fsl_ifc_nand: are blank pages protected by ECC?

Hi!

We have some problems with fsl_ifc_nand ... in the old kernels, but this one does not seem to be fixed in v4.11, either.

UBIFS complains:

UBIFS error (pid 931): ubifs_scan: corrupt empty space at LEB 282:252630 UBIFS error (pid 931): ubifs_scanned_corruption: corruption at LEB 282:252630 UBIFS error (pid 931): ubifs_scanned_corruption: first 1322 bytes from LEB 282:252630 UBIFS error (pid 931): ubifs_scan: LEB 282 scanning failed

Possible explanation is here:

https://e2e.ti.com/support/dsp/davinci_digital_media_processors/f/716/t/289605

# I see on the forum that this issue has been raised before - my

# understanding is that the omap2 nand driver does not perform ECC
# detection/correction on empty pages so when UBIFS checks the empty
# space data and doesn't read all 0xFF then it fails and mounts
# read-only. I didn't find any good solution - only a workaround to
# remove the UBIFS check..

So I checked fsl_ifc_nand.c in v4.11-rc, and yes, it seems to have the same problem:

			if (errors == 15) {
	                        /*
                                 * Uncorrectable error.
                                 * OK only if the whole page is blank.
                                 *
                                 * We disable ECCER reporting due to...
                                 * erratum IFC-A002770 -- so report it now if we
                                 * see an uncorrectable error in ECCSTAT.
                                 */
                                if (!is_blank(mtd, bufnum))
                                        ctrl->nand_stat |=
                                                IFC_NAND_EVTER_STAT_ECCER;
                                break;
                        }

is_blank() checks for all 0xff's, so single-bit 0xfe in the data will result in_blank() == 0 and uncorrectable error being signaled.

Should the driver be modified somehow?

Quoted 39 lines

We have some problems with fsl_ifc_nand ... in the old kernels, but this one does not seem to be fixed in v4.11, either.

UBIFS complains:

UBIFS error (pid 931): ubifs_scan: corrupt empty space at LEB 282:252630 UBIFS error (pid 931): ubifs_scanned_corruption: corruption at LEB 282:252630 UBIFS error (pid 931): ubifs_scanned_corruption: first 1322 bytes from LEB 282:252630 UBIFS error (pid 931): ubifs_scan: LEB 282 scanning failed

Possible explanation is here:

https://e2e.ti.com/support/dsp/davinci_digital_media_processors/f/716/t/289605

# I see on the forum that this issue has been raised before - my

# understanding is that the omap2 nand driver does not perform ECC
# detection/correction on empty pages so when UBIFS checks the empty
# space data and doesn't read all 0xFF then it fails and mounts
# read-only. I didn't find any good solution - only a workaround to
# remove the UBIFS check..

So I checked fsl_ifc_nand.c in v4.11-rc, and yes, it seems to have the same problem:

			if (errors == 15) {
	                        /*
                                 * Uncorrectable error.
                                 * OK only if the whole page is blank.
                                 *
                                 * We disable ECCER reporting due to...
                                 * erratum IFC-A002770 -- so report it now if we
                                 * see an uncorrectable error in ECCSTAT.
                                 */
                                if (!is_blank(mtd, bufnum))
                                        ctrl->nand_stat |=
                                                IFC_NAND_EVTER_STAT_ECCER;
                                break;
                        }

Quoted 4 lines

is_blank() checks for all 0xff's, so single-bit 0xfe in the data will result in_blank() == 0 and uncorrectable error being signaled.

Should the driver be modified somehow?

Yep, nand_check_erased_ecc_chunk() [1] is here to help you check this case, unfortunately, it's not directly applicable here, because this function takes regular pointers and not __iomem ones. You'll either have to copy the data in an intermediate buffer before calling nand_check_erased_ecc_chunk(), or cast the SRAM region to a void pointer (which is usually not a good idea). The last option would be to open code nand_check_erased_ecc_chunk(), but I'd really like to avoid that (for maintainability concerns).

Hi!

Quoted 44 lines

Quoted 39 lines

We have some problems with fsl_ifc_nand ... in the old kernels, but this one does not seem to be fixed in v4.11, either.

UBIFS complains:

UBIFS error (pid 931): ubifs_scan: corrupt empty space at LEB 282:252630 UBIFS error (pid 931): ubifs_scanned_corruption: corruption at LEB 282:252630 UBIFS error (pid 931): ubifs_scanned_corruption: first 1322 bytes from LEB 282:252630 UBIFS error (pid 931): ubifs_scan: LEB 282 scanning failed

Possible explanation is here:

https://e2e.ti.com/support/dsp/davinci_digital_media_processors/f/716/t/289605

# I see on the forum that this issue has been raised before - my

# understanding is that the omap2 nand driver does not perform ECC
# detection/correction on empty pages so when UBIFS checks the empty
# space data and doesn't read all 0xFF then it fails and mounts
# read-only. I didn't find any good solution - only a workaround to
# remove the UBIFS check..

So I checked fsl_ifc_nand.c in v4.11-rc, and yes, it seems to have the same problem:

			if (errors == 15) {
	                        /*
                                 * Uncorrectable error.
                                 * OK only if the whole page is blank.
                                 *
                                 * We disable ECCER reporting due to...
                                 * erratum IFC-A002770 -- so report it now if we
                                 * see an uncorrectable error in ECCSTAT.
                                 */
                                if (!is_blank(mtd, bufnum))
                                        ctrl->nand_stat |=
                                                IFC_NAND_EVTER_STAT_ECCER;
                                break;
                        }

Quoted 4 lines

is_blank() checks for all 0xff's, so single-bit 0xfe in the data will result in_blank() == 0 and uncorrectable error being signaled.

Should the driver be modified somehow?

Quoted 8 lines

Yep, nand_check_erased_ecc_chunk() [1] is here to help you check this case, unfortunately, it's not directly applicable here, because this function takes regular pointers and not __iomem ones. You'll either have to copy the data in an intermediate buffer before calling nand_check_erased_ecc_chunk(), or cast the SRAM region to a void pointer (which is usually not a good idea). The last option would be to open code nand_check_erased_ecc_chunk(), but I'd really like to avoid that (for maintainability concerns).

Ok, thanks a lot for the pointer, that should be doable.

Core of the code is:

1357 for (; len >= sizeof(long);
1358 len -= sizeof(long), bitmap += sizeof(long)) {
1359 weight = hweight_long(*((unsigned long
*)bitmap));
1360 bitflips += BITS_PER_LONG - weight;
1361 if (unlikely(bitflips > bitflips_threshold))
1362 return -EBADMSG;
1363 }

Someone clearly optimized this code (took care to do long accesses etc), but afaict hweight is quite a heavy operation:

_GLOBAL(__arch_hweight32)
BEGIN_FTR_SECTION

        b __sw_hweight32
        nop
        nop
        nop
        nop
	nop
	nop
FTR_SECTION_ELSE
  BEGIN_FTR_SECTION_NESTED(51)
	PPC_POPCNTB(R3,R3)
	srdi    r4,r3,16
        add     r3,r4,r3
        srdi    r4,r3,8
	add     r3,r4,r3
	clrldi  r3,r3,64-8
	blr
  FTR_SECTION_ELSE_NESTED(51)
	PPC_POPCNTW(R3,R3)
	clrldi  r3,r3,64-8
	blr

ALT_FTR_SECTION_END_NESTED_IFCLR(CPU_FTR_POPCNTD, 51)
ALT_FTR_SECTION_END_IFCLR(CPU_FTR_POPCNTB) EXPORT_SYMBOL(__arch_hweight32)

Would it make sense to only do hweight if *bitmap != ~0ULL ? Would it make sense to only check for bitflips > bitflips_threshold each 128 bytes or something like that?

