[RFC 0/3] Async vhost packed ring optimization

[RFC 0/3] Async vhost packed ring optimization

[Cheng Jiang]

To improve the performance of async vhost packed ring. We remove the
unnecessary data copy in async vhost packed ring. And add the batch
data path in both enqueue data path and dequeue data path.

Cheng Jiang (3):
  vhost: remove redundant copy for packed shadow used ring
  vhost: add batch enqueue in async vhost packed ring
  vhost: add batch dequeue in async vhost packed ring

 lib/vhost/virtio_net.c | 395 ++++++++++++++++++++++++++++++++++++-----
 1 file changed, 355 insertions(+), 40 deletions(-)

--
2.35.1

[RFC 1/3] vhost: remove redundant copy for packed shadow used ring

[Cheng Jiang]

In the packed ring enqueue data path of the current asynchronous
Vhost design, the shadow used ring is first copied to the sync
shadow used ring, and then it will be moved to the async shadow
used ring for some historical reasons. This is completely unnecessary.
This patch removes redundant copy for the shadow used ring. The async
shadow used ring will be updated directly.

Signed-off-by: Cheng Jiang <cheng1.jiang@intel.com>
---
 lib/vhost/virtio_net.c | 66 ++++++++++++++++++++----------------------
 1 file changed, 31 insertions(+), 35 deletions(-)

diff --git a/lib/vhost/virtio_net.c b/lib/vhost/virtio_net.c
index 4358899718..22f97d4f77 100644
--- a/lib/vhost/virtio_net.c
+++ b/lib/vhost/virtio_net.c
@@ -572,6 +572,26 @@ vhost_shadow_enqueue_packed(struct vhost_virtqueue *vq,
 	}
 }
 
+static __rte_always_inline void
+vhost_async_shadow_enqueue_packed(struct vhost_virtqueue *vq,
+				   uint32_t *len,
+				   uint16_t *id,
+				   uint16_t *count,
+				   uint16_t num_buffers)
+{
+	uint16_t i;
+	struct vhost_async *async = vq->async;
+
+	for (i = 0; i < num_buffers; i++) {
+		async->buffers_packed[async->buffer_idx_packed].id  = id[i];
+		async->buffers_packed[async->buffer_idx_packed].len = len[i];
+		async->buffers_packed[async->buffer_idx_packed].count = count[i];
+		async->buffer_idx_packed++;
+		if (async->buffer_idx_packed >= vq->size)
+			async->buffer_idx_packed -= vq->size;
+	}
+}
+
 static __rte_always_inline void
 vhost_shadow_enqueue_single_packed(struct virtio_net *dev,
 				   struct vhost_virtqueue *vq,
@@ -1647,23 +1667,6 @@ store_dma_desc_info_split(struct vring_used_elem *s_ring, struct vring_used_elem
 	}
 }
 
-static __rte_always_inline void
-store_dma_desc_info_packed(struct vring_used_elem_packed *s_ring,
-		struct vring_used_elem_packed *d_ring,
-		uint16_t ring_size, uint16_t s_idx, uint16_t d_idx, uint16_t count)
-{
-	size_t elem_size = sizeof(struct vring_used_elem_packed);
-
-	if (d_idx + count <= ring_size) {
-		rte_memcpy(d_ring + d_idx, s_ring + s_idx, count * elem_size);
-	} else {
-		uint16_t size = ring_size - d_idx;
-
-		rte_memcpy(d_ring + d_idx, s_ring + s_idx, size * elem_size);
-		rte_memcpy(d_ring, s_ring + s_idx + size, (count - size) * elem_size);
-	}
-}
-
 static __rte_noinline uint32_t
 virtio_dev_rx_async_submit_split(struct virtio_net *dev, struct vhost_virtqueue *vq,
 	struct rte_mbuf **pkts, uint32_t count, int16_t dma_id, uint16_t vchan_id)
@@ -1822,7 +1825,8 @@ vhost_enqueue_async_packed(struct virtio_net *dev,
 	if (unlikely(mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec, *nr_buffers, true) < 0))
 		return -1;
 
