[U-Boot] [PATCH 07/12 v2] ARM: OMAP3: Add memory and syslib common files, add NAND support

[Dirk Behme]

Scott Wood wrote:
> On Tue, Oct 07, 2008 at 06:25:11AM -0500, Nishanth Menon wrote:
> 
>>Dirk Behme said the following on 10/07/2008 04:42 AM:
>>
>>>It doesn't differ ;)
>>>
>>>So I removed this and tried to use default nand_read_buf16() instead:
>>>
>>>nand->read_buf = nand_read_buf16;
>>>
>>>in board_nand_init(). But this doesn't compile as nand_read_buf16() is
>>>static in nand_base.c. How do I use it in OMAP3 NAND driver? Marked it
>>>as FIXME in patch.
>>
>>Probably does not need an explicit initialization, mtd nand_scan should
>>populate that.
> 
> 
> Correct, NULL methods will be filled in with defaults if applicable.

ok, will do so, this is easy :)

>>>>>+/*
>>>>>+ *  omap_calculate_ecc - Generate non-inverted ECC bytes.
>>>>>+ *
>>>>>+ *  Using noninverted ECC can be considered ugly since writing a blank
>>>>>+ *  page ie. padding will clear the ECC bytes. This is no problem as
>>>>>+ *  long nobody is trying to write data on the seemingly unused page.
>>>>>+ *  Reading an erased page will produce an ECC mismatch between
>>>>>+ *  generated and read ECC bytes that has to be dealt with separately.
>>>>
>>>>
>>>>Is this a hardware limitation?  If so, say so in the comment.  If not,
>>>>why do it this way?
>>>
>>>Don't know.
>>>
>>>All: Any help?
>>
>>The issue is simple: assume we read a page of 0xFF's(fresh erased), IF
>>using H/w ECC engine within GPMC, the result of read will be 0x0 while
>>the ECC offsets of the spare area will be 0xFF which will result in an
>>ECC mismatch.
> 
> Right, I'd just like to see an explicit statement that this is the only way
> to do HW ECC that the hardware supports (following verification of that
> fact, of course), along with a pointer to where in the code the ECC error
> when reading an empty page is dealt with.

Will add Nishanth's explanation to comment and check code for this.

>>>>>+    .eccbytes = 12,
>>>>>+    .eccpos = {
>>>>>+           2, 3, 4, 5,
>>>>>+           6, 7, 8, 9,
>>>>>+           10, 11, 12, 13},
>>>>>+    .oobfree = { {20, 50} }    /* don't care */
>>>>
>>>>
>>>>Bytes 64-69 of a 64-byte OOB are free?
>>>>What don't we care about?
>>>
>>>+static struct nand_ecclayout hw_nand_oob_64 = {
>>>
>>>Don't know (or understand?).
>>>
>>>All: Any help?
>>
>>I do not get it either.. ECCPOS is in offset bytes, and oobfree should
>>be {.offset=20,.length=44} /*I always hated struct initialization done
>>as above..*/, but then,
> 
> 
> Why not offset 14, length 50?

Seems I need a closer look what we are talking about here ;)

>>>We need to be able to switch ECC at runtime cause some images have to
>>>be written to NAND with HW ECC and some with SW ECC. This depends on
>>>what user (reader) of these parts expect.
>>>
>>>OMAP3 has a boot ROM which is able to read a second level loader
>>>(called x-loader) from NAND and start/execute it. This 2nd level
>>>loader has to be written by U-Boot using HW ECC as ROM code does HW
>>>ECC to read the image. All other parts, e.g. Linux kernel, use SW ECC
>>>as default. For this we have to use SW ECC to write images, then.
>>>
>>>Therefore we add an additional user command in cmd_nand.c to switch
>>>ECC depending on what user wants to write.
>>
>>why not use h/w ecc which rom code understands in kernel and u-boot. H/w
>>ecc will be faster in comparison to doing s/w ecc and there is good
>>support in MTD to do it, then there would be no reason for s/w ecc IMHO..
> 
> 
> Agreed.

As already mentioned in previous post, I think for the the moment we 
have to go with both ways.

To summarize the open points I will look at:

a) Remove unnecessary nand->read_buf init
b) Add comment and check code for HW ecc issue with erased page
c) Fix offset 14, length 50 issue
d) Extend MTD API with a call to switch HW/SW ecc, remove 
platform-specific ifdefs in generic files

Nishanth M.: Would be nice to get some help for these changes. E.g. 
can you say if this HW ecc issue with erased page is handled correctly 
in exisiting code?

Thanks for the help

Dirk