vhost: fix async address mapping

From: Xuan Ding <[email protected]>

This patch renames the host_phys_addr to host_iova in guest_page struct. The host_phys_addr is iova, it depends on the DPDK IOVA mode.

Signed-off-by Xuan Ding <[email protected]>

---
  lib/vhost/vhost.h      | 10 +++++-----
  lib/vhost/vhost_user.c | 20 ++++++++++----------
  lib/vhost/virtio_net.c | 11 ++++++-----
  3 files changed, 21 insertions(+), 20 deletions(-)

From: Xuan Ding <[email protected]>

When choosing IOVA as PA mode, IOVA is likely to be discontinuous, which requires page by page mapping for DMA devices. To be consistent, this patch implements page by page mapping instead of mapping at the region granularity for both IOVA as VA and PA mode.

Fixes 7c61fa08b716 ("vhost: enable IOMMU for async vhost")

Signed-off-by Xuan Ding <[email protected]>

Signed-off-by Yuan Wang <[email protected]>

---
  lib/vhost/vhost.h      |   1 +
  lib/vhost/vhost_user.c | 116 ++++++++++++++++++++---------------------
  2 files changed, 57 insertions(+), 60 deletions(-)

lib/vhost/vhost.h +1 -0

index ca7f58039d..9521ae56da 100644
@@ -355,6 +355,7 @@ struct vring_packed_desc_event {
  struct guest_page {
  	uint64_t guest_phys_addr;
  	uint64_t host_iova;
+	uint64_t host_user_addr;
  	uint64_t size;
  };
 

lib/vhost/vhost_user.c +0 -1

index 95c9df697e..48c08716ba 100644
@@ -143,57 +143,56 @@ get_blk_size(int fd)
  	return ret == -1 ? (uint64_t)-1 : (uint64_t)stat.st_blksize;
  }
 
-static int

lib/vhost/vhost_user.c +42 -42

  	if (do_map) {
-		/* Add mapped region into the default container of DPDK. */
-		ret = rte_vfio_container_dma_map(RTE_VFIO_DEFAULT_CONTAINER_FD,
-						 region->host_user_addr,
-						 host_iova,
-						 region->size);
-		if (ret) {
-			/*
-			 * DMA device may bind with kernel driver, in this case,
-			 * we don't need to program IOMMU manually. However, if no
-			 * device is bound with vfio/uio in DPDK, and vfio kernel
-			 * module is loaded, the API will still be called and return
-			 * with ENODEV/ENOSUP.
-			 *
-			 * DPDK vfio only returns ENODEV/ENOSUP in very similar
-			 * situations(vfio either unsupported, or supported
-			 * but no devices found). Either way, no mappings could be
-			 * performed. We treat it as normal case in async path.
-			 */
-			if (rte_errno == ENODEV || rte_errno == ENOTSUP)
-				return 0;
-
-			VHOST_LOG_CONFIG(ERR, "DMA engine map failed\n");
-			/* DMA mapping errors won't stop VHST_USER_SET_MEM_TABLE. */
-			return 0;
+		for (i = 0; i < dev->nr_guest_pages; i++) {
+			page = &dev->guest_pages[i];
+			ret = rte_vfio_container_dma_map(RTE_VFIO_DEFAULT_CONTAINER_FD,
+							 page->host_user_addr,
+							 page->host_iova,
+							 page->size);
+			if (ret) {
+				/*
+				 * DMA device may bind with kernel driver, in this case,
+				 * we don't need to program IOMMU manually. However, if no
+				 * device is bound with vfio/uio in DPDK, and vfio kernel
+				 * module is loaded, the API will still be called and return
+				 * with ENODEV.
+				 *
+				 * DPDK vfio only returns ENODEV in very similar situations
+				 * (vfio either unsupported, or supported but no devices found).
+				 * Either way, no mappings could be performed. We treat it as
+				 * normal case in async path. This is a workaround.
+				 */
+				if (rte_errno == ENODEV)
+					return;
+
+				/* DMA mapping errors won't stop VHOST_USER_SET_MEM_TABLE. */
+				VHOST_LOG_CONFIG(ERR, "DMA engine map failed\n");
+			}
  		}
 
