vhost: support packed ring in async datapath
authorCheng Jiang <cheng1.jiang@intel.com>
Tue, 27 Apr 2021 08:03:33 +0000 (08:03 +0000)
committerMaxime Coquelin <maxime.coquelin@redhat.com>
Tue, 4 May 2021 08:21:52 +0000 (10:21 +0200)
For now async vhost data path only supports split ring. This patch
enables packed ring in async vhost data path to make async vhost
compatible with virtio 1.1 spec.

Signed-off-by: Cheng Jiang <cheng1.jiang@intel.com>
Reviewed-by: Jiayu Hu <jiayu.hu@intel.com>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
doc/guides/rel_notes/release_21_05.rst
lib/vhost/rte_vhost_async.h
lib/vhost/vhost.c
lib/vhost/vhost.h
lib/vhost/virtio_net.c

index 7ee8f66..885af4a 100644 (file)
@@ -196,6 +196,10 @@ New Features
 
   * Added support for preferred busy polling.
 
+* **Added support for vhost async packed ring data path.**
+
+  Added packed ring support for async vhost.
+
 * **Added support of multiple data-units in cryptodev API.**
 
   The cryptodev library has been enhanced to allow operations on multiple
index c855ff8..6faa31f 100644 (file)
@@ -89,6 +89,7 @@ struct rte_vhost_async_channel_ops {
 struct async_inflight_info {
        struct rte_mbuf *mbuf;
        uint16_t descs; /* num of descs inflight */
+       uint16_t nr_buffers; /* num of buffers inflight for packed ring */
 };
 
 /**
index ea38cf2..c9b6379 100644 (file)
@@ -340,17 +340,17 @@ cleanup_device(struct virtio_net *dev, int destroy)
 static void
 vhost_free_async_mem(struct vhost_virtqueue *vq)
 {
-       if (vq->async_pkts_info)
-               rte_free(vq->async_pkts_info);
-       if (vq->async_descs_split)
-               rte_free(vq->async_descs_split);
-       if (vq->it_pool)
-               rte_free(vq->it_pool);
-       if (vq->vec_pool)
-               rte_free(vq->vec_pool);
+       rte_free(vq->async_pkts_info);
 
-       vq->async_pkts_info = NULL;
+       rte_free(vq->async_buffers_packed);
+       vq->async_buffers_packed = NULL;
+       rte_free(vq->async_descs_split);
        vq->async_descs_split = NULL;
+
+       rte_free(vq->it_pool);
+       rte_free(vq->vec_pool);
+
+       vq->async_pkts_info = NULL;
        vq->it_pool = NULL;
        vq->vec_pool = NULL;
 }
@@ -360,10 +360,10 @@ free_vq(struct virtio_net *dev, struct vhost_virtqueue *vq)
 {
        if (vq_is_packed(dev))
                rte_free(vq->shadow_used_packed);
-       else {
+       else
                rte_free(vq->shadow_used_split);
-               vhost_free_async_mem(vq);
-       }
+
+       vhost_free_async_mem(vq);
        rte_free(vq->batch_copy_elems);
        if (vq->iotlb_pool)
                rte_mempool_free(vq->iotlb_pool);
@@ -1626,10 +1626,9 @@ int rte_vhost_async_channel_register(int vid, uint16_t queue_id,
        if (unlikely(vq == NULL || !dev->async_copy))
                return -1;
 