Quoted 3 lines

Would it make sense to only do hweight if *bitmap != ~0ULL ? Would it make sense to only check for bitflips > bitflips_threshold each 128 bytes or something like that?

I didn't go as far as you did and simply assumed hweight32/64() were already optimized. Feel free to propose extra improvements.

Hi!

Quoted 4 lines

Quoted 3 lines

Would it make sense to only do hweight if *bitmap != ~0ULL ? Would it make sense to only check for bitflips > bitflips_threshold each 128 bytes or something like that?

Quoted 2 lines

I didn't go as far as you did and simply assumed hweight32/64() were already optimized. Feel free to propose extra improvements.

I'd propose this one (only compile tested, sorry, not sure how to test this one). If we see ~0UL, there's no need for hweight, and no need to check number of bitflips. So this should be net win.

Signed-off-by Pavel Machek <[email protected]>

drivers/mtd/nand/nand_base.c +4 -1

index b0524f8..96c27ec 100644
@@ -1357,7 +1357,10 @@ static int nand_check_erased_buf(void *buf, int len, int bitflips_threshold)
 
 	for (; len >= sizeof(long);
 	     len -= sizeof(long), bitmap += sizeof(long)) {
-		weight = hweight_long(*((unsigned long *)bitmap));
+		unsigned long d = *((unsigned long *)bitmap);
+		if (d == ~0UL)
+			continue;
+		weight = hweight_long(d);
 		bitflips += BITS_PER_LONG - weight;
 		if (unlikely(bitflips > bitflips_threshold))
 			return -EBADMSG;

If we see ~0UL in flash, there's no need for hweight, and no need to check number of bitflips. So this should be net win.

Signed-off-by Pavel Machek <[email protected]>

drivers/mtd/nand/nand_base.c +0 -0

index b0524f8..96c27ec 100644

Quoted 10 lines · Pavel Machek

If we see ~0UL in flash, there's no need for hweight, and no need to check number of bitflips. So this should be net win.

Signed-off-by Pavel Machek <[email protected]>

drivers/mtd/nand/nand_base.c +0 -0

index b0524f8..96c27ec 100644

drivers/mtd/nand/nand_base.c static int nand_check_erased_buf(void *buf, int len, int bitflips_threshold) · line 1360

Quoted 4 lines

drivers/mtd/nand/nand_base.c +0 -1

@@ -1357,7 +1357,10 @@ static int nand_check_erased_buf(void *buf, int len, int bitflips_threshold)
 
 	for (; len >= sizeof(long);
 	     len -= sizeof(long), bitmap += sizeof(long)) {
-		weight = hweight_long(*((unsigned long *)bitmap));

I hadn't noticed this earlier. There is, obviously, an implicit requirement that 'buf' must be 4-byte aligned on 32-bit platforms, and 8-byte aligned on 64-bit platforms.

This is not true for my platform, as the ecc pointer is chip->oob_poi + 10

This is not true for my platform, as the ecc pointer is chip->oob_poi + 10

I suppose it's not a problem if the platform can handle unaligned loads, otherwise it spells trouble.

drivers/mtd/nand/nand_base.c static int nand_check_erased_buf(void *buf, int len, int bitflips_threshold) · line 1366

Quoted 8 lines

drivers/mtd/nand/nand_base.c +4 -0

+		unsigned long d = *((unsigned long *)bitmap);
+		if (d == ~0UL)
+			continue;
+		weight = hweight_long(d);
 		bitflips += BITS_PER_LONG - weight;
 		if (unlikely(bitflips > bitflips_threshold))
 			return -EBADMSG;

The optimization makes sense in itself, but given that it's on an error path (?) I'm not sure it will bring any tangible benefits?

If we see unrecoverable ECC error, we need to count number of bitflips from all-ones and report correctable/uncorrectable according to that. Otherwise we report ECC failed on erased flash with single bit error.

Signed-off-by Pavel Machek <[email protected]>

drivers/mtd/nand/fsl_ifc_nand.c +23 -33

index d1570f5..1c66696 100644
@@ -171,34 +171,6 @@ static void set_addr(struct mtd_info *mtd, int column, int page_addr, int oob)
 		ifc_nand_ctrl->index += mtd->writesize;
 }
 
-static int is_blank(struct mtd_info *mtd, unsigned int bufnum)
-{
-	struct nand_chip *chip = mtd_to_nand(mtd);
-	struct fsl_ifc_mtd *priv = nand_get_controller_data(chip);
-	u8 __iomem *addr = priv->vbase + bufnum * (mtd->writesize * 2);
-	u32 __iomem *mainarea = (u32 __iomem *)addr;
-	u8 __iomem *oob = addr + mtd->writesize;
-	struct mtd_oob_region oobregion = { };
-	int i, section = 0;
-
-	for (i = 0; i < mtd->writesize / 4; i++) {
-		if (__raw_readl(&mainarea[i]) != 0xffffffff)
-			return 0;
-	}
-
-	mtd_ooblayout_ecc(mtd, section++, &oobregion);
-	while (oobregion.length) {
-		for (i = 0; i < oobregion.length; i++) {
-			if (__raw_readb(&oob[oobregion.offset + i]) != 0xff)
-				return 0;
-		}
-
-		mtd_ooblayout_ecc(mtd, section++, &oobregion);
-	}
-
-	return 1;
-}
-
 /* returns nonzero if entire page is blank */
 static int check_read_ecc(struct mtd_info *mtd, struct fsl_ifc_ctrl *ctrl,
 			  u32 *eccstat, unsigned int bufnum)

@@ -274,16 +246,13 @@ static void fsl_ifc_run_command(struct mtd_info *mtd)
 			if (errors == 15) {
 				/*
 				 * Uncorrectable error.
-				 * OK only if the whole page is blank.
+				 * We'll check for blank pages later.
 				 *
 				 * We disable ECCER reporting due to...
 				 * erratum IFC-A002770 -- so report it now if we
 				 * see an uncorrectable error in ECCSTAT.
 				 */
-				if (!is_blank(mtd, bufnum))
-					ctrl->nand_stat |=
-						IFC_NAND_EVTER_STAT_ECCER;
-				break;
+				ctrl->nand_stat |= IFC_NAND_EVTER_STAT_ECCER;
 			}
 
 			mtd->ecc_stats.corrected += errors;

@@ -678,6 +648,41 @@ static int fsl_ifc_wait(struct mtd_info *mtd, struct nand_chip *chip)
 	return nand_fsr | NAND_STATUS_WP;
 }
 
+/*
+ * The controller does not check for bitflips in erased pages,
+ * therefore software must check instead.
+ */
+static int check_erased_page(struct nand_chip *chip, u8 *buf)
+{
+	struct mtd_info *mtd = nand_to_mtd(chip);
+	u8 *ecc = chip->oob_poi;
+	const int ecc_size = chip->ecc.bytes;
+	const int pkt_size = chip->ecc.size;
+	int i, res, bitflips = 0;
+	struct mtd_oob_region oobregion = { };
+
+	mtd_ooblayout_ecc(mtd, 0, &oobregion);
+	ecc += oobregion.offset;
+
+	for (i = 0; i < chip->ecc.steps; ++i) {
+		res = nand_check_erased_ecc_chunk(buf, pkt_size, ecc, ecc_size,
+						  NULL, 0,
+						  chip->ecc.strength);
+		if (res < 0)

drivers/mtd/nand/fsl_ifc_nand.c +19 -3

+
+	return bitflips;
+}
+
 static int fsl_ifc_read_page(struct mtd_info *mtd, struct nand_chip *chip,
 			     uint8_t *buf, int oob_required, int page)
 {