  	} else {
-		/* Remove mapped region from the default container of DPDK. */
-		ret = rte_vfio_container_dma_unmap(RTE_VFIO_DEFAULT_CONTAINER_FD,
-						   region->host_user_addr,
-						   host_iova,
-						   region->size);
-		if (ret) {
-			/* like DMA map, ignore the kernel driver case when unmap. */
-			if (rte_errno == EINVAL)
-				return 0;
-
-			VHOST_LOG_CONFIG(ERR, "DMA engine unmap failed\n");
-			return ret;
+		for (i = 0; i < dev->nr_guest_pages; i++) {
+			page = &dev->guest_pages[i];
+			ret = rte_vfio_container_dma_unmap(RTE_VFIO_DEFAULT_CONTAINER_FD,
+							   page->host_user_addr,
+							   page->host_iova,
+							   page->size);
+			if (ret) {
+				/* like DMA map, ignore the kernel driver case when unmap. */
+				if (rte_errno == EINVAL)
+					return;
+
+				VHOST_LOG_CONFIG(ERR, "DMA engine unmap failed\n");
+			}
  		}
  	}
-
-	return ret;
  }
 
  static void

@@ -205,12 +204,12 @@ free_mem_region(struct virtio_net *dev)
  	if (!dev || !dev->mem)
  		return;
 
+	if (dev->async_copy && rte_vfio_is_enabled("vfio"))
+		async_dma_map(dev, false);
+
  	for (i = 0; i < dev->mem->nregions; i++) {
  		reg = &dev->mem->regions[i];
  		if (reg->host_user_addr) {
-			if (dev->async_copy && rte_vfio_is_enabled("vfio"))
-				async_dma_map(reg, false);
-
  			munmap(reg->mmap_addr, reg->mmap_size);
  			close(reg->fd);
  		}

@@ -978,7 +977,7 @@ vhost_user_set_vring_base(struct virtio_net **pdev,
 
  static int
  add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
-		   uint64_t host_iova, uint64_t size)
+		   uint64_t host_iova, uint64_t host_user_addr, uint64_t size)
  {
  	struct guest_page *page, *last_page;
  	struct guest_page *old_pages;

lib/vhost/vhost_user.c +4 -11

  			last_page->size += size;
  			return 0;
  		}

@@ -1009,6 +1009,7 @@ add_one_guest_page(struct virtio_net *dev, uint64_t guest_phys_addr,
  	page = &dev->guest_pages[dev->nr_guest_pages++];
  	page->guest_phys_addr = guest_phys_addr;
  	page->host_iova  = host_iova;
+	page->host_user_addr = host_user_addr;
  	page->size = size;
 
  	return 0;

@@ -1028,7 +1029,8 @@ add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
  	size = page_size - (guest_phys_addr & (page_size - 1));
  	size = RTE_MIN(size, reg_size);
 
-	if (add_one_guest_page(dev, guest_phys_addr, host_iova, size) < 0)
+	if (add_one_guest_page(dev, guest_phys_addr, host_iova,
+			       host_user_addr, size) < 0)
  		return -1;
 
  	host_user_addr  += size;

@@ -1040,7 +1042,7 @@ add_guest_pages(struct virtio_net *dev, struct rte_vhost_mem_region *reg,
  		host_iova = rte_mem_virt2iova((void *)(uintptr_t)
  						  host_user_addr);
  		if (add_one_guest_page(dev, guest_phys_addr, host_iova,
-				size) < 0)
+				       host_user_addr, size) < 0)
  			return -1;
 
  		host_user_addr  += size;

@@ -1215,7 +1217,6 @@ vhost_user_mmap_region(struct virtio_net *dev,
  	uint64_t mmap_size;
  	uint64_t alignment;
  	int populate;
-	int ret;
 
  	/* Check for memory_size + mmap_offset overflow */
  	if (mmap_offset >= -region->size) {

@@ -1274,14 +1275,6 @@ vhost_user_mmap_region(struct virtio_net *dev,
  			VHOST_LOG_CONFIG(ERR, "adding guest pages to region failed.\n");
  			return -1;
  		}
-
-		if (rte_vfio_is_enabled("vfio")) {
-			ret = async_dma_map(region, true);
-			if (ret) {
-				VHOST_LOG_CONFIG(ERR, "Configure IOMMU for DMA engine failed\n");
-				return -1;
-			}
-		}
  	}
 
  	VHOST_LOG_CONFIG(INFO,

From: Xuan Ding <[email protected]>

When choosing IOVA as PA mode, IOVA is likely to be discontinuous, which requires page by page mapping for DMA devices. To be consistent, this patch implements page by page mapping instead of mapping at the region granularity for both IOVA as VA and PA mode.

Fixes 7c61fa08b716 ("vhost: enable IOMMU for async vhost")

Signed-off-by Xuan Ding <[email protected]>

Signed-off-by Yuan Wang <[email protected]>

---
  lib/vhost/vhost.h      |   1 +
  lib/vhost/vhost_user.c | 116 ++++++++++++++++++++---------------------
  2 files changed, 57 insertions(+), 60 deletions(-)

From: Xuan Ding <[email protected]>

This patchset fixes the issue of incorrect DMA mapping in PA mode. Due to the ambiguity of host_phys_addr naming in the guest page struct, rename it to host_iova.

v2:
* Change the order of patch.

