Hardware ECC in NAND flash driver

Hardware ECC in NAND flash driver

[Steven Hein]

Hello,

I'm working on the Linux port to the Samsung S3C2410 PDA controller
(ARM 920T core with many interfaces, including a NAND flash controller,
built-in).  The 2410's NAND flash controller generates ECC codes
(3 bytes ECC per each 512 bytes data).  Are there any plans to
incorporate hardware ECC support into this driver?  Specifically,
I guess that would mean adding "hooks" so the hardware-specific
driver could handle the ECC parts   (I only had to make
a few small changes to nand.c to get this working for the 2410.)
If there's an effort underway in this area, I'd be happy to
participate in the development.

Thanks,
Steve


-- 
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Steve Hein (ssh@sgi.com)              Engineering Diagnostics/Software
Silicon Graphics, Inc.                          
1168 Industrial Blvd.                 Phone: (715) 726-8410
Chippewa Falls, WI 54729              Fax:   (715) 726-6715
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Re: Hardware ECC in NAND flash driver

[Thomas Gleixner]

On Wed, 2002-08-07 at 15:06, Steven Hein wrote:
> Hello,
> 
> I'm working on the Linux port to the Samsung S3C2410 PDA controller
> (ARM 920T core with many interfaces, including a NAND flash controller,
> built-in).  The 2410's NAND flash controller generates ECC codes
> (3 bytes ECC per each 512 bytes data).  Are there any plans to
> incorporate hardware ECC support into this driver?  Specifically,
> I guess that would mean adding "hooks" so the hardware-specific
> driver could handle the ECC parts   (I only had to make
> a few small changes to nand.c to get this working for the 2410.)
> If there's an effort underway in this area, I'd be happy to
> participate in the development.
> 
We thougt about that already, but nobody made an attempt to do so,
because we had no hardware supporting this.
I was thinking about a similar hook as it is used for command function
and dev_ready function. So the hardware ecc support would be located in
the corresponding hardware driver.
If you send me a diff against current CVS, we can find a clean way
together to incorporate it into nand.c

-- 
Thomas 
____________________________________________________
linutronix - competence in embedded & realtime linux
http://www.linutronix.de
mail: tglx@linutronix.de

Re: Hardware ECC in NAND flash driver

[David Woodhouse]

tglx@linutronix.de said:
> We thougt about that already, but nobody made an attempt to do so,
> because we had no hardware supporting this.

Except the DiskOnChip, which probably ought to be using the generic NAND 
driver too.

--
dwmw2

Re: Hardware ECC in NAND flash driver

[Thomas Gleixner]

On Wed, 2002-08-07 at 15:28, David Woodhouse wrote:
> 
> tglx@linutronix.de said:
> > We thougt about that already, but nobody made an attempt to do so,
> > because we had no hardware supporting this.
> 
> Except the DiskOnChip, which probably ought to be using the generic NAND 
> driver too.
Yep, it could be handled the same way.
-- 
Thomas 
____________________________________________________
linutronix - competence in embedded & realtime linux
http://www.linutronix.de
mail: tglx@linutronix.de

Re: Hardware ECC in NAND flash driver

[Thomas Gleixner]

On Thu, 2002-08-08 at 15:50, Steven Hein wrote:
> Hi Thomas,
Hi Steven,
I put this back on the list, because it might be of public interest.

> I basically added 4 fields to the struct nand_chip:
>     eccsize:  size of ECC block  (512 in my case)
>     eccbytes: # bytes ECC per ECC block (3 in my case)
>     calculate_ecc:  pointer to ECC calculation (in my case,it reads HW ECC)
>     correct_data:  pointer to ECC checking/data fix function
> If these fields aren't set, by the hardware-specific driver, they
> all default to the software ECC functions and settings.
I'm not convinced yet, because I have to solve some compability issues.
See below.