@@ -689,11 +695,19 @@ static int fsl_ifc_read_page(struct mtd_info *mtd, struct nand_chip *chip,
 	if (oob_required)
 		fsl_ifc_read_buf(mtd, chip->oob_poi, mtd->oobsize);
 
-	if (ctrl->nand_stat & IFC_NAND_EVTER_STAT_ECCER)
-		dev_err(priv->dev, "NAND Flash ECC Uncorrectable Error\n");
-
 	if (ctrl->nand_stat != IFC_NAND_EVTER_STAT_OPC)
 		mtd->ecc_stats.failed++;
+
+	if (ctrl->nand_stat & IFC_NAND_EVTER_STAT_ECCER) {
+		int res;
+
+		if (!oob_required)
+			fsl_ifc_read_buf(mtd, chip->oob_poi, mtd->oobsize);
+
+		res = check_erased_page(chip, buf);
+		return res;
+	}
+
 
 	return nctrl->max_bitflips;
 }

@@ -904,6 +922,21 @@ static int fsl_ifc_chip_init(struct fsl_ifc_mtd *priv)
 		chip->ecc.algo = NAND_ECC_HAMMING;
 	}
 
+	{
+		struct mtd_oob_region oobregion = { };
+
+		mtd_ooblayout_ecc(mtd, 0, &oobregion);

drivers/mtd/nand/fsl_ifc_nand.c +1 -0

+
 	if (ctrl->version == FSL_IFC_VERSION_1_1_0)
 		fsl_ifc_sram_init(priv);
 

Quoted 35 lines · On Wed, 17 May 2017 14:22:24 +0200 Pavel Machek <[email protected]> wrote:

If we see unrecoverable ECC error, we need to count number of bitflips from all-ones and report correctable/uncorrectable according to that. Otherwise we report ECC failed on erased flash with single bit error.

Signed-off-by Pavel Machek <[email protected]>

drivers/mtd/nand/fsl_ifc_nand.c +25 -0

@@ -678,6 +648,41 @@ static int fsl_ifc_wait(struct mtd_info *mtd, struct nand_chip *chip)
 	return nand_fsr | NAND_STATUS_WP;
 }
 
+/*
+ * The controller does not check for bitflips in erased pages,
+ * therefore software must check instead.
+ */
+static int check_erased_page(struct nand_chip *chip, u8 *buf)
+{
+	struct mtd_info *mtd = nand_to_mtd(chip);
+	u8 *ecc = chip->oob_poi;
+	const int ecc_size = chip->ecc.bytes;
+	const int pkt_size = chip->ecc.size;
+	int i, res, bitflips = 0;
+	struct mtd_oob_region oobregion = { };
+
+	mtd_ooblayout_ecc(mtd, 0, &oobregion);
+	ecc += oobregion.offset;
+
+	for (i = 0; i < chip->ecc.steps; ++i) {
+		res = nand_check_erased_ecc_chunk(buf, pkt_size, ecc, ecc_size,
+						  NULL, 0,
+						  chip->ecc.strength);
+		if (res < 0)
+			mtd->ecc_stats.failed++;
+		else
+			mtd->ecc_stats.corrected += res;
+

drivers/mtd/nand/fsl_ifc_nand.c static int fsl_ifc_wait(struct mtd_info *mtd, struct nand_chip *chip) · line 681

Quoted 6 lines

drivers/mtd/nand/fsl_ifc_nand.c +6 -0

+		bitflips = max(res, bitflips);
+		buf += pkt_size;
+		ecc += ecc_size;
+	}
+
+	mtd_ooblayout_ecc(mtd, 1, &oobregion);

Why is this needed?

Quoted 14 lines

drivers/mtd/nand/fsl_ifc_nand.c +10 -0

@@ -904,6 +922,21 @@ static int fsl_ifc_chip_init(struct fsl_ifc_mtd *priv)
 		chip->ecc.algo = NAND_ECC_HAMMING;
 	}
 
+	{
+		struct mtd_oob_region oobregion = { };
+
+		mtd_ooblayout_ecc(mtd, 0, &oobregion);
+		if (!oobregion.length) {
+			dev_err(priv->dev, "No ECC in oobregion?\n");
+			return -EINVAL;
+		}
+		mtd_ooblayout_ecc(mtd, 1, &oobregion);
+		if (oobregion.length) {

drivers/mtd/nand/fsl_ifc_nand.c static int fsl_ifc_chip_init(struct fsl_ifc_mtd *priv) · line 938

Quoted 4 lines

drivers/mtd/nand/fsl_ifc_nand.c +4 -0

+			dev_err(priv->dev, "Extra data in oobregion?\n");
+			return -EINVAL;
+		}
+	}

This clearly doesn't belong in this patch. And if you really want to check that, please create a separate function instead of defining a non-conditional code block inside fsl_ifc_chip_init().

Quoted 15 lines

Quoted 15 lines

drivers/mtd/nand/fsl_ifc_nand.c +10 -0

@@ -904,6 +922,21 @@ static int fsl_ifc_chip_init(struct fsl_ifc_mtd *priv)
 		chip->ecc.algo = NAND_ECC_HAMMING;
 	}
 
+	{
+		struct mtd_oob_region oobregion = { };
+
+		mtd_ooblayout_ecc(mtd, 0, &oobregion);
+		if (!oobregion.length) {
+			dev_err(priv->dev, "No ECC in oobregion?\n");
+			return -EINVAL;
+		}
+		mtd_ooblayout_ecc(mtd, 1, &oobregion);
+		if (oobregion.length) {

Quoted 8 lines

drivers/mtd/nand/fsl_ifc_nand.c static int fsl_ifc_chip_init(struct fsl_ifc_mtd *priv) · line 938

Quoted 4 lines

drivers/mtd/nand/fsl_ifc_nand.c +4 -0

+			dev_err(priv->dev, "Extra data in oobregion?\n");
+			return -EINVAL;
+		}
+	}  

This clearly doesn't belong in this patch. And if you really want to check that, please create a separate function instead of defining a non-conditional code block inside fsl_ifc_chip_init().

I am not sure I want to check that. check_erased_page() can only handle layout with just one oobregion. If you think check is not needed, I'll happily remove the checking.

If we see unrecoverable ECC error, we need to count number of bitflips from all-ones and report correctable/uncorrectable according to that. Otherwise we report ECC failed on erased flash with single bit error.

Signed-off-by Pavel Machek <[email protected]>

Reported-by Darwin Dingel <[email protected]>

Hi Pavel,

Just a minor thing. Sorry about the late comment.

drivers/mtd/nand/fsl_ifc_nand.c static int fsl_ifc_read_page(struct mtd_info *mtd, struct nand_chip *chip, · line 702

Quoted 4 lines · Pavel Machek

drivers/mtd/nand/fsl_ifc_nand.c +4 -0

+
+		res = check_erased_page(chip, buf);
+		return res;
+	}

Can we just remove 'res' and change this line to:

	return check_erased_page(chip, buf);

Quoted 4 lines · Darwin Dingel

Hi Pavel,

Just a minor thing. Sorry about the late comment.

