diff mbox series

[v2,1/1] vhost: integrate dmadev in asynchronous datapath

Message ID 20220124164011.1402593-2-jiayu.hu@intel.com (mailing list archive)
State Superseded, archived
Delegated to: Maxime Coquelin
Headers show
Series integrate dmadev in vhost | expand

Checks

Context Check Description
ci/intel-Testing success Testing PASS
ci/Intel-compilation success Compilation OK
ci/iol-abi-testing success Testing PASS
ci/iol-aarch64-compile-testing success Testing PASS
ci/iol-aarch64-unit-testing success Testing PASS
ci/iol-intel-Performance success Performance Testing PASS
ci/iol-x86_64-unit-testing success Testing PASS
ci/iol-x86_64-compile-testing success Testing PASS
ci/iol-intel-Functional success Functional Testing PASS
ci/iol-broadcom-Functional success Functional Testing PASS
ci/github-robot: build success github build: passed
ci/iol-mellanox-Performance success Performance Testing PASS
ci/iol-broadcom-Performance success Performance Testing PASS
ci/checkpatch success coding style OK

Commit Message

Hu, Jiayu Jan. 24, 2022, 4:40 p.m. UTC
Since dmadev is introduced in 21.11, to avoid the overhead of vhost DMA
abstraction layer and simplify application logics, this patch integrates
dmadev in asynchronous data path.

Signed-off-by: Jiayu Hu <jiayu.hu@intel.com>
Signed-off-by: Sunil Pai G <sunil.pai.g@intel.com>
---
 doc/guides/prog_guide/vhost_lib.rst |  95 ++++-----
 examples/vhost/Makefile             |   2 +-
 examples/vhost/ioat.c               | 218 --------------------
 examples/vhost/ioat.h               |  63 ------
 examples/vhost/main.c               | 255 ++++++++++++++++++-----
 examples/vhost/main.h               |  11 +
 examples/vhost/meson.build          |   6 +-
 lib/vhost/meson.build               |   2 +-
 lib/vhost/rte_vhost.h               |   2 +
 lib/vhost/rte_vhost_async.h         | 132 +++++-------
 lib/vhost/version.map               |   3 +
 lib/vhost/vhost.c                   | 148 ++++++++++----
 lib/vhost/vhost.h                   |  64 +++++-
 lib/vhost/vhost_user.c              |   2 +
 lib/vhost/virtio_net.c              | 305 +++++++++++++++++++++++-----
 15 files changed, 744 insertions(+), 564 deletions(-)
 delete mode 100644 examples/vhost/ioat.c
 delete mode 100644 examples/vhost/ioat.h

Comments

Maxime Coquelin Feb. 3, 2022, 1:04 p.m. UTC | #1
Hi Jiayu,

On 1/24/22 17:40, Jiayu Hu wrote:
> Since dmadev is introduced in 21.11, to avoid the overhead of vhost DMA
> abstraction layer and simplify application logics, this patch integrates
> dmadev in asynchronous data path.
> 
> Signed-off-by: Jiayu Hu <jiayu.hu@intel.com>
> Signed-off-by: Sunil Pai G <sunil.pai.g@intel.com>
> ---
>   doc/guides/prog_guide/vhost_lib.rst |  95 ++++-----
>   examples/vhost/Makefile             |   2 +-
>   examples/vhost/ioat.c               | 218 --------------------
>   examples/vhost/ioat.h               |  63 ------
>   examples/vhost/main.c               | 255 ++++++++++++++++++-----
>   examples/vhost/main.h               |  11 +
>   examples/vhost/meson.build          |   6 +-
>   lib/vhost/meson.build               |   2 +-
>   lib/vhost/rte_vhost.h               |   2 +
>   lib/vhost/rte_vhost_async.h         | 132 +++++-------
>   lib/vhost/version.map               |   3 +
>   lib/vhost/vhost.c                   | 148 ++++++++++----
>   lib/vhost/vhost.h                   |  64 +++++-
>   lib/vhost/vhost_user.c              |   2 +
>   lib/vhost/virtio_net.c              | 305 +++++++++++++++++++++++-----
>   15 files changed, 744 insertions(+), 564 deletions(-)
>   delete mode 100644 examples/vhost/ioat.c
>   delete mode 100644 examples/vhost/ioat.h
> 

When you rebase to the next version, please ensure to rework all the 
logs to follow the new standard:
VHOST_LOG_CONFIG(ERR,"(%s) .....", dev->ifname, ...);

> git a/lib/vhost/rte_vhost_async.h b/lib/vhost/rte_vhost_async.h
> index a87ea6ba37..758a80f403 100644
> --- a/lib/vhost/rte_vhost_async.h
> +++ b/lib/vhost/rte_vhost_async.h
> @@ -26,73 +26,6 @@ struct rte_vhost_iov_iter {
>   	unsigned long nr_segs;
>   };
>   
> -/**
> - * dma transfer status
> - */
> -struct rte_vhost_async_status {
> -	/** An array of application specific data for source memory */
> -	uintptr_t *src_opaque_data;
> -	/** An array of application specific data for destination memory */
> -	uintptr_t *dst_opaque_data;
> -};
> -
> -/**
> - * dma operation callbacks to be implemented by applications
> - */
> -struct rte_vhost_async_channel_ops {
> -	/**
> -	 * instruct async engines to perform copies for a batch of packets
> -	 *
> -	 * @param vid
> -	 *  id of vhost device to perform data copies
> -	 * @param queue_id
> -	 *  queue id to perform data copies
> -	 * @param iov_iter
> -	 *  an array of IOV iterators
> -	 * @param opaque_data
> -	 *  opaque data pair sending to DMA engine
> -	 * @param count
> -	 *  number of elements in the "descs" array
> -	 * @return
> -	 *  number of IOV iterators processed, negative value means error
> -	 */
> -	int32_t (*transfer_data)(int vid, uint16_t queue_id,
> -		struct rte_vhost_iov_iter *iov_iter,
> -		struct rte_vhost_async_status *opaque_data,
> -		uint16_t count);
> -	/**
> -	 * check copy-completed packets from the async engine
> -	 * @param vid
> -	 *  id of vhost device to check copy completion
> -	 * @param queue_id
> -	 *  queue id to check copy completion
> -	 * @param opaque_data
> -	 *  buffer to receive the opaque data pair from DMA engine
> -	 * @param max_packets
> -	 *  max number of packets could be completed
> -	 * @return
> -	 *  number of async descs completed, negative value means error
> -	 */
> -	int32_t (*check_completed_copies)(int vid, uint16_t queue_id,
> -		struct rte_vhost_async_status *opaque_data,
> -		uint16_t max_packets);
> -};
> -
> -/**
> - *  async channel features
> - */
> -enum {
> -	RTE_VHOST_ASYNC_INORDER = 1U << 0,
> -};
> -
> -/**
> - *  async channel configuration
> - */
> -struct rte_vhost_async_config {
> -	uint32_t features;
> -	uint32_t rsvd[2];
> -};
> -
>   /**
>    * Register an async channel for a vhost queue
>    *
> @@ -100,17 +33,11 @@ struct rte_vhost_async_config {
>    *  vhost device id async channel to be attached to
>    * @param queue_id
>    *  vhost queue id async channel to be attached to
> - * @param config
> - *  Async channel configuration structure
> - * @param ops
> - *  Async channel operation callbacks
>    * @return
>    *  0 on success, -1 on failures
>    */
>   __rte_experimental
> -int rte_vhost_async_channel_register(int vid, uint16_t queue_id,
> -	struct rte_vhost_async_config config,
> -	struct rte_vhost_async_channel_ops *ops);
> +int rte_vhost_async_channel_register(int vid, uint16_t queue_id);
>   
>   /**
>    * Unregister an async channel for a vhost queue
> @@ -136,17 +63,11 @@ int rte_vhost_async_channel_unregister(int vid, uint16_t queue_id);
>    *  vhost device id async channel to be attached to
>    * @param queue_id
>    *  vhost queue id async channel to be attached to
> - * @param config
> - *  Async channel configuration
> - * @param ops
> - *  Async channel operation callbacks
>    * @return
>    *  0 on success, -1 on failures
>    */
>   __rte_experimental
> -int rte_vhost_async_channel_register_thread_unsafe(int vid, uint16_t queue_id,
> -	struct rte_vhost_async_config config,
> -	struct rte_vhost_async_channel_ops *ops);
> +int rte_vhost_async_channel_register_thread_unsafe(int vid, uint16_t queue_id);
>   
>   /**
>    * Unregister an async channel for a vhost queue without performing any
> @@ -179,12 +100,17 @@ int rte_vhost_async_channel_unregister_thread_unsafe(int vid,
>    *  array of packets to be enqueued
>    * @param count
>    *  packets num to be enqueued
> + * @param dma_id
> + *  the identifier of the DMA device
> + * @param vchan_id
> + *  the identifier of virtual DMA channel
>    * @return
>    *  num of packets enqueued
>    */
>   __rte_experimental
>   uint16_t rte_vhost_submit_enqueue_burst(int vid, uint16_t queue_id,
> -		struct rte_mbuf **pkts, uint16_t count);
> +		struct rte_mbuf **pkts, uint16_t count, int16_t dma_id,
> +		uint16_t vchan_id);
>   
>   /**
>    * This function checks async completion status for a specific vhost
> @@ -199,12 +125,17 @@ uint16_t rte_vhost_submit_enqueue_burst(int vid, uint16_t queue_id,
>    *  blank array to get return packet pointer
>    * @param count
>    *  size of the packet array
> + * @param dma_id
> + *  the identifier of the DMA device
> + * @param vchan_id
> + *  the identifier of virtual DMA channel
>    * @return
>    *  num of packets returned
>    */
>   __rte_experimental
>   uint16_t rte_vhost_poll_enqueue_completed(int vid, uint16_t queue_id,
> -		struct rte_mbuf **pkts, uint16_t count);
> +		struct rte_mbuf **pkts, uint16_t count, int16_t dma_id,
> +		uint16_t vchan_id);
>   
>   /**
>    * This function returns the amount of in-flight packets for the vhost
> @@ -235,11 +166,44 @@ int rte_vhost_async_get_inflight(int vid, uint16_t queue_id);
>    *  Blank array to get return packet pointer
>    * @param count
>    *  Size of the packet array
> + * @param dma_id
> + *  the identifier of the DMA device
> + * @param vchan_id
> + *  the identifier of virtual DMA channel
>    * @return
>    *  Number of packets returned
>    */
>   __rte_experimental
>   uint16_t rte_vhost_clear_queue_thread_unsafe(int vid, uint16_t queue_id,
> -		struct rte_mbuf **pkts, uint16_t count);
> +		struct rte_mbuf **pkts, uint16_t count, int16_t dma_id,
> +		uint16_t vchan_id);
> +/**
> + * The DMA vChannels used in asynchronous data path must be configured
> + * first. So this function needs to be called before enabling DMA
> + * acceleration for vring. If this function fails, asynchronous data path
> + * cannot be enabled for any vring further.
> + *
> + * DMA devices used in data-path must belong to DMA devices given in this
> + * function. But users are free to use DMA devices given in the function
> + * in non-vhost scenarios, only if guarantee no copies in vhost are
> + * offloaded to them at the same time.
> + *
> + * @param dmas_id
> + *  DMA ID array
> + * @param count
> + *  Element number of 'dmas_id'
> + * @param poll_factor
> + *  For large or scatter-gather packets, one packet would consist of
> + *  small buffers. In this case, vhost will issue several DMA copy
> + *  operations for the packet. Therefore, the number of copies to
> + *  check by rte_dma_completed() is calculated by "nb_pkts_to_poll *
> + *  poll_factor" andused in rte_vhost_poll_enqueue_completed(). The
> + *  default value of "poll_factor" is 1.
> + * @return
> + *  0 on success, and -1 on failure
> + */
> +__rte_experimental
> +int rte_vhost_async_dma_configure(int16_t *dmas_id, uint16_t count,
> +		uint16_t poll_factor);
>   
>   #endif /* _RTE_VHOST_ASYNC_H_ */
> diff --git a/lib/vhost/version.map b/lib/vhost/version.map
> index a7ef7f1976..1202ba9c1a 100644
> --- a/lib/vhost/version.map
> +++ b/lib/vhost/version.map
> @@ -84,6 +84,9 @@ EXPERIMENTAL {
>   
>   	# added in 21.11
>   	rte_vhost_get_monitor_addr;
> +
> +	# added in 22.03
> +	rte_vhost_async_dma_configure;
>   };
>   
>   INTERNAL {
> diff --git a/lib/vhost/vhost.c b/lib/vhost/vhost.c
> index 13a9bb9dd1..c408cee63e 100644
> --- a/lib/vhost/vhost.c
> +++ b/lib/vhost/vhost.c
> @@ -25,7 +25,7 @@
>   #include "vhost.h"
>   #include "vhost_user.h"
>   
> -struct virtio_net *vhost_devices[MAX_VHOST_DEVICE];
> +struct virtio_net *vhost_devices[RTE_MAX_VHOST_DEVICE];
>   pthread_mutex_t vhost_dev_lock = PTHREAD_MUTEX_INITIALIZER;
>   
>   /* Called with iotlb_lock read-locked */
> @@ -344,6 +344,7 @@ vhost_free_async_mem(struct vhost_virtqueue *vq)
>   		return;
>   
>   	rte_free(vq->async->pkts_info);
> +	rte_free(vq->async->pkts_cmpl_flag);
>   
>   	rte_free(vq->async->buffers_packed);
>   	vq->async->buffers_packed = NULL;
> @@ -667,12 +668,12 @@ vhost_new_device(void)
>   	int i;
>   
>   	pthread_mutex_lock(&vhost_dev_lock);
> -	for (i = 0; i < MAX_VHOST_DEVICE; i++) {
> +	for (i = 0; i < RTE_MAX_VHOST_DEVICE; i++) {
>   		if (vhost_devices[i] == NULL)
>   			break;
>   	}
>   
> -	if (i == MAX_VHOST_DEVICE) {
> +	if (i == RTE_MAX_VHOST_DEVICE) {
>   		VHOST_LOG_CONFIG(ERR,
>   			"Failed to find a free slot for new device.\n");
>   		pthread_mutex_unlock(&vhost_dev_lock);
> @@ -1626,8 +1627,7 @@ rte_vhost_extern_callback_register(int vid,
>   }
>   
>   static __rte_always_inline int
> -async_channel_register(int vid, uint16_t queue_id,
> -		struct rte_vhost_async_channel_ops *ops)
> +async_channel_register(int vid, uint16_t queue_id)
>   {
>   	struct virtio_net *dev = get_device(vid);
>   	struct vhost_virtqueue *vq = dev->virtqueue[queue_id];
> @@ -1656,6 +1656,14 @@ async_channel_register(int vid, uint16_t queue_id,
>   		goto out_free_async;
>   	}
>   
> +	async->pkts_cmpl_flag = rte_zmalloc_socket(NULL, vq->size * sizeof(bool),
> +			RTE_CACHE_LINE_SIZE, node);
> +	if (!async->pkts_cmpl_flag) {
> +		VHOST_LOG_CONFIG(ERR, "failed to allocate async pkts_cmpl_flag (vid %d, qid: %d)\n",
> +				vid, queue_id);
> +		goto out_free_async;
> +	}
> +
>   	if (vq_is_packed(dev)) {
>   		async->buffers_packed = rte_malloc_socket(NULL,
>   				vq->size * sizeof(struct vring_used_elem_packed),
> @@ -1676,9 +1684,6 @@ async_channel_register(int vid, uint16_t queue_id,
>   		}
>   	}
>   
> -	async->ops.check_completed_copies = ops->check_completed_copies;
> -	async->ops.transfer_data = ops->transfer_data;
> -
>   	vq->async = async;
>   
>   	return 0;
> @@ -1691,15 +1696,13 @@ async_channel_register(int vid, uint16_t queue_id,
>   }
>   
>   int
> -rte_vhost_async_channel_register(int vid, uint16_t queue_id,
> -		struct rte_vhost_async_config config,
> -		struct rte_vhost_async_channel_ops *ops)
> +rte_vhost_async_channel_register(int vid, uint16_t queue_id)
>   {
>   	struct vhost_virtqueue *vq;
>   	struct virtio_net *dev = get_device(vid);
>   	int ret;
>   
> -	if (dev == NULL || ops == NULL)
> +	if (dev == NULL)
>   		return -1;
>   
>   	if (queue_id >= VHOST_MAX_VRING)
> @@ -1710,33 +1713,20 @@ rte_vhost_async_channel_register(int vid, uint16_t queue_id,
>   	if (unlikely(vq == NULL || !dev->async_copy))
>   		return -1;
>   
> -	if (unlikely(!(config.features & RTE_VHOST_ASYNC_INORDER))) {
> -		VHOST_LOG_CONFIG(ERR,
> -			"async copy is not supported on non-inorder mode "
> -			"(vid %d, qid: %d)\n", vid, queue_id);
> -		return -1;
> -	}
> -
> -	if (unlikely(ops->check_completed_copies == NULL ||
> -		ops->transfer_data == NULL))
> -		return -1;
> -
>   	rte_spinlock_lock(&vq->access_lock);
> -	ret = async_channel_register(vid, queue_id, ops);
> +	ret = async_channel_register(vid, queue_id);
>   	rte_spinlock_unlock(&vq->access_lock);
>   
>   	return ret;
>   }
>   
>   int
> -rte_vhost_async_channel_register_thread_unsafe(int vid, uint16_t queue_id,
> -		struct rte_vhost_async_config config,
> -		struct rte_vhost_async_channel_ops *ops)
> +rte_vhost_async_channel_register_thread_unsafe(int vid, uint16_t queue_id)
>   {
>   	struct vhost_virtqueue *vq;
>   	struct virtio_net *dev = get_device(vid);
>   
> -	if (dev == NULL || ops == NULL)
> +	if (dev == NULL)
>   		return -1;
>   
>   	if (queue_id >= VHOST_MAX_VRING)
> @@ -1747,18 +1737,7 @@ rte_vhost_async_channel_register_thread_unsafe(int vid, uint16_t queue_id,
>   	if (unlikely(vq == NULL || !dev->async_copy))
>   		return -1;
>   
> -	if (unlikely(!(config.features & RTE_VHOST_ASYNC_INORDER))) {
> -		VHOST_LOG_CONFIG(ERR,
> -			"async copy is not supported on non-inorder mode "
> -			"(vid %d, qid: %d)\n", vid, queue_id);
> -		return -1;
> -	}
> -
> -	if (unlikely(ops->check_completed_copies == NULL ||
> -		ops->transfer_data == NULL))
> -		return -1;
> -
> -	return async_channel_register(vid, queue_id, ops);
> +	return async_channel_register(vid, queue_id);
>   }
>   
>   int
> @@ -1835,6 +1814,95 @@ rte_vhost_async_channel_unregister_thread_unsafe(int vid, uint16_t queue_id)
>   	return 0;
>   }
>   
> +static __rte_always_inline void
> +vhost_free_async_dma_mem(void)
> +{
> +	uint16_t i;
> +
> +	for (i = 0; i < RTE_DMADEV_DEFAULT_MAX; i++) {
> +		struct async_dma_info *dma = &dma_copy_track[i];
> +		int16_t j;
> +
> +		if (dma->max_vchans == 0)
> +			continue;
> +
> +		for (j = 0; j < dma->max_vchans; j++)
> +			rte_free(dma->vchans[j].pkts_completed_flag);
> +
> +		rte_free(dma->vchans);
> +		dma->vchans = NULL;
> +		dma->max_vchans = 0;
> +	}
> +}
> +
> +int
> +rte_vhost_async_dma_configure(int16_t *dmas_id, uint16_t count, uint16_t poll_factor)

I'm not fan of the poll_factor, I think it is too complex for the user
to know what value he should set.

Also, I would like that the API only registers one DMA channel at a
time and let the application call it multiple times. Dong that, user can
still use the DMA channels that could be registered.