-       /* packed queue is not supported */
-       if (unlikely(vq_is_packed(dev) || !f.async_inorder)) {
+       if (unlikely(!f.async_inorder)) {
                VHOST_LOG_CONFIG(ERR,
-                       "async copy is not supported on packed queue or non-inorder mode "
+                       "async copy is not supported on non-inorder mode "
                        "(vid %d, qid: %d)\n", vid, queue_id);
                return -1;
        }
@@ -1661,24 +1660,60 @@ int rte_vhost_async_channel_register(int vid, uint16_t queue_id,
        vq->async_pkts_info = rte_malloc_socket(NULL,
                        vq->size * sizeof(struct async_inflight_info),
                        RTE_CACHE_LINE_SIZE, node);
+       if (!vq->async_pkts_info) {
+               vhost_free_async_mem(vq);
+               VHOST_LOG_CONFIG(ERR,
+                       "async register failed: cannot allocate memory for async_pkts_info "
+                       "(vid %d, qid: %d)\n", vid, queue_id);
+               goto reg_out;
+       }
+
        vq->it_pool = rte_malloc_socket(NULL,
                        VHOST_MAX_ASYNC_IT * sizeof(struct rte_vhost_iov_iter),
                        RTE_CACHE_LINE_SIZE, node);
+       if (!vq->it_pool) {
+               vhost_free_async_mem(vq);
+               VHOST_LOG_CONFIG(ERR,
+                       "async register failed: cannot allocate memory for it_pool "
+                       "(vid %d, qid: %d)\n", vid, queue_id);
+               goto reg_out;
+       }
+
        vq->vec_pool = rte_malloc_socket(NULL,
                        VHOST_MAX_ASYNC_VEC * sizeof(struct iovec),
                        RTE_CACHE_LINE_SIZE, node);
-       vq->async_descs_split = rte_malloc_socket(NULL,
-                       vq->size * sizeof(struct vring_used_elem),
-                       RTE_CACHE_LINE_SIZE, node);
-       if (!vq->async_descs_split || !vq->async_pkts_info ||
-               !vq->it_pool || !vq->vec_pool) {
+       if (!vq->vec_pool) {
                vhost_free_async_mem(vq);
                VHOST_LOG_CONFIG(ERR,
-                               "async register failed: cannot allocate memory for vq data "
-                               "(vid %d, qid: %d)\n", vid, queue_id);
+                       "async register failed: cannot allocate memory for vec_pool "
+                       "(vid %d, qid: %d)\n", vid, queue_id);
                goto reg_out;
        }
 
+       if (vq_is_packed(dev)) {
+               vq->async_buffers_packed = rte_malloc_socket(NULL,
+                       vq->size * sizeof(struct vring_used_elem_packed),
+                       RTE_CACHE_LINE_SIZE, node);
+               if (!vq->async_buffers_packed) {
+                       vhost_free_async_mem(vq);
+                       VHOST_LOG_CONFIG(ERR,
+                               "async register failed: cannot allocate memory for async buffers "
+                               "(vid %d, qid: %d)\n", vid, queue_id);
+                       goto reg_out;
+               }
+       } else {
+               vq->async_descs_split = rte_malloc_socket(NULL,
+                       vq->size * sizeof(struct vring_used_elem),
+                       RTE_CACHE_LINE_SIZE, node);
+               if (!vq->async_descs_split) {
+                       vhost_free_async_mem(vq);
+                       VHOST_LOG_CONFIG(ERR,
+                               "async register failed: cannot allocate memory for async descs "
+                               "(vid %d, qid: %d)\n", vid, queue_id);
+                       goto reg_out;
+               }
+       }
+
        vq->async_ops.check_completed_copies = ops->check_completed_copies;
        vq->async_ops.transfer_data = ops->transfer_data;
 
index f628714..b303635 100644 (file)
@@ -201,9 +201,18 @@ struct vhost_virtqueue {
        uint16_t        async_pkts_idx;
        uint16_t        async_pkts_inflight_n;
        uint16_t        async_last_pkts_n;
-       struct vring_used_elem  *async_descs_split;
-       uint16_t async_desc_idx;
-       uint16_t last_async_desc_idx;
+       union {
+               struct vring_used_elem  *async_descs_split;
+               struct vring_used_elem_packed *async_buffers_packed;
+       };
+       union {
+               uint16_t async_desc_idx_split;
+               uint16_t async_buffer_idx_packed;
+       };
+       union {
+               uint16_t last_async_desc_idx_split;
+               uint16_t last_async_buffer_idx_packed;
+       };
 
        /* vq async features */
        bool            async_inorder;
index be12c09..f744675 100644 (file)
@@ -366,14 +366,14 @@ vhost_shadow_dequeue_single_packed_inorder(struct vhost_virtqueue *vq,
 }
 
 static __rte_always_inline void
-vhost_shadow_enqueue_single_packed(struct virtio_net *dev,
-                                  struct vhost_virtqueue *vq,
-                                  uint32_t len[],
-                                  uint16_t id[],
-                                  uint16_t count[],
+vhost_shadow_enqueue_packed(struct vhost_virtqueue *vq,
+                                  uint32_t *len,
+                                  uint16_t *id,
+                                  uint16_t *count,
                                   uint16_t num_buffers)
 {
        uint16_t i;
+
        for (i = 0; i < num_buffers; i++) {
                /* enqueue shadow flush action aligned with batch num */
                if (!vq->shadow_used_idx)
@@ -385,6 +385,17 @@ vhost_shadow_enqueue_single_packed(struct virtio_net *dev,
                vq->shadow_aligned_idx += count[i];
                vq->shadow_used_idx++;
        }
+}
+
+static __rte_always_inline void
+vhost_shadow_enqueue_single_packed(struct virtio_net *dev,
+                                  struct vhost_virtqueue *vq,
+                                  uint32_t *len,
+                                  uint16_t *id,
+                                  uint16_t *count,
+                                  uint16_t num_buffers)
+{
+       vhost_shadow_enqueue_packed(vq, len, id, count, num_buffers);
 