Quoted 8 lines

drivers/mtd/nand/fsl_ifc_nand.c static int fsl_ifc_read_page(struct mtd_info *mtd, struct nand_chip *chip, · line 702

Quoted 4 lines · Pavel Machek

drivers/mtd/nand/fsl_ifc_nand.c +4 -0

+
+		res = check_erased_page(chip, buf);
+		return res;
+	}  

Can we just remove 'res' and change this line to:

	return check_erased_page(chip, buf);  

Well... I originally had a printk there, and yes, it can be simplified. I can roll v3, if required, but I'd leave it as is -- it is still convenient place to add debugging to.

Quoted 14 lines · On Thu, 1 Jun 2017 15:12:32 +0200 Pavel Machek <[email protected]> wrote:

Quoted 12 lines · Darwin Dingel

Hi Pavel,

Just a minor thing. Sorry about the late comment.

drivers/mtd/nand/fsl_ifc_nand.c static int fsl_ifc_read_page(struct mtd_info *mtd, struct nand_chip *chip, · line 702

Quoted 4 lines · Pavel Machek

drivers/mtd/nand/fsl_ifc_nand.c +4 -0

+
+		res = check_erased_page(chip, buf);
+		return res;
+	}    

Can we just remove 'res' and change this line to:

	return check_erased_page(chip, buf);    

Quoted 3 lines

Well... I originally had a printk there, and yes, it can be simplified. I can roll v3, if required, but I'd leave it as is -- it is still convenient place to add debugging to.

I can fix it when applying, no need to send a new version.

Quoted 41 lines

Hi!

We have some problems with fsl_ifc_nand ... in the old kernels, but this one does not seem to be fixed in v4.11, either.

UBIFS complains:

UBIFS error (pid 931): ubifs_scan: corrupt empty space at LEB 282:252630 UBIFS error (pid 931): ubifs_scanned_corruption: corruption at LEB 282:252630 UBIFS error (pid 931): ubifs_scanned_corruption: first 1322 bytes from LEB 282:252630 UBIFS error (pid 931): ubifs_scan: LEB 282 scanning failed

Possible explanation is here:

https://e2e.ti.com/support/dsp/davinci_digital_media_processors/f/716/t/289605

# I see on the forum that this issue has been raised before - my

# understanding is that the omap2 nand driver does not perform ECC
# detection/correction on empty pages so when UBIFS checks the empty
# space data and doesn't read all 0xFF then it fails and mounts
# read-only. I didn't find any good solution - only a workaround to
# remove the UBIFS check..

So I checked fsl_ifc_nand.c in v4.11-rc, and yes, it seems to have the same problem:

			if (errors == 15) {
	                        /*
                                 * Uncorrectable error.
                                 * OK only if the whole page is blank.
                                 *
                                 * We disable ECCER reporting due to...
                                 * erratum IFC-A002770 -- so report it now if we
                                 * see an uncorrectable error in ECCSTAT.
                                 */
                                if (!is_blank(mtd, bufnum))
                                        ctrl->nand_stat |=
                                                IFC_NAND_EVTER_STAT_ECCER;
                                break;
                        }

Quoted 4 lines

is_blank() checks for all 0xff's, so single-bit 0xfe in the data will result in_blank() == 0 and uncorrectable error being signaled.

Should the driver be modified somehow?

Yep, nand_check_erased_ecc_chunk() [1] is here to help you check this case, unfortunately, it's not directly applicable here, because this function takes regular pointers and not __iomem ones. You'll either have to copy the data in an intermediate buffer before calling nand_check_erased_ecc_chunk(), or cast the SRAM region to a void pointer (which is usually not a good idea). The last option would be to open code nand_check_erased_ecc_chunk(), but I'd really like to avoid that (for maintainability concerns).

(BTW, switching to ecc.mode = ECC_SOFT will cause compatibility problems but should make the problem go away, right?)

Hi!

(Added driver author to the cc list, maybe he can help).

Quoted 46 lines

Quoted 41 lines

Hi!

We have some problems with fsl_ifc_nand ... in the old kernels, but this one does not seem to be fixed in v4.11, either.

UBIFS complains:

UBIFS error (pid 931): ubifs_scan: corrupt empty space at LEB 282:252630 UBIFS error (pid 931): ubifs_scanned_corruption: corruption at LEB 282:252630 UBIFS error (pid 931): ubifs_scanned_corruption: first 1322 bytes from LEB 282:252630 UBIFS error (pid 931): ubifs_scan: LEB 282 scanning failed

Possible explanation is here:

https://e2e.ti.com/support/dsp/davinci_digital_media_processors/f/716/t/289605

# I see on the forum that this issue has been raised before - my

# understanding is that the omap2 nand driver does not perform ECC
# detection/correction on empty pages so when UBIFS checks the empty
# space data and doesn't read all 0xFF then it fails and mounts
# read-only. I didn't find any good solution - only a workaround to
# remove the UBIFS check..

So I checked fsl_ifc_nand.c in v4.11-rc, and yes, it seems to have the same problem:

			if (errors == 15) {
	                        /*
                                 * Uncorrectable error.
                                 * OK only if the whole page is blank.
                                 *
                                 * We disable ECCER reporting due to...
                                 * erratum IFC-A002770 -- so report it now if we
                                 * see an uncorrectable error in ECCSTAT.
                                 */
                                if (!is_blank(mtd, bufnum))
                                        ctrl->nand_stat |=
                                                IFC_NAND_EVTER_STAT_ECCER;
                                break;
                        }

Quoted 4 lines

is_blank() checks for all 0xff's, so single-bit 0xfe in the data will result in_blank() == 0 and uncorrectable error being signaled.

Should the driver be modified somehow?

Quoted 8 lines

Yep, nand_check_erased_ecc_chunk() [1] is here to help you check this case, unfortunately, it's not directly applicable here, because this function takes regular pointers and not __iomem ones. You'll either have to copy the data in an intermediate buffer before calling nand_check_erased_ecc_chunk(), or cast the SRAM region to a void pointer (which is usually not a good idea). The last option would be to open code nand_check_erased_ecc_chunk(), but I'd really like to avoid that (for maintainability concerns).

Ok, took a look. __iomem is part of a problem, another part is that nand_check_erased_ecc_chunk() needs to actually write back 0xff's to undo the corruption, which would probably be bad idea to do in the iomem, and next one is that blank actually checks arbitrary number of regions, based on ecc.layout.

So this could be used to simplify the code (if nand_check_erased_buf was exported; it is not), but it does not fix the problem as we still need to undo the corruption.

Hints welcome, especially if you know right place where to put this checking.

(BTW, switching to ecc.mode = ECC_SOFT will cause compatibility problems but should make the problem go away, right?)

(Added driver author to the cc list, maybe he can help).

Quoted 6 lines

Quoted 6 lines

UBIFS complains:

UBIFS error (pid 931): ubifs_scan: corrupt empty space at LEB 282:252630 UBIFS error (pid 931): ubifs_scanned_corruption: corruption at LEB 282:252630 UBIFS error (pid 931): ubifs_scanned_corruption: first 1322 bytes from LEB 282:252630 UBIFS error (pid 931): ubifs_scan: LEB 282 scanning failed

Quoted 7 lines

Quoted 4 lines

is_blank() checks for all 0xff's, so single-bit 0xfe in the data will result in_blank() == 0 and uncorrectable error being signaled.

Should the driver be modified somehow?

Yep, nand_check_erased_ecc_chunk() [1] is here to help you check this case, unfortunately, it's not directly applicable here, because this

Hi!

Quoted 8 lines

(Added driver author to the cc list, maybe he can help).

Quoted 6 lines

Quoted 6 lines

UBIFS complains:

UBIFS error (pid 931): ubifs_scan: corrupt empty space at LEB 282:252630 UBIFS error (pid 931): ubifs_scanned_corruption: corruption at LEB 282:252630 UBIFS error (pid 931): ubifs_scanned_corruption: first 1322 bytes from LEB 282:252630 UBIFS error (pid 931): ubifs_scan: LEB 282 scanning failed

...