> +{
> +	uint16_t i;
> +
> +	if (!dmas_id) {
> +		VHOST_LOG_CONFIG(ERR, "Invalid DMA configuration parameter.\n");
> +		return -1;
> +	}
> +
> +	if (poll_factor == 0) {
> +		VHOST_LOG_CONFIG(ERR, "Invalid DMA poll factor %u\n", poll_factor);
> +		return -1;
> +	}
> +	dma_poll_factor = poll_factor;
> +
> +	for (i = 0; i < count; i++) {
> +		struct async_dma_vchan_info *vchans;
> +		struct rte_dma_info info;
> +		uint16_t max_vchans;
> +		uint16_t max_desc;
> +		uint16_t j;
> +
> +		if (!rte_dma_is_valid(dmas_id[i])) {
> +			VHOST_LOG_CONFIG(ERR, "DMA %d is not found. Cannot enable async"
> +				       " data-path\n.", dmas_id[i]);
> +			vhost_free_async_dma_mem();
> +			return -1;
> +		}
> +
> +		rte_dma_info_get(dmas_id[i], &info);
> +
> +		max_vchans = info.max_vchans;
> +		max_desc = info.max_desc;
> +
> +		if (!rte_is_power_of_2(max_desc))
> +			max_desc = rte_align32pow2(max_desc);
> +
> +		vchans = rte_zmalloc(NULL, sizeof(struct async_dma_vchan_info) * max_vchans,
> +				RTE_CACHE_LINE_SIZE);
> +		if (vchans == NULL) {
> +			VHOST_LOG_CONFIG(ERR, "Failed to allocate vchans for dma-%d."
> +					" Cannot enable async data-path.\n", dmas_id[i]);
> +			vhost_free_async_dma_mem();
> +			return -1;
> +		}
> +
> +		for (j = 0; j < max_vchans; j++) {
> +			vchans[j].pkts_completed_flag = rte_zmalloc(NULL, sizeof(bool *) * max_desc,
> +					RTE_CACHE_LINE_SIZE);
> +			if (!vchans[j].pkts_completed_flag) {
> +				VHOST_LOG_CONFIG(ERR, "Failed to allocate  pkts_completed_flag for "
> +						"dma-%d vchan-%u\n", dmas_id[i], j);
> +				vhost_free_async_dma_mem();
> +				return -1;
> +			}
> +
> +			vchans[j].ring_size = max_desc;
> +			vchans[j].ring_mask = max_desc - 1;
> +		}
> +
> +		dma_copy_track[dmas_id[i]].vchans = vchans;
> +		dma_copy_track[dmas_id[i]].max_vchans = max_vchans;
> +	}
> +
> +	return 0;
> +}
> +
>   int
>   rte_vhost_async_get_inflight(int vid, uint16_t queue_id)
>   {
> diff --git a/lib/vhost/vhost.h b/lib/vhost/vhost.h
> index 7085e0885c..475843fec0 100644
> --- a/lib/vhost/vhost.h
> +++ b/lib/vhost/vhost.h
> @@ -19,6 +19,7 @@
>   #include <rte_ether.h>
>   #include <rte_rwlock.h>
>   #include <rte_malloc.h>
> +#include <rte_dmadev.h>
>   
>   #include "rte_vhost.h"
>   #include "rte_vdpa.h"
> @@ -50,6 +51,7 @@
>   
>   #define VHOST_MAX_ASYNC_IT (MAX_PKT_BURST)
>   #define VHOST_MAX_ASYNC_VEC 2048
> +#define VHOST_ASYNC_DMA_BATCHING_SIZE 32
>   
>   #define PACKED_DESC_ENQUEUE_USED_FLAG(w)	\
>   	((w) ? (VRING_DESC_F_AVAIL | VRING_DESC_F_USED | VRING_DESC_F_WRITE) : \
> @@ -119,6 +121,42 @@ struct vring_used_elem_packed {
>   	uint32_t count;
>   };
>   
> +struct async_dma_vchan_info {
> +	/* circular array to track if packet copy completes */
> +	bool **pkts_completed_flag;
> +
> +	/* max elements in 'metadata' */
> +	uint16_t ring_size;
> +	/* ring index mask for 'metadata' */
> +	uint16_t ring_mask;
> +
> +	/* batching copies before a DMA doorbell */
> +	uint16_t nr_batching;
> +
> +	/**
> +	 * DMA virtual channel lock. Although it is able to bind DMA
> +	 * virtual channels to data plane threads, vhost control plane
> +	 * thread could call data plane functions too, thus causing
> +	 * DMA device contention.
> +	 *
> +	 * For example, in VM exit case, vhost control plane thread needs
> +	 * to clear in-flight packets before disable vring, but there could
> +	 * be anotther data plane thread is enqueuing packets to the same
> +	 * vring with the same DMA virtual channel. But dmadev PMD functions
> +	 * are lock-free, so the control plane and data plane threads
> +	 * could operate the same DMA virtual channel at the same time.
> +	 */
> +	rte_spinlock_t dma_lock;
> +};
> +
> +struct async_dma_info {
> +	uint16_t max_vchans;
> +	struct async_dma_vchan_info *vchans;
> +};
> +
> +extern struct async_dma_info dma_copy_track[RTE_DMADEV_DEFAULT_MAX];
> +extern uint16_t dma_poll_factor;
> +
>   /**
>    * inflight async packet information
>    */
> @@ -129,9 +167,6 @@ struct async_inflight_info {
>   };
>   
>   struct vhost_async {
> -	/* operation callbacks for DMA */
> -	struct rte_vhost_async_channel_ops ops;
> -
>   	struct rte_vhost_iov_iter iov_iter[VHOST_MAX_ASYNC_IT];
>   	struct rte_vhost_iovec iovec[VHOST_MAX_ASYNC_VEC];
>   	uint16_t iter_idx;
> @@ -139,6 +174,25 @@ struct vhost_async {
>   
>   	/* data transfer status */
>   	struct async_inflight_info *pkts_info;
> +	/**
> +	 * Packet reorder array. "true" indicates that DMA device
> +	 * completes all copies for the packet.
> +	 *
> +	 * Note that this array could be written by multiple threads
> +	 * simultaneously. For example, in the case of thread0 and
> +	 * thread1 RX packets from NIC and then enqueue packets to
> +	 * vring0 and vring1 with own DMA device DMA0 and DMA1, it's
> +	 * possible for thread0 to get completed copies belonging to
> +	 * vring1 from DMA0, while thread0 is calling rte_vhost_poll
> +	 * _enqueue_completed() for vring0 and thread1 is calling
> +	 * rte_vhost_submit_enqueue_burst() for vring1. In this case,
> +	 * vq->access_lock cannot protect pkts_cmpl_flag of vring1.
> +	 *
> +	 * However, since offloading is per-packet basis, each packet
> +	 * flag will only be written by one thread. And single byte
> +	 * write is atomic, so no lock for pkts_cmpl_flag is needed.
> +	 */
> +	bool *pkts_cmpl_flag;
>   	uint16_t pkts_idx;
>   	uint16_t pkts_inflight_n;
>   	union {
> @@ -198,6 +252,7 @@ struct vhost_virtqueue {
>   	/* Record packed ring first dequeue desc index */
>   	uint16_t		shadow_last_used_idx;
>   
> +	uint16_t		batch_copy_max_elems;
>   	uint16_t		batch_copy_nb_elems;
>   	struct batch_copy_elem	*batch_copy_elems;
>   	int			numa_node;
> @@ -568,8 +623,7 @@ extern int vhost_data_log_level;
>   #define PRINT_PACKET(device, addr, size, header) do {} while (0)
>   #endif
>   
> -#define MAX_VHOST_DEVICE	1024
> -extern struct virtio_net *vhost_devices[MAX_VHOST_DEVICE];
> +extern struct virtio_net *vhost_devices[RTE_MAX_VHOST_DEVICE];
>   
>   #define VHOST_BINARY_SEARCH_THRESH 256
>   
> diff --git a/lib/vhost/vhost_user.c b/lib/vhost/vhost_user.c
> index 5eb1dd6812..3147e72f04 100644
> --- a/lib/vhost/vhost_user.c
> +++ b/lib/vhost/vhost_user.c
> @@ -527,6 +527,8 @@ vhost_user_set_vring_num(struct virtio_net **pdev,
>   		return RTE_VHOST_MSG_RESULT_ERR;
>   	}
>   
> +	vq->batch_copy_max_elems = vq->size;
> +

I don't understand the point of this new field. But it can be removed
anyway if we agree to drop the SW fallback.

>   	return RTE_VHOST_MSG_RESULT_OK;
>   }
>   
> diff --git a/lib/vhost/virtio_net.c b/lib/vhost/virtio_net.c
> index b3d954aab4..305f6cd562 100644
> --- a/lib/vhost/virtio_net.c
> +++ b/lib/vhost/virtio_net.c
> @@ -11,6 +11,7 @@
>   #include <rte_net.h>
>   #include <rte_ether.h>
>   #include <rte_ip.h>
> +#include <rte_dmadev.h>
>   #include <rte_vhost.h>
>   #include <rte_tcp.h>
>   #include <rte_udp.h>
> @@ -25,6 +26,10 @@
>   
>   #define MAX_BATCH_LEN 256
>   
> +/* DMA device copy operation tracking array. */
> +struct async_dma_info dma_copy_track[RTE_DMADEV_DEFAULT_MAX];
> +uint16_t dma_poll_factor = 1;
> +
>   static  __rte_always_inline bool
>   rxvq_is_mergeable(struct virtio_net *dev)
>   {
> @@ -43,6 +48,140 @@ is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
>   	return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
>   }
>   
> +static __rte_always_inline uint16_t
> +vhost_async_dma_transfer(struct vhost_virtqueue *vq, int16_t dma_id,
> +		uint16_t vchan_id, uint16_t head_idx,
> +		struct rte_vhost_iov_iter *pkts, uint16_t nr_pkts)
> +{
> +	struct async_dma_vchan_info *dma_info = &dma_copy_track[dma_id].vchans[vchan_id];
> +	uint16_t ring_mask = dma_info->ring_mask;
> +	uint16_t pkt_idx, bce_idx = 0;
> +
> +	rte_spinlock_lock(&dma_info->dma_lock);
> +
> +	for (pkt_idx = 0; pkt_idx < nr_pkts; pkt_idx++) {
> +		struct rte_vhost_iovec *iov = pkts[pkt_idx].iov;
> +		int copy_idx, last_copy_idx = 0;
> +		uint16_t nr_segs = pkts[pkt_idx].nr_segs;
> +		uint16_t nr_sw_copy = 0;
> +		uint16_t i;
> +
> +		if (rte_dma_burst_capacity(dma_id, vchan_id) < nr_segs)
> +			goto out;

I would consider introducing a vhost_async_dma_transfer_one function to
avoid nesting too much loops and make the code cleaner.

> +		for (i = 0; i < nr_segs; i++) {
> +			/* Fallback to SW copy if error happens */
> +			copy_idx = rte_dma_copy(dma_id, vchan_id, (rte_iova_t)iov[i].src_addr,
> +					(rte_iova_t)iov[i].dst_addr, iov[i].len,
> +					RTE_DMA_OP_FLAG_LLC);
> +			if (unlikely(copy_idx < 0)) {

The DMA channel is protected by a lock, and we check the capacity before
initiating the copy.
So I don't expect rte_dma_copy() to fail because of lack of capacity. If
an error happens, that is a serious one.

So, I wonder whether having a SW fallback makes sense. Code would be
much simpler if we just exit early if an error happens. Logging an error
message instead would help debugging. Certainly with rate limiting not
to flood the log file.

> +				/* Find corresponding VA pair and do SW copy */
> +				rte_memcpy(vq->batch_copy_elems[bce_idx].dst,
> +						vq->batch_copy_elems[bce_idx].src,
> +						vq->batch_copy_elems[bce_idx].len);
> +				nr_sw_copy++;
> +
> +				/**
> +				 * All copies of the packet are performed
> +				 * by the CPU, set the packet completion flag
> +				 * to true, as all copies are done.
> +				 */

I think it would better be moved out of the loop to avoid doing the
check for every segment while only the last one has a chance to match.

> +				if (nr_sw_copy == nr_segs) {
> +					vq->async->pkts_cmpl_flag[head_idx % vq->size] = true;
> +					break;
> +				} else if (i == (nr_segs - 1)) {
> +					/**
> +					 * A part of copies of current packet
> +					 * are enqueued to the DMA successfully
> +					 * but the last copy fails, store the
> +					 * packet completion flag address
> +					 * in the last DMA copy slot.
> +					 */
> +					dma_info->pkts_completed_flag[last_copy_idx & ring_mask] =
> +						&vq->async->pkts_cmpl_flag[head_idx % vq->size];
> +					break;
> +				}
> +			} else
> +				last_copy_idx = copy_idx;

Braces on the else as you have braces for the if statement.

> +
> +			bce_idx++;
> +
> +			/**
> +			 * Only store packet completion flag address in the last copy's
> +			 * slot, and other slots are set to NULL.
> +			 */
> +			if (i == (nr_segs - 1)) {
> +				dma_info->pkts_completed_flag[copy_idx & ring_mask] =
> +					&vq->async->pkts_cmpl_flag[head_idx % vq->size];
> +			}
> +		}
> +
> +		dma_info->nr_batching += nr_segs;
> +		if (unlikely(dma_info->nr_batching >= VHOST_ASYNC_DMA_BATCHING_SIZE)) {
> +			rte_dma_submit(dma_id, vchan_id);
> +			dma_info->nr_batching = 0;
> +		}

I wonder whether we could just remove this submit.
I don't expect completions to happen between two packets as the DMA
channel is protected by a lock, so my understanding is once the DMA ring
is full, we just end-up exiting early because DMA channel capacity is 
checked for every packet.

Removing it will maybe improve performance a (very) little bit, but will
certainly make the code simpler to follow.

> +
> +		head_idx++;
> +	}
> +
> +out:
> +	if (dma_info->nr_batching > 0) {

if (likely(...))

> +		rte_dma_submit(dma_id, vchan_id);
> +		dma_info->nr_batching = 0;
> +	}
> +	rte_spinlock_unlock(&dma_info->dma_lock);
> +	vq->batch_copy_nb_elems = 0;
> +
> +	return pkt_idx;
> +}
> +
> +static __rte_always_inline uint16_t
> +vhost_async_dma_check_completed(int16_t dma_id, uint16_t vchan_id, uint16_t max_pkts)
> +{
> +	struct async_dma_vchan_info *dma_info = &dma_copy_track[dma_id].vchans[vchan_id];
> +	uint16_t ring_mask = dma_info->ring_mask;
> +	uint16_t last_idx = 0;
> +	uint16_t nr_copies;
> +	uint16_t copy_idx;
> +	uint16_t i;
> +	bool has_error = false;
> +
> +	rte_spinlock_lock(&dma_info->dma_lock);
> +
> +	/**
> +	 * Print error log for debugging, if DMA reports error during
> +	 * DMA transfer. We do not handle error in vhost level.
> +	 */
> +	nr_copies = rte_dma_completed(dma_id, vchan_id, max_pkts, &last_idx, &has_error);
> +	if (unlikely(has_error)) {
> +		VHOST_LOG_DATA(ERR, "dma %d vchannel %u reports error in rte_dma_completed()\n",
> +				dma_id, vchan_id);

I wonder if rate limiting would

> +	} else if (nr_copies == 0)
> +		goto out;
> +
> +	copy_idx = last_idx - nr_copies + 1;
> +	for (i = 0; i < nr_copies; i++) {
> +		bool *flag;
> +
> +		flag = dma_info->pkts_completed_flag[copy_idx & ring_mask];
> +		if (flag) {
> +			/**
> +			 * Mark the packet flag as received. The flag
> +			 * could belong to another virtqueue but write
> +			 * is atomic.
> +			 */
> +			*flag = true;
> +			dma_info->pkts_completed_flag[copy_idx & ring_mask] = NULL;
> +		}
> +		copy_idx++;
> +	}
> +
> +out:
> +	rte_spinlock_unlock(&dma_info->dma_lock);
> +	return nr_copies;
> +}
> +
>   static inline void
>   do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
>   {
> @@ -865,12 +1004,13 @@ async_iter_reset(struct vhost_async *async)
>   static __rte_always_inline int
>   async_mbuf_to_desc_seg(struct virtio_net *dev, struct vhost_virtqueue *vq,
>   		struct rte_mbuf *m, uint32_t mbuf_offset,
> -		uint64_t buf_iova, uint32_t cpy_len)
> +		uint64_t buf_addr, uint64_t buf_iova, uint32_t cpy_len)
>   {
>   	struct vhost_async *async = vq->async;
>   	uint64_t mapped_len;
>   	uint32_t buf_offset = 0;
>   	void *hpa;
> +	struct batch_copy_elem *bce = vq->batch_copy_elems;
>   
>   	while (cpy_len) {
>   		hpa = (void *)(uintptr_t)gpa_to_first_hpa(dev,
> @@ -886,6 +1026,31 @@ async_mbuf_to_desc_seg(struct virtio_net *dev, struct vhost_virtqueue *vq,
>   						hpa, (size_t)mapped_len)))
>   			return -1;
>   
> +		/**
> +		 * Keep VA for all IOVA segments for falling back to SW
> +		 * copy in case of rte_dma_copy() error.
> +		 */

As said below, I think we could get rid off the SW fallback.
But in case we didn't, I think it would be prefferable to change the
rte_vhost_iovec struct to have both the iova and the VA, that would make
the code simpler.

Also, while looking at this, I notice the structs rte_vhost_iov_iter and
rte_vhost_iovec are still part of the Vhost API, but it should not be
necessary now since application no more need to know about it.

> +		if (unlikely(vq->batch_copy_nb_elems >= vq->batch_copy_max_elems)) {
> +			struct batch_copy_elem *tmp;
> +			uint16_t nb_elems = 2 * vq->batch_copy_max_elems;
> +
> +			VHOST_LOG_DATA(DEBUG, "(%d) %s: run out of batch_copy_elems, "
> +					"and realloc double elements.\n", dev->vid, __func__);
> +			tmp = rte_realloc_socket(vq->batch_copy_elems, nb_elems * sizeof(*tmp),
> +					RTE_CACHE_LINE_SIZE, vq->numa_node);
> +			if (!tmp) {
> +				VHOST_LOG_DATA(ERR, "Failed to re-alloc batch_copy_elems\n");
> +				return -1;
> +			}
> +
> +			vq->batch_copy_max_elems = nb_elems;
> +			vq->batch_copy_elems = tmp;
> +			bce = tmp;
> +		}
> +		bce[vq->batch_copy_nb_elems].dst = (void *)((uintptr_t)(buf_addr + buf_offset));
> +		bce[vq->batch_copy_nb_elems].src = rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
> +		bce[vq->batch_copy_nb_elems++].len = mapped_len;
> +
>   		cpy_len -= (uint32_t)mapped_len;
>   		mbuf_offset += (uint32_t)mapped_len;
>   		buf_offset += (uint32_t)mapped_len;
> @@ -901,7 +1066,8 @@ sync_mbuf_to_desc_seg(struct virtio_net *dev, struct vhost_virtqueue *vq,
>   {
>   	struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
>   
> -	if (likely(cpy_len > MAX_BATCH_LEN || vq->batch_copy_nb_elems >= vq->size)) {
> +	if (likely(cpy_len > MAX_BATCH_LEN ||
> +				vq->batch_copy_nb_elems >= vq->batch_copy_max_elems)) {
>   		rte_memcpy((void *)((uintptr_t)(buf_addr)),
>   				rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
>   				cpy_len);
> @@ -1020,8 +1186,10 @@ mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
>   
>   		if (is_async) {
>   			if (async_mbuf_to_desc_seg(dev, vq, m, mbuf_offset,
> +						buf_addr + buf_offset,
>   						buf_iova + buf_offset, cpy_len) < 0)
>   				goto error;
> +

Remove new line.

>   		} else {
>   			sync_mbuf_to_desc_seg(dev, vq, m, mbuf_offset,
>   					buf_addr + buf_offset,
> @@ -1449,9 +1617,9 @@ store_dma_desc_info_packed(struct vring_used_elem_packed *s_ring,
>   }
>   

Regards,
Maxime
Hu, Jiayu Feb. 7, 2022, 1:34 a.m. UTC | #2
Hi Maxime,

Thanks for your comments. Please see replies inline.

> -----Original Message-----
> From: Maxime Coquelin <maxime.coquelin@redhat.com>
> Sent: Thursday, February 3, 2022 9:04 PM
> To: Hu, Jiayu <jiayu.hu@intel.com>; dev@dpdk.org
> Cc: i.maximets@ovn.org; Xia, Chenbo <chenbo.xia@intel.com>; Richardson,
> Bruce <bruce.richardson@intel.com>; Van Haaren, Harry
> <harry.van.haaren@intel.com>; Pai G, Sunil <sunil.pai.g@intel.com>;
> Mcnamara, John <john.mcnamara@intel.com>; Ding, Xuan
> <xuan.ding@intel.com>; Jiang, Cheng1 <cheng1.jiang@intel.com>;
> liangma@liangbit.com
> Subject: Re: [PATCH v2 1/1] vhost: integrate dmadev in asynchronous
> datapath
> 
> Hi Jiayu,
> 
> On 1/24/22 17:40, Jiayu Hu wrote:
> > Since dmadev is introduced in 21.11, to avoid the overhead of vhost
> > DMA abstraction layer and simplify application logics, this patch
> > integrates dmadev in asynchronous data path.
> >
> > Signed-off-by: Jiayu Hu <jiayu.hu@intel.com>
> > Signed-off-by: Sunil Pai G <sunil.pai.g@intel.com>
> > ---
> >   doc/guides/prog_guide/vhost_lib.rst |  95 ++++-----
> >   examples/vhost/Makefile             |   2 +-
> >   examples/vhost/ioat.c               | 218 --------------------
> >   examples/vhost/ioat.h               |  63 ------
> >   examples/vhost/main.c               | 255 ++++++++++++++++++-----
> >   examples/vhost/main.h               |  11 +
> >   examples/vhost/meson.build          |   6 +-
> >   lib/vhost/meson.build               |   2 +-
> >   lib/vhost/rte_vhost.h               |   2 +
> >   lib/vhost/rte_vhost_async.h         | 132 +++++-------
> >   lib/vhost/version.map               |   3 +
> >   lib/vhost/vhost.c                   | 148 ++++++++++----
> >   lib/vhost/vhost.h                   |  64 +++++-
> >   lib/vhost/vhost_user.c              |   2 +
> >   lib/vhost/virtio_net.c              | 305 +++++++++++++++++++++++-----
> >   15 files changed, 744 insertions(+), 564 deletions(-)
> >   delete mode 100644 examples/vhost/ioat.c
> >   delete mode 100644 examples/vhost/ioat.h
> >
> 
> When you rebase to the next version, please ensure to rework all the logs to
> follow the new standard:
> VHOST_LOG_CONFIG(ERR,"(%s) .....", dev->ifname, ...);

Sure, will do.

