Skip to contentSkip to conversation

libata: make ata_sg_setup_one() trim zero length sg

PATCH In review Tejun Heo 2 participants 9 messages Feb 17, 2006 – Feb 20, 2006

Subsystems (2)
lore
Share & tools
lore

Tejun Heo

author
#0 lore raw

PATCH 1/1 libata: make ata_sg_setup_one() trim zero length sg

This patch makes ata_sg_setup_one() trim sg entry (thus making qc->n_elem zero) if padding results in zero length sg entry.

---

Jeff, currently, in both sg_setup() and sg_setup_one() cases, zero length sg entry can reach low level driver fill_sg() function which could possibly cause weird problems. This and a following patch kill such cases.

 libata-core.c |   13 ++++++++-----
 1 file changed, 8 insertions(+), 5 deletions(-)
drivers/scsi/libata-core.c +8 -5
@@ -2583,7 +2583,7 @@ static void ata_sg_clean(struct ata_queu
 	WARN_ON(sg == NULL);
 
 	if (qc->flags & ATA_QCFLAG_SINGLE)
-		WARN_ON(qc->n_elem != 1);
+		WARN_ON(qc->n_elem > 1);
 
 	VPRINTK("unmapping %u sg elements\n", qc->n_elem);
 
@@ -2606,7 +2606,7 @@ static void ata_sg_clean(struct ata_queu
 			kunmap_atomic(addr, KM_IRQ0);
 		}
 	} else {
-		if (sg_dma_len(&sg[0]) > 0)
+		if (qc->n_elem)
 			dma_unmap_single(ap->host_set->dev,
 				sg_dma_address(&sg[0]), sg_dma_len(&sg[0]),
 				dir);
@@ -2784,6 +2784,7 @@ static int ata_sg_setup_one(struct ata_q
 	int dir = qc->dma_dir;
 	struct scatterlist *sg = qc->__sg;
 	dma_addr_t dma_address;
+	int trim_sg = 0;
 
 	/* we must lengthen transfers to end on a 32-bit boundary */
 	qc->pad_len = sg->length & 3;
@@ -2803,13 +2804,15 @@ static int ata_sg_setup_one(struct ata_q
 		sg_dma_len(psg) = ATA_DMA_PAD_SZ;
 		/* trim sg */
 		sg->length -= qc->pad_len;
+		if (sg->length == 0)
+			trim_sg = 1;
 
 		DPRINTK("padding done, sg->length=%u pad_len=%u\n",
 			sg->length, qc->pad_len);
 	}
 
-	if (!sg->length) {
-		sg_dma_address(sg) = 0;
+	if (trim_sg) {
+		qc->n_elem--;
 		goto skip_map;
 	}
 
@@ -2822,9 +2825,9 @@ static int ata_sg_setup_one(struct ata_q
 	}
 
 	sg_dma_address(sg) = dma_address;
-skip_map:
 	sg_dma_len(sg) = sg->length;
 
+skip_map:
 	DPRINTK("mapped buffer of %d bytes for %s\n", sg_dma_len(sg),
 		qc->tf.flags & ATA_TFLAG_WRITE ? "write" : "read");
 
#1 lore raw
Quoted 11 lines · Tejun Heo

This patch makes ata_sg_setup_one() trim sg entry (thus making qc->n_elem zero) if padding results in zero length sg entry.

---

Jeff, currently, in both sg_setup() and sg_setup_one() cases, zero length sg entry can reach low level driver fill_sg() function which could possibly cause weird problems. This and a following patch kill such cases.

Can you redo this series against vanilla upstream (and/or #upstream-fixes)?

Jeff

author
#2 lore raw
Quoted 9 lines · Jeff Garzik
Quoted 7 lines · Tejun Heo

This patch makes ata_sg_setup_one() trim sg entry (thus making qc->n_elem zero) if padding results in zero length sg entry.

---

Quoted 8 lines
Quoted 4 lines

Jeff, currently, in both sg_setup() and sg_setup_one() cases, zero length sg entry can reach low level driver fill_sg() function which could possibly cause weird problems. This and a following patch kill such cases.

Can you redo this series against vanilla upstream (and/or #upstream-fixes)?

Hello, Jeff.

