Re: [PATCH v4 4/6] mtd: rawnand: add NVIDIA Tegra NAND Flash controller driver

[Stefan Agner <stefan@agner.ch>]

On 12.06.2018 10:13, Boris Brezillon wrote:
> On Tue, 12 Jun 2018 10:02:12 +0200
> Stefan Agner <stefan@agner.ch> wrote:
> 
>> >> +static int tegra_nand_read_page_hwecc(struct mtd_info *mtd,
>> >> +				      struct nand_chip *chip,
>> >> +				      uint8_t *buf, int oob_required, int page)
>> >> +{
>> >> +	struct tegra_nand_controller *ctrl = to_tegra_ctrl(chip->controller);
>> >> +	struct tegra_nand_chip *nand = to_tegra_chip(chip);
>> >> +	void *oob_buf = oob_required ? chip->oob_poi : 0;
>> >> +	u32 dec_stat, max_corr_cnt;
>> >> +	unsigned long fail_sec_flag;
>> >> +	int ret;
>> >> +
>> >> +	tegra_nand_hw_ecc(ctrl, chip, true);
>> >> +	ret = tegra_nand_page_xfer(mtd, chip, buf, oob_buf, nand->tag.length,
>> >> +				   page, true);
>> >> +	tegra_nand_hw_ecc(ctrl, chip, false);
>> >> +	if (ret)
>> >> +		return ret;
>> >> +
>> >> +	/* No correctable or un-correctable errors, page must have 0 bitflips */
>> >> +	if (!ctrl->last_read_error)
>> >> +		return 0;
>> >> +
>> >> +	/*
>> >> +	 * Correctable or un-correctable errors occurred. Use DEC_STAT_BUF
>> >> +	 * which contains information for all ECC selections.
>> >> +	 *
>> >> +	 * Note that since we do not use Command Queues DEC_RESULT does not
>> >> +	 * state the number of pages we can read from the DEC_STAT_BUF. But
>> >> +	 * since CORRFAIL_ERR did occur during page read we do have a valid
>> >> +	 * result in DEC_STAT_BUF.
>> >> +	 */
>> >> +	ctrl->last_read_error = false;
>> >> +	dec_stat = readl_relaxed(ctrl->regs + DEC_STAT_BUF);
>> >> +
>> >> +	fail_sec_flag = (dec_stat & DEC_STAT_BUF_FAIL_SEC_FLAG_MASK) >>
>> >> +			DEC_STAT_BUF_FAIL_SEC_FLAG_SHIFT;
>> >> +
>> >> +	max_corr_cnt = (dec_stat & DEC_STAT_BUF_MAX_CORR_CNT_MASK) >>
>> >> +		       DEC_STAT_BUF_MAX_CORR_CNT_SHIFT;
>> >> +
>> >> +	if (fail_sec_flag) {
>> >> +		int bit, max_bitflips = 0;
>> >> +
>> >> +		/*
>> >> +		 * Check if all sectors in a page failed. If only some failed
>> >> +		 * its definitly not an erased page and we can return error
>> >> +		 * stats right away.
>> >> +		 *
>> >> +		 * E.g. controller might return fail_sec_flag with 0x4, which
>> >> +		 * would mean only the third sector failed to correct.
> 
> That works because you have NAND_NO_SUBPAGE_WRITE set (i.e. no partial
> page programming), probably something you should state here.
> 

Ok, will add a note.

>> >> +		 */
>> >> +		if (fail_sec_flag ^ GENMASK(chip->ecc.steps - 1, 0)) {
>> >> +			mtd->ecc_stats.failed += hweight8(fail_sec_flag);
>> >> +			return max_corr_cnt;
>> >> +		}
>> >> +
>> >> +		/*
>> >> +		 * All sectors failed to correct, but the ECC isn't smart
>> >> +		 * enough to figure out if a page is really completely erased.
>> >> +		 * We check the read data here to figure out if it's a
>> >> +		 * legitimate ECC error or only an erased page.
>> >> +		 */
>> >> +		for_each_set_bit(bit, &fail_sec_flag, chip->ecc.steps) {
>> >> +			u8 *data = buf + (chip->ecc.size * bit);
>> >> +
>> >> +			ret = nand_check_erased_ecc_chunk(data, chip->ecc.size,
>> >> +							  NULL, 0,
> 
> You should also check that the ECC bytes are 0xff here, otherwise you
> won't detect corruption of pages almost filled 0xff but with a few bits
> set to 0.
> 
> When you use nand_check_erased_ecc_chunk(), it's important to always
> pass the data along with its associated ECC bytes.
> 

Hm, I see this is important in case bitflips accumulate in OOB area
only.

>> >> +							  NULL, 0,
> 
> If you support writing extra OOB bytes, you should also pass them here.
> 

I see. OOB bytes handled together with the last subpage.

>> >> +							  chip->ecc.strength);
>> >> +			if (ret < 0)
>> >> +				mtd->ecc_stats.failed++;
>> >> +			else
>> >> +				max_bitflips = max(ret, max_bitflips);

I guess I should also increment ecc_stats.corrected here.

Is it correct that I increment for every step?

So if I have an ECC strength of 16, an empty page could have 8 bitflips
in the first step, and 12 in the second, I would increment
mtd->ecc_stats.corrected by 20 but return 12 (maximum number of bitflips
per step)?

--
Stefan

>> >> +		}
>> >> +
>> >> +		return max_t(unsigned int, max_corr_cnt, max_bitflips);
>> >> +	} else {
>> >> +		int corr_sec_flag;
>> >> +
>> >> +		corr_sec_flag = (dec_stat & DEC_STAT_BUF_CORR_SEC_FLAG_MASK) >>
>> >> +				DEC_STAT_BUF_CORR_SEC_FLAG_SHIFT;
>> >> +
>> >> +		/*
>> >> +		 * The value returned in the register is the maximum of
>> >> +		 * bitflips encountered in any of the ECC regions. As there is
>> >> +		 * no way to get the number of bitflips in a specific regions
>> >> +		 * we are not able to deliver correct stats but instead
>> >> +		 * overestimate the number of corrected bitflips by assuming
>> >> +		 * that all regions where errors have been corrected
>> >> +		 * encountered the maximum number of bitflips.
>> >> +		 */
>> >> +		mtd->ecc_stats.corrected += max_corr_cnt * hweight8(corr_sec_flag);
>> >> +
>> >> +		return max_corr_cnt;
>> >> +	}
>> >> +