> 
> > git a/lib/vhost/rte_vhost_async.h b/lib/vhost/rte_vhost_async.h index
> > a87ea6ba37..758a80f403 100644
> > --- a/lib/vhost/rte_vhost_async.h
> > +++ b/lib/vhost/rte_vhost_async.h
> > @@ -26,73 +26,6 @@ struct rte_vhost_iov_iter {
> >   	unsigned long nr_segs;
> >   };
> >
> > -/**
> > - * dma transfer status
> > - */
> > -struct rte_vhost_async_status {
> > -	/** An array of application specific data for source memory */
> > -	uintptr_t *src_opaque_data;
> > -	/** An array of application specific data for destination memory */
> > -	uintptr_t *dst_opaque_data;
> > -};
> > -
> > -/**
> > - * dma operation callbacks to be implemented by applications
> > - */
> > -struct rte_vhost_async_channel_ops {
> > -	/**
> > -	 * instruct async engines to perform copies for a batch of packets
> > -	 *
> > -	 * @param vid
> > -	 *  id of vhost device to perform data copies
> > -	 * @param queue_id
> > -	 *  queue id to perform data copies
> > -	 * @param iov_iter
> > -	 *  an array of IOV iterators
> > -	 * @param opaque_data
> > -	 *  opaque data pair sending to DMA engine
> > -	 * @param count
> > -	 *  number of elements in the "descs" array
> > -	 * @return
> > -	 *  number of IOV iterators processed, negative value means error
> > -	 */
> > -	int32_t (*transfer_data)(int vid, uint16_t queue_id,
> > -		struct rte_vhost_iov_iter *iov_iter,
> > -		struct rte_vhost_async_status *opaque_data,
> > -		uint16_t count);
> > -	/**
> > -	 * check copy-completed packets from the async engine
> > -	 * @param vid
> > -	 *  id of vhost device to check copy completion
> > -	 * @param queue_id
> > -	 *  queue id to check copy completion
> > -	 * @param opaque_data
> > -	 *  buffer to receive the opaque data pair from DMA engine
> > -	 * @param max_packets
> > -	 *  max number of packets could be completed
> > -	 * @return
> > -	 *  number of async descs completed, negative value means error
> > -	 */
> > -	int32_t (*check_completed_copies)(int vid, uint16_t queue_id,
> > -		struct rte_vhost_async_status *opaque_data,
> > -		uint16_t max_packets);
> > -};
> > -
> > -/**
> > - *  async channel features
> > - */
> > -enum {
> > -	RTE_VHOST_ASYNC_INORDER = 1U << 0,
> > -};
> > -
> > -/**
> > - *  async channel configuration
> > - */
> > -struct rte_vhost_async_config {
> > -	uint32_t features;
> > -	uint32_t rsvd[2];
> > -};
> > -
> >   /**
> >    * Register an async channel for a vhost queue
> >    *
> > @@ -100,17 +33,11 @@ struct rte_vhost_async_config {
> >    *  vhost device id async channel to be attached to
> >    * @param queue_id
> >    *  vhost queue id async channel to be attached to
> > - * @param config
> > - *  Async channel configuration structure
> > - * @param ops
> > - *  Async channel operation callbacks
> >    * @return
> >    *  0 on success, -1 on failures
> >    */
> >   __rte_experimental
> > -int rte_vhost_async_channel_register(int vid, uint16_t queue_id,
> > -	struct rte_vhost_async_config config,
> > -	struct rte_vhost_async_channel_ops *ops);
> > +int rte_vhost_async_channel_register(int vid, uint16_t queue_id);
> >
> >   /**
> >    * Unregister an async channel for a vhost queue @@ -136,17 +63,11
> > @@ int rte_vhost_async_channel_unregister(int vid, uint16_t queue_id);
> >    *  vhost device id async channel to be attached to
> >    * @param queue_id
> >    *  vhost queue id async channel to be attached to
> > - * @param config
> > - *  Async channel configuration
> > - * @param ops
> > - *  Async channel operation callbacks
> >    * @return
> >    *  0 on success, -1 on failures
> >    */
> >   __rte_experimental
> > -int rte_vhost_async_channel_register_thread_unsafe(int vid, uint16_t
> queue_id,
> > -	struct rte_vhost_async_config config,
> > -	struct rte_vhost_async_channel_ops *ops);
> > +int rte_vhost_async_channel_register_thread_unsafe(int vid, uint16_t
> > +queue_id);
> >
> >   /**
> >    * Unregister an async channel for a vhost queue without performing
> > any @@ -179,12 +100,17 @@ int
> rte_vhost_async_channel_unregister_thread_unsafe(int vid,
> >    *  array of packets to be enqueued
> >    * @param count
> >    *  packets num to be enqueued
> > + * @param dma_id
> > + *  the identifier of the DMA device
> > + * @param vchan_id
> > + *  the identifier of virtual DMA channel
> >    * @return
> >    *  num of packets enqueued
> >    */
> >   __rte_experimental
> >   uint16_t rte_vhost_submit_enqueue_burst(int vid, uint16_t queue_id,
> > -		struct rte_mbuf **pkts, uint16_t count);
> > +		struct rte_mbuf **pkts, uint16_t count, int16_t dma_id,
> > +		uint16_t vchan_id);
> >
> >   /**
> >    * This function checks async completion status for a specific vhost
> > @@ -199,12 +125,17 @@ uint16_t rte_vhost_submit_enqueue_burst(int
> vid, uint16_t queue_id,
> >    *  blank array to get return packet pointer
> >    * @param count
> >    *  size of the packet array
> > + * @param dma_id
> > + *  the identifier of the DMA device
> > + * @param vchan_id
> > + *  the identifier of virtual DMA channel
> >    * @return
> >    *  num of packets returned
> >    */
> >   __rte_experimental
> >   uint16_t rte_vhost_poll_enqueue_completed(int vid, uint16_t queue_id,
> > -		struct rte_mbuf **pkts, uint16_t count);
> > +		struct rte_mbuf **pkts, uint16_t count, int16_t dma_id,
> > +		uint16_t vchan_id);
> >
> >   /**
> >    * This function returns the amount of in-flight packets for the
> > vhost @@ -235,11 +166,44 @@ int rte_vhost_async_get_inflight(int vid,
> uint16_t queue_id);
> >    *  Blank array to get return packet pointer
> >    * @param count
> >    *  Size of the packet array
> > + * @param dma_id
> > + *  the identifier of the DMA device
> > + * @param vchan_id
> > + *  the identifier of virtual DMA channel
> >    * @return
> >    *  Number of packets returned
> >    */
> >   __rte_experimental
> >   uint16_t rte_vhost_clear_queue_thread_unsafe(int vid, uint16_t queue_id,
> > -		struct rte_mbuf **pkts, uint16_t count);
> > +		struct rte_mbuf **pkts, uint16_t count, int16_t dma_id,
> > +		uint16_t vchan_id);
> > +/**
> > + * The DMA vChannels used in asynchronous data path must be
> > +configured
> > + * first. So this function needs to be called before enabling DMA
> > + * acceleration for vring. If this function fails, asynchronous data
> > +path
> > + * cannot be enabled for any vring further.
> > + *
> > + * DMA devices used in data-path must belong to DMA devices given in
> > +this
> > + * function. But users are free to use DMA devices given in the
> > +function
> > + * in non-vhost scenarios, only if guarantee no copies in vhost are
> > + * offloaded to them at the same time.
> > + *
> > + * @param dmas_id
> > + *  DMA ID array
> > + * @param count
> > + *  Element number of 'dmas_id'
> > + * @param poll_factor
> > + *  For large or scatter-gather packets, one packet would consist of
> > + *  small buffers. In this case, vhost will issue several DMA copy
> > + *  operations for the packet. Therefore, the number of copies to
> > + *  check by rte_dma_completed() is calculated by "nb_pkts_to_poll *
> > + *  poll_factor" andused in rte_vhost_poll_enqueue_completed(). The
> > + *  default value of "poll_factor" is 1.
> > + * @return
> > + *  0 on success, and -1 on failure
> > + */
> > +__rte_experimental
> > +int rte_vhost_async_dma_configure(int16_t *dmas_id, uint16_t count,
> > +		uint16_t poll_factor);
> >
> >   #endif /* _RTE_VHOST_ASYNC_H_ */
> > diff --git a/lib/vhost/version.map b/lib/vhost/version.map index
> > a7ef7f1976..1202ba9c1a 100644
> > --- a/lib/vhost/version.map
> > +++ b/lib/vhost/version.map
> > @@ -84,6 +84,9 @@ EXPERIMENTAL {
> >
> >   	# added in 21.11
> >   	rte_vhost_get_monitor_addr;
> > +
> > +	# added in 22.03
> > +	rte_vhost_async_dma_configure;
> >   };
> >
> >   INTERNAL {
> > diff --git a/lib/vhost/vhost.c b/lib/vhost/vhost.c index
> > 13a9bb9dd1..c408cee63e 100644
> > --- a/lib/vhost/vhost.c
> > +++ b/lib/vhost/vhost.c
> > @@ -25,7 +25,7 @@
> >   #include "vhost.h"
> >   #include "vhost_user.h"
> >
> > -struct virtio_net *vhost_devices[MAX_VHOST_DEVICE];
> > +struct virtio_net *vhost_devices[RTE_MAX_VHOST_DEVICE];
> >   pthread_mutex_t vhost_dev_lock = PTHREAD_MUTEX_INITIALIZER;
> >
> >   /* Called with iotlb_lock read-locked */ @@ -344,6 +344,7 @@
> > vhost_free_async_mem(struct vhost_virtqueue *vq)
> >   		return;
> >
> >   	rte_free(vq->async->pkts_info);
> > +	rte_free(vq->async->pkts_cmpl_flag);
> >
> >   	rte_free(vq->async->buffers_packed);
> >   	vq->async->buffers_packed = NULL;
> > @@ -667,12 +668,12 @@ vhost_new_device(void)
> >   	int i;
> >
> >   	pthread_mutex_lock(&vhost_dev_lock);
> > -	for (i = 0; i < MAX_VHOST_DEVICE; i++) {
> > +	for (i = 0; i < RTE_MAX_VHOST_DEVICE; i++) {
> >   		if (vhost_devices[i] == NULL)
> >   			break;
> >   	}
> >
> > -	if (i == MAX_VHOST_DEVICE) {
> > +	if (i == RTE_MAX_VHOST_DEVICE) {
> >   		VHOST_LOG_CONFIG(ERR,
> >   			"Failed to find a free slot for new device.\n");
> >   		pthread_mutex_unlock(&vhost_dev_lock);
> > @@ -1626,8 +1627,7 @@ rte_vhost_extern_callback_register(int vid,
> >   }
> >
> >   static __rte_always_inline int
> > -async_channel_register(int vid, uint16_t queue_id,
> > -		struct rte_vhost_async_channel_ops *ops)
> > +async_channel_register(int vid, uint16_t queue_id)
> >   {
> >   	struct virtio_net *dev = get_device(vid);
> >   	struct vhost_virtqueue *vq = dev->virtqueue[queue_id]; @@ -1656,6
> > +1656,14 @@ async_channel_register(int vid, uint16_t queue_id,
> >   		goto out_free_async;
> >   	}
> >
> > +	async->pkts_cmpl_flag = rte_zmalloc_socket(NULL, vq->size *
> sizeof(bool),
> > +			RTE_CACHE_LINE_SIZE, node);
> > +	if (!async->pkts_cmpl_flag) {
> > +		VHOST_LOG_CONFIG(ERR, "failed to allocate async
> pkts_cmpl_flag (vid %d, qid: %d)\n",
> > +				vid, queue_id);
> > +		goto out_free_async;
> > +	}
> > +
> >   	if (vq_is_packed(dev)) {
> >   		async->buffers_packed = rte_malloc_socket(NULL,
> >   				vq->size * sizeof(struct
> vring_used_elem_packed), @@ -1676,9
> > +1684,6 @@ async_channel_register(int vid, uint16_t queue_id,
> >   		}
> >   	}
> >
> > -	async->ops.check_completed_copies = ops-
> >check_completed_copies;
> > -	async->ops.transfer_data = ops->transfer_data;
> > -
> >   	vq->async = async;
> >
> >   	return 0;
> > @@ -1691,15 +1696,13 @@ async_channel_register(int vid, uint16_t
> queue_id,
> >   }
> >
> >   int
> > -rte_vhost_async_channel_register(int vid, uint16_t queue_id,
> > -		struct rte_vhost_async_config config,
> > -		struct rte_vhost_async_channel_ops *ops)
> > +rte_vhost_async_channel_register(int vid, uint16_t queue_id)
> >   {
> >   	struct vhost_virtqueue *vq;
> >   	struct virtio_net *dev = get_device(vid);
> >   	int ret;
> >
> > -	if (dev == NULL || ops == NULL)
> > +	if (dev == NULL)
> >   		return -1;
> >
> >   	if (queue_id >= VHOST_MAX_VRING)
> > @@ -1710,33 +1713,20 @@ rte_vhost_async_channel_register(int vid,
> uint16_t queue_id,
> >   	if (unlikely(vq == NULL || !dev->async_copy))
> >   		return -1;
> >
> > -	if (unlikely(!(config.features & RTE_VHOST_ASYNC_INORDER))) {
> > -		VHOST_LOG_CONFIG(ERR,
> > -			"async copy is not supported on non-inorder mode "
> > -			"(vid %d, qid: %d)\n", vid, queue_id);
> > -		return -1;
> > -	}
> > -
> > -	if (unlikely(ops->check_completed_copies == NULL ||
> > -		ops->transfer_data == NULL))
> > -		return -1;
> > -
> >   	rte_spinlock_lock(&vq->access_lock);
> > -	ret = async_channel_register(vid, queue_id, ops);
> > +	ret = async_channel_register(vid, queue_id);
> >   	rte_spinlock_unlock(&vq->access_lock);
> >
> >   	return ret;
> >   }
> >
> >   int
> > -rte_vhost_async_channel_register_thread_unsafe(int vid, uint16_t
> queue_id,
> > -		struct rte_vhost_async_config config,
> > -		struct rte_vhost_async_channel_ops *ops)
> > +rte_vhost_async_channel_register_thread_unsafe(int vid, uint16_t
> > +queue_id)
> >   {
> >   	struct vhost_virtqueue *vq;
> >   	struct virtio_net *dev = get_device(vid);
> >
> > -	if (dev == NULL || ops == NULL)
> > +	if (dev == NULL)
> >   		return -1;
> >
> >   	if (queue_id >= VHOST_MAX_VRING)
> > @@ -1747,18 +1737,7 @@
> rte_vhost_async_channel_register_thread_unsafe(int vid, uint16_t queue_id,
> >   	if (unlikely(vq == NULL || !dev->async_copy))
> >   		return -1;
> >
> > -	if (unlikely(!(config.features & RTE_VHOST_ASYNC_INORDER))) {
> > -		VHOST_LOG_CONFIG(ERR,
> > -			"async copy is not supported on non-inorder mode "
> > -			"(vid %d, qid: %d)\n", vid, queue_id);
> > -		return -1;
> > -	}
> > -
> > -	if (unlikely(ops->check_completed_copies == NULL ||
> > -		ops->transfer_data == NULL))
> > -		return -1;
> > -
> > -	return async_channel_register(vid, queue_id, ops);
> > +	return async_channel_register(vid, queue_id);
> >   }
> >
> >   int
> > @@ -1835,6 +1814,95 @@
> rte_vhost_async_channel_unregister_thread_unsafe(int vid, uint16_t
> queue_id)
> >   	return 0;
> >   }
> >
> > +static __rte_always_inline void
> > +vhost_free_async_dma_mem(void)
> > +{
> > +	uint16_t i;
> > +
> > +	for (i = 0; i < RTE_DMADEV_DEFAULT_MAX; i++) {
> > +		struct async_dma_info *dma = &dma_copy_track[i];
> > +		int16_t j;
> > +
> > +		if (dma->max_vchans == 0)
> > +			continue;
> > +
> > +		for (j = 0; j < dma->max_vchans; j++)
> > +			rte_free(dma->vchans[j].pkts_completed_flag);
> > +
> > +		rte_free(dma->vchans);
> > +		dma->vchans = NULL;
> > +		dma->max_vchans = 0;
> > +	}
> > +}
> > +
> > +int
> > +rte_vhost_async_dma_configure(int16_t *dmas_id, uint16_t count,
> > +uint16_t poll_factor)
> 
> I'm not fan of the poll_factor, I think it is too complex for the user to know
> what value he should set.

It seems so, and users need to know DMA offloading details before setting it.
The simple way is setting the max copies to check by rte_dma_completed to
9728/2048*32, where 9728 is equal to VIRTIO_MAX_RX_PKTLEN and 32 is
MAX_PKT_BURST. It should be able to cover most of cases.

> 
> Also, I would like that the API only registers one DMA channel at a time and
> let the application call it multiple times. Dong that, user can still use the DMA
> channels that could be registered.