> I don't have any experience with how any other Hardware ECC generators
> for NAND flash might work (including DiskOnChip), but the one in the
> S3C2410 is pretty simple.  Before reading/writing data, you write a bit
> to an internal register to reset the ECC generator.  Then, after
> reading/writing the 512 bytes of data, you read a register that contains
> the 3 byte ECC code (this must be read *before* reading/writing OOB data).
3 byte for 512 byte data reduces the recovery probabilty by factor 2.
I'm wondering, why Samsung did it this way. They support 256 byte page
NAND chips and don't care about the ECC for these chips. Also they are
member of the SmartMedia Consortium and the spec's there use definitly 3
byte ECC per 256 Byte, which means 6 Byte per 512 Byte. So you cannot
read/write SmartMediaSpec compatible Cards, if you wan't to use the
onchip ECC generator. 
But maybe they have decided that the ugly SmartMedia DOS-FAT filesystem
is not reliable and give a sh.. on compability :). This is likely the
case as they support booting from NAND  in a way which is definitly not
compatible with SmartMediaDosFAT.

But nevertheless it breaks and influences existing code. But that should
be possible to handle.

> So, I had to make a few changes due to this implementation:
>   1) In nand_read_ecc(), I separated the read of data from read of OOB data
>      and put the calculate_ecc() call between them.  (This means that ECC
>      is calculated/read for blocks that weren't written with ECC--we can't
>      check that until we've read the OOB data).
Should be no problem.

>   2) In nand_write_page(), in order to generate a correct hardware ECC
>      for partial page writes, you need to write the actual data stored
>      in the page (rather than pre-padding with 0xff).
I see your need, but this breaks the datasheet specs. You have to pad
already written data with 0xff or in case of Samsung NAND you can write
only the bytes you want program to the chip and the chip takes care of
0xff padding itself. As far as I understand the physical insides of
NAND, it might lead to erroneous behaviour, if you reprogram contents.  
In that case, which should not happen with JFFS2 or any other
filesystem, which takes care of NAND specific issues, we should skip ECC
for this page in case we are using a 512 byte only hardware ECC
generator.

>   3) For NAND controllers like mine with 512-byte ECC, using ECC with
>      256-byte page NAND chips isn't really practical  (my changes currently
>      don't handle that case very gracefully, but I'm not using chips with
>      256-byte pages).
In this case you could use the software ECC. Your HW driver can specify
this after chipdetection. 

> In general, I don't think the implementation is *too* bad; I think the part
> that needs cleaning up yet is the handling of different ECC block sizes.
> In the cases were there are two operations due to 256 byte ECC in a
> 512 byte page, I put a (mtd->eccsize < 512) check in with the current
> (mtd->oobblock == 512) checks.  Maybe you can think of a better way
> to handle this.
It's definitly not *too* bad. I'm just thinking about a clean way to
seperate the resulting 3byte ECC/ 512 Byte and the standard of 3Byte ECC
/ 256 Byte. 

More comments later today.

-- 
Thomas 
____________________________________________________
linutronix - competence in embedded & realtime linux
http://www.linutronix.de
mail: tglx@linutronix.de

Re: Hardware ECC in NAND flash driver

[Thomas Gleixner]

On Sat, 2002-08-10 at 12:33, Thomas Gleixner wrote:
> More comments later today.
Hi Steven,

back again. I tried several possibilities to deal with all aspects. I
picked up most of your ideas and worked them into the existing code. My
final conclusion is as follows: 

struct nand_chip got following extensions:

calculate_ecc - function for ecc calculation or readback from ecc
hardware, defaults to nand_calculate_ecc for software ecc
correct_data - function for ecc correction, matching to ecc generator
(sw/hw), defaults to nand_correct_data for software ecc
enable_hwecc - function to enable (reset) hardware ecc generator
eccmod - mode of ecc: see constants
eccsize - databytes used per ecc-calculation, set by nand.c depending on
eccmod value

following eccmodes are defined at the moment

#define NAND_ECC_NONE		0 
is forced, if CONFIG_MTD_NAND_ECC is not set
#define NAND_ECC_SOFT		1 
is forced, if CONFIG_MTD_NAND_ECC is set and no hardware ecc is selected
#define NAND_ECC_HW3_256	2 
user selectable, if your ecc hardware supports 3 byte ecc for 256 byte
of data (DOC or passive SmartMedia adaptors)
#define NAND_ECC_HW3_512	3
user selectable, if your ecc hardware supports 3 byte ecc for 512 byte
of data (Samsung S3C2410)

If you select hardware ecc in your driver, you have to take care, that
calculate_ecc, correct_data and enable_hwecc are filled with the
relevant function pointers and eccmod is set to the appropriate type,
before you call nand_scan.

