[6/7] mtd: rawnand: sunxi: Update OOB layout to match hardware

Message ID 20221229181526.53766-7-samuel@sholland.org
State New
Headers
Series mtd: rawnand: sunxi: Bug fixes and cleanup |

Commit Message

Samuel Holland Dec. 29, 2022, 6:15 p.m. UTC
  When using the hardware ECC engine, the OOB data is made available in
the NFC_REG_USER_DATA registers, one 32-bit word per ECC step. Any
additional bytes are only accessible through raw reads and software
descrambling. For efficiency, and to match the vendor driver, ignore
these extra bytes.

Signed-off-by: Samuel Holland <samuel@sholland.org>
---

 drivers/mtd/nand/raw/sunxi_nand.c | 7 +++++++
 1 file changed, 7 insertions(+)
  

Comments

Miquel Raynal Jan. 2, 2023, 9:21 a.m. UTC | #1
Hi Samuel,

samuel@sholland.org wrote on Thu, 29 Dec 2022 12:15:25 -0600:

> When using the hardware ECC engine, the OOB data is made available in
> the NFC_REG_USER_DATA registers, one 32-bit word per ECC step. Any
> additional bytes are only accessible through raw reads and software
> descrambling. For efficiency, and to match the vendor driver, ignore
> these extra bytes.
> 
> Signed-off-by: Samuel Holland <samuel@sholland.org>
> ---
> 
>  drivers/mtd/nand/raw/sunxi_nand.c | 7 +++++++
>  1 file changed, 7 insertions(+)
> 
> diff --git a/drivers/mtd/nand/raw/sunxi_nand.c b/drivers/mtd/nand/raw/sunxi_nand.c
> index 8e873f4fec9a..a3bc9f7f9e5a 100644
> --- a/drivers/mtd/nand/raw/sunxi_nand.c
> +++ b/drivers/mtd/nand/raw/sunxi_nand.c
> @@ -1604,6 +1604,13 @@ static int sunxi_nand_ooblayout_free(struct mtd_info *mtd, int section,
>  		return 0;
>  	}
>  
> +	/*
> +	 * The controller does not provide access to OOB bytes
> +	 * past the end of the ECC data.
> +	 */
> +	if (section == ecc->steps && ecc->engine_type == NAND_ECC_ENGINE_TYPE_ON_HOST)
> +		return -ERANGE;

Again, I am sorry but I cannot take this change, it would typically
break jffs2 users (if any?) :(

>  	oobregion->offset = section * (ecc->bytes + 4);
>  
>  	if (section < ecc->steps)


Thanks,
Miquèl
  
Samuel Holland Jan. 2, 2023, 4:26 p.m. UTC | #2
Hi Miquèl,

On 1/2/23 03:21, Miquel Raynal wrote:
> Hi Samuel,
> 
> samuel@sholland.org wrote on Thu, 29 Dec 2022 12:15:25 -0600:
> 
>> When using the hardware ECC engine, the OOB data is made available in
>> the NFC_REG_USER_DATA registers, one 32-bit word per ECC step. Any
>> additional bytes are only accessible through raw reads and software
>> descrambling. For efficiency, and to match the vendor driver, ignore
>> these extra bytes.
>>
>> Signed-off-by: Samuel Holland <samuel@sholland.org>
>> ---
>>
>>  drivers/mtd/nand/raw/sunxi_nand.c | 7 +++++++
>>  1 file changed, 7 insertions(+)
>>
>> diff --git a/drivers/mtd/nand/raw/sunxi_nand.c b/drivers/mtd/nand/raw/sunxi_nand.c
>> index 8e873f4fec9a..a3bc9f7f9e5a 100644
>> --- a/drivers/mtd/nand/raw/sunxi_nand.c
>> +++ b/drivers/mtd/nand/raw/sunxi_nand.c
>> @@ -1604,6 +1604,13 @@ static int sunxi_nand_ooblayout_free(struct mtd_info *mtd, int section,
>>  		return 0;
>>  	}
>>  
>> +	/*
>> +	 * The controller does not provide access to OOB bytes
>> +	 * past the end of the ECC data.
>> +	 */
>> +	if (section == ecc->steps && ecc->engine_type == NAND_ECC_ENGINE_TYPE_ON_HOST)
>> +		return -ERANGE;
> 
> Again, I am sorry but I cannot take this change, it would typically
> break jffs2 users (if any?) :(

Considering the bug I fixed in the previous patch, and the fact that
mtd_ooblayout_free() zeroes out the structure before calling the .free
callback, that region was being reported with a length of zero already.
So I don't think anyone could have been using those bytes anyway.