Quoted 3 lines

Quoted 3 lines

Quoted 3 lines

is_blank() checks for all 0xff's, so single-bit 0xfe in the data will result in_blank() == 0 and uncorrectable error being signaled.

Quoted 4 lines

Quoted 4 lines

Quoted 1 line

Should the driver be modified somehow?

Yep, nand_check_erased_ecc_chunk() [1] is here to help you check this case, unfortunately, it's not directly applicable here, because this

Maybe I figured it out. Unfortunately, it is only compile tested. Does it look approximately right?

Thanks,
Pavel

drivers/mtd/nand/fsl_ifc_nand.c +15 -31

@@ -171,32 +171,6 @@ static void set_addr(struct mtd_info *mtd, int column, int page_addr, int oob)
 		ifc_nand_ctrl->index += mtd->writesize;
 }
 
-static int is_blank(struct mtd_info *mtd, unsigned int bufnum)
-{
-	struct nand_chip *chip = mtd_to_nand(mtd);
-	struct fsl_ifc_mtd *priv = nand_get_controller_data(chip);
-	u8 __iomem *addr = priv->vbase + bufnum * (mtd->writesize * 2);
-	u32 __iomem *mainarea = (u32 __iomem *)addr;
-	u8 __iomem *oob = addr + mtd->writesize;
-	struct mtd_oob_region oobregion = { };
-	int i, section = 0;
-
-	i = nand_check_erased_buf(&mainarea[i], mtd->writesize, 0);
-	if (i)
-		return 0;
-
-	mtd_ooblayout_ecc(mtd, section++, &oobregion);
-	while (oobregion.length) {
-		i = nand_check_erased_buf(&oob[oobregion.offset], oobregion.length, 0);
-		if (i)
-			return 0;
-
-		mtd_ooblayout_ecc(mtd, section++, &oobregion);
-	}
-
-	return 1;
-}
-
 /* returns nonzero if entire page is blank */
 static int check_read_ecc(struct mtd_info *mtd, struct fsl_ifc_ctrl *ctrl,
 			  u32 *eccstat, unsigned int bufnum)

@@ -272,16 +246,14 @@ static void fsl_ifc_run_command(struct mtd_info *mtd)
 			if (errors == 15) {
 				/*
 				 * Uncorrectable error.
-				 * OK only if the whole page is blank.
+				 * We'll check for blank pages later.
 				 *
 				 * We disable ECCER reporting due to...
 				 * erratum IFC-A002770 -- so report it now if we
 				 * see an uncorrectable error in ECCSTAT.
 				 */
-				if (!is_blank(mtd, bufnum))
-					ctrl->nand_stat |=
-						IFC_NAND_EVTER_STAT_ECCER;
-				break;
+				ctrl->nand_stat |= IFC_NAND_EVTER_STAT_ECCER;
+				printk("NAND flash read: error but continuing.\n");
 			}
 
 			mtd->ecc_stats.corrected += errors;

@@ -676,6 +648,40 @@ static int fsl_ifc_wait(struct mtd_info *mtd, struct nand_chip *chip)
 	return nand_fsr | NAND_STATUS_WP;
 }
 
+/*
+ * The controller does not check for bitflips in erased pages,
+ * therefore software must check instead.
+ */
+static int check_erased_page(struct nand_chip *chip, u8 *buf)
+{
+	struct mtd_info *mtd = nand_to_mtd(chip);
+	u8 *ecc = chip->oob_poi;
+	const int ecc_size = chip->ecc.bytes;
+	const int pkt_size = chip->ecc.size;
+	int i, res, bitflips = 0;
+	struct mtd_oob_region oobregion = { };

drivers/mtd/nand/fsl_ifc_nand.c +19 -3

+
+	return bitflips;
+}
+
 static int fsl_ifc_read_page(struct mtd_info *mtd, struct nand_chip *chip,
 			     uint8_t *buf, int oob_required, int page)
 {

@@ -687,11 +693,23 @@ static int fsl_ifc_read_page(struct mtd_info *mtd, struct nand_chip *chip,
 	if (oob_required)
 		fsl_ifc_read_buf(mtd, chip->oob_poi, mtd->oobsize);
 
-	if (ctrl->nand_stat & IFC_NAND_EVTER_STAT_ECCER)
-		dev_err(priv->dev, "NAND Flash ECC Uncorrectable Error\n");
-
 	if (ctrl->nand_stat != IFC_NAND_EVTER_STAT_OPC)
 		mtd->ecc_stats.failed++;
+	
+	if (ctrl->nand_stat & IFC_NAND_EVTER_STAT_ECCER) {
+		int res;
+
+		if (!oob_required)
+			fsl_ifc_read_buf(mtd, chip->oob_poi, mtd->oobsize);
+
+		printk("NAND flash: have error (%d)\n", nctrl->max_bitflips);
+		res = check_erased_page(chip, buf);
+		printk("NAND flash: but erased page says %d\n", res);
+
+		dev_err(priv->dev, "NAND Flash ECC Uncorrectable Error or erased page\n");
+		return res;
+	}
+
 
 	return nctrl->max_bitflips;
 }

Quoted 2 lines

Maybe I figured it out. Unfortunately, it is only compile tested. Does it look approximately right?

Yep that's definitely better. Just one thing missing (see below), otherwise it looks good.

drivers/mtd/nand/fsl_ifc_nand.c static int fsl_ifc_wait(struct mtd_info *mtd, struct nand_chip *chip) · line 672

Quoted 2 lines

drivers/mtd/nand/fsl_ifc_nand.c +2 -0

+		if (res < 0)
+			mtd->ecc_stats.failed++;

		else
			mtd->ecc_stats.corrected += res;

drivers/mtd/nand/fsl_ifc_nand.c static int fsl_ifc_wait(struct mtd_info *mtd, struct nand_chip *chip) · line 680

Quoted 8 lines

drivers/mtd/nand/fsl_ifc_nand.c +8 -0

+
+		bitflips = max(res, bitflips);
+		buf += pkt_size;
+		ecc += ecc_size;
+	}
+
+	mtd_ooblayout_ecc(mtd, 1, &oobregion);
+	BUG_ON(oobregion.length);

Probably something you should check at registration time only.

Fix ecc.stats_corrected in empty flash case.

Signed-off-by Pavel Machek <[email protected]>

---

This was suggested by Boris Brezillon in another context. Not tested; I don't have the hardware.

drivers/mtd/nand/tango_nand.c +0 -0

index 4a5e948..db4bff4 100644

[ Trimming CC list ]

Quoted 15 lines · Pavel Machek

Fix ecc.stats_corrected in empty flash case.

Signed-off-by Pavel Machek <[email protected]>

---

This was suggested by Boris Brezillon in another context. Not tested; I don't have the hardware.

drivers/mtd/nand/tango_nand.c +0 -0

index 4a5e948..db4bff4 100644

drivers/mtd/nand/tango_nand.c static int check_erased_page(struct nand_chip *chip, u8 *buf) · line 200

Quoted 8 lines

drivers/mtd/nand/tango_nand.c +2 -0

@@ -193,6 +193,8 @@ static int check_erased_page(struct nand_chip *chip, u8 *buf)
 						  chip->ecc.strength);
 		if (res < 0)
 			mtd->ecc_stats.failed++;
+		else
+			mtd->ecc_stats.corrected += res;
 
 		bitflips = max(res, bitflips);
 		buf += pkt_size;

Hello Pavel,

You may have noticed that ecc_stats.corrected is not updated in decode_error_report() which is the main code path, i.e. the path that will succeed 99.99% of the time (HW read).