        if (vq->shadow_aligned_idx >= PACKED_BATCH_SIZE) {
                do_data_copy_enqueue(dev, vq);
@@ -1474,6 +1485,23 @@ 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)
+{
+       uint16_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, uint16_t queue_id,
@@ -1556,12 +1584,12 @@ virtio_dev_rx_async_submit_split(struct virtio_net *dev,
                         * descriptors.
                         */
                        from = vq->shadow_used_idx - num_buffers;
-                       to = vq->async_desc_idx & (vq->size - 1);
+                       to = vq->async_desc_idx_split & (vq->size - 1);
 
                        store_dma_desc_info_split(vq->shadow_used_split,
                                        vq->async_descs_split, vq->size, from, to, num_buffers);
 
-                       vq->async_desc_idx += num_buffers;
+                       vq->async_desc_idx_split += num_buffers;
                        vq->shadow_used_idx -= num_buffers;
                } else
                        comp_pkts[num_done_pkts++] = pkts[pkt_idx];
@@ -1619,7 +1647,7 @@ virtio_dev_rx_async_submit_split(struct virtio_net *dev,
                        num_descs += pkts_info[slot_idx & (vq->size - 1)].descs;
                        slot_idx--;
                }
-               vq->async_desc_idx -= num_descs;
+               vq->async_desc_idx_split -= num_descs;
                /* recover shadow used ring and available ring */
                vq->shadow_used_idx -= (vq->last_avail_idx -
                                async_pkts_log[num_async_pkts].last_avail_idx -
@@ -1641,6 +1669,330 @@ virtio_dev_rx_async_submit_split(struct virtio_net *dev,
        return pkt_idx;
 }
 