I'm still not willing to write already written data to the device for a
second time. So I decided to skip hardware ecc for that case. If you use
a NAND aware filesystem e.g. JFFS2 this should not happen, otherwise you
will get in trouble anyway. 

The funtion, which does correct_data calculations for the specfic
hardware ecc should be implemented as a general function in nand_ecc.c
to make it usable for similar hardware drivers and to avoid duplicated
code in these drivers.

Please remove the ECC reset out of your command function and move it to
a seperate function, which will be called by enable_hwecc.

Get code from CVS and enjoy. Thanks for your suggestions. The standard
software ecc mode runs on my machine without complains. I tested HW3_256
mode with software ecc routines and a dummy enable_hwecc and found no
problems. Any suggestions and improvements are welcome. 
Happy testing.

-- 
Thomas 
____________________________________________________
linutronix - competence in embedded & realtime linux
http://www.linutronix.de
mail: tglx@linutronix.de

Re: Hardware ECC in NAND flash driver

[Steven Hein]

Thomas Gleixner wrote:

> > I don't have any experience with how any other Hardware ECC generators
> > for NAND flash might work (including DiskOnChip), but the one in the
> > S3C2410 is pretty simple.  Before reading/writing data, you write a bit
> > to an internal register to reset the ECC generator.  Then, after
> > reading/writing the 512 bytes of data, you read a register that contains
> > the 3 byte ECC code (this must be read *before* reading/writing OOB data).
> 3 byte for 512 byte data reduces the recovery probabilty by factor 2.
> I'm wondering, why Samsung did it this way. They support 256 byte page
> NAND chips and don't care about the ECC for these chips. Also they are
> member of the SmartMedia Consortium and the spec's there use definitly 3
> byte ECC per 256 Byte, which means 6 Byte per 512 Byte. So you cannot
> read/write SmartMediaSpec compatible Cards, if you wan't to use the
> onchip ECC generator.
> But maybe they have decided that the ugly SmartMedia DOS-FAT filesystem
> is not reliable and give a sh.. on compability :). This is likely the
> case as they support booting from NAND  in a way which is definitly not
> compatible with SmartMediaDosFAT.
> 
Interesting...... (I didn't have any background on the SmartMedia
spec.).
I just discovered that the next rev. of the Samsung 2410 (rev. 1) no
longer supports NAND parts with 256 byte pages (there was a problem with
nWAIT holding the bus during erase operations or something, so they
redefined the the pin used for 256/512 byte pages to be a R/nB signal
for NAND flash).


> 
> >   2) In nand_write_page(), in order to generate a correct hardware ECC
> >      for partial page writes, you need to write the actual data stored
> >      in the page (rather than pre-padding with 0xff).
> I see your need, but this breaks the datasheet specs. You have to pad
> already written data with 0xff or in case of Samsung NAND you can write
> only the bytes you want program to the chip and the chip takes care of
> 0xff padding itself. As far as I understand the physical insides of
> NAND, it might lead to erroneous behaviour, if you reprogram contents.
> In that case, which should not happen with JFFS2 or any other
> filesystem, which takes care of NAND specific issues, we should skip ECC
> for this page in case we are using a 512 byte only hardware ECC
> generator.

That sounds like the best approach to me also.   (Since I just have a
JFFS2 filesystems and a couple of partitions with a kernel and a
CRAMFS ramdisk, I don't expect to see any partial page programming
happening anyway......)

....As an alternative, if you *really* wanted to do ECC on when you
finally
program to the end of the page, couldn't you....
   -- write the entire 512 bytes out exactly as it should be in Flash
   -- capture/generate the correct hardware ECC
   -- reset (via the SEQIN command) and rewrite the block padded with
0xFF
   -- then issue the PAGE_PROGRAM command
The data doesn't actually get written to the flash until the
PAGE_PROGRAM
command, correct?  This would be slower (as you are writing out the data
twice), but it would allow the correct ECC to be generated.  (This may
not work for all hardware ECC generators I guess).  Again, just a
thought
but not worth pursuing.