I am looking for a solution here because the ECC/scrambling engine
really provides no way to access these bytes. Reading them requires
turning off the ECC engine, performing another read command, and then
descrambling in software. So we are sort of lying when we claim those
bytes are available with hardware ECC enabled.

If this change cannot be made as-is, is there any way the user could opt
in to the new layout, to get the improved performance?

Regards,
Samuel

>>  	oobregion->offset = section * (ecc->bytes + 4);
>>  
>>  	if (section < ecc->steps)
> 
> 
> Thanks,
> Miquèl
  
Miquel Raynal Jan. 2, 2023, 4:53 p.m. UTC | #3
Hi Samuel,

samuel@sholland.org wrote on Mon, 2 Jan 2023 10:26:48 -0600:

> Hi Miquèl,
> 
> On 1/2/23 03:21, Miquel Raynal wrote:
> > Hi Samuel,
> > 
> > samuel@sholland.org wrote on Thu, 29 Dec 2022 12:15:25 -0600:
> >   
> >> When using the hardware ECC engine, the OOB data is made available in
> >> the NFC_REG_USER_DATA registers, one 32-bit word per ECC step. Any
> >> additional bytes are only accessible through raw reads and software
> >> descrambling. For efficiency, and to match the vendor driver, ignore
> >> these extra bytes.
> >>
> >> Signed-off-by: Samuel Holland <samuel@sholland.org>
> >> ---
> >>
> >>  drivers/mtd/nand/raw/sunxi_nand.c | 7 +++++++
> >>  1 file changed, 7 insertions(+)
> >>
> >> diff --git a/drivers/mtd/nand/raw/sunxi_nand.c b/drivers/mtd/nand/raw/sunxi_nand.c
> >> index 8e873f4fec9a..a3bc9f7f9e5a 100644
> >> --- a/drivers/mtd/nand/raw/sunxi_nand.c
> >> +++ b/drivers/mtd/nand/raw/sunxi_nand.c
> >> @@ -1604,6 +1604,13 @@ static int sunxi_nand_ooblayout_free(struct mtd_info *mtd, int section,
> >>  		return 0;
> >>  	}
> >>  
> >> +	/*
> >> +	 * The controller does not provide access to OOB bytes
> >> +	 * past the end of the ECC data.
> >> +	 */
> >> +	if (section == ecc->steps && ecc->engine_type == NAND_ECC_ENGINE_TYPE_ON_HOST)
> >> +		return -ERANGE;  
> > 
> > Again, I am sorry but I cannot take this change, it would typically
> > break jffs2 users (if any?) :(  
> 
> Considering the bug I fixed in the previous patch, and the fact that
> mtd_ooblayout_free() zeroes out the structure before calling the .free
> callback, that region was being reported with a length of zero already.
> So I don't think anyone could have been using those bytes anyway.
> 
> I am looking for a solution here because the ECC/scrambling engine
> really provides no way to access these bytes. Reading them requires
> turning off the ECC engine, performing another read command, and then
> descrambling in software. So we are sort of lying when we claim those
> bytes are available with hardware ECC enabled.
> 
> If this change cannot be made as-is, is there any way the user could opt
> in to the new layout, to get the improved performance?

Actually that's true, you fixed the reporting of the free area which
was set to 0 until then, which means there cannot be any upstream user.
So knowing that, preventing the accesses to the end of the area seems
acceptable when using HW ECC. Please mention it in the commit log.

Thanks,
Miquèl
  

Patch

diff --git a/drivers/mtd/nand/raw/sunxi_nand.c b/drivers/mtd/nand/raw/sunxi_nand.c
index 8e873f4fec9a..a3bc9f7f9e5a 100644
--- a/drivers/mtd/nand/raw/sunxi_nand.c
+++ b/drivers/mtd/nand/raw/sunxi_nand.c
@@ -1604,6 +1604,13 @@  static int sunxi_nand_ooblayout_free(struct mtd_info *mtd, int section,
 		return 0;
 	}
 
+	/*
+	 * The controller does not provide access to OOB bytes
+	 * past the end of the ECC data.
+	 */
+	if (section == ecc->steps && ecc->engine_type == NAND_ECC_ENGINE_TYPE_ON_HOST)
+		return -ERANGE;
+
 	oobregion->offset = section * (ecc->bytes + 4);
 
 	if (section < ecc->steps)