+static __rte_always_inline void
+vhost_update_used_packed(struct vhost_virtqueue *vq,
+                       struct vring_used_elem_packed *shadow_ring,
+                       uint16_t count)
+{
+       int i;
+       uint16_t used_idx = vq->last_used_idx;
+       uint16_t head_idx = vq->last_used_idx;
+       uint16_t head_flags = 0;
+
+       if (count == 0)
+               return;
+
+       /* Split loop in two to save memory barriers */
+       for (i = 0; i < count; i++) {
+               vq->desc_packed[used_idx].id = shadow_ring[i].id;
+               vq->desc_packed[used_idx].len = shadow_ring[i].len;
+
+               used_idx += shadow_ring[i].count;
+               if (used_idx >= vq->size)
+                       used_idx -= vq->size;
+       }
+
+       /* The ordering for storing desc flags needs to be enforced. */
+       rte_atomic_thread_fence(__ATOMIC_RELEASE);
+
+       for (i = 0; i < count; i++) {
+               uint16_t flags;
+
+               if (vq->shadow_used_packed[i].len)
+                       flags = VRING_DESC_F_WRITE;
+               else
+                       flags = 0;
+
+               if (vq->used_wrap_counter) {
+                       flags |= VRING_DESC_F_USED;
+                       flags |= VRING_DESC_F_AVAIL;
+               } else {
+                       flags &= ~VRING_DESC_F_USED;
+                       flags &= ~VRING_DESC_F_AVAIL;
+               }
+
+               if (i > 0) {
+                       vq->desc_packed[vq->last_used_idx].flags = flags;
+               } else {
+                       head_idx = vq->last_used_idx;
+                       head_flags = flags;
+               }
+
+               vq_inc_last_used_packed(vq, shadow_ring[i].count);
+       }
+
+       vq->desc_packed[head_idx].flags = head_flags;
+}
+
+static __rte_always_inline int
+vhost_enqueue_async_single_packed(struct virtio_net *dev,
+                           struct vhost_virtqueue *vq,
+                           struct rte_mbuf *pkt,
+                           struct buf_vector *buf_vec,
+                           uint16_t *nr_descs,
+                           uint16_t *nr_buffers,
+                           struct vring_packed_desc *async_descs,
+                           struct iovec *src_iovec, struct iovec *dst_iovec,
+                           struct rte_vhost_iov_iter *src_it,
+                           struct rte_vhost_iov_iter *dst_it)
+{
+       uint16_t nr_vec = 0;
+       uint16_t avail_idx = vq->last_avail_idx;
+       uint16_t max_tries, tries = 0;
+       uint16_t buf_id = 0;
+       uint32_t len = 0;
+       uint16_t desc_count = 0;
+       uint32_t size = pkt->pkt_len + sizeof(struct virtio_net_hdr_mrg_rxbuf);
+       uint32_t buffer_len[vq->size];
+       uint16_t buffer_buf_id[vq->size];
+       uint16_t buffer_desc_count[vq->size];
+
+       if (rxvq_is_mergeable(dev))
+               max_tries = vq->size - 1;
+       else
+               max_tries = 1;
+
+       while (size > 0) {
+               /*
+                * if we tried all available ring items, and still
+                * can't get enough buf, it means something abnormal
+                * happened.
+                */
+               if (unlikely(++tries > max_tries))
+                       return -1;
+
+               if (unlikely(fill_vec_buf_packed(dev, vq, avail_idx, &desc_count, buf_vec, &nr_vec,
+                                               &buf_id, &len, VHOST_ACCESS_RW) < 0))
+                       return -1;
+
+               len = RTE_MIN(len, size);
+               size -= len;
+
+               buffer_len[*nr_buffers] = len;
+               buffer_buf_id[*nr_buffers] = buf_id;
+               buffer_desc_count[*nr_buffers] = desc_count;
+               *nr_buffers += 1;
+
+               *nr_descs += desc_count;
+               avail_idx += desc_count;
+               if (avail_idx >= vq->size)
+                       avail_idx -= vq->size;
+       }
+
+       if (async_mbuf_to_desc(dev, vq, pkt, buf_vec, nr_vec, *nr_buffers, src_iovec, dst_iovec,
+                       src_it, dst_it) < 0)
+               return -1;
+       /* store descriptors for DMA */
+       if (avail_idx >= *nr_descs) {
+               rte_memcpy(async_descs, &vq->desc_packed[vq->last_avail_idx],
+                       *nr_descs * sizeof(struct vring_packed_desc));
+       } else {
+               uint16_t nr_copy = vq->size - vq->last_avail_idx;
+
+               rte_memcpy(async_descs, &vq->desc_packed[vq->last_avail_idx],
+                       nr_copy * sizeof(struct vring_packed_desc));
+               rte_memcpy(async_descs + nr_copy, vq->desc_packed,
+                       (*nr_descs - nr_copy) * sizeof(struct vring_packed_desc));
+       }
+
+       vhost_shadow_enqueue_packed(vq, buffer_len, buffer_buf_id, buffer_desc_count, *nr_buffers);
+
+       return 0;
+}
+
+static __rte_always_inline int16_t
+virtio_dev_rx_async_single_packed(struct virtio_net *dev, struct vhost_virtqueue *vq,
+                           struct rte_mbuf *pkt, uint16_t *nr_descs, uint16_t *nr_buffers,
+                           struct vring_packed_desc *async_descs,
+                           struct iovec *src_iovec, struct iovec *dst_iovec,
+                           struct rte_vhost_iov_iter *src_it, struct rte_vhost_iov_iter *dst_it)
+{
+       struct buf_vector buf_vec[BUF_VECTOR_MAX];
+
+       if (unlikely(vhost_enqueue_async_single_packed(dev, vq, pkt, buf_vec, nr_descs, nr_buffers,
+                                                async_descs, src_iovec, dst_iovec,
+                                                src_it, dst_it) < 0)) {
+               VHOST_LOG_DATA(DEBUG, "(%d) failed to get enough desc from vring\n", dev->vid);
+               return -1;
+       }
+
+       VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
+                       dev->vid, vq->last_avail_idx, vq->last_avail_idx + *nr_descs);
+
+       return 0;
+}
+
+static __rte_always_inline void
+dma_error_handler_packed(struct vhost_virtqueue *vq, struct vring_packed_desc *async_descs,
+                       uint16_t async_descs_idx, uint16_t slot_idx, uint32_t nr_err,
+                       uint32_t *pkt_idx, uint32_t *num_async_pkts, uint32_t *num_done_pkts)
+{
+       uint16_t descs_err = 0;
+       uint16_t buffers_err = 0;
+       struct async_inflight_info *pkts_info = vq->async_pkts_info;
+
+       *num_async_pkts -= nr_err;
+       *pkt_idx -= nr_err;
+       /* calculate the sum of buffers and descs of DMA-error packets. */
+       while (nr_err-- > 0) {
+               descs_err += pkts_info[slot_idx % vq->size].descs;
+               buffers_err += pkts_info[slot_idx % vq->size].nr_buffers;
+               slot_idx--;
+       }
+