Sure, I will change it.
> 
> > +{
> > +	uint16_t i;
> > +
> > +	if (!dmas_id) {
> > +		VHOST_LOG_CONFIG(ERR, "Invalid DMA configuration
> parameter.\n");
> > +		return -1;
> > +	}
> > +
> > +	if (poll_factor == 0) {
> > +		VHOST_LOG_CONFIG(ERR, "Invalid DMA poll factor %u\n",
> poll_factor);
> > +		return -1;
> > +	}
> > +	dma_poll_factor = poll_factor;
> > +
> > +	for (i = 0; i < count; i++) {
> > +		struct async_dma_vchan_info *vchans;
> > +		struct rte_dma_info info;
> > +		uint16_t max_vchans;
> > +		uint16_t max_desc;
> > +		uint16_t j;
> > +
> > +		if (!rte_dma_is_valid(dmas_id[i])) {
> > +			VHOST_LOG_CONFIG(ERR, "DMA %d is not found.
> Cannot enable async"
> > +				       " data-path\n.", dmas_id[i]);
> > +			vhost_free_async_dma_mem();
> > +			return -1;
> > +		}
> > +
> > +		rte_dma_info_get(dmas_id[i], &info);
> > +
> > +		max_vchans = info.max_vchans;
> > +		max_desc = info.max_desc;
> > +
> > +		if (!rte_is_power_of_2(max_desc))
> > +			max_desc = rte_align32pow2(max_desc);
> > +
> > +		vchans = rte_zmalloc(NULL, sizeof(struct
> async_dma_vchan_info) * max_vchans,
> > +				RTE_CACHE_LINE_SIZE);
> > +		if (vchans == NULL) {
> > +			VHOST_LOG_CONFIG(ERR, "Failed to allocate vchans
> for dma-%d."
> > +					" Cannot enable async data-path.\n",
> dmas_id[i]);
> > +			vhost_free_async_dma_mem();
> > +			return -1;
> > +		}
> > +
> > +		for (j = 0; j < max_vchans; j++) {
> > +			vchans[j].pkts_completed_flag = rte_zmalloc(NULL,
> sizeof(bool *) * max_desc,
> > +					RTE_CACHE_LINE_SIZE);
> > +			if (!vchans[j].pkts_completed_flag) {
> > +				VHOST_LOG_CONFIG(ERR, "Failed to allocate
> pkts_completed_flag for "
> > +						"dma-%d vchan-%u\n",
> dmas_id[i], j);
> > +				vhost_free_async_dma_mem();
> > +				return -1;
> > +			}
> > +
> > +			vchans[j].ring_size = max_desc;
> > +			vchans[j].ring_mask = max_desc - 1;
> > +		}
> > +
> > +		dma_copy_track[dmas_id[i]].vchans = vchans;
> > +		dma_copy_track[dmas_id[i]].max_vchans = max_vchans;
> > +	}
> > +
> > +	return 0;
> > +}
> > +
> >   int
> >   rte_vhost_async_get_inflight(int vid, uint16_t queue_id)
> >   {
> > diff --git a/lib/vhost/vhost.h b/lib/vhost/vhost.h index
> > 7085e0885c..475843fec0 100644
> > --- a/lib/vhost/vhost.h
> > +++ b/lib/vhost/vhost.h
> > @@ -19,6 +19,7 @@
> >   #include <rte_ether.h>
> >   #include <rte_rwlock.h>
> >   #include <rte_malloc.h>
> > +#include <rte_dmadev.h>
> >
> >   #include "rte_vhost.h"
> >   #include "rte_vdpa.h"
> > @@ -50,6 +51,7 @@
> >
> >   #define VHOST_MAX_ASYNC_IT (MAX_PKT_BURST)
> >   #define VHOST_MAX_ASYNC_VEC 2048
> > +#define VHOST_ASYNC_DMA_BATCHING_SIZE 32
> >
> >   #define PACKED_DESC_ENQUEUE_USED_FLAG(w)	\
> >   	((w) ? (VRING_DESC_F_AVAIL | VRING_DESC_F_USED |
> > VRING_DESC_F_WRITE) : \ @@ -119,6 +121,42 @@ struct
> vring_used_elem_packed {
> >   	uint32_t count;
> >   };
> >
> > +struct async_dma_vchan_info {
> > +	/* circular array to track if packet copy completes */
> > +	bool **pkts_completed_flag;
> > +
> > +	/* max elements in 'metadata' */
> > +	uint16_t ring_size;
> > +	/* ring index mask for 'metadata' */
> > +	uint16_t ring_mask;
> > +
> > +	/* batching copies before a DMA doorbell */
> > +	uint16_t nr_batching;
> > +
> > +	/**
> > +	 * DMA virtual channel lock. Although it is able to bind DMA
> > +	 * virtual channels to data plane threads, vhost control plane
> > +	 * thread could call data plane functions too, thus causing
> > +	 * DMA device contention.
> > +	 *
> > +	 * For example, in VM exit case, vhost control plane thread needs
> > +	 * to clear in-flight packets before disable vring, but there could
> > +	 * be anotther data plane thread is enqueuing packets to the same
> > +	 * vring with the same DMA virtual channel. But dmadev PMD
> functions
> > +	 * are lock-free, so the control plane and data plane threads
> > +	 * could operate the same DMA virtual channel at the same time.
> > +	 */
> > +	rte_spinlock_t dma_lock;
> > +};
> > +
> > +struct async_dma_info {
> > +	uint16_t max_vchans;
> > +	struct async_dma_vchan_info *vchans; };
> > +
> > +extern struct async_dma_info
> dma_copy_track[RTE_DMADEV_DEFAULT_MAX];
> > +extern uint16_t dma_poll_factor;
> > +
> >   /**
> >    * inflight async packet information
> >    */
> > @@ -129,9 +167,6 @@ struct async_inflight_info {
> >   };
> >
> >   struct vhost_async {
> > -	/* operation callbacks for DMA */
> > -	struct rte_vhost_async_channel_ops ops;
> > -
> >   	struct rte_vhost_iov_iter iov_iter[VHOST_MAX_ASYNC_IT];
> >   	struct rte_vhost_iovec iovec[VHOST_MAX_ASYNC_VEC];
> >   	uint16_t iter_idx;
> > @@ -139,6 +174,25 @@ struct vhost_async {
> >
> >   	/* data transfer status */
> >   	struct async_inflight_info *pkts_info;
> > +	/**
> > +	 * Packet reorder array. "true" indicates that DMA device
> > +	 * completes all copies for the packet.
> > +	 *
> > +	 * Note that this array could be written by multiple threads
> > +	 * simultaneously. For example, in the case of thread0 and
> > +	 * thread1 RX packets from NIC and then enqueue packets to
> > +	 * vring0 and vring1 with own DMA device DMA0 and DMA1, it's
> > +	 * possible for thread0 to get completed copies belonging to
> > +	 * vring1 from DMA0, while thread0 is calling rte_vhost_poll
> > +	 * _enqueue_completed() for vring0 and thread1 is calling
> > +	 * rte_vhost_submit_enqueue_burst() for vring1. In this case,
> > +	 * vq->access_lock cannot protect pkts_cmpl_flag of vring1.
> > +	 *
> > +	 * However, since offloading is per-packet basis, each packet
> > +	 * flag will only be written by one thread. And single byte
> > +	 * write is atomic, so no lock for pkts_cmpl_flag is needed.
> > +	 */
> > +	bool *pkts_cmpl_flag;
> >   	uint16_t pkts_idx;
> >   	uint16_t pkts_inflight_n;
> >   	union {
> > @@ -198,6 +252,7 @@ struct vhost_virtqueue {
> >   	/* Record packed ring first dequeue desc index */
> >   	uint16_t		shadow_last_used_idx;
> >
> > +	uint16_t		batch_copy_max_elems;
> >   	uint16_t		batch_copy_nb_elems;
> >   	struct batch_copy_elem	*batch_copy_elems;
> >   	int			numa_node;
> > @@ -568,8 +623,7 @@ extern int vhost_data_log_level;
> >   #define PRINT_PACKET(device, addr, size, header) do {} while (0)
> >   #endif
> >
> > -#define MAX_VHOST_DEVICE	1024
> > -extern struct virtio_net *vhost_devices[MAX_VHOST_DEVICE];
> > +extern struct virtio_net *vhost_devices[RTE_MAX_VHOST_DEVICE];
> >
> >   #define VHOST_BINARY_SEARCH_THRESH 256
> >
> > diff --git a/lib/vhost/vhost_user.c b/lib/vhost/vhost_user.c index
> > 5eb1dd6812..3147e72f04 100644
> > --- a/lib/vhost/vhost_user.c
> > +++ b/lib/vhost/vhost_user.c
> > @@ -527,6 +527,8 @@ vhost_user_set_vring_num(struct virtio_net
> **pdev,
> >   		return RTE_VHOST_MSG_RESULT_ERR;
> >   	}
> >
> > +	vq->batch_copy_max_elems = vq->size;
> > +
> 
> I don't understand the point of this new field. But it can be removed anyway
> if we agree to drop the SW fallback.

This is for handling lacking of batch_copy elements in SW fallback.

> 
> >   	return RTE_VHOST_MSG_RESULT_OK;
> >   }
> >
> > diff --git a/lib/vhost/virtio_net.c b/lib/vhost/virtio_net.c index
> > b3d954aab4..305f6cd562 100644
> > --- a/lib/vhost/virtio_net.c
> > +++ b/lib/vhost/virtio_net.c
> > @@ -11,6 +11,7 @@
> >   #include <rte_net.h>
> >   #include <rte_ether.h>
> >   #include <rte_ip.h>
> > +#include <rte_dmadev.h>
> >   #include <rte_vhost.h>
> >   #include <rte_tcp.h>
> >   #include <rte_udp.h>
> > @@ -25,6 +26,10 @@
> >
> >   #define MAX_BATCH_LEN 256
> >
> > +/* DMA device copy operation tracking array. */ struct async_dma_info
> > +dma_copy_track[RTE_DMADEV_DEFAULT_MAX];
> > +uint16_t dma_poll_factor = 1;
> > +
> >   static  __rte_always_inline bool
> >   rxvq_is_mergeable(struct virtio_net *dev)
> >   {
> > @@ -43,6 +48,140 @@ is_valid_virt_queue_idx(uint32_t idx, int is_tx,
> uint32_t nr_vring)
> >   	return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
> >   }
> >
> > +static __rte_always_inline uint16_t
> > +vhost_async_dma_transfer(struct vhost_virtqueue *vq, int16_t dma_id,
> > +		uint16_t vchan_id, uint16_t head_idx,
> > +		struct rte_vhost_iov_iter *pkts, uint16_t nr_pkts) {
> > +	struct async_dma_vchan_info *dma_info =
> &dma_copy_track[dma_id].vchans[vchan_id];
> > +	uint16_t ring_mask = dma_info->ring_mask;
> > +	uint16_t pkt_idx, bce_idx = 0;
> > +
> > +	rte_spinlock_lock(&dma_info->dma_lock);
> > +
> > +	for (pkt_idx = 0; pkt_idx < nr_pkts; pkt_idx++) {
> > +		struct rte_vhost_iovec *iov = pkts[pkt_idx].iov;
> > +		int copy_idx, last_copy_idx = 0;
> > +		uint16_t nr_segs = pkts[pkt_idx].nr_segs;
> > +		uint16_t nr_sw_copy = 0;
> > +		uint16_t i;
> > +
> > +		if (rte_dma_burst_capacity(dma_id, vchan_id) < nr_segs)
> > +			goto out;
> 
> I would consider introducing a vhost_async_dma_transfer_one function to
> avoid nesting too much loops and make the code cleaner.

Sure, I will add it.

> 
> > +		for (i = 0; i < nr_segs; i++) {
> > +			/* Fallback to SW copy if error happens */
> > +			copy_idx = rte_dma_copy(dma_id, vchan_id,
> (rte_iova_t)iov[i].src_addr,
> > +					(rte_iova_t)iov[i].dst_addr, iov[i].len,
> > +					RTE_DMA_OP_FLAG_LLC);
> > +			if (unlikely(copy_idx < 0)) {
> 
> The DMA channel is protected by a lock, and we check the capacity before
> initiating the copy.
> So I don't expect rte_dma_copy() to fail because of lack of capacity. If an
> error happens, that is a serious one.
> 
> So, I wonder whether having a SW fallback makes sense. Code would be
> much simpler if we just exit early if an error happens. Logging an error
> message instead would help debugging. Certainly with rate limiting not to
> flood the log file.

That's correct. If error really happens in this case, it means DMA definitely
has something wrong. Better to stop async data-path and debug. SW fallback
may hide this serious issue.

If no objections, I will remove SW fallback but adding an error log.

> 
> > +				/* Find corresponding VA pair and do SW
> copy */
> > +				rte_memcpy(vq-
> >batch_copy_elems[bce_idx].dst,
> > +						vq-
> >batch_copy_elems[bce_idx].src,
> > +						vq-
> >batch_copy_elems[bce_idx].len);
> > +				nr_sw_copy++;
> > +
> > +				/**
> > +				 * All copies of the packet are performed
> > +				 * by the CPU, set the packet completion flag
> > +				 * to true, as all copies are done.
> > +				 */
> 
> I think it would better be moved out of the loop to avoid doing the check for
> every segment while only the last one has a chance to match.

I didn't get the point. How to get rid of the loop? Do you suggest to do SW copy
for all left copies once one DMA error happens (if SW copy is kept)?

> 
> > +				if (nr_sw_copy == nr_segs) {
> > +					vq->async->pkts_cmpl_flag[head_idx %
> vq->size] = true;
> > +					break;
> > +				} else if (i == (nr_segs - 1)) {
> > +					/**
> > +					 * A part of copies of current packet
> > +					 * are enqueued to the DMA
> successfully
> > +					 * but the last copy fails, store the
> > +					 * packet completion flag address
> > +					 * in the last DMA copy slot.
> > +					 */
> > +					dma_info-
> >pkts_completed_flag[last_copy_idx & ring_mask] =
> > +						&vq->async-
> >pkts_cmpl_flag[head_idx % vq->size];
> > +					break;
> > +				}
> > +			} else
> > +				last_copy_idx = copy_idx;
> 
> Braces on the else as you have braces for the if statement.
> 
> > +
> > +			bce_idx++;
> > +
> > +			/**
> > +			 * Only store packet completion flag address in the
> last copy's
> > +			 * slot, and other slots are set to NULL.
> > +			 */
> > +			if (i == (nr_segs - 1)) {
> > +				dma_info->pkts_completed_flag[copy_idx &
> ring_mask] =
> > +					&vq->async-
> >pkts_cmpl_flag[head_idx % vq->size];
> > +			}
> > +		}
> > +
> > +		dma_info->nr_batching += nr_segs;
> > +		if (unlikely(dma_info->nr_batching >=
> VHOST_ASYNC_DMA_BATCHING_SIZE)) {
> > +			rte_dma_submit(dma_id, vchan_id);
> > +			dma_info->nr_batching = 0;
> > +		}
> 
> I wonder whether we could just remove this submit.
> I don't expect completions to happen between two packets as the DMA
> channel is protected by a lock, so my understanding is once the DMA ring is
> full, we just end-up exiting early because DMA channel capacity is checked
> for every packet.
> 
> Removing it will maybe improve performance a (very) little bit, but will
> certainly make the code simpler to follow.

Good suggestion. I will remove it.

> 
> > +
> > +		head_idx++;
> > +	}
> > +
> > +out:
> > +	if (dma_info->nr_batching > 0) {
> 
> if (likely(...))

Sure, I will add later.

> 
> > +		rte_dma_submit(dma_id, vchan_id);
> > +		dma_info->nr_batching = 0;
> > +	}
> > +	rte_spinlock_unlock(&dma_info->dma_lock);
> > +	vq->batch_copy_nb_elems = 0;
> > +
> > +	return pkt_idx;
> > +}
> > +
> > +static __rte_always_inline uint16_t
> > +vhost_async_dma_check_completed(int16_t dma_id, uint16_t vchan_id,
> > +uint16_t max_pkts) {
> > +	struct async_dma_vchan_info *dma_info =
> &dma_copy_track[dma_id].vchans[vchan_id];
> > +	uint16_t ring_mask = dma_info->ring_mask;
> > +	uint16_t last_idx = 0;
> > +	uint16_t nr_copies;
> > +	uint16_t copy_idx;
> > +	uint16_t i;
> > +	bool has_error = false;
> > +
> > +	rte_spinlock_lock(&dma_info->dma_lock);
> > +
> > +	/**
> > +	 * Print error log for debugging, if DMA reports error during
> > +	 * DMA transfer. We do not handle error in vhost level.
> > +	 */
> > +	nr_copies = rte_dma_completed(dma_id, vchan_id, max_pkts,
> &last_idx, &has_error);
> > +	if (unlikely(has_error)) {
> > +		VHOST_LOG_DATA(ERR, "dma %d vchannel %u reports error
> in rte_dma_completed()\n",
> > +				dma_id, vchan_id);
> 
> I wonder if rate limiting would

Sure, I will avoid log flooding.

> 
> > +	} else if (nr_copies == 0)
> > +		goto out;
> > +
> > +	copy_idx = last_idx - nr_copies + 1;
> > +	for (i = 0; i < nr_copies; i++) {
> > +		bool *flag;
> > +
> > +		flag = dma_info->pkts_completed_flag[copy_idx &
> ring_mask];
> > +		if (flag) {
> > +			/**
> > +			 * Mark the packet flag as received. The flag
> > +			 * could belong to another virtqueue but write
> > +			 * is atomic.
> > +			 */
> > +			*flag = true;
> > +			dma_info->pkts_completed_flag[copy_idx &
> ring_mask] = NULL;
> > +		}
> > +		copy_idx++;
> > +	}
> > +
> > +out:
> > +	rte_spinlock_unlock(&dma_info->dma_lock);
> > +	return nr_copies;
> > +}
> > +
> >   static inline void
> >   do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
> >   {
> > @@ -865,12 +1004,13 @@ async_iter_reset(struct vhost_async *async)
> >   static __rte_always_inline int
> >   async_mbuf_to_desc_seg(struct virtio_net *dev, struct vhost_virtqueue
> *vq,
> >   		struct rte_mbuf *m, uint32_t mbuf_offset,
> > -		uint64_t buf_iova, uint32_t cpy_len)
> > +		uint64_t buf_addr, uint64_t buf_iova, uint32_t cpy_len)
> >   {
> >   	struct vhost_async *async = vq->async;
> >   	uint64_t mapped_len;
> >   	uint32_t buf_offset = 0;
> >   	void *hpa;
> > +	struct batch_copy_elem *bce = vq->batch_copy_elems;
> >
> >   	while (cpy_len) {
> >   		hpa = (void *)(uintptr_t)gpa_to_first_hpa(dev,
> > @@ -886,6 +1026,31 @@ async_mbuf_to_desc_seg(struct virtio_net *dev,
> struct vhost_virtqueue *vq,
> >   						hpa, (size_t)mapped_len)))
> >   			return -1;
> >
> > +		/**
> > +		 * Keep VA for all IOVA segments for falling back to SW
> > +		 * copy in case of rte_dma_copy() error.
> > +		 */
> 
> As said below, I think we could get rid off the SW fallback.

Like the replies above, a better way for me is to remove SW fallback.

> But in case we didn't, I think it would be prefferable to change the
> rte_vhost_iovec struct to have both the iova and the VA, that would make
> the code simpler.
> 
> Also, while looking at this, I notice the structs rte_vhost_iov_iter and
> rte_vhost_iovec are still part of the Vhost API, but it should not be necessary
> now since application no more need to know about it.

Good catch, and I will change it later.

> 
> > +		if (unlikely(vq->batch_copy_nb_elems >= vq-
> >batch_copy_max_elems)) {
> > +			struct batch_copy_elem *tmp;
> > +			uint16_t nb_elems = 2 * vq->batch_copy_max_elems;
> > +
> > +			VHOST_LOG_DATA(DEBUG, "(%d) %s: run out of
> batch_copy_elems, "
> > +					"and realloc double elements.\n",
> dev->vid, __func__);
> > +			tmp = rte_realloc_socket(vq->batch_copy_elems,
> nb_elems * sizeof(*tmp),
> > +					RTE_CACHE_LINE_SIZE, vq-
> >numa_node);
> > +			if (!tmp) {
> > +				VHOST_LOG_DATA(ERR, "Failed to re-alloc
> batch_copy_elems\n");
> > +				return -1;
> > +			}
> > +
> > +			vq->batch_copy_max_elems = nb_elems;
> > +			vq->batch_copy_elems = tmp;
> > +			bce = tmp;
> > +		}
> > +		bce[vq->batch_copy_nb_elems].dst = (void
> *)((uintptr_t)(buf_addr + buf_offset));
> > +		bce[vq->batch_copy_nb_elems].src =
> rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
> > +		bce[vq->batch_copy_nb_elems++].len = mapped_len;
> > +
> >   		cpy_len -= (uint32_t)mapped_len;
> >   		mbuf_offset += (uint32_t)mapped_len;
> >   		buf_offset += (uint32_t)mapped_len; @@ -901,7 +1066,8
> @@
> > sync_mbuf_to_desc_seg(struct virtio_net *dev, struct vhost_virtqueue *vq,
> >   {
> >   	struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
> >
> > -	if (likely(cpy_len > MAX_BATCH_LEN || vq->batch_copy_nb_elems >=
> vq->size)) {
> > +	if (likely(cpy_len > MAX_BATCH_LEN ||
> > +				vq->batch_copy_nb_elems >= vq-
> >batch_copy_max_elems)) {
> >   		rte_memcpy((void *)((uintptr_t)(buf_addr)),
> >   				rte_pktmbuf_mtod_offset(m, void *,
> mbuf_offset),
> >   				cpy_len);
> > @@ -1020,8 +1186,10 @@ mbuf_to_desc(struct virtio_net *dev, struct
> > vhost_virtqueue *vq,
> >
> >   		if (is_async) {
> >   			if (async_mbuf_to_desc_seg(dev, vq, m, mbuf_offset,
> > +						buf_addr + buf_offset,
> >   						buf_iova + buf_offset,
> cpy_len) < 0)
> >   				goto error;
> > +
> 
> Remove new line.

I will remove it later.

Thanks,
Jiayu
diff mbox series

Patch

diff --git a/doc/guides/prog_guide/vhost_lib.rst b/doc/guides/prog_guide/vhost_lib.rst
index 76f5d303c9..acc10ea851 100644
--- a/doc/guides/prog_guide/vhost_lib.rst
+++ b/doc/guides/prog_guide/vhost_lib.rst
@@ -106,12 +106,11 @@  The following is an overview of some key Vhost API functions:
   - ``RTE_VHOST_USER_ASYNC_COPY``
 
     Asynchronous data path will be enabled when this flag is set. Async data
-    path allows applications to register async copy devices (typically
-    hardware DMA channels) to the vhost queues. Vhost leverages the copy
-    device registered to free CPU from memory copy operations. A set of
-    async data path APIs are defined for DPDK applications to make use of
-    the async capability. Only packets enqueued/dequeued by async APIs are
-    processed through the async data path.
+    path allows applications to register DMA channels to the vhost queues.
+    Vhost leverages the registered DMA devices to free CPU from memory copy
+    operations. A set of async data path APIs are defined for DPDK applications
+    to make use of the async capability. Only packets enqueued/dequeued by
+    async APIs are processed through the async data path.
 
     Currently this feature is only implemented on split ring enqueue data
     path.
@@ -218,52 +217,30 @@  The following is an overview of some key Vhost API functions:
 
   Enable or disable zero copy feature of the vhost crypto backend.
 