-	vhost_shadow_enqueue_packed(vq, buffer_len, buffer_buf_id, buffer_desc_count, *nr_buffers);
+	vhost_async_shadow_enqueue_packed(vq, buffer_len, buffer_buf_id,
+				buffer_desc_count, *nr_buffers);
 
 	return 0;
 }
@@ -1852,6 +1856,7 @@ dma_error_handler_packed(struct vhost_virtqueue *vq, uint16_t slot_idx,
 {
 	uint16_t descs_err = 0;
 	uint16_t buffers_err = 0;
+	struct vhost_async *async = vq->async;
 	struct async_inflight_info *pkts_info = vq->async->pkts_info;
 
 	*pkt_idx -= nr_err;
@@ -1869,7 +1874,10 @@ dma_error_handler_packed(struct vhost_virtqueue *vq, uint16_t slot_idx,
 		vq->avail_wrap_counter ^= 1;
 	}
 
-	vq->shadow_used_idx -= buffers_err;
+	if (async->buffer_idx_packed >= buffers_err)
+		async->buffer_idx_packed -= buffers_err;
+	else
+		async->buffer_idx_packed = async->buffer_idx_packed + vq->size - buffers_err;
 }
 
 static __rte_noinline uint32_t
@@ -1923,23 +1931,11 @@ virtio_dev_rx_async_submit_packed(struct virtio_net *dev, struct vhost_virtqueue
 		dma_error_handler_packed(vq, slot_idx, pkt_err, &pkt_idx);
 	}
 
-	if (likely(vq->shadow_used_idx)) {
-		/* keep used descriptors. */
-		store_dma_desc_info_packed(vq->shadow_used_packed, async->buffers_packed,
-					vq->size, 0, async->buffer_idx_packed,
-					vq->shadow_used_idx);
-
-		async->buffer_idx_packed += vq->shadow_used_idx;
-		if (async->buffer_idx_packed >= vq->size)
-			async->buffer_idx_packed -= vq->size;
-
-		async->pkts_idx += pkt_idx;
-		if (async->pkts_idx >= vq->size)
-			async->pkts_idx -= vq->size;
+	async->pkts_idx += pkt_idx;
+	if (async->pkts_idx >= vq->size)
+		async->pkts_idx -= vq->size;
 
-		vq->shadow_used_idx = 0;
-		async->pkts_inflight_n += pkt_idx;
-	}
+	async->pkts_inflight_n += pkt_idx;
 
 	return pkt_idx;
 }
-- 
2.35.1

[RFC 2/3] vhost: add batch enqueue in async vhost packed ring

[Cheng Jiang]

Add batch enqueue function in asynchronous vhost packed ring to
improve the performance. Chained mbufs are not supported, it will
be handled in single enqueue function.

Signed-off-by: Cheng Jiang <cheng1.jiang@intel.com>
---
 lib/vhost/virtio_net.c | 159 ++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 157 insertions(+), 2 deletions(-)

diff --git a/lib/vhost/virtio_net.c b/lib/vhost/virtio_net.c
index 22f97d4f77..b87405ba54 100644
--- a/lib/vhost/virtio_net.c
+++ b/lib/vhost/virtio_net.c
@@ -432,6 +432,24 @@ vhost_flush_enqueue_batch_packed(struct virtio_net *dev,
 	vq_inc_last_used_packed(vq, PACKED_BATCH_SIZE);
 }
 
+static __rte_always_inline void
+vhost_async_shadow_enqueue_packed_batch(struct vhost_virtqueue *vq,
+				 uint64_t *lens,
+				 uint16_t *ids)
+{
+	uint16_t i;
+	struct vhost_async *async = vq->async;
+
+	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+		async->buffers_packed[async->buffer_idx_packed].id  = ids[i];
+		async->buffers_packed[async->buffer_idx_packed].len = lens[i];
+		async->buffers_packed[async->buffer_idx_packed].count = 1;
+		async->buffer_idx_packed++;
+		if (async->buffer_idx_packed >= vq->size)
+			async->buffer_idx_packed -= vq->size;
+	}
+}
+
 static __rte_always_inline void
 vhost_shadow_dequeue_batch_packed_inorder(struct vhost_virtqueue *vq,
 					  uint16_t id)
@@ -1451,6 +1469,58 @@ virtio_dev_rx_sync_batch_check(struct virtio_net *dev,
 	return 0;
 }
 
