Re: [PATCH v2 10/53] mtd: nand: denali: fix erased page checking

From: Masahiro Yamada
Date: Thu Mar 23 2017 - 22:44:09 EST


Hi Boris,


2017-03-23 16:56 GMT+09:00 Boris Brezillon <boris.brezillon@xxxxxxxxxxxxxxxxxx>:
> On Thu, 23 Mar 2017 14:04:44 +0900
> Masahiro Yamada <yamada.masahiro@xxxxxxxxxxxxx> wrote:
>
>> Hi Boris,
>>
>> 2017-03-23 5:56 GMT+09:00 Boris Brezillon <boris.brezillon@xxxxxxxxxxxxxxxxxx>:
>> > On Wed, 22 Mar 2017 23:07:17 +0900
>> > Masahiro Yamada <yamada.masahiro@xxxxxxxxxxxxx> wrote:
>> >> dev_err(denali->dev,
>> >> @@ -1148,12 +1136,15 @@ static int denali_read_page(struct mtd_info *mtd, struct nand_chip *chip,
>> >> if (check_erased_page) {
>> >> read_oob_data(mtd, chip->oob_poi, denali->page);
>> >>
>> >> - /* check ECC failures that may have occurred on erased pages */
>> >> - if (check_erased_page) {
>> >> - if (!is_erased(buf, mtd->writesize))
>> >> - mtd->ecc_stats.failed++;
>> >> - if (!is_erased(buf, mtd->oobsize))
>> >> - mtd->ecc_stats.failed++;
>> >> + stat = nand_check_erased_ecc_chunk(
>> >> + buf, mtd->writesize,
>> >> + chip->oob_poi, mtd->oobsize,
>> >> + NULL, 0,
>> >> + chip->ecc.strength * chip->ecc.steps);
>> >
>> > That's not how it's supposed to be done. Each chunk should be checked
>> > independently. Here is a simple example explaining why this is
>> > important:
>> >
>> > Let's consider the following setup:
>> > - 4k pages
>> > - 16bits/1024bytes ECC
>> >
>> > With your approach, you turn this into:
>> > - 4k pages
>> > - 64bits/4096bytes ECC
>> >
>> > Now suppose you have 32 bitflips in the first 1024 bytes. The real ECC
>> > config is expected to report uncorrectable errors, but your approach
>> > will just report that 32 bits have been fixed, which is wrong.
>>
>>
>> OK. How about adding a helper like follows:
>>
>> static int denali_check_erased_page(struct mtd_info *mtd,
>> struct nand_chip *chip, uint8_t *buf)
>> {
>> uint8_t *ecc_code = chip->buffers->ecccode;
>> int ecc_steps = chip->ecc.steps;
>> int ecc_size = chip->ecc.size;
>> int ecc_bytes = chip->ecc.bytes;
>> int i, ret;
>>
>> ret = mtd_ooblayout_get_eccbytes(mtd, ecc_code, chip->oob_poi, 0,
>> chip->ecc.total);
>> if (ret)
>> return ret;
>>
>> for (i = 0; i < ecc_steps; i++) {
>> ret = nand_check_erased_ecc_chunk(buf, ecc_size,
>> ecc_code, ecc_bytes,
>> NULL, 0,
>> chip->ecc.strength);
>> if (ret < 0)
>> return ret;
>> buf += ecc_size;
>> ecc_code += ecc_bytes;
>> }
>>
>> return 0;
>> }
>>
>>
>>
>> Then,
>>
>> stat = denali_check_erased_page(mtd, chip, buf);
>> if (stat < 0) {
>> mtd->ecc_stats.failed++;
>> /* return 0 for uncorrectable bitflips */
>> stat = 0;
>> }
>
> What's the point of checking all ECC chunks if only one contains ECC
> errors? I really recommend to put the nand_check_erased_ecc_chunk()
> call next to the per-ECC-block correction test.


OK. I can fix it for software ECC fixup.


What should I do for hardware ECC fixup case?
http://patchwork.ozlabs.org/patch/742321/


If at least one ECC sector fails to correct bit-flips,
the controller sets INTR__ECC_UNCOR_ERR flag.


In this case, we can not know the number of uncorrectable errors.

Possible solutions are:

- Increment ecc_stats.failed only by one (compromised solution)

- If the controller IP supports sub-page read,
transfer sectors once again, one by one, checking the register
flag each time.


As far as I see, there are three cases.

[1] SW ECC fixup (Intel)
This can be fixed

[2] HW ECC fixup is supported, but sub-page read is not supported
(old UniPhier SoCs, probably SOCFPGA too)

[3] HW ECC fixup is supported, and sub-page read is supported as well
(new UniPhier SoCs)


I do not know how to precisely increment
ecc_stats.failed and ecc_stats.corrected for [2].


As for [3], we can solve the issue by making more efforts,
but I am not sure this effort is worthwhile.




> Also, mtd->ecc_stats.failed is supposed to be incremented each time an
> uncorrectable error is detected. In your denali_sw_ecc_fixup()
> implementation you can detect errors at the ECC chunk level, so you
> should increment ecc_stats.failed for each failure and not once if at
> least one chunk is faulty.


Yes, I can do this for denali_sw_ecc_fixup().

Can I ask what disadvantage would happen
if ecc_stats.failed / .corrected is incremented only by one,
where actually errors happen in multiple sectors.




--
Best Regards
Masahiro Yamada