-* ``rte_vhost_async_channel_register(vid, queue_id, config, ops)``
+* ``rte_vhost_async_dma_configure(dmas_id, count, poll_factor)``
 
-  Register an async copy device channel for a vhost queue after vring
-  is enabled. Following device ``config`` must be specified together
-  with the registration:
+  Tell vhost what DMA devices are going to use. This function needs to
+  be called before register async data-path for vring.
 
-  * ``features``
+* ``rte_vhost_async_channel_register(vid, queue_id)``
 
-    This field is used to specify async copy device features.
+  Register async DMA acceleration for a vhost queue after vring is enabled.
 
-    ``RTE_VHOST_ASYNC_INORDER`` represents the async copy device can
-    guarantee the order of copy completion is the same as the order
-    of copy submission.
+* ``rte_vhost_async_channel_register_thread_unsafe(vid, queue_id)``
 
-    Currently, only ``RTE_VHOST_ASYNC_INORDER`` capable device is
-    supported by vhost.
-
-  Applications must provide following ``ops`` callbacks for vhost lib to
-  work with the async copy devices:
-
-  * ``transfer_data(vid, queue_id, descs, opaque_data, count)``
-
-    vhost invokes this function to submit copy data to the async devices.
-    For non-async_inorder capable devices, ``opaque_data`` could be used
-    for identifying the completed packets.
-
-  * ``check_completed_copies(vid, queue_id, opaque_data, max_packets)``
-
-    vhost invokes this function to get the copy data completed by async
-    devices.
-
-* ``rte_vhost_async_channel_register_thread_unsafe(vid, queue_id, config, ops)``
-
-  Register an async copy device channel for a vhost queue without
-  performing any locking.
+  Register async DMA acceleration for a vhost queue without performing
+  any locking.
 
   This function is only safe to call in vhost callback functions
   (i.e., struct rte_vhost_device_ops).
 
 * ``rte_vhost_async_channel_unregister(vid, queue_id)``
 
-  Unregister the async copy device channel from a vhost queue.
+  Unregister the async DMA acceleration from a vhost queue.
   Unregistration will fail, if the vhost queue has in-flight
   packets that are not completed.
 
-  Unregister async copy devices in vring_state_changed() may
+  Unregister async DMA acceleration in vring_state_changed() may
   fail, as this API tries to acquire the spinlock of vhost
   queue. The recommended way is to unregister async copy
   devices for all vhost queues in destroy_device(), when a
@@ -271,24 +248,19 @@  The following is an overview of some key Vhost API functions:
 
 * ``rte_vhost_async_channel_unregister_thread_unsafe(vid, queue_id)``
 
-  Unregister the async copy device channel for a vhost queue without
-  performing any locking.
+  Unregister async DMA acceleration for a vhost queue without performing
+  any locking.
 
   This function is only safe to call in vhost callback functions
   (i.e., struct rte_vhost_device_ops).
 
-* ``rte_vhost_submit_enqueue_burst(vid, queue_id, pkts, count, comp_pkts, comp_count)``
+* ``rte_vhost_submit_enqueue_burst(vid, queue_id, pkts, count, dma_id, vchan_id)``
 
   Submit an enqueue request to transmit ``count`` packets from host to guest
-  by async data path. Successfully enqueued packets can be transfer completed
-  or being occupied by DMA engines; transfer completed packets are returned in
-  ``comp_pkts``, but others are not guaranteed to finish, when this API
-  call returns.
+  by async data path. Applications must not free the packets submitted for
+  enqueue until the packets are completed.
 
-  Applications must not free the packets submitted for enqueue until the
-  packets are completed.
-
-* ``rte_vhost_poll_enqueue_completed(vid, queue_id, pkts, count)``
+* ``rte_vhost_poll_enqueue_completed(vid, queue_id, pkts, count, dma_id, vchan_id)``
 
   Poll enqueue completion status from async data path. Completed packets
   are returned to applications through ``pkts``.
@@ -298,7 +270,7 @@  The following is an overview of some key Vhost API functions:
   This function returns the amount of in-flight packets for the vhost
   queue using async acceleration.
 
-* ``rte_vhost_clear_queue_thread_unsafe(vid, queue_id, **pkts, count)``
+* ``rte_vhost_clear_queue_thread_unsafe(vid, queue_id, **pkts, count, dma_id, vchan_id)``
 
   Clear inflight packets which are submitted to DMA engine in vhost async data
   path. Completed packets are returned to applications through ``pkts``.
@@ -442,3 +414,26 @@  Finally, a set of device ops is defined for device specific operations:
 * ``get_notify_area``
 
   Called to get the notify area info of the queue.
+
+Vhost asynchronous data path
+----------------------------
+
+Vhost asynchronous data path leverages DMA devices to offload memory
+copies from the CPU and it is implemented in an asynchronous way. It
+enables applications, like OVS, to save CPU cycles and hide memory copy
+overhead, thus achieving higher throughput.
+
+Vhost doesn't manage DMA devices and applications, like OVS, need to
+manage and configure DMA devices. Applications need to tell vhost what
+DMA devices to use in every data path function call. This design enables
+the flexibility for applications to dynamically use DMA channels in
+different function modules, not limited in vhost.
+
+In addition, vhost supports M:N mapping between vrings and DMA virtual
+channels. Specifically, one vring can use multiple different DMA channels
+and one DMA channel can be shared by multiple vrings at the same time.
+The reason of enabling one vring to use multiple DMA channels is that
+it's possible that more than one dataplane threads enqueue packets to
+the same vring with their own DMA virtual channels. Besides, the number
+of DMA devices is limited. For the purpose of scaling, it's necessary to
+support sharing DMA channels among vrings.
diff --git a/examples/vhost/Makefile b/examples/vhost/Makefile
index 587ea2ab47..975a5dfe40 100644
--- a/examples/vhost/Makefile
+++ b/examples/vhost/Makefile
@@ -5,7 +5,7 @@ 
 APP = vhost-switch
 
 # all source are stored in SRCS-y
-SRCS-y := main.c virtio_net.c ioat.c
+SRCS-y := main.c virtio_net.c
 
 PKGCONF ?= pkg-config
 
diff --git a/examples/vhost/ioat.c b/examples/vhost/ioat.c
deleted file mode 100644
index 9aeeb12fd9..0000000000
--- a/examples/vhost/ioat.c
+++ /dev/null
@@ -1,218 +0,0 @@ 
-/* SPDX-License-Identifier: BSD-3-Clause
- * Copyright(c) 2010-2020 Intel Corporation
- */
-
-#include <sys/uio.h>
-#ifdef RTE_RAW_IOAT
-#include <rte_rawdev.h>
-#include <rte_ioat_rawdev.h>
-
-#include "ioat.h"
-#include "main.h"
-
-struct dma_for_vhost dma_bind[MAX_VHOST_DEVICE];
-
-struct packet_tracker {
-	unsigned short size_track[MAX_ENQUEUED_SIZE];
-	unsigned short next_read;
-	unsigned short next_write;
-	unsigned short last_remain;
-	unsigned short ioat_space;
-};
-
-struct packet_tracker cb_tracker[MAX_VHOST_DEVICE];
-
-int
-open_ioat(const char *value)
-{
-	struct dma_for_vhost *dma_info = dma_bind;
-	char *input = strndup(value, strlen(value) + 1);
-	char *addrs = input;
-	char *ptrs[2];
-	char *start, *end, *substr;
-	int64_t vid, vring_id;
-	struct rte_ioat_rawdev_config config;
-	struct rte_rawdev_info info = { .dev_private = &config };
-	char name[32];
-	int dev_id;
-	int ret = 0;
-	uint16_t i = 0;
-	char *dma_arg[MAX_VHOST_DEVICE];
-	int args_nr;
-
-	while (isblank(*addrs))
-		addrs++;
-	if (*addrs == '\0') {
-		ret = -1;
-		goto out;
-	}
-
-	/* process DMA devices within bracket. */
-	addrs++;
-	substr = strtok(addrs, ";]");
-	if (!substr) {
-		ret = -1;
-		goto out;
-	}
-	args_nr = rte_strsplit(substr, strlen(substr),
-			dma_arg, MAX_VHOST_DEVICE, ',');
-	if (args_nr <= 0) {
-		ret = -1;
-		goto out;
-	}
-	while (i < args_nr) {
-		char *arg_temp = dma_arg[i];
-		uint8_t sub_nr;
-		sub_nr = rte_strsplit(arg_temp, strlen(arg_temp), ptrs, 2, '@');
-		if (sub_nr != 2) {
-			ret = -1;
-			goto out;
-		}
-
-		start = strstr(ptrs[0], "txd");
-		if (start == NULL) {
-			ret = -1;
-			goto out;
-		}
-
-		start += 3;
-		vid = strtol(start, &end, 0);
-		if (end == start) {
-			ret = -1;
-			goto out;
-		}
-
-		vring_id = 0 + VIRTIO_RXQ;
-		if (rte_pci_addr_parse(ptrs[1],
-				&(dma_info + vid)->dmas[vring_id].addr) < 0) {
-			ret = -1;
-			goto out;
-		}
-
-		rte_pci_device_name(&(dma_info + vid)->dmas[vring_id].addr,
-				name, sizeof(name));
-		dev_id = rte_rawdev_get_dev_id(name);
-		if (dev_id == (uint16_t)(-ENODEV) ||
-		dev_id == (uint16_t)(-EINVAL)) {
-			ret = -1;
-			goto out;
-		}
-
-		if (rte_rawdev_info_get(dev_id, &info, sizeof(config)) < 0 ||
-		strstr(info.driver_name, "ioat") == NULL) {
-			ret = -1;
-			goto out;
-		}
-
-		(dma_info + vid)->dmas[vring_id].dev_id = dev_id;
-		(dma_info + vid)->dmas[vring_id].is_valid = true;
-		config.ring_size = IOAT_RING_SIZE;
-		config.hdls_disable = true;
-		if (rte_rawdev_configure(dev_id, &info, sizeof(config)) < 0) {
-			ret = -1;
-			goto out;
-		}
-		rte_rawdev_start(dev_id);
-		cb_tracker[dev_id].ioat_space = IOAT_RING_SIZE - 1;
-		dma_info->nr++;
-		i++;
-	}
-out:
-	free(input);
-	return ret;
-}
-
-int32_t
-ioat_transfer_data_cb(int vid, uint16_t queue_id,
-		struct rte_vhost_iov_iter *iov_iter,
-		struct rte_vhost_async_status *opaque_data, uint16_t count)
-{
-	uint32_t i_iter;
-	uint16_t dev_id = dma_bind[vid].dmas[queue_id * 2 + VIRTIO_RXQ].dev_id;
-	struct rte_vhost_iov_iter *iter = NULL;
-	unsigned long i_seg;
-	unsigned short mask = MAX_ENQUEUED_SIZE - 1;
-	unsigned short write = cb_tracker[dev_id].next_write;
-
-	if (!opaque_data) {
-		for (i_iter = 0; i_iter < count; i_iter++) {
-			iter = iov_iter + i_iter;
-			i_seg = 0;
-			if (cb_tracker[dev_id].ioat_space < iter->nr_segs)
-				break;
-			while (i_seg < iter->nr_segs) {
-				rte_ioat_enqueue_copy(dev_id,
-					(uintptr_t)(iter->iov[i_seg].src_addr),
-					(uintptr_t)(iter->iov[i_seg].dst_addr),
-					iter->iov[i_seg].len,
-					0,
-					0);
-				i_seg++;
-			}
-			write &= mask;
-			cb_tracker[dev_id].size_track[write] = iter->nr_segs;
-			cb_tracker[dev_id].ioat_space -= iter->nr_segs;
-			write++;
-		}
-	} else {
-		/* Opaque data is not supported */
-		return -1;
-	}
-	/* ring the doorbell */
-	rte_ioat_perform_ops(dev_id);
-	cb_tracker[dev_id].next_write = write;
-	return i_iter;
-}
-
-int32_t
-ioat_check_completed_copies_cb(int vid, uint16_t queue_id,
-		struct rte_vhost_async_status *opaque_data,
-		uint16_t max_packets)
-{
-	if (!opaque_data) {
-		uintptr_t dump[255];
-		int n_seg;
-		unsigned short read, write;
-		unsigned short nb_packet = 0;
-		unsigned short mask = MAX_ENQUEUED_SIZE - 1;
-		unsigned short i;
-
-		uint16_t dev_id = dma_bind[vid].dmas[queue_id * 2
-				+ VIRTIO_RXQ].dev_id;
-		n_seg = rte_ioat_completed_ops(dev_id, 255, NULL, NULL, dump, dump);
-		if (n_seg < 0) {
-			RTE_LOG(ERR,
-				VHOST_DATA,
-				"fail to poll completed buf on IOAT device %u",
-				dev_id);
-			return 0;
-		}
-		if (n_seg == 0)
-			return 0;
-
-		cb_tracker[dev_id].ioat_space += n_seg;
-		n_seg += cb_tracker[dev_id].last_remain;
-
-		read = cb_tracker[dev_id].next_read;
-		write = cb_tracker[dev_id].next_write;
-		for (i = 0; i < max_packets; i++) {
-			read &= mask;
-			if (read == write)
-				break;
-			if (n_seg >= cb_tracker[dev_id].size_track[read]) {
-				n_seg -= cb_tracker[dev_id].size_track[read];
-				read++;
-				nb_packet++;
-			} else {
-				break;
-			}
-		}
-		cb_tracker[dev_id].next_read = read;
-		cb_tracker[dev_id].last_remain = n_seg;
-		return nb_packet;
-	}
-	/* Opaque data is not supported */
-	return -1;
-}
-
-#endif /* RTE_RAW_IOAT */
diff --git a/examples/vhost/ioat.h b/examples/vhost/ioat.h
deleted file mode 100644
index d9bf717e8d..0000000000
--- a/examples/vhost/ioat.h
+++ /dev/null
@@ -1,63 +0,0 @@ 
-/* SPDX-License-Identifier: BSD-3-Clause
- * Copyright(c) 2010-2020 Intel Corporation
- */
-
-#ifndef _IOAT_H_
-#define _IOAT_H_
-
-#include <rte_vhost.h>
-#include <rte_pci.h>
-#include <rte_vhost_async.h>
-
-#define MAX_VHOST_DEVICE 1024
-#define IOAT_RING_SIZE 4096
-#define MAX_ENQUEUED_SIZE 4096
-
-struct dma_info {
-	struct rte_pci_addr addr;
-	uint16_t dev_id;
-	bool is_valid;
-};
-
-struct dma_for_vhost {
-	struct dma_info dmas[RTE_MAX_QUEUES_PER_PORT * 2];
-	uint16_t nr;
-};
-
-#ifdef RTE_RAW_IOAT
-int open_ioat(const char *value);
-
-int32_t
-ioat_transfer_data_cb(int vid, uint16_t queue_id,
-		struct rte_vhost_iov_iter *iov_iter,
-		struct rte_vhost_async_status *opaque_data, uint16_t count);
-
-int32_t
-ioat_check_completed_copies_cb(int vid, uint16_t queue_id,
-		struct rte_vhost_async_status *opaque_data,
-		uint16_t max_packets);
-#else
-static int open_ioat(const char *value __rte_unused)
-{
-	return -1;
-}
-
-static int32_t
-ioat_transfer_data_cb(int vid __rte_unused, uint16_t queue_id __rte_unused,
-		struct rte_vhost_iov_iter *iov_iter __rte_unused,
-		struct rte_vhost_async_status *opaque_data __rte_unused,
-		uint16_t count __rte_unused)
-{
-	return -1;
-}
-
-static int32_t
-ioat_check_completed_copies_cb(int vid __rte_unused,
-		uint16_t queue_id __rte_unused,
-		struct rte_vhost_async_status *opaque_data __rte_unused,
-		uint16_t max_packets __rte_unused)
-{
-	return -1;
-}
-#endif
-#endif /* _IOAT_H_ */
diff --git a/examples/vhost/main.c b/examples/vhost/main.c
index 590a77c723..b2c272059e 100644
--- a/examples/vhost/main.c
+++ b/examples/vhost/main.c
@@ -24,8 +24,9 @@ 
 #include <rte_ip.h>
 #include <rte_tcp.h>
 #include <rte_pause.h>
+#include <rte_dmadev.h>
+#include <rte_vhost_async.h>
 
-#include "ioat.h"
 #include "main.h"
 
 #ifndef MAX_QUEUES
@@ -56,6 +57,13 @@ 
 #define RTE_TEST_TX_DESC_DEFAULT 512
 
 #define INVALID_PORT_ID 0xFF
+#define INVALID_DMA_ID -1
+
+#define DMA_RING_SIZE 4096
+
+struct dma_for_vhost dma_bind[RTE_MAX_VHOST_DEVICE];
+int16_t dmas_id[RTE_DMADEV_DEFAULT_MAX];
+static int dma_count;
 
 /* mask of enabled ports */
 static uint32_t enabled_port_mask = 0;
@@ -94,10 +102,6 @@  static int client_mode;
 
 static int builtin_net_driver;
 
-static int async_vhost_driver;
-
-static char *dma_type;
-
 /* Specify timeout (in useconds) between retries on RX. */
 static uint32_t burst_rx_delay_time = BURST_RX_WAIT_US;
 /* Specify the number of retries on RX. */
@@ -191,18 +195,150 @@  struct mbuf_table lcore_tx_queue[RTE_MAX_LCORE];
  * Every data core maintains a TX buffer for every vhost device,
  * which is used for batch pkts enqueue for higher performance.
  */
-struct vhost_bufftable *vhost_txbuff[RTE_MAX_LCORE * MAX_VHOST_DEVICE];
+struct vhost_bufftable *vhost_txbuff[RTE_MAX_LCORE * RTE_MAX_VHOST_DEVICE];
 
 #define MBUF_TABLE_DRAIN_TSC	((rte_get_tsc_hz() + US_PER_S - 1) \
 				 / US_PER_S * BURST_TX_DRAIN_US)
 
+static inline bool
+is_dma_configured(int16_t dev_id)
+{
+	int i;
+
+	for (i = 0; i < dma_count; i++)
+		if (dmas_id[i] == dev_id)
+			return true;
+	return false;
+}
+
 static inline int
 open_dma(const char *value)
 {
-	if (dma_type != NULL && strncmp(dma_type, "ioat", 4) == 0)
-		return open_ioat(value);
+	struct dma_for_vhost *dma_info = dma_bind;
+	char *input = strndup(value, strlen(value) + 1);
+	char *addrs = input;
+	char *ptrs[2];
+	char *start, *end, *substr;
+	int64_t vid;
+
+	struct rte_dma_info info;
+	struct rte_dma_conf dev_config = { .nb_vchans = 1 };
+	struct rte_dma_vchan_conf qconf = {
+		.direction = RTE_DMA_DIR_MEM_TO_MEM,
+		.nb_desc = DMA_RING_SIZE
+	};
+
+	int dev_id;
+	int ret = 0;
+	uint16_t i = 0;
+	char *dma_arg[RTE_MAX_VHOST_DEVICE];
+	int args_nr;
+
+	while (isblank(*addrs))
+		addrs++;
+	if (*addrs == '\0') {
+		ret = -1;
+		goto out;
+	}
+
+	/* process DMA devices within bracket. */
+	addrs++;
+	substr = strtok(addrs, ";]");
+	if (!substr) {
+		ret = -1;
+		goto out;
+	}
+
+	args_nr = rte_strsplit(substr, strlen(substr), dma_arg, RTE_MAX_VHOST_DEVICE, ',');
+	if (args_nr <= 0) {
+		ret = -1;
+		goto out;
+	}
+
+	while (i < args_nr) {
+		char *arg_temp = dma_arg[i];
+		uint8_t sub_nr;
+
+		sub_nr = rte_strsplit(arg_temp, strlen(arg_temp), ptrs, 2, '@');
+		if (sub_nr != 2) {
+			ret = -1;
+			goto out;
+		}
+
+		start = strstr(ptrs[0], "txd");
+		if (start == NULL) {
+			ret = -1;
+			goto out;
+		}
+
+		start += 3;
+		vid = strtol(start, &end, 0);
+		if (end == start) {
+			ret = -1;
+			goto out;
+		}
+
+		dev_id = rte_dma_get_dev_id_by_name(ptrs[1]);
+		if (dev_id < 0) {
+			RTE_LOG(ERR, VHOST_CONFIG, "Fail to find DMA %s.\n", ptrs[1]);
+			ret = -1;
+			goto out;
+		}
+
+		/* DMA device is already configured, so skip */
+		if (is_dma_configured(dev_id))
+			goto done;
+
+		if (rte_dma_info_get(dev_id, &info) != 0) {
+			RTE_LOG(ERR, VHOST_CONFIG, "Error with rte_dma_info_get()\n");
+			ret = -1;
+			goto out;
+		}
+
+		if (info.max_vchans < 1) {
+			RTE_LOG(ERR, VHOST_CONFIG, "No channels available on device %d\n", dev_id);
+			ret = -1;
+			goto out;
+		}
 