Actually, this series is already against the vanilla upstream[1]. The order is...

libata: fix WARN_ON() condition in *_fill_sg() [2]
libata: make ata_sg_setup_one() trim zero length sg [3]
libata: fix qc->n_elem == 0 case handling in ata_qc_next_sg [4]

Signature · 6 lines
tejun

[1] f1b318793dcd2d9ff6b5ac06e7762098fa079cee
[2] http://article.gmane.org/gmane.linux.ide/8090
[3] http://article.gmane.org/gmane.linux.ide/8091
[4] http://article.gmane.org/gmane.linux.ide/8092
#3 lore raw
Quoted 21 lines · Tejun Heo
Quoted 9 lines · Jeff Garzik
Quoted 7 lines · Tejun Heo

This patch makes ata_sg_setup_one() trim sg entry (thus making qc->n_elem zero) if padding results in zero length sg entry.

---

Quoted 9 lines
Quoted 4 lines

Jeff, currently, in both sg_setup() and sg_setup_one() cases, zero length sg entry can reach low level driver fill_sg() function which could possibly cause weird problems. This and a following patch kill such cases.

Can you redo this series against vanilla upstream (and/or #upstream-fixes)?

Quoted 7 lines

Hello, Jeff.

Actually, this series is already against the vanilla upstream[1]. The order is...

libata: fix WARN_ON() condition in *_fill_sg() [2]
libata: make ata_sg_setup_one() trim zero length sg [3]
libata: fix qc->n_elem == 0 case handling in ata_qc_next_sg [4]

Sorry, I should have been more specific.

These changes should go into 2.6.16-rc, and should be based off the vanilla Linus tree. Too many "vanilla upstreams" running around...

Jeff

author
#4 lore raw

PATCH 1/3 libata: fix WARN_ON() condition in *_fill_sg()

For ATAPI commands, padding can reduce qc->n_elem by one and thus to zero making assert(qc->n_elem > 0)'s in ata_fill_sg() and qs_fill_sg() fail for legal commands. This patch fixes the assert()'s to take qc->pad_len into account.

Although the condition check seems a bit excessive, as this part of code isn't still stable yet, I think it's worth to keep those.

---

 drivers/scsi/libata-core.c |    2 +-
 drivers/scsi/sata_qstor.c  |    2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

cc1feb52b7a9a10c2438759d39e8649834e668fb

drivers/scsi/libata-core.c +1 -1
index 7ddd5a6..bbac87a 100644
@@ -2570,7 +2570,7 @@ static void ata_fill_sg(struct ata_queue
 	unsigned int idx;
 
 	assert(qc->__sg != NULL);
-	assert(qc->n_elem > 0);
+	assert(qc->n_elem > 0 || qc->pad_len > 0);
 
 	idx = 0;
 	ata_for_each_sg(sg, qc) {
drivers/scsi/sata_qstor.c +1 -1
index de05e28..80480f0 100644
@@ -277,7 +277,7 @@ static unsigned int qs_fill_sg(struct at
 	u8 *prd = pp->pkt + QS_CPB_BYTES;
 
 	assert(qc->__sg != NULL);
-	assert(qc->n_elem > 0);
+	assert(qc->n_elem > 0 || qc->pad_len > 0);
 
 	nelem = 0;
 	ata_for_each_sg(sg, qc) {
Signature · 1 line
1.1.5
#5 lore raw

Re: [PATCH 1/3] libata: fix WARN_ON() condition in *_fill_sg()

Quoted 9 lines · Tejun Heo

For ATAPI commands, padding can reduce qc->n_elem by one and thus to zero making assert(qc->n_elem > 0)'s in ata_fill_sg() and qs_fill_sg() fail for legal commands. This patch fixes the assert()'s to take qc->pad_len into account.

Although the condition check seems a bit excessive, as this part of code isn't still stable yet, I think it's worth to keep those.

applied 1-3

author
#6 lore raw

PATCH 1/1 libata: sg corner case fixes against Linus tree

In reply to Jeff Garzik

Hello, Jeff.

This is the three sg corner case fix patches[1] regenerated against the current Linus #master[2]. The assert()->WARN_ON() patch in libata-dev #upstream will collide with this one. It should be easy to resolve though.

--
tejun

[1] http://article.gmane.org/gmane.linux.ide/8090

    http://article.gmane.org/gmane.linux.ide/8091
    http://article.gmane.org/gmane.linux.ide/8092

[2] bd71c2b17468a2531fb4c81ec1d73520845e97e1

author
#7 lore raw

PATCH 2/3 libata: fix qc->n_elem == 0 case handling in ata_qc_next_sg

In reply to Jeff Garzik

This patch makes ata_for_each_sg() start with pad_sgent when qc->n_elem is zero. Previously, ata_for_each_sg() unconditionally started with qc->__sg, handling the first sg to fill_sg() routines even when the entry was invalid. And while at it, unwind ?: in ata_qc_next_sg() into if statement.

