diff mbox series

[v3,2/2] vhost: enable IOMMU for async vhost

Message ID 20210925100358.61995-3-xuan.ding@intel.com (mailing list archive)
State Superseded, archived
Headers
Series support IOMMU for DMA device |

Checks

Context Check Description
ci/checkpatch warning coding style issues

Commit Message

Ding, Xuan Sept. 25, 2021, 10:03 a.m. UTC 
  The use of IOMMU has many advantages, such as isolation and address
translation. This patch extends the capbility of DMA engine to use
IOMMU if the DMA engine is bound to vfio.

When set memory table, the guest memory will be mapped
into the default container of DPDK.

Signed-off-by: Xuan Ding <xuan.ding@intel.com>
---
 lib/vhost/vhost.h      |   4 ++
 lib/vhost/vhost_user.c | 112 ++++++++++++++++++++++++++++++++++++++++-
 2 files changed, 114 insertions(+), 2 deletions(-)

Comments

Hu, Jiayu Sept. 27, 2021, 4:17 a.m. UTC | #1 
Hi Xuan,

> -----Original Message-----
> From: Ding, Xuan <xuan.ding@intel.com>
> Sent: Saturday, September 25, 2021 6:04 PM
> To: dev@dpdk.org; Burakov, Anatoly <anatoly.burakov@intel.com>;
> maxime.coquelin@redhat.com; Xia, Chenbo <chenbo.xia@intel.com>
> Cc: Hu, Jiayu <jiayu.hu@intel.com>; Jiang, Cheng1 <cheng1.jiang@intel.com>;
> Richardson, Bruce <bruce.richardson@intel.com>; Pai G, Sunil
> <sunil.pai.g@intel.com>; Wang, Yinan <yinan.wang@intel.com>; Yang,
> YvonneX <yvonnex.yang@intel.com>; Ding, Xuan <xuan.ding@intel.com>
> Subject: [PATCH v3 2/2] vhost: enable IOMMU for async vhost
> 
> The use of IOMMU has many advantages, such as isolation and address
> translation. This patch extends the capbility of DMA engine to use IOMMU if
> the DMA engine is bound to vfio.
> 
> When set memory table, the guest memory will be mapped into the default
> container of DPDK.
> 
> Signed-off-by: Xuan Ding <xuan.ding@intel.com>
> ---
>  lib/vhost/vhost.h      |   4 ++
>  lib/vhost/vhost_user.c | 112
> ++++++++++++++++++++++++++++++++++++++++-
>  2 files changed, 114 insertions(+), 2 deletions(-)
> 
> diff --git a/lib/vhost/vhost.h b/lib/vhost/vhost.h index
> 89a31e4ca8..bc5695e899 100644
> --- a/lib/vhost/vhost.h
> +++ b/lib/vhost/vhost.h
> @@ -370,6 +370,10 @@ struct virtio_net {
>  	int16_t			broadcast_rarp;
>  	uint32_t		nr_vring;
>  	int			async_copy;
> +
> +	/* Record the dma map status for each region. */
> +	bool			*async_map_status;
> +
>  	int			extbuf;
>  	int			linearbuf;
>  	struct vhost_virtqueue	*virtqueue[VHOST_MAX_QUEUE_PAIRS * 2];
> diff --git a/lib/vhost/vhost_user.c b/lib/vhost/vhost_user.c index
> 29a4c9af60..3990e9b057 100644
> --- a/lib/vhost/vhost_user.c
> +++ b/lib/vhost/vhost_user.c
> @@ -45,6 +45,8 @@
>  #include <rte_common.h>
>  #include <rte_malloc.h>
>  #include <rte_log.h>
> +#include <rte_vfio.h>
> +#include <rte_errno.h>
> 
>  #include "iotlb.h"
>  #include "vhost.h"
> @@ -141,6 +143,63 @@ get_blk_size(int fd)
>  	return ret == -1 ? (uint64_t)-1 : (uint64_t)stat.st_blksize;  }
> 
> +static int
> +async_dma_map(struct rte_vhost_mem_region *region, bool
> +*dma_map_success, bool do_map) {
> +	uint64_t host_iova;
> +	int ret = 0;
> +
> +	host_iova = rte_mem_virt2iova((void *)(uintptr_t)region-
