mbox series

[0/3] mtd: rawnand: micron: Get the actual number of bitflips

Message ID 20180703122009.29914-1-boris.brezillon@bootlin.com
Headers show
Series mtd: rawnand: micron: Get the actual number of bitflips | expand

Message

Boris Brezillon July 3, 2018, 12:20 p.m. UTC 
Hello,

Back when support for Micron 4bit/512byte on-die ECC was added, we
decided that it wasn't worth retrieving the real number of bitflips
when the chip was returning WRITE_RECOMMENDED and instead decided to
always return the maximum value (ECC strength).

This decision leads to UBI moving eraseblocks around as soon as 1
bitflip is present, which is far from optimal, and might wear the NAND
out faster than if we get the actual number of bitflips by re-reading
the page in raw mode and comparing its content to the corrected
version.

IIRC, Bean warned us about that, but it seems we didn't listen, so now
is time to revisit the implementation and implement what Bean initially
suggested.

This implementation has been tested on an MT29F2G08ABAEAH4, and seems
to work as expected (nandbiterrs works fine, and the real number of
bitflips is now returned).

Here are some details about these patches:
Patch 1 is just a cleanup to avoid passing parameters we don't need to
the ecc_status() functions. Patch 2 is preparing things for the actual
changes by reworking the ordering in the read function, and patch 3 is
implementing the read-in-raw-mode-and-compare logic.

Regards,

Boris

Boris Brezillon (3):
  mtd: rawnand: micron: Stop passing mtd to ecc_status() funcs
  mtd: rawnand: micron: Disable ECC earlier in the read path
  mtd: rawnand: micron: Get the actual number of bitflips

 drivers/mtd/nand/raw/nand_micron.c | 145 ++++++++++++++++++++++++++++++-------
 1 file changed, 120 insertions(+), 25 deletions(-)

Comments

Miquel Raynal July 8, 2018, 9:48 p.m. UTC | #1 
Hi Boris,

Boris Brezillon <boris.brezillon@bootlin.com> wrote on Tue,  3 Jul 2018
14:20:06 +0200:

> Hello,
> 
> Back when support for Micron 4bit/512byte on-die ECC was added, we
> decided that it wasn't worth retrieving the real number of bitflips
> when the chip was returning WRITE_RECOMMENDED and instead decided to
> always return the maximum value (ECC strength).
> 
> This decision leads to UBI moving eraseblocks around as soon as 1
> bitflip is present, which is far from optimal, and might wear the NAND
> out faster than if we get the actual number of bitflips by re-reading
> the page in raw mode and comparing its content to the corrected
> version.
> 
> IIRC, Bean warned us about that, but it seems we didn't listen, so now
> is time to revisit the implementation and implement what Bean initially
> suggested.
> 
> This implementation has been tested on an MT29F2G08ABAEAH4, and seems
> to work as expected (nandbiterrs works fine, and the real number of
> bitflips is now returned).
> 
> Here are some details about these patches:
> Patch 1 is just a cleanup to avoid passing parameters we don't need to
> the ecc_status() functions. Patch 2 is preparing things for the actual
> changes by reworking the ordering in the read function, and patch 3 is
> implementing the read-in-raw-mode-and-compare logic.
> 
> Regards,
> 
> Boris
> 
> Boris Brezillon (3):
>   mtd: rawnand: micron: Stop passing mtd to ecc_status() funcs
>   mtd: rawnand: micron: Disable ECC earlier in the read path
>   mtd: rawnand: micron: Get the actual number of bitflips
> 
>  drivers/mtd/nand/raw/nand_micron.c | 145 ++++++++++++++++++++++++++++++-------
>  1 file changed, 120 insertions(+), 25 deletions(-)
> 


Series applied to nand/next.

Thanks,
Miquèl