NAND hw ECC (Was: HZ value change groundings)

Gennady Kupava gb at bsdmn.com
Mon Nov 8 18:13:00 CET 2010


В Пнд, 25/10/2010 в 22:35 +0200, Lars-Peter Clausen пишет:
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
> 
> Gennady Kupava wrote:
> > В Пнд, 25/10/2010 в 18:24 +0300, Vasily Khoruzhick пишет:
> >> On Monday 25 October 2010 18:13:13 Martix wrote:
> >>
> >>> I read on #openmoko-cdevel few months ago that hardware NAND ECC is
> >>> buggy. 
> > 
> > Here is archive of openmoko-cdevel:
> > 
> >> What changed? Was it only rumor or SW bug?
> >>
> >> It's buggy on some s3c2410 CPUs, it's bug free on s3c2442, but ECC layout is a 
> >> bit weird :)
> > 
> > But ECC calculations for HW ECC and SW ECC are same? In-kernel
> > implementation is ok? So... need to investigate how to turn it on.
> > 
> > Gennady.
> > 
> > 
> 
> Pretty simple: boot with hardware_ecc=1
> 
> But I think the ecc layout is different between HW and SW ECC, so you'll probably
> need to turn on HW ECC in u-boot as well.
> 

Thanks for suggestion, i tested it. Seem it works very well.

And yes, it's just bootloader that do not support it. So, i just backup
my qtmoko, enabled ecc in kernel in bootloader, reformat my / partition
on flash, copy qtmoko back, and it works very well. In-kernel support
seem fine, i am using it for 2 days and it's working well.

I did some tests, seem ubifs with zlib, in reading urandom data (same
for each test of course) test speed is increased from 3.9Mb/s to 4.2Mb/s
while using hw ecc.

My plan is to add support of hw ecc to u-boot: may be port from linux
kernel of use some patches for u-boot floating around.

Btw, also i've found that ubifs contains additional crc check, which can
be disabled (while reading only), raising speed from 3.7Ms/s to 3.9Mb/s.
(mount option no_chk_data_crc or kernel option
'rootflags=no_chk_data_crc')
Suggest to remove that crc, as we already have ecc and write verify
turned on.

Such sad read speed in general caused by zlib, according to this tab:
http://www.linux-mtd.infradead.org/misc.html#L_ubifs_compr it should be
much better with lzo, but i can understand value of free space :)
 
regards,
Gennady.





More information about the openmoko-kernel mailing list