Skip to contentSkip to conversation

SW ECC - double bit flip detection on old NAND devices

DISCUSSION Lukasz Majewski 2 participants 3 messages May 5, 2017 – May 7, 2017 Memory technology devices (MTD)

lore
Share & tools
lore
#0 lore raw

Dear All,

I've a problem with pretty old Flash NAND memory (Samsung 128Mx8) [1]

It doesn't support On-Chip ECC - one needs to calculate ECC manually.

The Yaffs2 FS (for this version) uses "1bit correction ECC" (yaffs_ecc.c). It calculates ECC for 256 bytes -> we have got 22 bits for ECC (rounded up to 3 bytes).

For 2048 bytes page we do have 8 such ECC blocks -> 24 ECC bytes in total in OOB.

This code (as noted in yaffs_ecc.* header) is able to correct one single bit flip.

I've also looked into Linux kernel code for SW ECC calculation:

http://elixir.free-electrons.com/linux/latest/source/drivers/mtd/nand/nand_ecc.c#L523

And here it is also explicitly said that we can correct one bit in such chunk.

Please correct me if I'm wrong but when we have two bit-flips in such 256 bytes chunk, the ECC will be still correct and such obviously broken page will not be "retired".

What one can do to prevent such situation?

My idea, if the above holds, would be to implement better ECC scheme as proposed in "Error Correction Code (ECC) in Micron" doc [2].

Maybe somebody knows better/simpler solution?

Side note: newer NANDs support On-Chip ECC with algorithms allowing correction of up to 4 bits in 512B chunks of data.

[1] -
http://www.sst-ic.com/File/Seriea/PDF/100923163334f47b2ce6-13b7-43d0-9b8c-ec1c75ca2c6f.pdf
[2] -
https://www.micron.com/~/media/documents/products/technical-note/nand-flash/tn2963_ecc_in_slc_nand.pdf

Best regards,

Lukasz Majewski

--

DENX Software Engineering GmbH, Managing Director: Wolfgang Denk
HRB 165235 Munich, Office: Kirchenstr.5, D-82194 Groebenzell, Germany
Phone: (+49)-8142-66989-10 Fax: (+49)-8142-66989-80 Email: [email protected]

#1 lore raw

Lukasz,

Quoted 13 lines · Lukasz Majewski

Dear All,

I've a problem with pretty old Flash NAND memory (Samsung 128Mx8) [1]

It doesn't support On-Chip ECC - one needs to calculate ECC manually.

The Yaffs2 FS (for this version) uses "1bit correction ECC" (yaffs_ecc.c). It calculates ECC for 256 bytes -> we have got 22 bits for ECC (rounded up to 3 bytes).

For 2048 bytes page we do have 8 such ECC blocks -> 24 ECC bytes in total in OOB.

Quoted 6 lines

This code (as noted in yaffs_ecc.* header) is able to correct one single bit flip.

I've also looked into Linux kernel code for SW ECC calculation:

http://elixir.free-electrons.com/linux/latest/source/drivers/mtd/nand/nand_ecc.c#L523

This is for ecc->algo == NAND_ECC_HAMMING.

Quoted 7 lines

And here it is also explicitly said that we can correct one bit in such chunk.

Please correct me if I'm wrong but when we have two bit-flips in such 256 bytes chunk, the ECC will be still correct and such obviously broken page will not be "retired".

Quoted 6 lines

What one can do to prevent such situation?

My idea, if the above holds, would be to implement better ECC scheme as proposed in "Error Correction Code (ECC) in Micron" doc [2].

Maybe somebody knows better/simpler solution?

I'd suggest to use BCH instead of Hamming.
Please see NAND_ECC_BCH.

Signature · 2 lines
Thanks,
//richard
#2 lore raw

Hi Richard,

Thanks for your support.

Quoted 18 lines

Lukasz,

Quoted 15 lines · Lukasz Majewski

Dear All,

I've a problem with pretty old Flash NAND memory (Samsung 128Mx8)
[1]

It doesn't support On-Chip ECC - one needs to calculate ECC manually.

The Yaffs2 FS (for this version) uses "1bit correction ECC" (yaffs_ecc.c). It calculates ECC for 256 bytes -> we have got 22 bits for ECC (rounded up to 3 bytes).

For 2048 bytes page we do have 8 such ECC blocks -> 24 ECC bytes in total in OOB.

Quoted 8 lines
Quoted 6 lines

This code (as noted in yaffs_ecc.* header) is able to correct one single bit flip.

I've also looked into Linux kernel code for SW ECC calculation:

http://elixir.free-electrons.com/linux/latest/source/drivers/mtd/nand/nand_ecc.c#L523

This is for ecc->algo == NAND_ECC_HAMMING.

In my current 2.6.27 kernel it is called:

NAND_ECC_SOFT -> But this is the same.

the

ecc.correct = nand_correct_data()

, which has following statement in the function description:

*
* Detect and correct a 1 bit error for 256 byte block */

So now it is clear that two+ bit flips happening in the 256 B chunk cannot be detected.

Quoted 10 lines
Quoted 10 lines

And here it is also explicitly said that we can correct one bit in such chunk.

Please correct me if I'm wrong but when we have two bit-flips in such 256 bytes chunk, the ECC will be still correct and such obviously broken page will not be "retired".

What one can do to prevent such situation?

Quoted 8 lines
Quoted 5 lines

My idea, if the above holds, would be to implement better ECC scheme as proposed in "Error Correction Code (ECC) in Micron" doc [2].

Maybe somebody knows better/simpler solution?

I'd suggest to use BCH instead of Hamming.
Please see NAND_ECC_BCH.

Thanks for your tip. I will try to backport this feature ...

Quoted 1 line

Best regards,

Lukasz Majewski

--

DENX Software Engineering GmbH, Managing Director: Wolfgang Denk
HRB 165235 Munich, Office: Kirchenstr.5, D-82194 Groebenzell, Germany
Phone: (+49)-8142-66989-10 Fax: (+49)-8142-66989-80 Email: [email protected]