> >host_user_addr);
> +	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);
> +		*dma_map_success = ret == 0;
> +
> +		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;
> +			} else {
> +				VHOST_LOG_CONFIG(ERR, "DMA engine map
> failed\n");
> +				return ret;
> +			}
> +		}
> +
> +	} else {
> +		/* No need to do vfio unmap if the map failed. */
> +		if (!*dma_map_success)
> +			return 0;
> +
> +		/* 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) {
> +			VHOST_LOG_CONFIG(ERR, "DMA engine unmap
> failed\n");
> +			return ret;
> +		}
> +		/* Clear the flag once the unmap succeeds. */
> +		*dma_map_success = 0;
> +	}
> +
> +	return ret;
> +}
> +
>  static void
>  free_mem_region(struct virtio_net *dev)  { @@ -153,6 +212,9 @@
> free_mem_region(struct virtio_net *dev)
>  	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, &dev-
> >async_map_status[i], false);
> +
>  			munmap(reg->mmap_addr, reg->mmap_size);
>  			close(reg->fd);
>  		}
> @@ -203,6 +265,11 @@ vhost_backend_cleanup(struct virtio_net *dev)
>  	}
> 
>  	dev->postcopy_listening = 0;
> +
> +	if (dev->async_map_status) {
> +		rte_free(dev->async_map_status);
> +		dev->async_map_status = NULL;
> +	}
>  }
> 
>  static void
> @@ -621,6 +688,17 @@ numa_realloc(struct virtio_net *dev, int index)
>  	}
>  	dev->mem = mem;
> 
> +	if (dev->async_copy && rte_vfio_is_enabled("vfio")) {
> +		dev->async_map_status = rte_zmalloc_socket("async-dma-
> map-status",
> +					sizeof(bool) * dev->mem->nregions,
> 0, node);
> +		if (!dev->async_map_status) {
> +			VHOST_LOG_CONFIG(ERR,
> +				"(%d) failed to realloc dma mapping status on
> node\n",
> +				dev->vid);
> +			return dev;
> +		}
> +	}
> +
>  	gp = rte_realloc_socket(dev->guest_pages, dev->max_guest_pages *
> sizeof(*gp),
>  			RTE_CACHE_LINE_SIZE, node);
>  	if (!gp) {
> @@ -1151,12 +1229,14 @@ vhost_user_postcopy_register(struct virtio_net
> *dev, int main_fd,  static int  vhost_user_mmap_region(struct virtio_net *dev,
>  		struct rte_vhost_mem_region *region,
> +		uint32_t region_index,
>  		uint64_t mmap_offset)
>  {
>  	void *mmap_addr;
>  	uint64_t mmap_size;
>  	uint64_t alignment;
>  	int populate;
> +	int ret;
> 
>  	/* Check for memory_size + mmap_offset overflow */
>  	if (mmap_offset >= -region->size) {
> @@ -1210,13 +1290,25 @@ vhost_user_mmap_region(struct virtio_net *dev,
>  	region->mmap_size = mmap_size;
>  	region->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
> mmap_offset;
> 
> -	if (dev->async_copy)
> +	if (dev->async_copy) {
>  		if (add_guest_pages(dev, region, alignment) < 0) {
>  			VHOST_LOG_CONFIG(ERR,
>  					"adding guest pages to region
> failed.\n");
>  			return -1;
>  		}
> 
> +		if (rte_vfio_is_enabled("vfio")) {
> +			ret = async_dma_map(region, &dev-
> >async_map_status[region_index], true);
> +			if (ret) {
> +				VHOST_LOG_CONFIG(ERR, "Configure
> IOMMU for DMA "
> +							"engine failed\n");
> +				rte_free(dev->async_map_status);
> +				dev->async_map_status = NULL;

The freed dev->async_map_status is accessed in free_mem_region() later.
You need to free it after calling free_mem_region().