It turns out that the HW does not report the number of errors corrected in a page... Instead it reports two values:
1) U = number of errors corrected in the first packet/step
2) V = max number of errors corrected in other packets/steps

Thus, it is not possible to determine the actual number of errors corrected in a page (unless V is 0). Otherwise, we just have an interval; let n be the number of packets/steps:

U + V <= corrected errors count <= U + (n-1)*V

In my opinion, it is better to provide no information than to provide incorrect information. Therefore, I did not update ecc_stats.corrected in decode_error_report().

One could argue that updating ecc_stats.corrected in check_erased_page() sets the correct value, since the error counts are computed in software for each step. But updating the value here is IMO pointless if we can't do it in the main code path.

[ Trimming CC list ]

Quoted 15 lines · Pavel Machek

Fix ecc.stats_corrected in empty flash case.

Signed-off-by Pavel Machek <[email protected]>

---

This was suggested by Boris Brezillon in another context. Not tested; I don't have the hardware.

drivers/mtd/nand/tango_nand.c +0 -0

index 4a5e948..db4bff4 100644

drivers/mtd/nand/tango_nand.c static int check_erased_page(struct nand_chip *chip, u8 *buf) · line 200

Quoted 8 lines

drivers/mtd/nand/tango_nand.c +2 -0

@@ -193,6 +193,8 @@ static int check_erased_page(struct nand_chip *chip, u8 *buf)
 						  chip->ecc.strength);
 		if (res < 0)
 			mtd->ecc_stats.failed++;
+		else
+			mtd->ecc_stats.corrected += res;
 
 		bitflips = max(res, bitflips);
 		buf += pkt_size;

Hello Pavel,

You may have noticed that ecc_stats.corrected is not updated in decode_error_report() which is the main code path, i.e. the path that will succeed 99.99% of the time (HW read).

It turns out that the HW does not report the number of errors corrected in a page... Instead it reports two values:
1) U = number of errors corrected in the first packet/step
2) V = max number of errors corrected in other packets/steps

Thus, it is not possible to determine the actual number of errors corrected in a page (unless V is 0). Otherwise, we just have an interval; let n be the number of packets/steps:

U + V <= corrected errors count <= U + (n-1)*V

In my opinion, it is better to provide no information than to provide incorrect information. Therefore, I did not update ecc_stats.corrected in decode_error_report().

One could argue that updating ecc_stats.corrected in check_erased_page() sets the correct value, since the error counts are computed in software for each step. But updating the value here is IMO pointless if we can't do it in the main code path.

Hi Marc,

Quoted 43 lines · On Mon, 24 Apr 2017 10:58:47 +0200 Marc Gonzalez…

[ Trimming CC list ]

Quoted 15 lines · Pavel Machek

Fix ecc.stats_corrected in empty flash case.

Signed-off-by Pavel Machek <[email protected]>

---

This was suggested by Boris Brezillon in another context. Not tested; I don't have the hardware.

drivers/mtd/nand/tango_nand.c +0 -0

index 4a5e948..db4bff4 100644

drivers/mtd/nand/tango_nand.c static int check_erased_page(struct nand_chip *chip, u8 *buf) · line 200

Quoted 8 lines

drivers/mtd/nand/tango_nand.c +2 -0

@@ -193,6 +193,8 @@ static int check_erased_page(struct nand_chip *chip, u8 *buf)
 						  chip->ecc.strength);
 		if (res < 0)
 			mtd->ecc_stats.failed++;
+		else
+			mtd->ecc_stats.corrected += res;
 
 		bitflips = max(res, bitflips);
 		buf += pkt_size;

Hello Pavel,

You may have noticed that ecc_stats.corrected is not updated in decode_error_report() which is the main code path, i.e. the path that will succeed 99.99% of the time (HW read).

It turns out that the HW does not report the number of errors corrected in a page... Instead it reports two values:
1) U = number of errors corrected in the first packet/step
2) V = max number of errors corrected in other packets/steps

Thus, it is not possible to determine the actual number of errors corrected in a page (unless V is 0). Otherwise, we just have an interval; let n be the number of packets/steps:

Quoted 5 lines

U + V <= corrected errors count <= U + (n-1)*V

In my opinion, it is better to provide no information than to provide incorrect information. Therefore, I did not update ecc_stats.corrected in decode_error_report().

Hm, not sure I agree with that. The situation is far from ideal, but some userspace tools query the number of corrected bits before and after doing a read operation to report the number of bitflips that have been correcte. Letting users think there were no bitflips at all is not a good thing IMO.

Quoted 5 lines

One could argue that updating ecc_stats.corrected in check_erased_page() sets the correct value, since the error counts are computed in software for each step. But updating the value here is IMO pointless if we can't do it in the main code path.

Well, it's not pointless if you want to inform the user that some bitflips were present on the NAND. Yes, the numbers you report won't be accurate in your case, but at least the user can tell when bitflips were discovered.

Note that ideally, we should have per-page (or per-block) max-bitflips information (where max-bitflips is the number returned by ecc->read_page()), because counting the total number of corrected bits is certainly not helping when it comes to checking how reliable a NAND page/eraseblock is.

Quoted 54 lines · Boris Brezillon

Hi Marc,

Quoted 43 lines · On Mon, 24 Apr 2017 10:58:47 +0200 Marc Gonzalez…

[ Trimming CC list ]

Quoted 15 lines · Pavel Machek

Fix ecc.stats_corrected in empty flash case.

Signed-off-by Pavel Machek <[email protected]>

---

This was suggested by Boris Brezillon in another context. Not tested; I don't have the hardware.

drivers/mtd/nand/tango_nand.c +0 -0

index 4a5e948..db4bff4 100644

drivers/mtd/nand/tango_nand.c static int check_erased_page(struct nand_chip *chip, u8 *buf) · line 200

Quoted 8 lines

drivers/mtd/nand/tango_nand.c +2 -0

@@ -193,6 +193,8 @@ static int check_erased_page(struct nand_chip *chip, u8 *buf)
 						  chip->ecc.strength);
 		if (res < 0)
 			mtd->ecc_stats.failed++;
+		else
+			mtd->ecc_stats.corrected += res;
 
 		bitflips = max(res, bitflips);
 		buf += pkt_size;

Hello Pavel,

You may have noticed that ecc_stats.corrected is not updated in decode_error_report() which is the main code path, i.e. the path that will succeed 99.99% of the time (HW read).

It turns out that the HW does not report the number of errors corrected in a page... Instead it reports two values:
1) U = number of errors corrected in the first packet/step
2) V = max number of errors corrected in other packets/steps

Thus, it is not possible to determine the actual number of errors corrected in a page (unless V is 0). Otherwise, we just have an interval; let n be the number of packets/steps:

Quoted 5 lines

U + V <= corrected errors count <= U + (n-1)*V

In my opinion, it is better to provide no information than to provide incorrect information. Therefore, I did not update ecc_stats.corrected in decode_error_report().

Quoted 5 lines

Hm, not sure I agree with that. The situation is far from ideal, but some userspace tools query the number of corrected bits before and after doing a read operation to report the number of bitflips that have been correcte. Letting users think there were no bitflips at all is not a good thing IMO.

I (still) find the API of ecc->read_page somewhat confusing ;-) Let me try to work through the assumptions, as I remember.