+       vq->async_buffer_idx_packed -= buffers_err;
+
+       if (vq->last_avail_idx >= descs_err) {
+               vq->last_avail_idx -= descs_err;
+
+               rte_memcpy(&vq->desc_packed[vq->last_avail_idx],
+                       &async_descs[async_descs_idx - descs_err],
+                       descs_err * sizeof(struct vring_packed_desc));
+       } else {
+               uint16_t nr_copy;
+
+               vq->last_avail_idx = vq->last_avail_idx + vq->size - descs_err;
+               nr_copy = vq->size - vq->last_avail_idx;
+               rte_memcpy(&vq->desc_packed[vq->last_avail_idx],
+                       &async_descs[async_descs_idx - descs_err],
+                       nr_copy * sizeof(struct vring_packed_desc));
+               descs_err -= nr_copy;
+               rte_memcpy(&vq->desc_packed[0], &async_descs[async_descs_idx - descs_err],
+                       descs_err * sizeof(struct vring_packed_desc));
+               vq->avail_wrap_counter ^= 1;
+       }
+
+       *num_done_pkts = *pkt_idx - *num_async_pkts;
+}
+
+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,
+       struct rte_mbuf **comp_pkts, uint32_t *comp_count)
+{
+       uint32_t pkt_idx = 0, pkt_burst_idx = 0;
+       uint16_t async_descs_idx = 0;
+       uint16_t num_buffers;
+       uint16_t num_descs;
+
+       struct rte_vhost_iov_iter *it_pool = vq->it_pool;
+       struct iovec *vec_pool = vq->vec_pool;
+       struct rte_vhost_async_desc tdes[MAX_PKT_BURST];
+       struct iovec *src_iovec = vec_pool;
+       struct iovec *dst_iovec = vec_pool + (VHOST_MAX_ASYNC_VEC >> 1);
+       uint16_t slot_idx = 0;
+       uint16_t segs_await = 0;
+       uint16_t iovec_idx = 0, it_idx = 0;
+       struct async_inflight_info *pkts_info = vq->async_pkts_info;
+       uint32_t n_pkts = 0, pkt_err = 0;
+       uint32_t num_async_pkts = 0, num_done_pkts = 0;
+       struct vring_packed_desc async_descs[vq->size];
+
+       rte_prefetch0(&vq->desc_packed[vq->last_avail_idx]);
+
+       for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
+               num_buffers = 0;
+               num_descs = 0;
+
+               if (unlikely(virtio_dev_rx_async_single_packed(dev, vq, pkts[pkt_idx],
+                                               &num_descs, &num_buffers,
+                                               &async_descs[async_descs_idx],
+                                               &src_iovec[iovec_idx], &dst_iovec[iovec_idx],
+                                               &it_pool[it_idx], &it_pool[it_idx + 1]) < 0))
+                       break;
+
+               VHOST_LOG_DATA(DEBUG, "(%d) current index %d | end index %d\n",
+                       dev->vid, vq->last_avail_idx,
+                       vq->last_avail_idx + num_descs);
+
+               slot_idx = (vq->async_pkts_idx + num_async_pkts) % vq->size;
+               if (it_pool[it_idx].count) {
+                       uint16_t from, to;
+
+                       async_descs_idx += num_descs;
+                       async_fill_desc(&tdes[pkt_burst_idx++],
+                               &it_pool[it_idx], &it_pool[it_idx + 1]);
+                       pkts_info[slot_idx].descs = num_descs;
+                       pkts_info[slot_idx].nr_buffers = num_buffers;
+                       pkts_info[slot_idx].mbuf = pkts[pkt_idx];
+                       num_async_pkts++;
+                       iovec_idx += it_pool[it_idx].nr_segs;
+                       it_idx += 2;
+
+                       segs_await += it_pool[it_idx].nr_segs;
+
+                       /**
+                        * recover shadow used ring and keep DMA-occupied
+                        * descriptors.
+                        */
+                       from = vq->shadow_used_idx - num_buffers;
+                       to = vq->async_buffer_idx_packed % vq->size;
+                       store_dma_desc_info_packed(vq->shadow_used_packed,
+                                       vq->async_buffers_packed, vq->size, from, to, num_buffers);
+
+                       vq->async_buffer_idx_packed += num_buffers;
+                       vq->shadow_used_idx -= num_buffers;
+               } else {
+                       comp_pkts[num_done_pkts++] = pkts[pkt_idx];
+               }
+
+               vq_inc_last_avail_packed(vq, num_descs);
+
+               /*
+                * conditions to trigger async device transfer:
+                * - buffered packet number reaches transfer threshold
+                * - unused async iov number is less than max vhost vector
+                */
+               if (unlikely(pkt_burst_idx >= VHOST_ASYNC_BATCH_THRESHOLD ||
+                       ((VHOST_MAX_ASYNC_VEC >> 1) - segs_await < BUF_VECTOR_MAX))) {
+                       n_pkts = vq->async_ops.transfer_data(dev->vid, queue_id,
+                               tdes, 0, pkt_burst_idx);
+                       iovec_idx = 0;
+                       it_idx = 0;
+                       segs_await = 0;
+                       vq->async_pkts_inflight_n += n_pkts;
+
+                       if (unlikely(n_pkts < pkt_burst_idx)) {
+                               /*
+                                * log error packets number here and do actual
+                                * error processing when applications poll
+                                * completion
+                                */
+                               pkt_err = pkt_burst_idx - n_pkts;
+                               pkt_burst_idx = 0;
+                               pkt_idx++;
+                               break;
+                       }
+
+                       pkt_burst_idx = 0;
+               }
+       }
+
+       if (pkt_burst_idx) {
+               n_pkts = vq->async_ops.transfer_data(dev->vid, queue_id, tdes, 0, pkt_burst_idx);
+               vq->async_pkts_inflight_n += n_pkts;
+
+               if (unlikely(n_pkts < pkt_burst_idx))
+                       pkt_err = pkt_burst_idx - n_pkts;
+       }
+
+       do_data_copy_enqueue(dev, vq);
+
+       if (unlikely(pkt_err))
+               dma_error_handler_packed(vq, async_descs, async_descs_idx, slot_idx, pkt_err,
+                                       &pkt_idx, &num_async_pkts, &num_done_pkts);
+       vq->async_pkts_idx += num_async_pkts;
+       *comp_count = num_done_pkts;
+
+       if (likely(vq->shadow_used_idx)) {
+               vhost_flush_enqueue_shadow_packed(dev, vq);
+               vhost_vring_call_packed(dev, vq);
+       }
+
+       return pkt_idx;
+}
+
 static __rte_always_inline void
 write_back_completed_descs_split(struct vhost_virtqueue *vq, uint16_t n_descs)
 {
@@ -1649,7 +2001,7 @@ write_back_completed_descs_split(struct vhost_virtqueue *vq, uint16_t n_descs)
        uint16_t to, from;
 