-	return -1;
+		if (rte_dma_configure(dev_id, &dev_config) != 0) {
+			RTE_LOG(ERR, VHOST_CONFIG, "Fail to configure DMA %d.\n", dev_id);
+			ret = -1;
+			goto out;
+		}
+
+		/* Check the max desc supported by DMA device */
+		rte_dma_info_get(dev_id, &info);
+		if (info.nb_vchans != 1) {
+			RTE_LOG(ERR, VHOST_CONFIG, "No configured queues reported by DMA %d.\n",
+					dev_id);
+			ret = -1;
+			goto out;
+		}
+
+		qconf.nb_desc = RTE_MIN(DMA_RING_SIZE, info.max_desc);
+
+		if (rte_dma_vchan_setup(dev_id, 0, &qconf) != 0) {
+			RTE_LOG(ERR, VHOST_CONFIG, "Fail to set up DMA %d.\n", dev_id);
+			ret = -1;
+			goto out;
+		}
+
+		if (rte_dma_start(dev_id) != 0) {
+			RTE_LOG(ERR, VHOST_CONFIG, "Fail to start DMA %u.\n", dev_id);
+			ret = -1;
+			goto out;
+		}
+
+		dmas_id[dma_count++] = dev_id;
+
+done:
+		(dma_info + vid)->dmas[VIRTIO_RXQ].dev_id = dev_id;
+		i++;
+	}
+out:
+	free(input);
+	return ret;
 }
 
 /*
@@ -500,8 +636,6 @@  enum {
 	OPT_CLIENT_NUM,
 #define OPT_BUILTIN_NET_DRIVER  "builtin-net-driver"
 	OPT_BUILTIN_NET_DRIVER_NUM,
-#define OPT_DMA_TYPE            "dma-type"
-	OPT_DMA_TYPE_NUM,
 #define OPT_DMAS                "dmas"
 	OPT_DMAS_NUM,
 };
@@ -539,8 +673,6 @@  us_vhost_parse_args(int argc, char **argv)
 				NULL, OPT_CLIENT_NUM},
 		{OPT_BUILTIN_NET_DRIVER, no_argument,
 				NULL, OPT_BUILTIN_NET_DRIVER_NUM},
-		{OPT_DMA_TYPE, required_argument,
-				NULL, OPT_DMA_TYPE_NUM},
 		{OPT_DMAS, required_argument,
 				NULL, OPT_DMAS_NUM},
 		{NULL, 0, 0, 0},
@@ -661,10 +793,6 @@  us_vhost_parse_args(int argc, char **argv)
 			}
 			break;
 
-		case OPT_DMA_TYPE_NUM:
-			dma_type = optarg;
-			break;
-
 		case OPT_DMAS_NUM:
 			if (open_dma(optarg) == -1) {
 				RTE_LOG(INFO, VHOST_CONFIG,
@@ -672,7 +800,6 @@  us_vhost_parse_args(int argc, char **argv)
 				us_vhost_usage(prgname);
 				return -1;
 			}
-			async_vhost_driver = 1;
 			break;
 
 		case OPT_CLIENT_NUM:
@@ -841,9 +968,10 @@  complete_async_pkts(struct vhost_dev *vdev)
 {
 	struct rte_mbuf *p_cpl[MAX_PKT_BURST];
 	uint16_t complete_count;
+	int16_t dma_id = dma_bind[vdev->vid].dmas[VIRTIO_RXQ].dev_id;
 
 	complete_count = rte_vhost_poll_enqueue_completed(vdev->vid,
-					VIRTIO_RXQ, p_cpl, MAX_PKT_BURST);
+					VIRTIO_RXQ, p_cpl, MAX_PKT_BURST, dma_id, 0);
 	if (complete_count) {
 		free_pkts(p_cpl, complete_count);
 		__atomic_sub_fetch(&vdev->pkts_inflight, complete_count, __ATOMIC_SEQ_CST);
@@ -877,17 +1005,18 @@  static __rte_always_inline void
 drain_vhost(struct vhost_dev *vdev)
 {
 	uint16_t ret;
-	uint32_t buff_idx = rte_lcore_id() * MAX_VHOST_DEVICE + vdev->vid;
+	uint32_t buff_idx = rte_lcore_id() * RTE_MAX_VHOST_DEVICE + vdev->vid;
 	uint16_t nr_xmit = vhost_txbuff[buff_idx]->len;
 	struct rte_mbuf **m = vhost_txbuff[buff_idx]->m_table;
 
 	if (builtin_net_driver) {
 		ret = vs_enqueue_pkts(vdev, VIRTIO_RXQ, m, nr_xmit);
-	} else if (async_vhost_driver) {
+	} else if (dma_bind[vdev->vid].dmas[VIRTIO_RXQ].async_enabled) {
 		uint16_t enqueue_fail = 0;
+		int16_t dma_id = dma_bind[vdev->vid].dmas[VIRTIO_RXQ].dev_id;
 
 		complete_async_pkts(vdev);
-		ret = rte_vhost_submit_enqueue_burst(vdev->vid, VIRTIO_RXQ, m, nr_xmit);
+		ret = rte_vhost_submit_enqueue_burst(vdev->vid, VIRTIO_RXQ, m, nr_xmit, dma_id, 0);
 		__atomic_add_fetch(&vdev->pkts_inflight, ret, __ATOMIC_SEQ_CST);
 
 		enqueue_fail = nr_xmit - ret;
@@ -905,7 +1034,7 @@  drain_vhost(struct vhost_dev *vdev)
 				__ATOMIC_SEQ_CST);
 	}
 
-	if (!async_vhost_driver)
+	if (!dma_bind[vdev->vid].dmas[VIRTIO_RXQ].async_enabled)
 		free_pkts(m, nr_xmit);
 }
 
@@ -921,7 +1050,7 @@  drain_vhost_table(void)
 		if (unlikely(vdev->remove == 1))
 			continue;
 
-		vhost_txq = vhost_txbuff[lcore_id * MAX_VHOST_DEVICE
+		vhost_txq = vhost_txbuff[lcore_id * RTE_MAX_VHOST_DEVICE
 						+ vdev->vid];
 
 		cur_tsc = rte_rdtsc();
@@ -970,7 +1099,7 @@  virtio_tx_local(struct vhost_dev *vdev, struct rte_mbuf *m)
 		return 0;
 	}
 
-	vhost_txq = vhost_txbuff[lcore_id * MAX_VHOST_DEVICE + dst_vdev->vid];
+	vhost_txq = vhost_txbuff[lcore_id * RTE_MAX_VHOST_DEVICE + dst_vdev->vid];
 	vhost_txq->m_table[vhost_txq->len++] = m;
 
 	if (enable_stats) {
@@ -1211,12 +1340,13 @@  drain_eth_rx(struct vhost_dev *vdev)
 	if (builtin_net_driver) {
 		enqueue_count = vs_enqueue_pkts(vdev, VIRTIO_RXQ,
 						pkts, rx_count);
-	} else if (async_vhost_driver) {
+	} else if (dma_bind[vdev->vid].dmas[VIRTIO_RXQ].async_enabled) {
 		uint16_t enqueue_fail = 0;
+		int16_t dma_id = dma_bind[vdev->vid].dmas[VIRTIO_RXQ].dev_id;
 
 		complete_async_pkts(vdev);
 		enqueue_count = rte_vhost_submit_enqueue_burst(vdev->vid,
-					VIRTIO_RXQ, pkts, rx_count);
+					VIRTIO_RXQ, pkts, rx_count, dma_id, 0);
 		__atomic_add_fetch(&vdev->pkts_inflight, enqueue_count, __ATOMIC_SEQ_CST);
 
 		enqueue_fail = rx_count - enqueue_count;
@@ -1235,7 +1365,7 @@  drain_eth_rx(struct vhost_dev *vdev)
 				__ATOMIC_SEQ_CST);
 	}
 
-	if (!async_vhost_driver)
+	if (!dma_bind[vdev->vid].dmas[VIRTIO_RXQ].async_enabled)
 		free_pkts(pkts, rx_count);
 }
 
@@ -1357,7 +1487,7 @@  destroy_device(int vid)
 	}
 
 	for (i = 0; i < RTE_MAX_LCORE; i++)
-		rte_free(vhost_txbuff[i * MAX_VHOST_DEVICE + vid]);
+		rte_free(vhost_txbuff[i * RTE_MAX_VHOST_DEVICE + vid]);
 
 	if (builtin_net_driver)
 		vs_vhost_net_remove(vdev);
@@ -1387,18 +1517,20 @@  destroy_device(int vid)
 		"(%d) device has been removed from data core\n",
 		vdev->vid);
 
-	if (async_vhost_driver) {
+	if (dma_bind[vid].dmas[VIRTIO_RXQ].async_enabled) {
 		uint16_t n_pkt = 0;
+		int16_t dma_id = dma_bind[vid].dmas[VIRTIO_RXQ].dev_id;
 		struct rte_mbuf *m_cpl[vdev->pkts_inflight];
 
 		while (vdev->pkts_inflight) {
 			n_pkt = rte_vhost_clear_queue_thread_unsafe(vid, VIRTIO_RXQ,
-						m_cpl, vdev->pkts_inflight);
+						m_cpl, vdev->pkts_inflight, dma_id, 0);
 			free_pkts(m_cpl, n_pkt);
 			__atomic_sub_fetch(&vdev->pkts_inflight, n_pkt, __ATOMIC_SEQ_CST);
 		}
 
 		rte_vhost_async_channel_unregister(vid, VIRTIO_RXQ);
+		dma_bind[vid].dmas[VIRTIO_RXQ].async_enabled = false;
 	}
 
 	rte_free(vdev);
@@ -1425,12 +1557,12 @@  new_device(int vid)
 	vdev->vid = vid;
 
 	for (i = 0; i < RTE_MAX_LCORE; i++) {
-		vhost_txbuff[i * MAX_VHOST_DEVICE + vid]
+		vhost_txbuff[i * RTE_MAX_VHOST_DEVICE + vid]
 			= rte_zmalloc("vhost bufftable",
 				sizeof(struct vhost_bufftable),
 				RTE_CACHE_LINE_SIZE);
 
-		if (vhost_txbuff[i * MAX_VHOST_DEVICE + vid] == NULL) {
+		if (vhost_txbuff[i * RTE_MAX_VHOST_DEVICE + vid] == NULL) {
 			RTE_LOG(INFO, VHOST_DATA,
 			  "(%d) couldn't allocate memory for vhost TX\n", vid);
 			return -1;
@@ -1468,20 +1600,13 @@  new_device(int vid)
 		"(%d) device has been added to data core %d\n",
 		vid, vdev->coreid);
 
-	if (async_vhost_driver) {
-		struct rte_vhost_async_config config = {0};
-		struct rte_vhost_async_channel_ops channel_ops;
-
-		if (dma_type != NULL && strncmp(dma_type, "ioat", 4) == 0) {
-			channel_ops.transfer_data = ioat_transfer_data_cb;
-			channel_ops.check_completed_copies =
-				ioat_check_completed_copies_cb;
-
-			config.features = RTE_VHOST_ASYNC_INORDER;
+	if (dma_bind[vid].dmas[VIRTIO_RXQ].dev_id != INVALID_DMA_ID) {
+		int ret;
 
-			return rte_vhost_async_channel_register(vid, VIRTIO_RXQ,
-				config, &channel_ops);
-		}
+		ret = rte_vhost_async_channel_register(vid, VIRTIO_RXQ);
+		if (ret == 0)
+			dma_bind[vid].dmas[VIRTIO_RXQ].async_enabled = true;
+		return ret;
 	}
 
 	return 0;
@@ -1502,14 +1627,15 @@  vring_state_changed(int vid, uint16_t queue_id, int enable)
 	if (queue_id != VIRTIO_RXQ)
 		return 0;
 
-	if (async_vhost_driver) {
+	if (dma_bind[vid].dmas[queue_id].async_enabled) {
 		if (!enable) {
 			uint16_t n_pkt = 0;
+			int16_t dma_id = dma_bind[vid].dmas[VIRTIO_RXQ].dev_id;
 			struct rte_mbuf *m_cpl[vdev->pkts_inflight];
 
 			while (vdev->pkts_inflight) {
 				n_pkt = rte_vhost_clear_queue_thread_unsafe(vid, queue_id,
-							m_cpl, vdev->pkts_inflight);
+							m_cpl, vdev->pkts_inflight, dma_id, 0);
 				free_pkts(m_cpl, n_pkt);
 				__atomic_sub_fetch(&vdev->pkts_inflight, n_pkt, __ATOMIC_SEQ_CST);
 			}
@@ -1657,6 +1783,24 @@  create_mbuf_pool(uint16_t nr_port, uint32_t nr_switch_core, uint32_t mbuf_size,
 		rte_exit(EXIT_FAILURE, "Cannot create mbuf pool\n");
 }
 
+static void
+reset_dma(void)
+{
+	int i;
+
+	for (i = 0; i < RTE_MAX_VHOST_DEVICE; i++) {
+		int j;
+
+		for (j = 0; j < RTE_MAX_QUEUES_PER_PORT * 2; j++) {
+			dma_bind[i].dmas[j].dev_id = INVALID_DMA_ID;
+			dma_bind[i].dmas[j].async_enabled = false;
+		}
+	}
+
+	for (i = 0; i < RTE_DMADEV_DEFAULT_MAX; i++)
+		dmas_id[i] = INVALID_DMA_ID;
+}
+
 /*
  * Main function, does initialisation and calls the per-lcore functions.
  */
@@ -1679,6 +1823,9 @@  main(int argc, char *argv[])
 	argc -= ret;
 	argv += ret;
 
+	/* initialize dma structures */
+	reset_dma();
+
 	/* parse app arguments */
 	ret = us_vhost_parse_args(argc, argv);
 	if (ret < 0)
@@ -1754,11 +1901,21 @@  main(int argc, char *argv[])
 	if (client_mode)
 		flags |= RTE_VHOST_USER_CLIENT;
 
+	if (dma_count) {
+		if (rte_vhost_async_dma_configure(dmas_id, dma_count, 1) < 0) {
+			RTE_LOG(ERR, VHOST_PORT, "Failed to configure DMA in vhost.\n");
+			for (i = 0; i < dma_count; i++)
+				if (dmas_id[i] >= 0)
+					rte_dma_stop(dmas_id[i]);
+			rte_exit(EXIT_FAILURE, "Cannot use given DMA devices\n");
+		}
+	}
+
 	/* Register vhost user driver to handle vhost messages. */
 	for (i = 0; i < nb_sockets; i++) {
 		char *file = socket_files + i * PATH_MAX;
 
-		if (async_vhost_driver)
+		if (dma_count)
 			flags = flags | RTE_VHOST_USER_ASYNC_COPY;
 
 		ret = rte_vhost_driver_register(file, flags);
diff --git a/examples/vhost/main.h b/examples/vhost/main.h
index e7b1ac60a6..b4a453e77e 100644
--- a/examples/vhost/main.h
+++ b/examples/vhost/main.h
@@ -8,6 +8,7 @@ 
 #include <sys/queue.h>
 
 #include <rte_ether.h>
+#include <rte_pci.h>
 
 /* Macros for printing using RTE_LOG */
 #define RTE_LOGTYPE_VHOST_CONFIG RTE_LOGTYPE_USER1
@@ -79,6 +80,16 @@  struct lcore_info {
 	struct vhost_dev_tailq_list vdev_list;
 };
 
+struct dma_info {
+	struct rte_pci_addr addr;
+	int16_t dev_id;
+	bool async_enabled;
+};
+
+struct dma_for_vhost {
+	struct dma_info dmas[RTE_MAX_QUEUES_PER_PORT * 2];
+};
+
 /* we implement non-extra virtio net features */
 #define VIRTIO_NET_FEATURES	0
 
diff --git a/examples/vhost/meson.build b/examples/vhost/meson.build
index 3efd5e6540..87a637f83f 100644
--- a/examples/vhost/meson.build
+++ b/examples/vhost/meson.build
@@ -12,13 +12,9 @@  if not is_linux
 endif
 
 deps += 'vhost'
+deps += 'dmadev'
 allow_experimental_apis = true
 sources = files(
         'main.c',
         'virtio_net.c',
 )
-
-if dpdk_conf.has('RTE_RAW_IOAT')
-    deps += 'raw_ioat'
-    sources += files('ioat.c')
-endif
diff --git a/lib/vhost/meson.build b/lib/vhost/meson.build
index cdb37a4814..bc7272053b 100644
--- a/lib/vhost/meson.build
+++ b/lib/vhost/meson.build
@@ -36,4 +36,4 @@  headers = files(
 driver_sdk_headers = files(
         'vdpa_driver.h',
 )
-deps += ['ethdev', 'cryptodev', 'hash', 'pci']
+deps += ['ethdev', 'cryptodev', 'hash', 'pci', 'dmadev']
diff --git a/lib/vhost/rte_vhost.h b/lib/vhost/rte_vhost.h
index b454c05868..15c37dd26e 100644
--- a/lib/vhost/rte_vhost.h
+++ b/lib/vhost/rte_vhost.h
@@ -113,6 +113,8 @@  extern "C" {
 #define VHOST_USER_F_PROTOCOL_FEATURES	30
 #endif
 
+#define RTE_MAX_VHOST_DEVICE	1024
+
 struct rte_vdpa_device;
 
 /**
diff --git a/lib/vhost/rte_vhost_async.h b/lib/vhost/rte_vhost_async.h
index a87ea6ba37..758a80f403 100644
--- a/lib/vhost/rte_vhost_async.h
+++ b/lib/vhost/rte_vhost_async.h
@@ -26,73 +26,6 @@  struct rte_vhost_iov_iter {
 	unsigned long nr_segs;
 };
 
-/**
- * dma transfer status
- */
-struct rte_vhost_async_status {
-	/** An array of application specific data for source memory */
-	uintptr_t *src_opaque_data;
-	/** An array of application specific data for destination memory */
-	uintptr_t *dst_opaque_data;
-};
-
-/**
- * dma operation callbacks to be implemented by applications
- */
-struct rte_vhost_async_channel_ops {
-	/**
-	 * instruct async engines to perform copies for a batch of packets
-	 *
-	 * @param vid
-	 *  id of vhost device to perform data copies
-	 * @param queue_id
-	 *  queue id to perform data copies
-	 * @param iov_iter
-	 *  an array of IOV iterators
-	 * @param opaque_data
-	 *  opaque data pair sending to DMA engine
-	 * @param count
-	 *  number of elements in the "descs" array
-	 * @return
-	 *  number of IOV iterators processed, negative value means error
-	 */
-	int32_t (*transfer_data)(int vid, uint16_t queue_id,
-		struct rte_vhost_iov_iter *iov_iter,
-		struct rte_vhost_async_status *opaque_data,
-		uint16_t count);
-	/**
-	 * check copy-completed packets from the async engine
-	 * @param vid
-	 *  id of vhost device to check copy completion
-	 * @param queue_id
-	 *  queue id to check copy completion
-	 * @param opaque_data
-	 *  buffer to receive the opaque data pair from DMA engine
-	 * @param max_packets
-	 *  max number of packets could be completed
-	 * @return
-	 *  number of async descs completed, negative value means error
-	 */
-	int32_t (*check_completed_copies)(int vid, uint16_t queue_id,
-		struct rte_vhost_async_status *opaque_data,
-		uint16_t max_packets);
-};
-
-/**
- *  async channel features
- */
-enum {
-	RTE_VHOST_ASYNC_INORDER = 1U << 0,
-};
-
-/**
- *  async channel configuration
- */
-struct rte_vhost_async_config {
-	uint32_t features;
-	uint32_t rsvd[2];
-};
-
 /**
  * Register an async channel for a vhost queue
  *
@@ -100,17 +33,11 @@  struct rte_vhost_async_config {
  *  vhost device id async channel to be attached to
  * @param queue_id
  *  vhost queue id async channel to be attached to
- * @param config
- *  Async channel configuration structure
- * @param ops
- *  Async channel operation callbacks
  * @return
  *  0 on success, -1 on failures
  */
 __rte_experimental
-int rte_vhost_async_channel_register(int vid, uint16_t queue_id,
-	struct rte_vhost_async_config config,
-	struct rte_vhost_async_channel_ops *ops);
+int rte_vhost_async_channel_register(int vid, uint16_t queue_id);
 
 /**
  * Unregister an async channel for a vhost queue
@@ -136,17 +63,11 @@  int rte_vhost_async_channel_unregister(int vid, uint16_t queue_id);
  *  vhost device id async channel to be attached to
  * @param queue_id
  *  vhost queue id async channel to be attached to
- * @param config
- *  Async channel configuration
- * @param ops
- *  Async channel operation callbacks
  * @return
  *  0 on success, -1 on failures
  */
 __rte_experimental
-int rte_vhost_async_channel_register_thread_unsafe(int vid, uint16_t queue_id,
-	struct rte_vhost_async_config config,
-	struct rte_vhost_async_channel_ops *ops);
+int rte_vhost_async_channel_register_thread_unsafe(int vid, uint16_t queue_id);
 