> +				return -1;
> +			}
> +		}
> +	}
> +
>  	VHOST_LOG_CONFIG(INFO,
>  			"guest memory region size: 0x%" PRIx64 "\n"
>  			"\t guest physical addr: 0x%" PRIx64 "\n"
> @@ -1291,6 +1383,11 @@ vhost_user_set_mem_table(struct virtio_net
> **pdev, struct VhostUserMsg *msg,
>  		dev->mem = NULL;
>  	}
> 
> +	if (dev->async_map_status) {
> +		rte_free(dev->async_map_status);
> +		dev->async_map_status = NULL;
> +	}

To handle the gust memory hot-plug case, you need to un-map
iommu tables before program iommu for new memory. But you
seem only free the old dev->async_map_status.

Thanks,
Jiayu

> +
>  	/* Flush IOTLB cache as previous HVAs are now invalid */
>  	if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
>  		for (i = 0; i < dev->nr_vring; i++)
> @@ -1329,6 +1426,17 @@ vhost_user_set_mem_table(struct virtio_net
> **pdev, struct VhostUserMsg *msg,
>  		goto free_guest_pages;
>  	}
> 
> +	if (dev->async_copy) {
> +		dev->async_map_status = rte_zmalloc_socket("async-dma-
> map-status",
> +					sizeof(bool) * memory->nregions, 0,
> numa_node);
> +		if (!dev->async_map_status) {
> +			VHOST_LOG_CONFIG(ERR,
> +				"(%d) failed to allocate memory for dma
> mapping status\n",
> +				dev->vid);
> +			goto free_guest_pages;
> +		}
> +	}
> +
>  	for (i = 0; i < memory->nregions; i++) {
>  		reg = &dev->mem->regions[i];
> 
> @@ -1345,7 +1453,7 @@ vhost_user_set_mem_table(struct virtio_net
> **pdev, struct VhostUserMsg *msg,
> 
>  		mmap_offset = memory->regions[i].mmap_offset;
> 
> -		if (vhost_user_mmap_region(dev, reg, mmap_offset) < 0) {
> +		if (vhost_user_mmap_region(dev, reg, i, mmap_offset) < 0) {
>  			VHOST_LOG_CONFIG(ERR, "Failed to mmap
> region %u\n", i);
>  			goto free_mem_table;
>  		}
> --
> 2.17.1
Ding, Xuan Sept. 27, 2021, 4:55 a.m. UTC | #2 
Hi Jiayu,