+static __rte_always_inline int
+virtio_dev_rx_async_batch_check(struct vhost_virtqueue *vq,
+			   struct rte_mbuf **pkts,
+			   uint64_t *desc_addrs,
+			   uint64_t *lens,
+			   int16_t dma_id,
+			   uint16_t vchan_id)
+{
+	bool wrap_counter = vq->avail_wrap_counter;
+	struct vring_packed_desc *descs = vq->desc_packed;
+	uint16_t avail_idx = vq->last_avail_idx;
+	uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
+	uint16_t i;
+
+	if (unlikely(avail_idx & PACKED_BATCH_MASK))
+		return -1;
+
+	if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
+		return -1;
+
+	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+		if (unlikely(pkts[i]->next != NULL))
+			return -1;
+		if (unlikely(!desc_is_avail(&descs[avail_idx + i],
+					    wrap_counter)))
+			return -1;
+	}
+
+	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
+		lens[i] = descs[avail_idx + i].len;
+
+	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+		if (unlikely(pkts[i]->pkt_len > (lens[i] - buf_offset)))
+			return -1;
+	}
+
+	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
+		desc_addrs[i] =  descs[avail_idx + i].addr;
+
+	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+		if (unlikely(!desc_addrs[i]))
+			return -1;
+		if (unlikely(lens[i] != descs[avail_idx + i].len))
+			return -1;
+	}
+
+	if (rte_dma_burst_capacity(dma_id, vchan_id) < PACKED_BATCH_SIZE)
+		return -1;
+
+	return 0;
+}
+
 static __rte_always_inline void
 virtio_dev_rx_batch_packed_copy(struct virtio_net *dev,
 			   struct vhost_virtqueue *vq,
@@ -1850,6 +1920,78 @@ virtio_dev_rx_async_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 	return 0;
 }
 
+static __rte_always_inline void
+virtio_dev_rx_async_packed_batch_enqueue(struct virtio_net *dev,
+			   struct vhost_virtqueue *vq,
+			   struct rte_mbuf **pkts,
+			   uint64_t *desc_addrs,
+			   uint64_t *lens)
+{
+	uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
+	struct virtio_net_hdr_mrg_rxbuf *hdrs[PACKED_BATCH_SIZE];
+	struct vring_packed_desc *descs = vq->desc_packed;
+	struct vhost_async *async = vq->async;
+	uint16_t avail_idx = vq->last_avail_idx;
+	uint32_t mbuf_offset = 0;
+	uint16_t ids[PACKED_BATCH_SIZE];
+	uint64_t mapped_len[PACKED_BATCH_SIZE];
+	void *host_iova[PACKED_BATCH_SIZE];
+	uintptr_t desc;
+	uint16_t i;
+
+	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+		rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
+		desc = vhost_iova_to_vva(dev, vq, desc_addrs[i], &lens[i], VHOST_ACCESS_RW);
+		hdrs[i] = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)desc;
+		lens[i] = pkts[i]->pkt_len +
+			sizeof(struct virtio_net_hdr_mrg_rxbuf);
+	}
+
+	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
+		virtio_enqueue_offload(pkts[i], &hdrs[i]->hdr);
+
+	vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
+
+	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+		host_iova[i] = (void *)(uintptr_t)gpa_to_first_hpa(dev,
+			desc_addrs[i] + buf_offset, lens[i], &mapped_len[i]);
+	}
+
+	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+		async_iter_initialize(dev, async);
+		async_iter_add_iovec(dev, async,
+				(void *)(uintptr_t)rte_pktmbuf_iova_offset(pkts[i], mbuf_offset),
+				host_iova[i],
+				mapped_len[i]);
+		async->iter_idx++;
+	}
+
+	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
+		vhost_log_cache_write_iova(dev, vq, descs[avail_idx + i].addr, lens[i]);
+
+	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
+		ids[i] = descs[avail_idx + i].id;
+
+	vhost_async_shadow_enqueue_packed_batch(vq, lens, ids);
+}
+
+static __rte_always_inline int
+virtio_dev_rx_async_packed_batch(struct virtio_net *dev,
+			   struct vhost_virtqueue *vq,
+			   struct rte_mbuf **pkts,
+			   int16_t dma_id, uint16_t vchan_id)
+{
+	uint64_t desc_addrs[PACKED_BATCH_SIZE];
+	uint64_t lens[PACKED_BATCH_SIZE];
+
+	if (virtio_dev_rx_async_batch_check(vq, pkts, desc_addrs, lens, dma_id, vchan_id) == -1)
+		return -1;
+
+	virtio_dev_rx_async_packed_batch_enqueue(dev, vq, pkts, desc_addrs, lens);
+
+	return 0;
+}
+
 static __rte_always_inline void
 dma_error_handler_packed(struct vhost_virtqueue *vq, uint16_t slot_idx,
 			uint32_t nr_err, uint32_t *pkt_idx)
@@ -1885,7 +2027,6 @@ virtio_dev_rx_async_submit_packed(struct virtio_net *dev, struct vhost_virtqueue
 	struct rte_mbuf **pkts, uint32_t count, int16_t dma_id, uint16_t vchan_id)
 {
 	uint32_t pkt_idx = 0;
-	uint32_t remained = count;
 	uint16_t n_xfer;
 	uint16_t num_buffers;
 	uint16_t num_descs;
@@ -1894,10 +2035,25 @@ virtio_dev_rx_async_submit_packed(struct virtio_net *dev, struct vhost_virtqueue
 	struct async_inflight_info *pkts_info = async->pkts_info;
 	uint32_t pkt_err = 0;
 	uint16_t slot_idx = 0;
+	uint16_t i;
 