 /**
  * Unregister an async channel for a vhost queue without performing any
@@ -179,12 +100,17 @@  int rte_vhost_async_channel_unregister_thread_unsafe(int vid,
  *  array of packets to be enqueued
  * @param count
  *  packets num to be enqueued
+ * @param dma_id
+ *  the identifier of the DMA device
+ * @param vchan_id
+ *  the identifier of virtual DMA channel
  * @return
  *  num of packets enqueued
  */
 __rte_experimental
 uint16_t rte_vhost_submit_enqueue_burst(int vid, uint16_t queue_id,
-		struct rte_mbuf **pkts, uint16_t count);
+		struct rte_mbuf **pkts, uint16_t count, int16_t dma_id,
+		uint16_t vchan_id);
 
 /**
  * This function checks async completion status for a specific vhost
@@ -199,12 +125,17 @@  uint16_t rte_vhost_submit_enqueue_burst(int vid, uint16_t queue_id,
  *  blank array to get return packet pointer
  * @param count
  *  size of the packet array
+ * @param dma_id
+ *  the identifier of the DMA device
+ * @param vchan_id
+ *  the identifier of virtual DMA channel
  * @return
  *  num of packets returned
  */
 __rte_experimental
 uint16_t rte_vhost_poll_enqueue_completed(int vid, uint16_t queue_id,
-		struct rte_mbuf **pkts, uint16_t count);
+		struct rte_mbuf **pkts, uint16_t count, int16_t dma_id,
+		uint16_t vchan_id);
 
 /**
  * This function returns the amount of in-flight packets for the vhost
@@ -235,11 +166,44 @@  int rte_vhost_async_get_inflight(int vid, uint16_t queue_id);
  *  Blank array to get return packet pointer
  * @param count
  *  Size of the packet array
+ * @param dma_id
+ *  the identifier of the DMA device
+ * @param vchan_id
+ *  the identifier of virtual DMA channel
  * @return
  *  Number of packets returned
  */
 __rte_experimental
 uint16_t rte_vhost_clear_queue_thread_unsafe(int vid, uint16_t queue_id,
-		struct rte_mbuf **pkts, uint16_t count);
+		struct rte_mbuf **pkts, uint16_t count, int16_t dma_id,
+		uint16_t vchan_id);
+/**
+ * The DMA vChannels used in asynchronous data path must be configured
+ * first. So this function needs to be called before enabling DMA
+ * acceleration for vring. If this function fails, asynchronous data path
+ * cannot be enabled for any vring further.
+ *
+ * DMA devices used in data-path must belong to DMA devices given in this
+ * function. But users are free to use DMA devices given in the function
+ * in non-vhost scenarios, only if guarantee no copies in vhost are
+ * offloaded to them at the same time.
+ *
+ * @param dmas_id
+ *  DMA ID array
+ * @param count
+ *  Element number of 'dmas_id'
+ * @param poll_factor
+ *  For large or scatter-gather packets, one packet would consist of
+ *  small buffers. In this case, vhost will issue several DMA copy
+ *  operations for the packet. Therefore, the number of copies to
+ *  check by rte_dma_completed() is calculated by "nb_pkts_to_poll *
+ *  poll_factor" andused in rte_vhost_poll_enqueue_completed(). The
+ *  default value of "poll_factor" is 1.
+ * @return
+ *  0 on success, and -1 on failure
+ */
+__rte_experimental
+int rte_vhost_async_dma_configure(int16_t *dmas_id, uint16_t count,
+		uint16_t poll_factor);
 
 #endif /* _RTE_VHOST_ASYNC_H_ */
diff --git a/lib/vhost/version.map b/lib/vhost/version.map
index a7ef7f1976..1202ba9c1a 100644
--- a/lib/vhost/version.map
+++ b/lib/vhost/version.map
@@ -84,6 +84,9 @@  EXPERIMENTAL {
 
 	# added in 21.11
 	rte_vhost_get_monitor_addr;
+
+	# added in 22.03
+	rte_vhost_async_dma_configure;
 };
 
 INTERNAL {
diff --git a/lib/vhost/vhost.c b/lib/vhost/vhost.c
index 13a9bb9dd1..c408cee63e 100644
--- a/lib/vhost/vhost.c
+++ b/lib/vhost/vhost.c
@@ -25,7 +25,7 @@ 
 #include "vhost.h"
 #include "vhost_user.h"
 
-struct virtio_net *vhost_devices[MAX_VHOST_DEVICE];
+struct virtio_net *vhost_devices[RTE_MAX_VHOST_DEVICE];
 pthread_mutex_t vhost_dev_lock = PTHREAD_MUTEX_INITIALIZER;
 
 /* Called with iotlb_lock read-locked */
@@ -344,6 +344,7 @@  vhost_free_async_mem(struct vhost_virtqueue *vq)
 		return;
 
 	rte_free(vq->async->pkts_info);
+	rte_free(vq->async->pkts_cmpl_flag);
 
 	rte_free(vq->async->buffers_packed);
 	vq->async->buffers_packed = NULL;
@@ -667,12 +668,12 @@  vhost_new_device(void)
 	int i;
 
 	pthread_mutex_lock(&vhost_dev_lock);
-	for (i = 0; i < MAX_VHOST_DEVICE; i++) {
+	for (i = 0; i < RTE_MAX_VHOST_DEVICE; i++) {
 		if (vhost_devices[i] == NULL)
 			break;
 	}
 
-	if (i == MAX_VHOST_DEVICE) {
+	if (i == RTE_MAX_VHOST_DEVICE) {
 		VHOST_LOG_CONFIG(ERR,
 			"Failed to find a free slot for new device.\n");
 		pthread_mutex_unlock(&vhost_dev_lock);
@@ -1626,8 +1627,7 @@  rte_vhost_extern_callback_register(int vid,
 }
 
 static __rte_always_inline int
-async_channel_register(int vid, uint16_t queue_id,
-		struct rte_vhost_async_channel_ops *ops)
+async_channel_register(int vid, uint16_t queue_id)
 {
 	struct virtio_net *dev = get_device(vid);
 	struct vhost_virtqueue *vq = dev->virtqueue[queue_id];
@@ -1656,6 +1656,14 @@  async_channel_register(int vid, uint16_t queue_id,
 		goto out_free_async;
 	}
 
+	async->pkts_cmpl_flag = rte_zmalloc_socket(NULL, vq->size * sizeof(bool),
+			RTE_CACHE_LINE_SIZE, node);
+	if (!async->pkts_cmpl_flag) {
+		VHOST_LOG_CONFIG(ERR, "failed to allocate async pkts_cmpl_flag (vid %d, qid: %d)\n",
+				vid, queue_id);
+		goto out_free_async;
+	}
+
 	if (vq_is_packed(dev)) {
 		async->buffers_packed = rte_malloc_socket(NULL,
 				vq->size * sizeof(struct vring_used_elem_packed),
@@ -1676,9 +1684,6 @@  async_channel_register(int vid, uint16_t queue_id,
 		}
 	}
 
-	async->ops.check_completed_copies = ops->check_completed_copies;
-	async->ops.transfer_data = ops->transfer_data;
-
 	vq->async = async;
 
 	return 0;
@@ -1691,15 +1696,13 @@  async_channel_register(int vid, uint16_t queue_id,
 }
 
 int
-rte_vhost_async_channel_register(int vid, uint16_t queue_id,
-		struct rte_vhost_async_config config,
-		struct rte_vhost_async_channel_ops *ops)
+rte_vhost_async_channel_register(int vid, uint16_t queue_id)
 {
 	struct vhost_virtqueue *vq;
 	struct virtio_net *dev = get_device(vid);
 	int ret;
 
-	if (dev == NULL || ops == NULL)
+	if (dev == NULL)
 		return -1;
 
 	if (queue_id >= VHOST_MAX_VRING)
@@ -1710,33 +1713,20 @@  rte_vhost_async_channel_register(int vid, uint16_t queue_id,
 	if (unlikely(vq == NULL || !dev->async_copy))
 		return -1;
 
-	if (unlikely(!(config.features & RTE_VHOST_ASYNC_INORDER))) {
-		VHOST_LOG_CONFIG(ERR,
-			"async copy is not supported on non-inorder mode "
-			"(vid %d, qid: %d)\n", vid, queue_id);
-		return -1;
-	}
-
-	if (unlikely(ops->check_completed_copies == NULL ||
-		ops->transfer_data == NULL))
-		return -1;
-
 	rte_spinlock_lock(&vq->access_lock);
-	ret = async_channel_register(vid, queue_id, ops);
+	ret = async_channel_register(vid, queue_id);
 	rte_spinlock_unlock(&vq->access_lock);
 
 	return ret;
 }
 
 int
-rte_vhost_async_channel_register_thread_unsafe(int vid, uint16_t queue_id,
-		struct rte_vhost_async_config config,
-		struct rte_vhost_async_channel_ops *ops)
+rte_vhost_async_channel_register_thread_unsafe(int vid, uint16_t queue_id)
 {
 	struct vhost_virtqueue *vq;
 	struct virtio_net *dev = get_device(vid);
 
-	if (dev == NULL || ops == NULL)
+	if (dev == NULL)
 		return -1;
 
 	if (queue_id >= VHOST_MAX_VRING)
@@ -1747,18 +1737,7 @@  rte_vhost_async_channel_register_thread_unsafe(int vid, uint16_t queue_id,
 	if (unlikely(vq == NULL || !dev->async_copy))
 		return -1;
 
-	if (unlikely(!(config.features & RTE_VHOST_ASYNC_INORDER))) {
-		VHOST_LOG_CONFIG(ERR,
-			"async copy is not supported on non-inorder mode "
-			"(vid %d, qid: %d)\n", vid, queue_id);
-		return -1;
-	}
-
-	if (unlikely(ops->check_completed_copies == NULL ||
-		ops->transfer_data == NULL))
-		return -1;
-
-	return async_channel_register(vid, queue_id, ops);
+	return async_channel_register(vid, queue_id);
 }
 
 int
@@ -1835,6 +1814,95 @@  rte_vhost_async_channel_unregister_thread_unsafe(int vid, uint16_t queue_id)
 	return 0;
 }
 
+static __rte_always_inline void
+vhost_free_async_dma_mem(void)
+{
+	uint16_t i;
+
+	for (i = 0; i < RTE_DMADEV_DEFAULT_MAX; i++) {
+		struct async_dma_info *dma = &dma_copy_track[i];
+		int16_t j;
+
+		if (dma->max_vchans == 0)
+			continue;
+
+		for (j = 0; j < dma->max_vchans; j++)
+			rte_free(dma->vchans[j].pkts_completed_flag);
+
+		rte_free(dma->vchans);
+		dma->vchans = NULL;
+		dma->max_vchans = 0;
+	}
+}
+
+int
+rte_vhost_async_dma_configure(int16_t *dmas_id, uint16_t count, uint16_t poll_factor)
+{
+	uint16_t i;
+
+	if (!dmas_id) {
+		VHOST_LOG_CONFIG(ERR, "Invalid DMA configuration parameter.\n");
+		return -1;
+	}
+
+	if (poll_factor == 0) {
+		VHOST_LOG_CONFIG(ERR, "Invalid DMA poll factor %u\n", poll_factor);
+		return -1;
+	}
+	dma_poll_factor = poll_factor;
+
+	for (i = 0; i < count; i++) {
+		struct async_dma_vchan_info *vchans;
+		struct rte_dma_info info;
+		uint16_t max_vchans;
+		uint16_t max_desc;
+		uint16_t j;
+
+		if (!rte_dma_is_valid(dmas_id[i])) {
+			VHOST_LOG_CONFIG(ERR, "DMA %d is not found. Cannot enable async"
+				       " data-path\n.", dmas_id[i]);
+			vhost_free_async_dma_mem();
+			return -1;
+		}
+
+		rte_dma_info_get(dmas_id[i], &info);
+
+		max_vchans = info.max_vchans;
+		max_desc = info.max_desc;
+
+		if (!rte_is_power_of_2(max_desc))
+			max_desc = rte_align32pow2(max_desc);
+
+		vchans = rte_zmalloc(NULL, sizeof(struct async_dma_vchan_info) * max_vchans,
+				RTE_CACHE_LINE_SIZE);
+		if (vchans == NULL) {
+			VHOST_LOG_CONFIG(ERR, "Failed to allocate vchans for dma-%d."
+					" Cannot enable async data-path.\n", dmas_id[i]);
+			vhost_free_async_dma_mem();
+			return -1;
+		}
+
+		for (j = 0; j < max_vchans; j++) {
+			vchans[j].pkts_completed_flag = rte_zmalloc(NULL, sizeof(bool *) * max_desc,
+					RTE_CACHE_LINE_SIZE);
+			if (!vchans[j].pkts_completed_flag) {
+				VHOST_LOG_CONFIG(ERR, "Failed to allocate  pkts_completed_flag for "
+						"dma-%d vchan-%u\n", dmas_id[i], j);
+				vhost_free_async_dma_mem();
+				return -1;
+			}
+
+			vchans[j].ring_size = max_desc;
+			vchans[j].ring_mask = max_desc - 1;
+		}
+
+		dma_copy_track[dmas_id[i]].vchans = vchans;
+		dma_copy_track[dmas_id[i]].max_vchans = max_vchans;
+	}
+
+	return 0;
+}
+
 int
 rte_vhost_async_get_inflight(int vid, uint16_t queue_id)
 {
diff --git a/lib/vhost/vhost.h b/lib/vhost/vhost.h
index 7085e0885c..475843fec0 100644
--- a/lib/vhost/vhost.h
+++ b/lib/vhost/vhost.h
@@ -19,6 +19,7 @@ 
 #include <rte_ether.h>
 #include <rte_rwlock.h>
 #include <rte_malloc.h>
+#include <rte_dmadev.h>
 
 #include "rte_vhost.h"
 #include "rte_vdpa.h"
@@ -50,6 +51,7 @@ 
 
 #define VHOST_MAX_ASYNC_IT (MAX_PKT_BURST)
 #define VHOST_MAX_ASYNC_VEC 2048
+#define VHOST_ASYNC_DMA_BATCHING_SIZE 32
 
 #define PACKED_DESC_ENQUEUE_USED_FLAG(w)	\
 	((w) ? (VRING_DESC_F_AVAIL | VRING_DESC_F_USED | VRING_DESC_F_WRITE) : \
@@ -119,6 +121,42 @@  struct vring_used_elem_packed {
 	uint32_t count;
 };
 
+struct async_dma_vchan_info {
+	/* circular array to track if packet copy completes */
+	bool **pkts_completed_flag;
+
+	/* max elements in 'metadata' */
+	uint16_t ring_size;
+	/* ring index mask for 'metadata' */
+	uint16_t ring_mask;
+
+	/* batching copies before a DMA doorbell */
+	uint16_t nr_batching;
+
+	/**
+	 * DMA virtual channel lock. Although it is able to bind DMA
+	 * virtual channels to data plane threads, vhost control plane
+	 * thread could call data plane functions too, thus causing
+	 * DMA device contention.
+	 *
+	 * For example, in VM exit case, vhost control plane thread needs
+	 * to clear in-flight packets before disable vring, but there could
+	 * be anotther data plane thread is enqueuing packets to the same
+	 * vring with the same DMA virtual channel. But dmadev PMD functions
+	 * are lock-free, so the control plane and data plane threads
+	 * could operate the same DMA virtual channel at the same time.
+	 */
+	rte_spinlock_t dma_lock;
+};
+
+struct async_dma_info {
+	uint16_t max_vchans;
+	struct async_dma_vchan_info *vchans;
+};
+
+extern struct async_dma_info dma_copy_track[RTE_DMADEV_DEFAULT_MAX];
+extern uint16_t dma_poll_factor;
+
 /**
  * inflight async packet information
  */
@@ -129,9 +167,6 @@  struct async_inflight_info {
 };
 
 struct vhost_async {
-	/* operation callbacks for DMA */
-	struct rte_vhost_async_channel_ops ops;
-
 	struct rte_vhost_iov_iter iov_iter[VHOST_MAX_ASYNC_IT];
 	struct rte_vhost_iovec iovec[VHOST_MAX_ASYNC_VEC];
 	uint16_t iter_idx;
@@ -139,6 +174,25 @@  struct vhost_async {
 
 	/* data transfer status */
 	struct async_inflight_info *pkts_info;
+	/**
+	 * Packet reorder array. "true" indicates that DMA device
+	 * completes all copies for the packet.
+	 *
+	 * Note that this array could be written by multiple threads
+	 * simultaneously. For example, in the case of thread0 and
+	 * thread1 RX packets from NIC and then enqueue packets to
+	 * vring0 and vring1 with own DMA device DMA0 and DMA1, it's
+	 * possible for thread0 to get completed copies belonging to
+	 * vring1 from DMA0, while thread0 is calling rte_vhost_poll
+	 * _enqueue_completed() for vring0 and thread1 is calling
+	 * rte_vhost_submit_enqueue_burst() for vring1. In this case,
+	 * vq->access_lock cannot protect pkts_cmpl_flag of vring1.
+	 *
+	 * However, since offloading is per-packet basis, each packet
+	 * flag will only be written by one thread. And single byte
+	 * write is atomic, so no lock for pkts_cmpl_flag is needed.
+	 */
+	bool *pkts_cmpl_flag;
 	uint16_t pkts_idx;
 	uint16_t pkts_inflight_n;
 	union {
@@ -198,6 +252,7 @@  struct vhost_virtqueue {
 	/* Record packed ring first dequeue desc index */
 	uint16_t		shadow_last_used_idx;
 
+	uint16_t		batch_copy_max_elems;
 	uint16_t		batch_copy_nb_elems;
 	struct batch_copy_elem	*batch_copy_elems;
 	int			numa_node;
@@ -568,8 +623,7 @@  extern int vhost_data_log_level;
 #define PRINT_PACKET(device, addr, size, header) do {} while (0)
 #endif
 
-#define MAX_VHOST_DEVICE	1024
-extern struct virtio_net *vhost_devices[MAX_VHOST_DEVICE];
+extern struct virtio_net *vhost_devices[RTE_MAX_VHOST_DEVICE];
 
 #define VHOST_BINARY_SEARCH_THRESH 256
 
diff --git a/lib/vhost/vhost_user.c b/lib/vhost/vhost_user.c
index 5eb1dd6812..3147e72f04 100644
--- a/lib/vhost/vhost_user.c
+++ b/lib/vhost/vhost_user.c
@@ -527,6 +527,8 @@  vhost_user_set_vring_num(struct virtio_net **pdev,
 		return RTE_VHOST_MSG_RESULT_ERR;
 	}
 
+	vq->batch_copy_max_elems = vq->size;
+
 	return RTE_VHOST_MSG_RESULT_OK;
 }
 
diff --git a/lib/vhost/virtio_net.c b/lib/vhost/virtio_net.c
index b3d954aab4..305f6cd562 100644
--- a/lib/vhost/virtio_net.c
+++ b/lib/vhost/virtio_net.c
@@ -11,6 +11,7 @@ 
 #include <rte_net.h>
 #include <rte_ether.h>
 #include <rte_ip.h>
+#include <rte_dmadev.h>
 #include <rte_vhost.h>
 #include <rte_tcp.h>
 #include <rte_udp.h>
@@ -25,6 +26,10 @@ 
 
 #define MAX_BATCH_LEN 256
 
+/* DMA device copy operation tracking array. */
+struct async_dma_info dma_copy_track[RTE_DMADEV_DEFAULT_MAX];
+uint16_t dma_poll_factor = 1;
+
 static  __rte_always_inline bool
 rxvq_is_mergeable(struct virtio_net *dev)
 {
@@ -43,6 +48,140 @@  is_valid_virt_queue_idx(uint32_t idx, int is_tx, uint32_t nr_vring)
 	return (is_tx ^ (idx & 1)) == 0 && idx < nr_vring;
 }
 
