Skip to contentSkip to conversation

aic94xx: Don't call pci_map_sg for already-mapped scatterlists

PATCH In review Darrick J. Wong 2 participants 3 messages Nov 11, 2006 – Dec 5, 2006

Subsystems (2)
SCSI SCSI libsas
lore
Share & tools
lore

Darrick J. Wong

author
#0 lore raw

PATCH 1/1 aic94xx: Don't call pci_map_sg for already-mapped scatterlists

It turns out that libata has already dma_map_sg'd the scatterlist entries that go with an ata_queued_cmd by the time it calls sas_ata_qc_issue. sas_ata_qc_issue passes this scatterlist to aic94xx. Unfortunately, aic94xx assumes that any scatterlist passed to it needs to be pci_map_sg'd... which blows away the mapping that libata created! This causes (on a x260) Calgary IOMMU table leaks and duplicate frees when aic94xx and libata try to {pci,dma}_unmap_sg the scatterlist.

Since dma_map_sg and pci_map_sg are fed the same struct device, I think it's safe to add a flag to sas_task that tells aic94xx that it need not map the scatterlist. It didn't break anything on the x260, though I don't have any SATAPI devices to test with. Is this the correct approach to fixing this problem?

--

Signed-off-by Darrick J. Wong <[email protected]>
drivers/scsi/aic94xx/aic94xx_task.c +11 -6
index 9840708..eaffaa6 100644
@@ -74,8 +74,11 @@ static inline int asd_map_scatterlist(st
 		return 0;
 	}
 
-	num_sg = pci_map_sg(asd_ha->pcidev, task->scatter, task->num_scatter,
-			    task->data_dir);
+	if (task->external_sg)
+		num_sg = task->num_scatter;
+	else
+		num_sg = pci_map_sg(asd_ha->pcidev, task->scatter,
+				    task->num_scatter, task->data_dir);
 	if (num_sg == 0)
 		return -ENOMEM;
 

@@ -120,8 +123,9 @@ static inline int asd_map_scatterlist(st
 
 	return 0;
 err_unmap:
-	pci_unmap_sg(asd_ha->pcidev, task->scatter, task->num_scatter,
-		     task->data_dir);
+	if (!task->external_sg)
+		pci_unmap_sg(asd_ha->pcidev, task->scatter, task->num_scatter,
+			     task->data_dir);
 	return res;
 }
 

@@ -142,8 +146,9 @@ static inline void asd_unmap_scatterlist
 	}
 
 	asd_free_coherent(asd_ha, ascb->sg_arr);
-	pci_unmap_sg(asd_ha->pcidev, task->scatter, task->num_scatter,
-		     task->data_dir);
+	if (!task->external_sg)
+		pci_unmap_sg(asd_ha->pcidev, task->scatter, task->num_scatter,
+			     task->data_dir);
 }
 
 /* ---------- Task complete tasklet ---------- */
drivers/scsi/libsas/sas_ata.c +1 -0
index 8f8cd40..04115c3 100644
@@ -167,6 +167,7 @@ static unsigned int sas_ata_qc_issue(str
 		task->num_scatter = num;
 	}
 
+	task->external_sg = 1;
 	task->data_dir = qc->dma_dir;
 	task->scatter = qc->__sg;
 	task->ata_task.retry_count = 1;
include/scsi/libsas.h +1 -0
index 1263bd5..25a1fb3 100644
@@ -546,6 +546,7 @@ struct sas_task {
 	void   *uldd_task;
 
 	struct work_struct abort_work;
+	int    external_sg;
 };
#1 lore raw
Quoted 8 lines · Darrick J. Wong

It turns out that libata has already dma_map_sg'd the scatterlist entries that go with an ata_queued_cmd by the time it calls sas_ata_qc_issue. sas_ata_qc_issue passes this scatterlist to aic94xx. Unfortunately, aic94xx assumes that any scatterlist passed to it needs to be pci_map_sg'd... which blows away the mapping that libata created! This causes (on a x260) Calgary IOMMU table leaks and duplicate frees when aic94xx and libata try to {pci,dma}_unmap_sg the scatterlist.

Quoted 5 lines

Since dma_map_sg and pci_map_sg are fed the same struct device, I think it's safe to add a flag to sas_task that tells aic94xx that it need not map the scatterlist. It didn't break anything on the x260, though I don't have any SATAPI devices to test with. Is this the correct approach to fixing this problem?

Rather than introduce an extra flag, I think we can key of the protocol
flag: libata is the only thing that initiates STP tasks. How does this look?

James

drivers/scsi/aic94xx/aic94xx_task.c +13 -6
Index: BUILD-2.6/drivers/scsi/aic94xx/aic94xx_task.c
===================================================================
@@ -74,8 +74,13 @@ static inline int asd_map_scatterlist(st
 		return 0;
 	}
 
-	num_sg = pci_map_sg(asd_ha->pcidev, task->scatter, task->num_scatter,
-			    task->data_dir);
+	/* STP tasks come from libata which has already mapped
+	 * the SG list */
+	if (task->task_proto == SAS_PROTOCOL_STP)
+		num_sg = task->num_scatter;
+	else
+		num_sg = pci_map_sg(asd_ha->pcidev, task->scatter,
+				    task->num_scatter, task->data_dir);
 	if (num_sg == 0)
 		return -ENOMEM;
 

@@ -120,8 +125,9 @@ static inline int asd_map_scatterlist(st
 
 	return 0;
 err_unmap:
-	pci_unmap_sg(asd_ha->pcidev, task->scatter, task->num_scatter,
-		     task->data_dir);
+	if (task->task_proto != SAS_PROTOCOL_STP)
+		pci_unmap_sg(asd_ha->pcidev, task->scatter, task->num_scatter,
+			     task->data_dir);
 	return res;
 }
 

@@ -142,8 +148,9 @@ static inline void asd_unmap_scatterlist
 	}
 
 	asd_free_coherent(asd_ha, ascb->sg_arr);
-	pci_unmap_sg(asd_ha->pcidev, task->scatter, task->num_scatter,
-		     task->data_dir);
+	if (task->task_proto != SAS_PROTOCOL_STP)
+		pci_unmap_sg(asd_ha->pcidev, task->scatter, task->num_scatter,
+			     task->data_dir);
 }
 
 /* ---------- Task complete tasklet ---------- */
author
#2 lore raw
Quoted 3 lines · James Bottomley

Rather than introduce an extra flag, I think we can key of the protocol
flag: libata is the only thing that initiates STP tasks. How does this look?

ACK.

--D