 	do {
 		rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
 
+		if (count - pkt_idx >= PACKED_BATCH_SIZE) {
+			if (!virtio_dev_rx_async_packed_batch(dev, vq, &pkts[pkt_idx],
+					dma_id, vchan_id)) {
+				for (i = 0; i < PACKED_BATCH_SIZE; i++) {
+					slot_idx = (async->pkts_idx + pkt_idx) % vq->size;
+					pkts_info[slot_idx].descs = 1;
+					pkts_info[slot_idx].nr_buffers = 1;
+					pkts_info[slot_idx].mbuf = pkts[pkt_idx];
+					pkt_idx++;
+				}
+				continue;
+			}
+		}
+
 		num_buffers = 0;
 		num_descs = 0;
 		if (unlikely(virtio_dev_rx_async_packed(dev, vq, pkts[pkt_idx],
@@ -1911,7 +2067,6 @@ virtio_dev_rx_async_submit_packed(struct virtio_net *dev, struct vhost_virtqueue
 		pkts_info[slot_idx].mbuf = pkts[pkt_idx];
 
 		pkt_idx++;
-		remained--;
 		vq_inc_last_avail_packed(vq, num_descs);
 	} while (pkt_idx < count);
 
-- 
2.35.1

[RFC 3/3] vhost: add batch dequeue in async vhost packed ring

[Cheng Jiang]

Add batch dequeue function in asynchronous vhost packed ring to
improve the performance. Chained mbufs are not supported, it will
be handled in single enqueue function.

Signed-off-by: Cheng Jiang <cheng1.jiang@intel.com>
Signed-off-by: Yuan Wang <yuanx.wang@intel.com>
---
 lib/vhost/virtio_net.c | 170 ++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 167 insertions(+), 3 deletions(-)

diff --git a/lib/vhost/virtio_net.c b/lib/vhost/virtio_net.c
index b87405ba54..24307a4ec9 100644
--- a/lib/vhost/virtio_net.c
+++ b/lib/vhost/virtio_net.c
@@ -450,6 +450,23 @@ vhost_async_shadow_enqueue_packed_batch(struct vhost_virtqueue *vq,
 	}
 }
 
+static __rte_always_inline void
+vhost_async_shadow_dequeue_packed_batch(struct vhost_virtqueue *vq, uint16_t *ids)
+{
+	uint16_t i;
+	struct vhost_async *async = vq->async;
+
+	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+		async->buffers_packed[async->buffer_idx_packed].id  = ids[i];
+		async->buffers_packed[async->buffer_idx_packed].len = 0;
+		async->buffers_packed[async->buffer_idx_packed].count = 1;
+
+		async->buffer_idx_packed++;
+		if (async->buffer_idx_packed >= vq->size)
+			async->buffer_idx_packed -= vq->size;
+	}
+}
+
 static __rte_always_inline void
 vhost_shadow_dequeue_batch_packed_inorder(struct vhost_virtqueue *vq,
 					  uint16_t id)
@@ -3193,6 +3210,80 @@ vhost_reserve_avail_batch_packed(struct virtio_net *dev,
 	return -1;
 }
 