1) If the driver was not able to read the page,
then we return an error code < 0

2) If the driver was able to read the page, and there were no bitflips, then we return 0

3) If the driver was able to read the page, and there were less than 'strength' bitflips in each step, then we return the *max* of all bitflips across all steps

4) If the driver was able to read the page, but there was at least one step with too many bitflips, then we are expected to increment mtd->ecc_stats.failed, and return the max of all bitflips across successful steps

AFAICT, the NAND framework does not use or update ecc_stats.corrected?

What about the MTD layer?

Quoted 8 lines

Quoted 7 lines

One could argue that updating ecc_stats.corrected in check_erased_page() sets the correct value, since the error counts are computed in software for each step. But updating the value here is IMO pointless if we can't do it in the main code path.

Quoted 4 lines

Well, it's not pointless if you want to inform the user that some bitflips were present on the NAND. Yes, the numbers you report won't be accurate in your case, but at least the user can tell when bitflips were discovered.

I do report some bitflips value: it's the return value from ecc->read_page -- which is the max bitflips per step. But perhaps that value is not available to user-space?

In other words,
the NAND and MTD layers do not forward it to user-space, as this is left as a responsibility of individual drivers?

Quoted 6 lines

Note that ideally, we should have per-page (or per-block) max-bitflips information (where max-bitflips is the number returned by ecc->read_page()), because counting the total number of corrected bits is certainly not helping when it comes to checking how reliable a NAND page/eraseblock is.

That would be quite large an array for a 1 TB NAND chip?

Anyway, I think we went too far. I was just arguing that never updating stats->corrected just because you don't know the exact number of bitflips is not necessarily a better idea than updating corrected with a potentially invalid number of bitflips. At least the 2nd approach shows that bitflips were present on the media.

[ Trimming CC list ]

Quoted 15 lines · Pavel Machek

Fix ecc.stats_corrected in empty flash case.

Signed-off-by Pavel Machek <[email protected]>

---

This was suggested by Boris Brezillon in another context. Not tested; I don't have the hardware.

drivers/mtd/nand/tango_nand.c +0 -0

index 4a5e948..db4bff4 100644

drivers/mtd/nand/tango_nand.c static int check_erased_page(struct nand_chip *chip, u8 *buf) · line 200

Quoted 8 lines

drivers/mtd/nand/tango_nand.c +2 -0

@@ -193,6 +193,8 @@ static int check_erased_page(struct nand_chip *chip, u8 *buf)
 						  chip->ecc.strength);
 		if (res < 0)
 			mtd->ecc_stats.failed++;
+		else
+			mtd->ecc_stats.corrected += res;
 
 		bitflips = max(res, bitflips);
 		buf += pkt_size;

Hello Pavel,

You may have noticed that ecc_stats.corrected is not updated in decode_error_report() which is the main code path, i.e. the path that will succeed 99.99% of the time (HW read).

It turns out that the HW does not report the number of errors corrected in a page... Instead it reports two values:
1) U = number of errors corrected in the first packet/step
2) V = max number of errors corrected in other packets/steps

Thus, it is not possible to determine the actual number of errors corrected in a page (unless V is 0). Otherwise, we just have an interval; let n be the number of packets/steps:

U + V <= corrected errors count <= U + (n-1)*V

In my opinion, it is better to provide no information than to provide incorrect information. Therefore, I did not update ecc_stats.corrected in decode_error_report().

Hi!

Quoted 43 lines · Marc Gonzalez

[ Trimming CC list ]

Quoted 15 lines · Pavel Machek

Fix ecc.stats_corrected in empty flash case.

Signed-off-by Pavel Machek <[email protected]>

---

This was suggested by Boris Brezillon in another context. Not tested; I don't have the hardware.

drivers/mtd/nand/tango_nand.c +0 -0

index 4a5e948..db4bff4 100644

drivers/mtd/nand/tango_nand.c static int check_erased_page(struct nand_chip *chip, u8 *buf) · line 200

Quoted 8 lines

drivers/mtd/nand/tango_nand.c +2 -0

@@ -193,6 +193,8 @@ static int check_erased_page(struct nand_chip *chip, u8 *buf)
 						  chip->ecc.strength);
 		if (res < 0)
 			mtd->ecc_stats.failed++;
+		else
+			mtd->ecc_stats.corrected += res;
 
 		bitflips = max(res, bitflips);
 		buf += pkt_size;

Hello Pavel,

You may have noticed that ecc_stats.corrected is not updated in decode_error_report() which is the main code path, i.e. the path that will succeed 99.99% of the time (HW read).

It turns out that the HW does not report the number of errors corrected in a page... Instead it reports two values:
1) U = number of errors corrected in the first packet/step
2) V = max number of errors corrected in other packets/steps

Thus, it is not possible to determine the actual number of errors corrected in a page (unless V is 0). Otherwise, we just have an interval; let n be the number of packets/steps:

Quoted 5 lines

U + V <= corrected errors count <= U + (n-1)*V

In my opinion, it is better to provide no information than to provide incorrect information. Therefore, I did not update ecc_stats.corrected in decode_error_report().

Well... Having corrected ECC errors is pretty rare, right?

So one
solution would be to re-compute ECCs in software if we see U or V > 0...

According to Boris, some user-space tools expect MTD drivers to update ecc_stats.corrected, and it's better to provide a lower bound than to provide no information at all.

Reported-by Pavel Machek <[email protected]>

Signed-off-by Marc Gonzalez <[email protected]>

---
NB: this patch is based on 4.9; if it looks good, I'll rebase it on v4.12-rc1 next week.

---
 drivers/mtd/nand/tango_nand.c | 21 ++++++++++++++-------
 1 file changed, 14 insertions(+), 7 deletions(-)

drivers/mtd/nand/tango_nand.c +4 -2

index 4a5e948c62df..471348462e42 100644
@@ -55,10 +55,10 @@
  * byte 1 for other packets in the page (PKT_N, for N > 0)
  * ERR_COUNT_PKT_N is the max error count over all but the first packet.
  */
-#define DECODE_OK_PKT_0(v)	((v) & BIT(7))
-#define DECODE_OK_PKT_N(v)	((v) & BIT(15))
 #define ERR_COUNT_PKT_0(v)	(((v) >> 0) & 0x3f)
 #define ERR_COUNT_PKT_N(v)	(((v) >> 8) & 0x3f)
+#define DECODE_FAIL_PKT_0(v)	(((v) & BIT(7)) == 0)
+#define DECODE_FAIL_PKT_N(v)	(((v) & BIT(15)) == 0)
 
 /* Offsets relative to pbus_base */
 #define PBUS_CS_CTRL	0x83c

@@ -193,6 +193,8 @@ static int check_erased_page(struct nand_chip *chip, u8 *buf)
 						  chip->ecc.strength);
 		if (res < 0)
 			mtd->ecc_stats.failed++;
+		else
+			mtd->ecc_stats.corrected += res;
 
 		bitflips = max(res, bitflips);
 		buf += pkt_size;

drivers/mtd/nand/tango_nand.c +9 -4

 {
 	u32 status, res;
+	struct mtd_info *mtd = nand_to_mtd(chip);
 
 	status = readl_relaxed(nfc->reg_base + NFC_XFER_STATUS);
 	if (status & PAGE_IS_EMPTY)

@@ -212,10 +215,14 @@ static int decode_error_report(struct tango_nfc *nfc)
 
 	res = readl_relaxed(nfc->mem_base + ERROR_REPORT);
 
-	if (DECODE_OK_PKT_0(res) && DECODE_OK_PKT_N(res))
-		return max(ERR_COUNT_PKT_0(res), ERR_COUNT_PKT_N(res));
+	if (DECODE_FAIL_PKT_0(res) || DECODE_FAIL_PKT_N(res))
+		return -EBADMSG;
+
+	/* ERR_COUNT_PKT_N is max, not sum, but that's all we have */
+	mtd->ecc_stats.corrected +=
+		ERR_COUNT_PKT_0(res) + ERR_COUNT_PKT_N(res);
 
-	return -EBADMSG;
+	return max(ERR_COUNT_PKT_0(res), ERR_COUNT_PKT_N(res));
 }
 
 static void tango_dma_callback(void *arg)

@@ -280,7 +287,7 @@ static int tango_read_page(struct mtd_info *mtd, struct nand_chip *chip,
 	if (err)
 		return err;
 
-	res = decode_error_report(nfc);
+	res = decode_error_report(chip, nfc);
 	if (res < 0) {
 		chip->ecc.read_oob_raw(mtd, chip, page);
 		res = check_erased_page(chip, buf);

According to Boris, some user-space tools expect MTD drivers to update ecc_stats.corrected, and it's better to provide a lower bound than to provide no information at all.

