Skip to contentSkip to conversation

Adding offset keeping option when mapping data via SWIOTLB.

PATCH In review Jianxiong Gao 3 participants 3 messages Nov 23, 2020 – Nov 24, 2020

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Adding offset keeping option when mapping data via SWIOTLB.

NVMe driver and other applications may depend on the data offset to operate correctly. Currently when unaligned data is mapped via SWIOTLB, the data is mapped as slab aligned with the SWIOTLB. When booting with --swiotlb=force option and using NVMe as interface, running mkfs.xfs on Rhel fails because of the unalignment issue. This patch adds an option to make sure the mapped data preserves its offset of the orginal addrss. Tested on latest kernel that this patch fixes the issue.

---
 drivers/nvme/host/pci.c     |  3 ++-
 include/linux/dma-mapping.h |  8 ++++++++
 kernel/dma/swiotlb.c        | 13 +++++++++++++
 3 files changed, 23 insertions(+), 1 deletion(-)
drivers/nvme/host/pci.c +2 -1
index 0578ff253c47..a366fb8a1ff0 100644
@@ -833,7 +833,8 @@ static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req,
 				iod->nents, rq_dma_dir(req), DMA_ATTR_NO_WARN);
 	else
 		nr_mapped = dma_map_sg_attrs(dev->dev, iod->sg, iod->nents,
-					     rq_dma_dir(req), DMA_ATTR_NO_WARN);
+				rq_dma_dir(req),
+			       	DMA_ATTR_NO_WARN|DMA_ATTR_SWIOTLB_KEEP_OFFSET);
 	if (!nr_mapped)
 		goto out;
 
include/linux/dma-mapping.h +8 -0
index 956151052d45..e46d23d9fa20 100644
@@ -61,6 +61,14 @@
  */
 #define DMA_ATTR_PRIVILEGED		(1UL << 9)
 
+/*
+ * DMA_ATTR_SWIOTLB_KEEP_OFFSET: used to indicate that the buffer has to keep
+ * its offset when mapped via SWIOTLB. Some application functionality depends
+ * on the address offset, thus when buffers are mapped via SWIOTLB, the offset
+ * needs to be preserved.
+ */
+#define DMA_ATTR_SWIOTLB_KEEP_OFFSET	(1UL << 10)
+
 /*
  * A dma_addr_t can hold any valid DMA or bus address for the platform.  It can
  * be given to a device to use as a DMA source or target.  It is specific to a
kernel/dma/swiotlb.c +13 -0
index 781b9dca197c..f43d7be1342d 100644
@@ -483,6 +483,13 @@ phys_addr_t swiotlb_tbl_map_single(struct device *hwdev, phys_addr_t orig_addr,
 	max_slots = mask + 1
 		    ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
 		    : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
+ 
+	/*
+	 * If we need to keep the offset when mapping, we need to add the offset
+	 * to the total set we need to allocate in SWIOTLB
+	 */
+	if (attrs & DMA_ATTR_SWIOTLB_KEEP_OFFSET)
+		alloc_size += offset_in_page(orig_addr);
 
 	/*
 	 * For mappings greater than or equal to a page, we limit the stride
@@ -567,6 +574,12 @@ phys_addr_t swiotlb_tbl_map_single(struct device *hwdev, phys_addr_t orig_addr,
 	 */
 	for (i = 0; i < nslots; i++)
 		io_tlb_orig_addr[index+i] = orig_addr + (i << IO_TLB_SHIFT);
+	/*
+	 * When keeping the offset of the original data, we need to advance
+	 * the tlb_addr by the offset of orig_addr.
+	 */
+	if (attrs & DMA_ATTR_SWIOTLB_KEEP_OFFSET)
+		tlb_addr += orig_addr & (PAGE_SIZE - 1);
 	if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
 	    (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
 		swiotlb_bounce(orig_addr, tlb_addr, mapping_size, DMA_TO_DEVICE);
Signature · 1 line
2.27.0
#1 lore raw
Quoted 11 lines · Jianxiong Gao

NVMe driver and other applications may depend on the data offset to operate correctly. Currently when unaligned data is mapped via SWIOTLB, the data is mapped as slab aligned with the SWIOTLB. When booting with --swiotlb=force option and using NVMe as interface, running mkfs.xfs on Rhel fails because of the unalignment issue. This patch adds an option to make sure the mapped data preserves its offset of the orginal addrss. Tested on latest kernel that this patch fixes the issue.