+static __rte_always_inline int
+vhost_async_tx_batch_packed_check(struct virtio_net *dev,
+				 struct vhost_virtqueue *vq,
+				 struct rte_mbuf **pkts,
+				 uint16_t avail_idx,
+				 uintptr_t *desc_addrs,
+				 uint64_t *lens,
+				 uint16_t *ids,
+				 int16_t dma_id,
+				 uint16_t vchan_id)
+{
+	bool wrap = vq->avail_wrap_counter;
+	struct vring_packed_desc *descs = vq->desc_packed;
+	uint64_t buf_lens[PACKED_BATCH_SIZE];
+	uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
+	uint16_t flags, i;
+
+	if (unlikely(avail_idx & PACKED_BATCH_MASK))
+		return -1;
+	if (unlikely((avail_idx + PACKED_BATCH_SIZE) > vq->size))
+		return -1;
+
+	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+		flags = descs[avail_idx + i].flags;
+		if (unlikely((wrap != !!(flags & VRING_DESC_F_AVAIL)) ||
+			     (wrap == !!(flags & VRING_DESC_F_USED))  ||
+			     (flags & PACKED_DESC_SINGLE_DEQUEUE_FLAG)))
+			return -1;
+	}
+
+	rte_atomic_thread_fence(__ATOMIC_ACQUIRE);
+
+	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
+		lens[i] = descs[avail_idx + i].len;
+
+	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+		desc_addrs[i] = descs[avail_idx + i].addr;
+	}
+
+	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+		if (unlikely(!desc_addrs[i]))
+			return -1;
+		if (unlikely((lens[i] != descs[avail_idx + i].len)))
+			return -1;
+	}
+
+	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+		if (virtio_dev_pktmbuf_prep(dev, pkts[i], lens[i]))
+			goto err;
+	}
+
+	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
+		buf_lens[i] = pkts[i]->buf_len - pkts[i]->data_off;
+
+	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+		if (unlikely(buf_lens[i] < (lens[i] - buf_offset)))
+			goto err;
+	}
+
+	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+		pkts[i]->pkt_len = lens[i] - buf_offset;
+		pkts[i]->data_len = pkts[i]->pkt_len;
+		ids[i] = descs[avail_idx + i].id;
+	}
+
+	if (rte_dma_burst_capacity(dma_id, vchan_id) < PACKED_BATCH_SIZE)
+		return -1;
+
+	return 0;
+
+err:
+	return -1;
+}
+
 static __rte_always_inline int
 virtio_dev_tx_batch_packed(struct virtio_net *dev,
 			   struct vhost_virtqueue *vq,
@@ -3769,16 +3860,74 @@ virtio_dev_tx_async_single_packed(struct virtio_net *dev,
 	return err;
 }
 