> -----Original Message-----
> From: Hu, Jiayu <jiayu.hu@intel.com>
> Sent: Monday, September 27, 2021 12:18 PM
> To: Ding, Xuan <xuan.ding@intel.com>; dev@dpdk.org; Burakov, Anatoly
> <anatoly.burakov@intel.com>; maxime.coquelin@redhat.com; Xia, Chenbo
> <chenbo.xia@intel.com>
> Cc: Jiang, Cheng1 <cheng1.jiang@intel.com>; Richardson, Bruce
> <bruce.richardson@intel.com>; Pai G, Sunil <sunil.pai.g@intel.com>; Wang,
> Yinan <yinan.wang@intel.com>; Yang, YvonneX <yvonnex.yang@intel.com>
> Subject: RE: [PATCH v3 2/2] vhost: enable IOMMU for async vhost
> 
> Hi Xuan,
> 
> > -----Original Message-----
> > From: Ding, Xuan <xuan.ding@intel.com>
> > Sent: Saturday, September 25, 2021 6:04 PM
> > To: dev@dpdk.org; Burakov, Anatoly <anatoly.burakov@intel.com>;
> > maxime.coquelin@redhat.com; Xia, Chenbo <chenbo.xia@intel.com>
> > Cc: Hu, Jiayu <jiayu.hu@intel.com>; Jiang, Cheng1 <cheng1.jiang@intel.com>;
> > Richardson, Bruce <bruce.richardson@intel.com>; Pai G, Sunil
> > <sunil.pai.g@intel.com>; Wang, Yinan <yinan.wang@intel.com>; Yang,
> > YvonneX <yvonnex.yang@intel.com>; Ding, Xuan <xuan.ding@intel.com>
> > Subject: [PATCH v3 2/2] vhost: enable IOMMU for async vhost
> >
> > The use of IOMMU has many advantages, such as isolation and address
> > translation. This patch extends the capbility of DMA engine to use IOMMU if
> > the DMA engine is bound to vfio.
> >
> > When set memory table, the guest memory will be mapped into the default
> > container of DPDK.
> >
> > Signed-off-by: Xuan Ding <xuan.ding@intel.com>
> > ---
> >  lib/vhost/vhost.h      |   4 ++
> >  lib/vhost/vhost_user.c | 112
> > ++++++++++++++++++++++++++++++++++++++++-
> >  2 files changed, 114 insertions(+), 2 deletions(-)
> >
> > diff --git a/lib/vhost/vhost.h b/lib/vhost/vhost.h index
> > 89a31e4ca8..bc5695e899 100644
> > --- a/lib/vhost/vhost.h
> > +++ b/lib/vhost/vhost.h
> > @@ -370,6 +370,10 @@ struct virtio_net {
> >  int16_tbroadcast_rarp;
> >  uint32_tnr_vring;
> >  intasync_copy;
> > +
> > +/* Record the dma map status for each region. */
> > +bool*async_map_status;
> > +
> >  intextbuf;
> >  intlinearbuf;
> >  struct vhost_virtqueue*virtqueue[VHOST_MAX_QUEUE_PAIRS * 2];
> > diff --git a/lib/vhost/vhost_user.c b/lib/vhost/vhost_user.c index
> > 29a4c9af60..3990e9b057 100644
> > --- a/lib/vhost/vhost_user.c
> > +++ b/lib/vhost/vhost_user.c
> > @@ -45,6 +45,8 @@
> >  #include <rte_common.h>
> >  #include <rte_malloc.h>
> >  #include <rte_log.h>
> > +#include <rte_vfio.h>
> > +#include <rte_errno.h>
> >
> >  #include "iotlb.h"
> >  #include "vhost.h"
> > @@ -141,6 +143,63 @@ get_blk_size(int fd)
> >  return ret == -1 ? (uint64_t)-1 : (uint64_t)stat.st_blksize;  }
> >
> > +static int
> > +async_dma_map(struct rte_vhost_mem_region *region, bool
> > +*dma_map_success, bool do_map) {
> > +uint64_t host_iova;
> > +int ret = 0;
> > +
> > +host_iova = rte_mem_virt2iova((void *)(uintptr_t)region-
> > >host_user_addr);
> > +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);
> > +*dma_map_success = ret == 0;
> > +
