
> reading is exactly the same, just reading the data register multiple
> times. The ECC engine in both SLC and MLC modes works just like S3C2440.
> The only differnce is that one has to set ECC data direction bit prior
> to operation.

AFAIR the MLC ECC generator is different in its handling. At least it is much 
slower than the 1 bit ECC generator and so the routines must wait the result 
to continue on the block data.

> >> S3C6410 has the same controller as well.
> >
> > No it hasn't. Believe me. Using the old 1 bit ECC (and also the 4 bit
> > ECC) makes no sense any more. Even the built-in iROM forces the 8 bit ECC
> > mode if you want to boot it from NAND. And I guess it is the same on the
> > S5P CPU.
>
> Hm, looks like my board is not the case. Anyway, this seems to be just
> adding one more HW specific ECC function and not a complete rewrite.
>
> There is a Linux driver from Samsung, named "s3c_nand.c", found in
> S5PVxx kernel patches here and there (not in mainstream). It works for
> S3C2416 and up and supports SLC and MLC. I use it for reference.

I know a similar driver for the S3C6410 CPU.

> I suggest to support both new and old hardware in the same code. Why
> not? It is 95% the same.

Instead of making the S3c2440 NAND driver more and more complicated (what is 
the benefit of all in one driver?) I would vote for fading it out (as its 
hardware do not change anymore). And creating a new driver for all more 
recent CPUs with MLC support.

My idea was also to not support simple ECC on these newer CPUs anymore. Using 
the 8 bit reed-solomon checksum would be an improvement for SLCs as well (and 
also their OOB sizes are large enough to store the 8 bit checksums). And at 
least if we want to boot from NAND we cannot continue to use simple ECC 
checksums anymore.

Regards,
Juergen

-- 
Pengutronix e.K.                              | Juergen Beisert             |
Linux Solutions for Science and Industry      | http://www.pengutronix.de/  |



More information about the barebox mailing list