---

 include/linux/libata.h |   16 ++++++++++++++--
 1 files changed, 14 insertions(+), 2 deletions(-)

940e8d008d8e53b532de8de994477baf78e7007a

include/linux/libata.h +14 -2
index 9e5db29..c91be5e 100644
@@ -557,17 +557,29 @@ ata_sg_is_last(struct scatterlist *sg, s
 }
 
 static inline struct scatterlist *
+ata_qc_first_sg(struct ata_queued_cmd *qc)
+{
+	if (qc->n_elem)
+		return qc->__sg;
+	if (qc->pad_len)
+		return &qc->pad_sgent;
+	return NULL;
+}
+
+static inline struct scatterlist *
 ata_qc_next_sg(struct scatterlist *sg, struct ata_queued_cmd *qc)
 {
 	if (sg == &qc->pad_sgent)
 		return NULL;
 	if (++sg - qc->__sg < qc->n_elem)
 		return sg;
-	return qc->pad_len ? &qc->pad_sgent : NULL;
+	if (qc->pad_len)
+		return &qc->pad_sgent;
+	return NULL;
 }
 
 #define ata_for_each_sg(sg, qc) \
-	for (sg = qc->__sg; sg; sg = ata_qc_next_sg(sg, qc))
+	for (sg = ata_qc_first_sg(qc); sg; sg = ata_qc_next_sg(sg, qc))
 
 static inline unsigned int ata_tag_valid(unsigned int tag)
 {
Signature · 1 line
1.1.5
author
#8 lore raw

PATCH 3/3 libata: make ata_sg_setup_one() trim zero length sg

In reply to Jeff Garzik

This patch makes ata_sg_setup_one() trim sg entry (thus making qc->n_elem zero) if padding results in zero length sg entry.

---

 drivers/scsi/libata-core.c |   13 ++++++++-----
 1 files changed, 8 insertions(+), 5 deletions(-)

600e098c7092d0c0ba4c316e2cec2d1c8edc495c

drivers/scsi/libata-core.c +8 -5
index bbac87a..5f1d758 100644
@@ -2514,7 +2514,7 @@ static void ata_sg_clean(struct ata_queu
 	assert(sg != NULL);
 
 	if (qc->flags & ATA_QCFLAG_SINGLE)
-		assert(qc->n_elem == 1);
+		assert(qc->n_elem <= 1);
 
 	VPRINTK("unmapping %u sg elements\n", qc->n_elem);
 
@@ -2537,7 +2537,7 @@ static void ata_sg_clean(struct ata_queu
 			kunmap_atomic(addr, KM_IRQ0);
 		}
 	} else {
-		if (sg_dma_len(&sg[0]) > 0)
+		if (qc->n_elem)
 			dma_unmap_single(ap->host_set->dev,
 				sg_dma_address(&sg[0]), sg_dma_len(&sg[0]),
 				dir);
@@ -2715,6 +2715,7 @@ static int ata_sg_setup_one(struct ata_q
 	int dir = qc->dma_dir;
 	struct scatterlist *sg = qc->__sg;
 	dma_addr_t dma_address;
+	int trim_sg = 0;
 
 	/* we must lengthen transfers to end on a 32-bit boundary */
 	qc->pad_len = sg->length & 3;
@@ -2734,13 +2735,15 @@ static int ata_sg_setup_one(struct ata_q
 		sg_dma_len(psg) = ATA_DMA_PAD_SZ;
 		/* trim sg */
 		sg->length -= qc->pad_len;
+		if (sg->length == 0)
+			trim_sg = 1;
 
 		DPRINTK("padding done, sg->length=%u pad_len=%u\n",
 			sg->length, qc->pad_len);
 	}
 
-	if (!sg->length) {
-		sg_dma_address(sg) = 0;
+	if (trim_sg) {
+		qc->n_elem--;
 		goto skip_map;
 	}
 
@@ -2753,9 +2756,9 @@ static int ata_sg_setup_one(struct ata_q
 	}
 
 	sg_dma_address(sg) = dma_address;
-skip_map:
 	sg_dma_len(sg) = sg->length;
 
+skip_map:
 	DPRINTK("mapped buffer of %d bytes for %s\n", sg_dma_len(sg),
 		qc->tf.flags & ATA_TFLAG_WRITE ? "write" : "read");
 
Signature · 1 line
1.1.5