+static __rte_always_inline uint16_t
+vhost_async_dma_transfer(struct vhost_virtqueue *vq, int16_t dma_id,
+		uint16_t vchan_id, uint16_t head_idx,
+		struct rte_vhost_iov_iter *pkts, uint16_t nr_pkts)
+{
+	struct async_dma_vchan_info *dma_info = &dma_copy_track[dma_id].vchans[vchan_id];
+	uint16_t ring_mask = dma_info->ring_mask;
+	uint16_t pkt_idx, bce_idx = 0;
+
+	rte_spinlock_lock(&dma_info->dma_lock);
+
+	for (pkt_idx = 0; pkt_idx < nr_pkts; pkt_idx++) {
+		struct rte_vhost_iovec *iov = pkts[pkt_idx].iov;
+		int copy_idx, last_copy_idx = 0;
+		uint16_t nr_segs = pkts[pkt_idx].nr_segs;
+		uint16_t nr_sw_copy = 0;
+		uint16_t i;
+
+		if (rte_dma_burst_capacity(dma_id, vchan_id) < nr_segs)
+			goto out;
+
+		for (i = 0; i < nr_segs; i++) {
+			/* Fallback to SW copy if error happens */
+			copy_idx = rte_dma_copy(dma_id, vchan_id, (rte_iova_t)iov[i].src_addr,
+					(rte_iova_t)iov[i].dst_addr, iov[i].len,
+					RTE_DMA_OP_FLAG_LLC);
+			if (unlikely(copy_idx < 0)) {
+				/* Find corresponding VA pair and do SW copy */
+				rte_memcpy(vq->batch_copy_elems[bce_idx].dst,
+						vq->batch_copy_elems[bce_idx].src,
+						vq->batch_copy_elems[bce_idx].len);
+				nr_sw_copy++;
+
+				/**
+				 * All copies of the packet are performed
+				 * by the CPU, set the packet completion flag
+				 * to true, as all copies are done.
+				 */
+				if (nr_sw_copy == nr_segs) {
+					vq->async->pkts_cmpl_flag[head_idx % vq->size] = true;
+					break;
+				} else if (i == (nr_segs - 1)) {
+					/**
+					 * A part of copies of current packet
+					 * are enqueued to the DMA successfully
+					 * but the last copy fails, store the
+					 * packet completion flag address
+					 * in the last DMA copy slot.
+					 */
+					dma_info->pkts_completed_flag[last_copy_idx & ring_mask] =
+						&vq->async->pkts_cmpl_flag[head_idx % vq->size];
+					break;
+				}
+			} else
+				last_copy_idx = copy_idx;
+
+			bce_idx++;
+
+			/**
+			 * Only store packet completion flag address in the last copy's
+			 * slot, and other slots are set to NULL.
+			 */
+			if (i == (nr_segs - 1)) {
+				dma_info->pkts_completed_flag[copy_idx & ring_mask] =
+					&vq->async->pkts_cmpl_flag[head_idx % vq->size];
+			}
+		}
+
+		dma_info->nr_batching += nr_segs;
+		if (unlikely(dma_info->nr_batching >= VHOST_ASYNC_DMA_BATCHING_SIZE)) {
+			rte_dma_submit(dma_id, vchan_id);
+			dma_info->nr_batching = 0;
+		}
+
+		head_idx++;
+	}
+
+out:
+	if (dma_info->nr_batching > 0) {
+		rte_dma_submit(dma_id, vchan_id);
+		dma_info->nr_batching = 0;
+	}
+	rte_spinlock_unlock(&dma_info->dma_lock);
+	vq->batch_copy_nb_elems = 0;
+
+	return pkt_idx;
+}
+
+static __rte_always_inline uint16_t
+vhost_async_dma_check_completed(int16_t dma_id, uint16_t vchan_id, uint16_t max_pkts)
+{
+	struct async_dma_vchan_info *dma_info = &dma_copy_track[dma_id].vchans[vchan_id];
+	uint16_t ring_mask = dma_info->ring_mask;
+	uint16_t last_idx = 0;
+	uint16_t nr_copies;
+	uint16_t copy_idx;
+	uint16_t i;
+	bool has_error = false;
+
+	rte_spinlock_lock(&dma_info->dma_lock);
+
+	/**
+	 * Print error log for debugging, if DMA reports error during
+	 * DMA transfer. We do not handle error in vhost level.
+	 */
+	nr_copies = rte_dma_completed(dma_id, vchan_id, max_pkts, &last_idx, &has_error);
+	if (unlikely(has_error)) {
+		VHOST_LOG_DATA(ERR, "dma %d vchannel %u reports error in rte_dma_completed()\n",
+				dma_id, vchan_id);
+	} else if (nr_copies == 0)
+		goto out;
+
+	copy_idx = last_idx - nr_copies + 1;
+	for (i = 0; i < nr_copies; i++) {
+		bool *flag;
+
+		flag = dma_info->pkts_completed_flag[copy_idx & ring_mask];
+		if (flag) {
+			/**
+			 * Mark the packet flag as received. The flag
+			 * could belong to another virtqueue but write
+			 * is atomic.
+			 */
+			*flag = true;
+			dma_info->pkts_completed_flag[copy_idx & ring_mask] = NULL;
+		}
+		copy_idx++;
+	}
+
+out:
+	rte_spinlock_unlock(&dma_info->dma_lock);
+	return nr_copies;
+}
+
 static inline void
 do_data_copy_enqueue(struct virtio_net *dev, struct vhost_virtqueue *vq)
 {
@@ -865,12 +1004,13 @@  async_iter_reset(struct vhost_async *async)
 static __rte_always_inline int
 async_mbuf_to_desc_seg(struct virtio_net *dev, struct vhost_virtqueue *vq,
 		struct rte_mbuf *m, uint32_t mbuf_offset,
-		uint64_t buf_iova, uint32_t cpy_len)
+		uint64_t buf_addr, uint64_t buf_iova, uint32_t cpy_len)
 {
 	struct vhost_async *async = vq->async;
 	uint64_t mapped_len;
 	uint32_t buf_offset = 0;
 	void *hpa;
+	struct batch_copy_elem *bce = vq->batch_copy_elems;
 
 	while (cpy_len) {
 		hpa = (void *)(uintptr_t)gpa_to_first_hpa(dev,
@@ -886,6 +1026,31 @@  async_mbuf_to_desc_seg(struct virtio_net *dev, struct vhost_virtqueue *vq,
 						hpa, (size_t)mapped_len)))
 			return -1;
 
+		/**
+		 * Keep VA for all IOVA segments for falling back to SW
+		 * copy in case of rte_dma_copy() error.
+		 */
+		if (unlikely(vq->batch_copy_nb_elems >= vq->batch_copy_max_elems)) {
+			struct batch_copy_elem *tmp;
+			uint16_t nb_elems = 2 * vq->batch_copy_max_elems;
+
+			VHOST_LOG_DATA(DEBUG, "(%d) %s: run out of batch_copy_elems, "
+					"and realloc double elements.\n", dev->vid, __func__);
+			tmp = rte_realloc_socket(vq->batch_copy_elems, nb_elems * sizeof(*tmp),
+					RTE_CACHE_LINE_SIZE, vq->numa_node);
+			if (!tmp) {
+				VHOST_LOG_DATA(ERR, "Failed to re-alloc batch_copy_elems\n");
+				return -1;
+			}
+
+			vq->batch_copy_max_elems = nb_elems;
+			vq->batch_copy_elems = tmp;
+			bce = tmp;
+		}
+		bce[vq->batch_copy_nb_elems].dst = (void *)((uintptr_t)(buf_addr + buf_offset));
+		bce[vq->batch_copy_nb_elems].src = rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
+		bce[vq->batch_copy_nb_elems++].len = mapped_len;
+
 		cpy_len -= (uint32_t)mapped_len;
 		mbuf_offset += (uint32_t)mapped_len;
 		buf_offset += (uint32_t)mapped_len;
@@ -901,7 +1066,8 @@  sync_mbuf_to_desc_seg(struct virtio_net *dev, struct vhost_virtqueue *vq,
 {
 	struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 
-	if (likely(cpy_len > MAX_BATCH_LEN || vq->batch_copy_nb_elems >= vq->size)) {
+	if (likely(cpy_len > MAX_BATCH_LEN ||
+				vq->batch_copy_nb_elems >= vq->batch_copy_max_elems)) {
 		rte_memcpy((void *)((uintptr_t)(buf_addr)),
 				rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
 				cpy_len);
@@ -1020,8 +1186,10 @@  mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 
 		if (is_async) {
 			if (async_mbuf_to_desc_seg(dev, vq, m, mbuf_offset,
+						buf_addr + buf_offset,
 						buf_iova + buf_offset, cpy_len) < 0)
 				goto error;
+
 		} else {
 			sync_mbuf_to_desc_seg(dev, vq, m, mbuf_offset,
 					buf_addr + buf_offset,
@@ -1449,9 +1617,9 @@  store_dma_desc_info_packed(struct vring_used_elem_packed *s_ring,
 }
 
 static __rte_noinline uint32_t
-virtio_dev_rx_async_submit_split(struct virtio_net *dev,
-	struct vhost_virtqueue *vq, uint16_t queue_id,
-	struct rte_mbuf **pkts, uint32_t count)
+virtio_dev_rx_async_submit_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
+		uint16_t queue_id, struct rte_mbuf **pkts, uint32_t count,
+		int16_t dma_id, uint16_t vchan_id)
 {
 	struct buf_vector buf_vec[BUF_VECTOR_MAX];
 	uint32_t pkt_idx = 0;
@@ -1503,17 +1671,16 @@  virtio_dev_rx_async_submit_split(struct virtio_net *dev,
 	if (unlikely(pkt_idx == 0))
 		return 0;
 
-	n_xfer = async->ops.transfer_data(dev->vid, queue_id, async->iov_iter, 0, pkt_idx);
-	if (unlikely(n_xfer < 0)) {
-		VHOST_LOG_DATA(ERR, "(%d) %s: failed to transfer data for queue id %d.\n",
-				dev->vid, __func__, queue_id);
-		n_xfer = 0;
-	}
+	n_xfer = vhost_async_dma_transfer(vq, dma_id, vchan_id, async->pkts_idx, async->iov_iter,
+			pkt_idx);
 
 	pkt_err = pkt_idx - n_xfer;
 	if (unlikely(pkt_err)) {
 		uint16_t num_descs = 0;
 
+		VHOST_LOG_DATA(DEBUG, "(%d) %s: failed to transfer %u packets for queue %u.\n",
+				dev->vid, __func__, pkt_err, queue_id);
+
 		/* update number of completed packets */
 		pkt_idx = n_xfer;
 
@@ -1656,13 +1823,13 @@  dma_error_handler_packed(struct vhost_virtqueue *vq, uint16_t slot_idx,
 }
 
 static __rte_noinline uint32_t
-virtio_dev_rx_async_submit_packed(struct virtio_net *dev,
-	struct vhost_virtqueue *vq, uint16_t queue_id,
-	struct rte_mbuf **pkts, uint32_t count)
+virtio_dev_rx_async_submit_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
+		uint16_t queue_id, struct rte_mbuf **pkts, uint32_t count,
+		int16_t dma_id, uint16_t vchan_id)
 {
 	uint32_t pkt_idx = 0;
 	uint32_t remained = count;
-	int32_t n_xfer;
+	uint16_t n_xfer;
 	uint16_t num_buffers;
 	uint16_t num_descs;
 
@@ -1670,6 +1837,7 @@  virtio_dev_rx_async_submit_packed(struct virtio_net *dev,
 	struct async_inflight_info *pkts_info = async->pkts_info;
 	uint32_t pkt_err = 0;
 	uint16_t slot_idx = 0;
+	uint16_t head_idx = async->pkts_idx % vq->size;
 
 	do {
 		rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
@@ -1694,19 +1862,17 @@  virtio_dev_rx_async_submit_packed(struct virtio_net *dev,
 	if (unlikely(pkt_idx == 0))
 		return 0;
 
-	n_xfer = async->ops.transfer_data(dev->vid, queue_id, async->iov_iter, 0, pkt_idx);
-	if (unlikely(n_xfer < 0)) {
-		VHOST_LOG_DATA(ERR, "(%d) %s: failed to transfer data for queue id %d.\n",
-				dev->vid, __func__, queue_id);
-		n_xfer = 0;
-	}
-
-	pkt_err = pkt_idx - n_xfer;
+	n_xfer = vhost_async_dma_transfer(vq, dma_id, vchan_id, head_idx,
+			async->iov_iter, pkt_idx);
 
 	async_iter_reset(async);
 
-	if (unlikely(pkt_err))
+	pkt_err = pkt_idx - n_xfer;
+	if (unlikely(pkt_err)) {
+		VHOST_LOG_DATA(DEBUG, "(%d) %s: failed to transfer %u packets for queue %u.\n",
+				dev->vid, __func__, pkt_err, queue_id);
 		dma_error_handler_packed(vq, slot_idx, pkt_err, &pkt_idx);
+	}
 
 	if (likely(vq->shadow_used_idx)) {
 		/* keep used descriptors. */
@@ -1826,28 +1992,43 @@  write_back_completed_descs_packed(struct vhost_virtqueue *vq,
 
 static __rte_always_inline uint16_t
 vhost_poll_enqueue_completed(struct virtio_net *dev, uint16_t queue_id,
-		struct rte_mbuf **pkts, uint16_t count)
+		struct rte_mbuf **pkts, uint16_t count, int16_t dma_id,
+		uint16_t vchan_id)
 {
 	struct vhost_virtqueue *vq = dev->virtqueue[queue_id];
 	struct vhost_async *async = vq->async;
 	struct async_inflight_info *pkts_info = async->pkts_info;
-	int32_t n_cpl;
+	uint32_t max_count;
+	uint16_t nr_cpl_pkts = 0;
 	uint16_t n_descs = 0, n_buffers = 0;
 	uint16_t start_idx, from, i;
 
-	n_cpl = async->ops.check_completed_copies(dev->vid, queue_id, 0, count);
-	if (unlikely(n_cpl < 0)) {
-		VHOST_LOG_DATA(ERR, "(%d) %s: failed to check completed copies for queue id %d.\n",
-				dev->vid, __func__, queue_id);
-		return 0;
+	/* Check completed copies for the given DMA vChannel */
+	max_count = count * dma_poll_factor;
+	vhost_async_dma_check_completed(dma_id, vchan_id, max_count <= UINT16_MAX ? max_count :
+			UINT16_MAX);
+
+	start_idx = async_get_first_inflight_pkt_idx(vq);
+
+	/**
+	 * Calculate the number of copy completed packets.
+	 * Note that there may be completed packets even if
+	 * no copies are reported done by the given DMA vChannel,
+	 * as DMA vChannels could be shared by other threads.
+	 */
+	from = start_idx;
+	while (vq->async->pkts_cmpl_flag[from] && count--) {
+		vq->async->pkts_cmpl_flag[from] = false;
+		from++;
+		if (from >= vq->size)
+			from -= vq->size;
+		nr_cpl_pkts++;
 	}
 
-	if (n_cpl == 0)
+	if (nr_cpl_pkts == 0)
 		return 0;
 
-	start_idx = async_get_first_inflight_pkt_idx(vq);
-
-	for (i = 0; i < n_cpl; i++) {
+	for (i = 0; i < nr_cpl_pkts; i++) {
 		from = (start_idx + i) % vq->size;
 		/* Only used with packed ring */
 		n_buffers += pkts_info[from].nr_buffers;
@@ -1856,7 +2037,7 @@  vhost_poll_enqueue_completed(struct virtio_net *dev, uint16_t queue_id,
 		pkts[i] = pkts_info[from].mbuf;
 	}
 
-	async->pkts_inflight_n -= n_cpl;
+	async->pkts_inflight_n -= nr_cpl_pkts;
 
 	if (likely(vq->enabled && vq->access_ok)) {
 		if (vq_is_packed(dev)) {
@@ -1877,12 +2058,13 @@  vhost_poll_enqueue_completed(struct virtio_net *dev, uint16_t queue_id,
 		}
 	}
 
-	return n_cpl;
+	return nr_cpl_pkts;
 }
 
 uint16_t
 rte_vhost_poll_enqueue_completed(int vid, uint16_t queue_id,
-		struct rte_mbuf **pkts, uint16_t count)
+		struct rte_mbuf **pkts, uint16_t count, int16_t dma_id,
+		uint16_t vchan_id)
 {
 	struct virtio_net *dev = get_device(vid);
 	struct vhost_virtqueue *vq;
@@ -1906,9 +2088,20 @@  rte_vhost_poll_enqueue_completed(int vid, uint16_t queue_id,
 		return 0;
 	}
 
-	rte_spinlock_lock(&vq->access_lock);
+	if (unlikely(!dma_copy_track[dma_id].vchans ||
+				vchan_id > dma_copy_track[dma_id].max_vchans)) {
+		VHOST_LOG_DATA(ERR, "(%d) %s: invalid DMA %d vchan %u.\n",
+			       dev->vid, __func__, dma_id, vchan_id);
+		return 0;
+	}
 
-	n_pkts_cpl = vhost_poll_enqueue_completed(dev, queue_id, pkts, count);
+	if (!rte_spinlock_trylock(&vq->access_lock)) {
+		VHOST_LOG_CONFIG(DEBUG, "Failed to poll completed packets from queue id %u. "
+			"virt queue busy.\n", queue_id);
+		return 0;
+	}
+
+	n_pkts_cpl = vhost_poll_enqueue_completed(dev, queue_id, pkts, count, dma_id, vchan_id);
 
 	rte_spinlock_unlock(&vq->access_lock);
 
@@ -1917,7 +2110,8 @@  rte_vhost_poll_enqueue_completed(int vid, uint16_t queue_id,
 
 uint16_t
 rte_vhost_clear_queue_thread_unsafe(int vid, uint16_t queue_id,
-		struct rte_mbuf **pkts, uint16_t count)
+		struct rte_mbuf **pkts, uint16_t count, int16_t dma_id,
+		uint16_t vchan_id)
 {
 	struct virtio_net *dev = get_device(vid);
 	struct vhost_virtqueue *vq;
@@ -1941,14 +2135,21 @@  rte_vhost_clear_queue_thread_unsafe(int vid, uint16_t queue_id,
 		return 0;
 	}
 
-	n_pkts_cpl = vhost_poll_enqueue_completed(dev, queue_id, pkts, count);
+	if (unlikely(!dma_copy_track[dma_id].vchans ||
+				vchan_id > dma_copy_track[dma_id].max_vchans)) {
+		VHOST_LOG_DATA(ERR, "(%d) %s: invalid DMA %d vchan %u.\n",
+			       dev->vid, __func__, dma_id, vchan_id);
+		return 0;
+	}
+
+	n_pkts_cpl = vhost_poll_enqueue_completed(dev, queue_id, pkts, count, dma_id, vchan_id);
 
 	return n_pkts_cpl;
 }
 
 static __rte_always_inline uint32_t
 virtio_dev_rx_async_submit(struct virtio_net *dev, uint16_t queue_id,
-	struct rte_mbuf **pkts, uint32_t count)
+	struct rte_mbuf **pkts, uint32_t count, int16_t dma_id, uint16_t vchan_id)
 {
 	struct vhost_virtqueue *vq;
 	uint32_t nb_tx = 0;
@@ -1960,6 +2161,13 @@  virtio_dev_rx_async_submit(struct virtio_net *dev, uint16_t queue_id,
 		return 0;
 	}
 
+	if (unlikely(!dma_copy_track[dma_id].vchans ||
+				vchan_id > dma_copy_track[dma_id].max_vchans)) {
+		VHOST_LOG_DATA(ERR, "(%d) %s: invalid DMA %d vchan %u.\n", dev->vid, __func__,
+				dma_id, vchan_id);
+		return 0;
+	}
+
 	vq = dev->virtqueue[queue_id];
 
 	rte_spinlock_lock(&vq->access_lock);
@@ -1980,10 +2188,10 @@  virtio_dev_rx_async_submit(struct virtio_net *dev, uint16_t queue_id,
 
 	if (vq_is_packed(dev))
 		nb_tx = virtio_dev_rx_async_submit_packed(dev, vq, queue_id,
-				pkts, count);
+				pkts, count, dma_id, vchan_id);
 	else
 		nb_tx = virtio_dev_rx_async_submit_split(dev, vq, queue_id,
-				pkts, count);
+				pkts, count, dma_id, vchan_id);
 
 out:
 	if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
@@ -1997,7 +2205,8 @@  virtio_dev_rx_async_submit(struct virtio_net *dev, uint16_t queue_id,
 
 uint16_t
 rte_vhost_submit_enqueue_burst(int vid, uint16_t queue_id,
-		struct rte_mbuf **pkts, uint16_t count)
+		struct rte_mbuf **pkts, uint16_t count, int16_t dma_id,
+		uint16_t vchan_id)
 {
 	struct virtio_net *dev = get_device(vid);
 
@@ -2011,7 +2220,7 @@  rte_vhost_submit_enqueue_burst(int vid, uint16_t queue_id,
 		return 0;
 	}
 
-	return virtio_dev_rx_async_submit(dev, queue_id, pkts, count);
+	return virtio_dev_rx_async_submit(dev, queue_id, pkts, count, dma_id, vchan_id);
 }
 
 static inline bool
@@ -2369,7 +2578,7 @@  copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 		cpy_len = RTE_MIN(buf_avail, mbuf_avail);
 
 		if (likely(cpy_len > MAX_BATCH_LEN ||
-					vq->batch_copy_nb_elems >= vq->size ||
+					vq->batch_copy_nb_elems >= vq->batch_copy_max_elems ||
 					(hdr && cur == m))) {
 			rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
 						mbuf_offset),