> > +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;
> > +} else {
> > +VHOST_LOG_CONFIG(ERR, "DMA engine map
> > failed\n");
> > +return ret;
> > +}
> > +}
> > +
> > +} else {
> > +/* No need to do vfio unmap if the map failed. */
> > +if (!*dma_map_success)
> > +return 0;
> > +
> > +/* 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) {
> > +VHOST_LOG_CONFIG(ERR, "DMA engine unmap
> > failed\n");
> > +return ret;
> > +}
> > +/* Clear the flag once the unmap succeeds. */
> > +*dma_map_success = 0;
> > +}
> > +
> > +return ret;
> > +}
> > +
> >  static void
> >  free_mem_region(struct virtio_net *dev)  { @@ -153,6 +212,9 @@
> > free_mem_region(struct virtio_net *dev)
> >  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, &dev-
> > >async_map_status[i], false);
> > +
> >  munmap(reg->mmap_addr, reg->mmap_size);
> >  close(reg->fd);
> >  }
> > @@ -203,6 +265,11 @@ vhost_backend_cleanup(struct virtio_net *dev)
> >  }
> >
> >  dev->postcopy_listening = 0;
> > +
> > +if (dev->async_map_status) {
> > +rte_free(dev->async_map_status);
> > +dev->async_map_status = NULL;
> > +}
> >  }
> >
> >  static void
> > @@ -621,6 +688,17 @@ numa_realloc(struct virtio_net *dev, int index)
> >  }
> >  dev->mem = mem;
> >
> > +if (dev->async_copy && rte_vfio_is_enabled("vfio")) {
> > +dev->async_map_status = rte_zmalloc_socket("async-dma-
> > map-status",
> > +sizeof(bool) * dev->mem->nregions,
> > 0, node);
> > +if (!dev->async_map_status) {
> > +VHOST_LOG_CONFIG(ERR,
> > +"(%d) failed to realloc dma mapping status on
> > node\n",
> > +dev->vid);
> > +return dev;
> > +}
> > +}
> > +
> >  gp = rte_realloc_socket(dev->guest_pages, dev->max_guest_pages *
> > sizeof(*gp),
> >  RTE_CACHE_LINE_SIZE, node);
> >  if (!gp) {
> > @@ -1151,12 +1229,14 @@ vhost_user_postcopy_register(struct virtio_net
> > *dev, int main_fd,  static int  vhost_user_mmap_region(struct virtio_net *dev,
> >  struct rte_vhost_mem_region *region,
> > +uint32_t region_index,
> >  uint64_t mmap_offset)
> >  {
> >  void *mmap_addr;
> >  uint64_t mmap_size;
> >  uint64_t alignment;
> >  int populate;
> > +int ret;
> >
> >  /* Check for memory_size + mmap_offset overflow */
> >  if (mmap_offset >= -region->size) {
> > @@ -1210,13 +1290,25 @@ vhost_user_mmap_region(struct virtio_net *dev,
> >  region->mmap_size = mmap_size;
> >  region->host_user_addr = (uint64_t)(uintptr_t)mmap_addr +
> > mmap_offset;
> >
> > -if (dev->async_copy)
> > +if (dev->async_copy) {
> >  if (add_guest_pages(dev, region, alignment) < 0) {
> >  VHOST_LOG_CONFIG(ERR,
> >  "adding guest pages to region
> > failed.\n");
> >  return -1;
> >  }
> >
> > +if (rte_vfio_is_enabled("vfio")) {
> > +ret = async_dma_map(region, &dev-
> > >async_map_status[region_index], true);
> > +if (ret) {
> > +VHOST_LOG_CONFIG(ERR, "Configure
> > IOMMU for DMA "
> > +"engine failed\n");
> > +rte_free(dev->async_map_status);
> > +dev->async_map_status = NULL;
> 
> The freed dev->async_map_status is accessed in free_mem_region() later.
> You need to free it after calling free_mem_region().