+static __rte_always_inline int
+virtio_dev_tx_async_packed_batch(struct virtio_net *dev,
+			   struct vhost_virtqueue *vq,
+			   struct rte_mbuf **pkts, uint16_t slot_idx,
+			   uint16_t dma_id, uint16_t vchan_id)
+{
+	uint16_t avail_idx = vq->last_avail_idx;
+	uint32_t buf_offset = sizeof(struct virtio_net_hdr_mrg_rxbuf);
+	struct vhost_async *async = vq->async;
+	struct async_inflight_info *pkts_info = async->pkts_info;
+	struct virtio_net_hdr *hdr;
+	uint32_t mbuf_offset = 0;
+	uintptr_t desc_addrs[PACKED_BATCH_SIZE];
+	uint64_t desc_vva;
+	uint64_t lens[PACKED_BATCH_SIZE];
+	void *host_iova[PACKED_BATCH_SIZE];
+	uint64_t mapped_len[PACKED_BATCH_SIZE];
+	uint16_t ids[PACKED_BATCH_SIZE];
+	uint16_t i;
+
+	if (vhost_async_tx_batch_packed_check(dev, vq, pkts, avail_idx,
+					     desc_addrs, lens, ids, dma_id, vchan_id))
+		return -1;
+
+	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE)
+		rte_prefetch0((void *)(uintptr_t)desc_addrs[i]);
+
+	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+		host_iova[i] = (void *)(uintptr_t)gpa_to_first_hpa(dev,
+			desc_addrs[i] + buf_offset, pkts[i]->pkt_len, &mapped_len[i]);
+	}
+
+	vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+		async_iter_initialize(dev, async);
+		async_iter_add_iovec(dev, async,
+		host_iova[i],
+		(void *)(uintptr_t)rte_pktmbuf_iova_offset(pkts[i], mbuf_offset),
+		mapped_len[i]);
+		async->iter_idx++;
+	}
+
+	if (virtio_net_with_host_offload(dev)) {
+		vhost_for_each_try_unroll(i, 0, PACKED_BATCH_SIZE) {
+			desc_vva = vhost_iova_to_vva(dev, vq, desc_addrs[i], &lens[i],
+							VHOST_ACCESS_RO);
+			hdr = (struct virtio_net_hdr *)(uintptr_t)desc_vva;
+			pkts_info[slot_idx + i].nethdr = *hdr;
+		}
+	}
+
+	vq_inc_last_avail_packed(vq, PACKED_BATCH_SIZE);
+
+	vhost_async_shadow_dequeue_packed_batch(vq, ids);
+
+	return 0;
+}
+
 static __rte_always_inline uint16_t
 virtio_dev_tx_async_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 		struct rte_mempool *mbuf_pool, struct rte_mbuf **pkts,
 		uint16_t count, uint16_t dma_id, uint16_t vchan_id, bool legacy_ol_flags)
 {
-	uint16_t pkt_idx;
+	uint32_t pkt_idx = 0;
 	uint16_t slot_idx = 0;
 	uint16_t nr_done_pkts = 0;
 	uint16_t pkt_err = 0;
 	uint32_t n_xfer;
+	uint16_t i;
 	struct vhost_async *async = vq->async;
 	struct async_inflight_info *pkts_info = async->pkts_info;
 	struct rte_mbuf *pkts_prealloc[MAX_PKT_BURST];
@@ -3790,12 +3939,26 @@ virtio_dev_tx_async_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 	if (rte_pktmbuf_alloc_bulk(mbuf_pool, pkts_prealloc, count))
 		goto out;
 
-	for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
+	do {
 		struct rte_mbuf *pkt = pkts_prealloc[pkt_idx];
 
 		rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
 
 		slot_idx = (async->pkts_idx + pkt_idx) % vq->size;
+		if (count - pkt_idx >= PACKED_BATCH_SIZE) {
+			if (!virtio_dev_tx_async_packed_batch(dev, vq, &pkts_prealloc[pkt_idx],
+						slot_idx, dma_id, vchan_id)) {
+				for (i = 0; i < PACKED_BATCH_SIZE; i++) {
+					slot_idx = (async->pkts_idx + pkt_idx) % vq->size;
+					pkts_info[slot_idx].descs = 1;
+					pkts_info[slot_idx].nr_buffers = 1;
+					pkts_info[slot_idx].mbuf = pkts_prealloc[pkt_idx];
+					pkt_idx++;
+				}
+				continue;
+			}
+		}
+
 		if (unlikely(virtio_dev_tx_async_single_packed(dev, vq, mbuf_pool, pkt,
 				slot_idx, legacy_ol_flags))) {
 			rte_pktmbuf_free_bulk(&pkts_prealloc[pkt_idx], count - pkt_idx);
@@ -3809,7 +3972,8 @@ virtio_dev_tx_async_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
 		}
 
 		pkts_info[slot_idx].mbuf = pkt;
-	}
+		pkt_idx++;
+	} while (pkt_idx < count);
 
 	n_xfer = vhost_async_dma_transfer(dev, vq, dma_id, vchan_id, async->pkts_idx,
 					async->iov_iter, pkt_idx);
-- 
2.35.1