Skip to contentSkip to conversation

scsi: esp_scsi: report the residual byte count

PATCH Applied Tim Douglas 3 participants 4 messages Nov 18 – Nov 20 SCSI

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 scsi: esp_scsi: report the residual byte count

The ESP core tracks the number of bytes left to transfer in tot_residue but does not report that count to the SCSI midlayer. As a result, SG_IO returns a residual of zero even when a target transfers fewer bytes than requested, as can happen with INQUIRY and MODE SENSE.

Report tot_residue with scsi_set_resid() when a command with a data buffer completes with DID_OK. Initialize the scatterlist DMA length on the identity-mapped PIO/pseudo-DMA path so the residual starts from the buffer length on architectures with a separate DMA length.

On a Performa 6200CD with an out-of-tree ESP front end, an SG_IO INQUIRY with an allocation length of 255 returns 36 bytes from the internal CD-ROM drive. The residual is 219 with the fix and 0 before it.

AI assistance was used to investigate ESP transfer accounting and draft the fix.

---
 drivers/scsi/esp_scsi.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
drivers/scsi/esp_scsi.c +5 -1
index 9993b0ae9..162f8fe14 100644
@@ -388,6 +388,7 @@ static void esp_map_dma(struct esp *esp, struct scsi_cmnd *cmd)
 
 		scsi_for_each_sg(cmd, s, spriv->num_sg, i) {
 			s->dma_address = (uintptr_t)sg_virt(s);
+			sg_dma_len(s) = s->length;
 			total += sg_dma_len(s);
 		}
 	} else {
@@ -907,8 +908,11 @@ static void esp_cmd_is_done(struct esp *esp, struct esp_cmd_entry *ent,
 	esp_free_lun_tag(ent, dev->hostdata);
 	cmd->result = 0;
 	set_host_byte(cmd, host_byte);
-	if (host_byte == DID_OK)
+	if (host_byte == DID_OK) {
 		set_status_byte(cmd, ent->status);
+		if (cmd->sc_data_direction != DMA_NONE)
+			scsi_set_resid(cmd, ESP_CMD_PRIV(cmd)->tot_residue);
+	}
 
 	if (ent->eh_done) {
 		complete(ent->eh_done);

base-commit: aebb7274476403040d582c61478df8f4a7bf9cd3

Signature · 1 line
2.39.5
Tested-by on v1 1/1
#1 lore raw
Quoted 10 lines · Tim Douglas

The ESP core tracks the number of bytes left to transfer in tot_residue but does not report that count to the SCSI midlayer. As a result, SG_IO returns a residual of zero even when a target transfers fewer bytes than requested, as can happen with INQUIRY and MODE SENSE.

Report tot_residue with scsi_set_resid() when a command with a data buffer completes with DID_OK. Initialize the scatterlist DMA length on the identity-mapped PIO/pseudo-DMA path so the residual starts from the buffer length on architectures with a separate DMA length.

Quoted 8 lines

On a Performa 6200CD with an out-of-tree ESP front end, an SG_IO INQUIRY with an allocation length of 255 returns 36 bytes from the internal CD-ROM drive. The residual is 219 with the fix and 0 before it.

AI assistance was used to investigate ESP transfer accounting and draft the fix.

Regarding that front-end for the esp_scsi driver, do you think it can replace the mac53c94 driver? Has your driver been published yet?

Quoted 32 lines
---
 drivers/scsi/esp_scsi.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
drivers/scsi/esp_scsi.c +5 -1
index 9993b0ae9..162f8fe14 100644
@@ -388,6 +388,7 @@ static void esp_map_dma(struct esp *esp, struct scsi_cmnd *cmd)
 
 		scsi_for_each_sg(cmd, s, spriv->num_sg, i) {
 			s->dma_address = (uintptr_t)sg_virt(s);
+			sg_dma_len(s) = s->length;
 			total += sg_dma_len(s);
 		}
 	} else {
@@ -907,8 +908,11 @@ static void esp_cmd_is_done(struct esp *esp, struct esp_cmd_entry *ent,
 	esp_free_lun_tag(ent, dev->hostdata);
 	cmd->result = 0;
 	set_host_byte(cmd, host_byte);
-	if (host_byte == DID_OK)
+	if (host_byte == DID_OK) {
 		set_status_byte(cmd, ent->status);
+		if (cmd->sc_data_direction != DMA_NONE)
+			scsi_set_resid(cmd, ESP_CMD_PRIV(cmd)->tot_residue);
+	}
 
 	if (ent->eh_done) {
 		complete(ent->eh_done);

base-commit: aebb7274476403040d582c61478df8f4a7bf9cd3

#2 lore raw
Quoted 2 lines · Finn Thain

Regarding that front-end for the esp_scsi driver, do you think it can replace the mac53c94 driver? Has your driver been published yet?

Thanks for testing! Probably not. The front end uses PrimeTime pseudo-DMA like mac_esp, and a replacement for mac53c94 would need DBDMA support. (The "cordyceps" Performa 52xx/62xx family is a bit of an odd duck, essentially a PowerPC bolted onto Quadra-style IO.)

The driver is still in my local tree, happy to share the patch if you’d like to take a look.

-Tim

maintainer Applied
#3 lore raw
In reply to Tim Douglas

Tim,

Quoted 4 lines

The ESP core tracks the number of bytes left to transfer in tot_residue but does not report that count to the SCSI midlayer. As a result, SG_IO returns a residual of zero even when a target transfers fewer bytes than requested, as can happen with INQUIRY and MODE SENSE.

Applied to 7.5/scsi-staging, thanks!

Signature · 1 line
Martin K. Petersen