> 
> >   3) For NAND controllers like mine with 512-byte ECC, using ECC with
> >      256-byte page NAND chips isn't really practical  (my changes currently
> >      don't handle that case very gracefully, but I'm not using chips with
> >      256-byte pages).
> In this case you could use the software ECC. Your HW driver can specify
> this after chipdetection.

I guess I don't have to worry about this anymore   :)



Thanks,
Steve


-- 
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Steve Hein (ssh@sgi.com)              Engineering Diagnostics/Software
Silicon Graphics, Inc.                          
1168 Industrial Blvd.                 Phone: (715) 726-8410
Chippewa Falls, WI 54729              Fax:   (715) 726-6715
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Re: Hardware ECC in NAND flash driver

[Steven Hein]

Thomas Gleixner wrote:
> 
> On Sat, 2002-08-10 at 12:33, Thomas Gleixner wrote:
> > More comments later today.
> Hi Steven,
> 
> back again. I tried several possibilities to deal with all aspects. I
> picked up most of your ideas and worked them into the existing code. My
> final conclusion is as follows:
> 
> struct nand_chip got following extensions:
> 
> calculate_ecc - function for ecc calculation or readback from ecc
> hardware, defaults to nand_calculate_ecc for software ecc
> correct_data - function for ecc correction, matching to ecc generator
> (sw/hw), defaults to nand_correct_data for software ecc
> enable_hwecc - function to enable (reset) hardware ecc generator
> eccmod - mode of ecc: see constants
> eccsize - databytes used per ecc-calculation, set by nand.c depending on
> eccmod value
> 
> following eccmodes are defined at the moment
> 
> #define NAND_ECC_NONE           0
> is forced, if CONFIG_MTD_NAND_ECC is not set
> #define NAND_ECC_SOFT           1
> is forced, if CONFIG_MTD_NAND_ECC is set and no hardware ecc is selected
> #define NAND_ECC_HW3_256        2
> user selectable, if your ecc hardware supports 3 byte ecc for 256 byte
> of data (DOC or passive SmartMedia adaptors)
> #define NAND_ECC_HW3_512        3
> user selectable, if your ecc hardware supports 3 byte ecc for 512 byte
> of data (Samsung S3C2410)
> 
> If you select hardware ecc in your driver, you have to take care, that
> calculate_ecc, correct_data and enable_hwecc are filled with the
> relevant function pointers and eccmod is set to the appropriate type,
> before you call nand_scan.

This all looks great!

> 
> I'm still not willing to write already written data to the device for a
> second time. So I decided to skip hardware ecc for that case. If you use
> a NAND aware filesystem e.g. JFFS2 this should not happen, otherwise you
> will get in trouble anyway.

Sounds good.

> 
> The funtion, which does correct_data calculations for the specfic
> hardware ecc should be implemented as a general function in nand_ecc.c
> to make it usable for similar hardware drivers and to avoid duplicated
> code in these drivers.

Dumb question--is ECC calculation "generic" enough to be applicable from
one hardware ECC generator to the next, even if they do the same number
of ECC bytes for the same block size?  Is it very likely that the
algorithm
is actually going to be the same between platforms?


> 
> Get code from CVS and enjoy. Thanks for your suggestions. The standard
> software ecc mode runs on my machine without complains. I tested HW3_256
> mode with software ecc routines and a dummy enable_hwecc and found no
> problems. Any suggestions and improvements are welcome.
> Happy testing.
> 
THANKS for adding this!!  I'll try it today/tomorrow.

Another question--how do you typically handle adding new HW drivers
to the NAND flash code?  I know that in my case, the HW implementation
uses access to specific S3C2410 processor registers, and those
register definitions live in include/asm-arm/arch-s3c2410 directory
in the Linux kernel tree.  I would expect that these arch-specific
files would not live in the MTD tree.  Let me know--I'd like to get
this HW driver incorporated into the standard tree as soon as I
get it working!


Thanks again,
Steve

-- 
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Steve Hein (ssh@sgi.com)              Engineering Diagnostics/Software
Silicon Graphics, Inc.                          
1168 Industrial Blvd.                 Phone: (715) 726-8410
Chippewa Falls, WI 54729              Fax:   (715) 726-6715
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~