Thanks for the catch! Will fix it in next version.

> 
> > +return -1;
> > +}
> > +}
> > +}
> > +
> >  VHOST_LOG_CONFIG(INFO,
> >  "guest memory region size: 0x%" PRIx64 "\n"
> >  "\t guest physical addr: 0x%" PRIx64 "\n"
> > @@ -1291,6 +1383,11 @@ vhost_user_set_mem_table(struct virtio_net
> > **pdev, struct VhostUserMsg *msg,
> >  dev->mem = NULL;
> >  }
> >
> > +if (dev->async_map_status) {
> > +rte_free(dev->async_map_status);
> > +dev->async_map_status = NULL;
> > +}
> 
> To handle the gust memory hot-plug case, you need to un-map
> iommu tables before program iommu for new memory. But you
> seem only free the old dev->async_map_status.

Yes, you are right. Will unmap the region in iommu table in hot-plug scenario.

Regards,
Xuan

> 
> Thanks,
> Jiayu
> 
> > +
> >  /* Flush IOTLB cache as previous HVAs are now invalid */
> >  if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
> >  for (i = 0; i < dev->nr_vring; i++)
> > @@ -1329,6 +1426,17 @@ vhost_user_set_mem_table(struct virtio_net
> > **pdev, struct VhostUserMsg *msg,
> >  goto free_guest_pages;
> >  }
> >
> > +if (dev->async_copy) {
> > +dev->async_map_status = rte_zmalloc_socket("async-dma-
> > map-status",
> > +sizeof(bool) * memory->nregions, 0,
> > numa_node);
> > +if (!dev->async_map_status) {
> > +VHOST_LOG_CONFIG(ERR,
> > +"(%d) failed to allocate memory for dma
> > mapping status\n",
> > +dev->vid);
> > +goto free_guest_pages;
> > +}
> > +}
> > +
> >  for (i = 0; i < memory->nregions; i++) {
> >  reg = &dev->mem->regions[i];
> >
> > @@ -1345,7 +1453,7 @@ vhost_user_set_mem_table(struct virtio_net
> > **pdev, struct VhostUserMsg *msg,
> >
> >  mmap_offset = memory->regions[i].mmap_offset;
> >
> > -if (vhost_user_mmap_region(dev, reg, mmap_offset) < 0) {
> > +if (vhost_user_mmap_region(dev, reg, i, mmap_offset) < 0) {
> >  VHOST_LOG_CONFIG(ERR, "Failed to mmap
> > region %u\n", i);
> >  goto free_mem_table;
> >  }
> > --
> > 2.17.1
>
diff mbox series

Patch

diff --git a/lib/vhost/vhost.h b/lib/vhost/vhost.h
index 89a31e4ca8..bc5695e899 100644
--- a/lib/vhost/vhost.h
+++ b/lib/vhost/vhost.h
@@ -370,6 +370,10 @@  struct virtio_net {
 	int16_t			broadcast_rarp;
 	uint32_t		nr_vring;
 	int			async_copy;
+
+	/* Record the dma map status for each region. */
+	bool			*async_map_status;
+
 	int			extbuf;
 	int			linearbuf;
 	struct vhost_virtqueue	*virtqueue[VHOST_MAX_QUEUE_PAIRS * 2];
diff --git a/lib/vhost/vhost_user.c b/lib/vhost/vhost_user.c
index 29a4c9af60..3990e9b057 100644
--- a/lib/vhost/vhost_user.c
+++ b/lib/vhost/vhost_user.c
@@ -45,6 +45,8 @@ 
 #include <rte_common.h>
 #include <rte_malloc.h>
 #include <rte_log.h>
+#include <rte_vfio.h>
+#include <rte_errno.h>
 
 #include "iotlb.h"
 #include "vhost.h"
@@ -141,6 +143,63 @@  get_blk_size(int fd)
 	return ret == -1 ? (uint64_t)-1 : (uint64_t)stat.st_blksize;
 }
 
+static int
+async_dma_map(struct rte_vhost_mem_region *region, bool *dma_map_success, bool do_map)
+{
+	uint64_t host_iova;
+	int ret = 0;
+
+	host_iova = rte_mem_virt2iova((void *)(uintptr_t)region->host_user_addr);
+	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);
+		*dma_map_success = ret == 0;
+
+		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;
+			} else {
+				VHOST_LOG_CONFIG(ERR, "DMA engine map failed\n");
+				return ret;
+			}
+		}
+
+	} else {
+		/* No need to do vfio unmap if the map failed. */
+		if (!*dma_map_success)
+			return 0;
+
+		/* 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) {
+			VHOST_LOG_CONFIG(ERR, "DMA engine unmap failed\n");
+			return ret;
+		}
+		/* Clear the flag once the unmap succeeds. */
+		*dma_map_success = 0;
+	}
+
+	return ret;
+}
+
 static void
 free_mem_region(struct virtio_net *dev)
 {
@@ -153,6 +212,9 @@  free_mem_region(struct virtio_net *dev)
 	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, &dev->async_map_status[i], false);
+
 			munmap(reg->mmap_addr, reg->mmap_size);
 			close(reg->fd);
 		}
@@ -203,6 +265,11 @@  vhost_backend_cleanup(struct virtio_net *dev)
 	}
 
 	dev->postcopy_listening = 0;
+
+	if (dev->async_map_status) {
+		rte_free(dev->async_map_status);
+		dev->async_map_status = NULL;
+	}
 }
 
 static void
@@ -621,6 +688,17 @@  numa_realloc(struct virtio_net *dev, int index)
 	}
 	dev->mem = mem;
 