Reported-by Pavel Machek <[email protected]>

Signed-off-by Marc Gonzalez <[email protected]>

---
NB: this patch is based on 4.9; if it looks good, I'll rebase it on v4.12-rc1 next week.

---
 drivers/mtd/nand/tango_nand.c | 21 ++++++++++++++-------
 1 file changed, 14 insertions(+), 7 deletions(-)

drivers/mtd/nand/tango_nand.c +14 -7

index 4a5e948c62df..471348462e42 100644
@@ -55,10 +55,10 @@
  * byte 1 for other packets in the page (PKT_N, for N > 0)
  * ERR_COUNT_PKT_N is the max error count over all but the first packet.
  */
-#define DECODE_OK_PKT_0(v)	((v) & BIT(7))
-#define DECODE_OK_PKT_N(v)	((v) & BIT(15))
 #define ERR_COUNT_PKT_0(v)	(((v) >> 0) & 0x3f)
 #define ERR_COUNT_PKT_N(v)	(((v) >> 8) & 0x3f)
+#define DECODE_FAIL_PKT_0(v)	(((v) & BIT(7)) == 0)
+#define DECODE_FAIL_PKT_N(v)	(((v) & BIT(15)) == 0)
 
 /* Offsets relative to pbus_base */
 #define PBUS_CS_CTRL	0x83c

@@ -193,6 +193,8 @@ static int check_erased_page(struct nand_chip *chip, u8 *buf)
 						  chip->ecc.strength);
 		if (res < 0)
 			mtd->ecc_stats.failed++;
+		else
+			mtd->ecc_stats.corrected += res;
 
 		bitflips = max(res, bitflips);
 		buf += pkt_size;

@@ -202,9 +204,10 @@ static int check_erased_page(struct nand_chip *chip, u8 *buf)
 	return bitflips;
 }
 
-static int decode_error_report(struct tango_nfc *nfc)
+static int decode_error_report(struct nand_chip *chip, struct tango_nfc *nfc)
 {
 	u32 status, res;
+	struct mtd_info *mtd = nand_to_mtd(chip);
 
 	status = readl_relaxed(nfc->reg_base + NFC_XFER_STATUS);
 	if (status & PAGE_IS_EMPTY)

@@ -212,10 +215,14 @@ static int decode_error_report(struct tango_nfc *nfc)
 
 	res = readl_relaxed(nfc->mem_base + ERROR_REPORT);
 
-	if (DECODE_OK_PKT_0(res) && DECODE_OK_PKT_N(res))
-		return max(ERR_COUNT_PKT_0(res), ERR_COUNT_PKT_N(res));
+	if (DECODE_FAIL_PKT_0(res) || DECODE_FAIL_PKT_N(res))
+		return -EBADMSG;
+
+	/* ERR_COUNT_PKT_N is max, not sum, but that's all we have */
+	mtd->ecc_stats.corrected +=
+		ERR_COUNT_PKT_0(res) + ERR_COUNT_PKT_N(res);
 
-	return -EBADMSG;
+	return max(ERR_COUNT_PKT_0(res), ERR_COUNT_PKT_N(res));
 }
 
 static void tango_dma_callback(void *arg)

@@ -280,7 +287,7 @@ static int tango_read_page(struct mtd_info *mtd, struct nand_chip *chip,
 	if (err)
 		return err;
 
-	res = decode_error_report(nfc);
+	res = decode_error_report(chip, nfc);
 	if (res < 0) {
 		chip->ecc.read_oob_raw(mtd, chip, page);
 		res = check_erased_page(chip, buf);

Quoted 15 lines

Quoted 15 lines

Hello Pavel,

You may have noticed that ecc_stats.corrected is not updated in decode_error_report() which is the main code path, i.e. the path that will succeed 99.99% of the time (HW read).

It turns out that the HW does not report the number of errors corrected in a page... Instead it reports two values:
1) U = number of errors corrected in the first packet/step
2) V = max number of errors corrected in other packets/steps

Thus, it is not possible to determine the actual number of errors corrected in a page (unless V is 0). Otherwise, we just have an interval; let n be the number of packets/steps:

Quoted 7 lines

Quoted 5 lines

U + V <= corrected errors count <= U + (n-1)*V

In my opinion, it is better to provide no information than to provide incorrect information. Therefore, I did not update ecc_stats.corrected in decode_error_report().

Well... Having corrected ECC errors is pretty rare, right?

Depends on the NAND chip. On modern SLC NAND chips requiring ECC of 8bits/512bytes are likely to have frequent bitflips.

Quoted 3 lines

So one
solution would be to re-compute ECCs in software if we see U or V > 0...

Hm, not sure it's worth the trouble for statistics that are anyway rarely used, and when they are, are only used has a metric to determine how worn the NAND is.

I'd prefer to see a better user-space interface returning the max_bitflips information when someone reads from an MTD device (see [1]) rather than trying to fix drivers to return the exact number of corrected bitflips (which might be impossible for some of them anyway).

Quoted 1 line

[1]http://lists.infradead.org/pipermail/linux-mtd/2016-April/067187.html

Quoted 2 lines

Maybe I figured it out. Unfortunately, it is only compile tested. Does it look approximately right?

Yep that's definitely better. Just one thing missing (see below), otherwise it looks good.

Hi!

Quoted 5 lines

Quoted 2 lines

Maybe I figured it out. Unfortunately, it is only compile tested. Does it look approximately right?

Yep that's definitely better. Just one thing missing (see below), otherwise it looks good.

I'm copying from tango_nand, therefore I had to check tango_nand, too.

static int check_erased_page(struct nand_chip *chip, u8 *buf) {
...

                res = nand_check_erased_ecc_chunk(buf, pkt_size, ecc, ecc_size,
                                                  meta, meta_len,
                                                  chip->ecc.strength);
                if (res < 0)
                        mtd->ecc_stats.failed++;
                else
                        mtd->ecc_stats.corrected += res;

                bitflips = max(res, bitflips);
...
        return bitflips;
}

static int tango_read_page(struct mtd_info *mtd, struct nand_chip *chip,
u8 *buf, int oob_required, int page)
{
...

        res = decode_error_report(nfc);
        if (res < 0) {
                chip->ecc.read_oob_raw(mtd, chip, page);
                res = check_erased_page(chip, buf);
        }

        return res;
}

So nand_check_erased_ecc_chunk() returns < 0 (failed ECC), but then we perform max() with bitflips (lets say 1, correctable ECC) and return 1? tango_read_page then returns 1 (correctable ECC) forgetting about failed ECC...?