I think we actually need to do this by default. There are plenty of other hardware designs that rely on dma mapping not adding offsets that did not exist, e.g. ahci and various RDMA NICs.
_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu

#2 lore raw
In reply to Jianxiong Gao
Quoted 5 lines · Jianxiong Gao

NVMe driver and other applications may depend on the data offset to operate correctly. Currently when unaligned data is mapped via SWIOTLB, the data is mapped as slab aligned with the SWIOTLB. When booting with --swiotlb=force option and using NVMe as interface, running mkfs.xfs on Rhel fails because of the unalignment issue.

RHEL? So a specific RHEL kernel. Is there a Red Hat bug created for this that can be linked to this patch to make it easier for folks to figure this?

Why would you be using swiotlb=force?
Ah, you are using AMD SEV!

Quoted 2 lines

This patch adds an option to make sure the mapped data preserves its offset of the orginal addrss. Tested on latest kernel that

s/addrss/address/

Quoted 78 lines

this patch fixes the issue.

---
 drivers/nvme/host/pci.c     |  3 ++-
 include/linux/dma-mapping.h |  8 ++++++++
 kernel/dma/swiotlb.c        | 13 +++++++++++++
 3 files changed, 23 insertions(+), 1 deletion(-)
drivers/nvme/host/pci.c +2 -1
index 0578ff253c47..a366fb8a1ff0 100644
@@ -833,7 +833,8 @@ static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req,
 				iod->nents, rq_dma_dir(req), DMA_ATTR_NO_WARN);
 	else
 		nr_mapped = dma_map_sg_attrs(dev->dev, iod->sg, iod->nents,
-					     rq_dma_dir(req), DMA_ATTR_NO_WARN);
+				rq_dma_dir(req),
+			       	DMA_ATTR_NO_WARN|DMA_ATTR_SWIOTLB_KEEP_OFFSET);
 	if (!nr_mapped)
 		goto out;
 
include/linux/dma-mapping.h +8 -0
index 956151052d45..e46d23d9fa20 100644
@@ -61,6 +61,14 @@
  */
 #define DMA_ATTR_PRIVILEGED		(1UL << 9)
 
+/*
+ * DMA_ATTR_SWIOTLB_KEEP_OFFSET: used to indicate that the buffer has to keep
+ * its offset when mapped via SWIOTLB. Some application functionality depends
+ * on the address offset, thus when buffers are mapped via SWIOTLB, the offset
+ * needs to be preserved.
+ */
+#define DMA_ATTR_SWIOTLB_KEEP_OFFSET	(1UL << 10)
+
 /*
  * A dma_addr_t can hold any valid DMA or bus address for the platform.  It can
  * be given to a device to use as a DMA source or target.  It is specific to a
kernel/dma/swiotlb.c +13 -0
index 781b9dca197c..f43d7be1342d 100644
@@ -483,6 +483,13 @@ phys_addr_t swiotlb_tbl_map_single(struct device *hwdev, phys_addr_t orig_addr,
 	max_slots = mask + 1
 		    ? ALIGN(mask + 1, 1 << IO_TLB_SHIFT) >> IO_TLB_SHIFT
 		    : 1UL << (BITS_PER_LONG - IO_TLB_SHIFT);
+ 
+	/*
+	 * If we need to keep the offset when mapping, we need to add the offset
+	 * to the total set we need to allocate in SWIOTLB
+	 */
+	if (attrs & DMA_ATTR_SWIOTLB_KEEP_OFFSET)
+		alloc_size += offset_in_page(orig_addr);
 
 	/*
 	 * For mappings greater than or equal to a page, we limit the stride
@@ -567,6 +574,12 @@ phys_addr_t swiotlb_tbl_map_single(struct device *hwdev, phys_addr_t orig_addr,
 	 */
 	for (i = 0; i < nslots; i++)
 		io_tlb_orig_addr[index+i] = orig_addr + (i << IO_TLB_SHIFT);
+	/*
+	 * When keeping the offset of the original data, we need to advance
+	 * the tlb_addr by the offset of orig_addr.
+	 */
+	if (attrs & DMA_ATTR_SWIOTLB_KEEP_OFFSET)
+		tlb_addr += orig_addr & (PAGE_SIZE - 1);
 	if (!(attrs & DMA_ATTR_SKIP_CPU_SYNC) &&
 	    (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL))
 		swiotlb_bounce(orig_addr, tlb_addr, mapping_size, DMA_TO_DEVICE);
Signature · 1 line
2.27.0

_______________________________________________
iommu mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/iommu