+	if (dev->async_copy && rte_vfio_is_enabled("vfio")) {
+		dev->async_map_status = rte_zmalloc_socket("async-dma-map-status",
+					sizeof(bool) * dev->mem->nregions, 0, node);
+		if (!dev->async_map_status) {
+			VHOST_LOG_CONFIG(ERR,
+				"(%d) failed to realloc dma mapping status on node\n",
+				dev->vid);
+			return dev;
+		}
+	}
+
 	gp = rte_realloc_socket(dev->guest_pages, dev->max_guest_pages * sizeof(*gp),
 			RTE_CACHE_LINE_SIZE, node);
 	if (!gp) {
@@ -1151,12 +1229,14 @@  vhost_user_postcopy_register(struct virtio_net *dev, int main_fd,
 static int
 vhost_user_mmap_region(struct virtio_net *dev,
 		struct rte_vhost_mem_region *region,
+		uint32_t region_index,
 		uint64_t mmap_offset)
 {
 	void *mmap_addr;
 	uint64_t mmap_size;
 	uint64_t alignment;
 	int populate;
+	int ret;
 
 	/* Check for memory_size + mmap_offset overflow */
 	if (mmap_offset >= -region->size) {
@@ -1210,13 +1290,25 @@  vhost_user_mmap_region(struct virtio_net *dev,
 	region->mmap_size = mmap_size;
 	region->host_user_addr = (uint64_t)(uintptr_t)mmap_addr + mmap_offset;
 
-	if (dev->async_copy)
+	if (dev->async_copy) {
 		if (add_guest_pages(dev, region, alignment) < 0) {
 			VHOST_LOG_CONFIG(ERR,
 					"adding guest pages to region failed.\n");
 			return -1;
 		}
 
+		if (rte_vfio_is_enabled("vfio")) {
+			ret = async_dma_map(region, &dev->async_map_status[region_index], true);
+			if (ret) {
+				VHOST_LOG_CONFIG(ERR, "Configure IOMMU for DMA "
+							"engine failed\n");
+				rte_free(dev->async_map_status);
+				dev->async_map_status = NULL;
+				return -1;
+			}
+		}
+	}
+
 	VHOST_LOG_CONFIG(INFO,
 			"guest memory region size: 0x%" PRIx64 "\n"
 			"\t guest physical addr: 0x%" PRIx64 "\n"
@@ -1291,6 +1383,11 @@  vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
 		dev->mem = NULL;
 	}
 
+	if (dev->async_map_status) {
+		rte_free(dev->async_map_status);
+		dev->async_map_status = NULL;
+	}
+
 	/* Flush IOTLB cache as previous HVAs are now invalid */
 	if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 		for (i = 0; i < dev->nr_vring; i++)
@@ -1329,6 +1426,17 @@  vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
 		goto free_guest_pages;
 	}
 
+	if (dev->async_copy) {
+		dev->async_map_status = rte_zmalloc_socket("async-dma-map-status",
+					sizeof(bool) * memory->nregions, 0, numa_node);
+		if (!dev->async_map_status) {
+			VHOST_LOG_CONFIG(ERR,
+				"(%d) failed to allocate memory for dma mapping status\n",
+				dev->vid);
+			goto free_guest_pages;
+		}
+	}
+
 	for (i = 0; i < memory->nregions; i++) {
 		reg = &dev->mem->regions[i];
 
@@ -1345,7 +1453,7 @@  vhost_user_set_mem_table(struct virtio_net **pdev, struct VhostUserMsg *msg,
 
 		mmap_offset = memory->regions[i].mmap_offset;
 
-		if (vhost_user_mmap_region(dev, reg, mmap_offset) < 0) {
+		if (vhost_user_mmap_region(dev, reg, i, mmap_offset) < 0) {
 			VHOST_LOG_CONFIG(ERR, "Failed to mmap region %u\n", i);
 			goto free_mem_table;
 		}