Xuan Ding (2):
vhost: rename field in guest page struct
vhost: fix physical address mapping

  lib/vhost/vhost.h      |  11 ++--
  lib/vhost/vhost_user.c | 130 ++++++++++++++++++++---------------------
  lib/vhost/virtio_net.c |  11 ++--
  3 files changed, 75 insertions(+), 77 deletions(-)

Quoted 8 lines · On 1/19/22 16:10, [email protected] wrote:

From: Xuan Ding <[email protected]>

This patchset fixes the issue of incorrect DMA mapping in PA mode. Due to the ambiguity of host_phys_addr naming in the guest page struct, rename it to host_iova.

v2:
* Change the order of patch.

I'm not sure why you changed the order of the patches.
Now, the second one is the fix, so it will make the backport more difficult. Either both are considered to be fixes. I think it can make sense as the renaming does not introduce risk of regression and will make backporting patches easier in the future.

Other solution is to reverse the order again, but I think tagging the renaming as a fix is OK for me here.

What do you think?

Regards,
Maxime

Quoted 10 lines

Xuan Ding (2):
vhost: rename field in guest page struct
vhost: fix physical address mapping

   lib/vhost/vhost.h      |  11 ++--
   lib/vhost/vhost_user.c | 130 ++++++++++++++++++++---------------------
   lib/vhost/virtio_net.c |  11 ++--
   3 files changed, 75 insertions(+), 77 deletions(-)

Quoted 18 lines · Maxime Coquelin

Quoted 8 lines · On 1/19/22 16:10, [email protected] wrote:

From: Xuan Ding <[email protected]>

This patchset fixes the issue of incorrect DMA mapping in PA mode. Due to the ambiguity of host_phys_addr naming in the guest page struct, rename it to host_iova.

v2:
* Change the order of patch.

I'm not sure why you changed the order of the patches.
Now, the second one is the fix, so it will make the backport more difficult. Either both are considered to be fixes. I think it can make sense as the renaming does not introduce risk of regression and will make backporting patches easier in the future.

Quoted 5 lines

Other solution is to reverse the order again, but I think tagging the renaming as a fix is OK for me here.

What do you think?

Either way sounds ok to me, but can you also add stable tag(s). There isn't a stable tag on either patch at present. Thanks.

Quoted 15 lines

Regards,
Maxime

Quoted 11 lines

Xuan Ding (2):
    vhost: rename field in guest page struct
    vhost: fix physical address mapping

   lib/vhost/vhost.h      |  11 ++--
   lib/vhost/vhost_user.c | 130
++++++++++++++++++++---------------------
   lib/vhost/virtio_net.c |  11 ++--

   3 files changed, 75 insertions(+), 77 deletions(-)

From: Xuan Ding <[email protected]>

This patchset fixes the issue of incorrect DMA mapping in PA mode. Due to the ambiguity of host_phys_addr naming in the guest page struct, rename it to host_iova.

v2:
* Change the order of patch.

Xuan Ding (2):
vhost: rename field in guest page struct
vhost: fix physical address mapping

  lib/vhost/vhost.h      |  11 ++--
  lib/vhost/vhost_user.c | 130 ++++++++++++++++++++---------------------
  lib/vhost/virtio_net.c |  11 ++--
  3 files changed, 75 insertions(+), 77 deletions(-)

-----Original Message-----
From: Maxime Coquelin <[email protected]>
Sent: 2022年2月4日 18:57
To: Ding, Xuan <[email protected]>; Xia, Chenbo <[email protected]>
Cc: [email protected]; Hu, Jiayu <[email protected]>; Wang, YuanX <[email protected]>
Subject: Re: [PATCH v2 0/2] vhost: fix async address mapping

Quoted 8 lines · On 1/19/22 16:10, [email protected] wrote:

From: Xuan Ding <[email protected]>

This patchset fixes the issue of incorrect DMA mapping in PA mode. Due to the ambiguity of host_phys_addr naming in the guest page struct, rename it to host_iova.

v2:
* Change the order of patch.

Quoted 5 lines

Xuan Ding (2):
vhost: rename field in guest page struct
vhost: fix physical address mapping

Quoted 5 lines

  lib/vhost/vhost.h      |  11 ++--
  lib/vhost/vhost_user.c | 130 ++++++++++++++++++++---------------------
  lib/vhost/virtio_net.c |  11 ++--
  3 files changed, 75 insertions(+), 77 deletions(-)

I was willing to apply the series, but it does not apply. Could you please rebase with taking our comments into account?

Thanks,
Maxime