        do {
-               from = vq->last_async_desc_idx & (vq->size - 1);
+               from = vq->last_async_desc_idx_split & (vq->size - 1);
                nr_copy = nr_left + from <= vq->size ? nr_left : vq->size - from;
                to = vq->last_used_idx & (vq->size - 1);
 
@@ -1665,18 +2017,41 @@ write_back_completed_descs_split(struct vhost_virtqueue *vq, uint16_t n_descs)
                                        (nr_copy - size) * sizeof(struct vring_used_elem));
                }
 
-               vq->last_async_desc_idx += nr_copy;
+               vq->last_async_desc_idx_split += nr_copy;
                vq->last_used_idx += nr_copy;
                nr_left -= nr_copy;
        } while (nr_left > 0);
 }
 
+static __rte_always_inline void
+write_back_completed_descs_packed(struct vhost_virtqueue *vq,
+                               uint16_t n_buffers)
+{
+       uint16_t nr_left = n_buffers;
+       uint16_t from, to;
+
+       do {
+               from = vq->last_async_buffer_idx_packed % vq->size;
+               to = (from + nr_left) % vq->size;
+               if (to > from) {
+                       vhost_update_used_packed(vq, vq->async_buffers_packed + from, to - from);
+                       vq->last_async_buffer_idx_packed += nr_left;
+                       nr_left = 0;
+               } else {
+                       vhost_update_used_packed(vq, vq->async_buffers_packed + from,
+                               vq->size - from);
+                       vq->last_async_buffer_idx_packed += vq->size - from;
+                       nr_left -= vq->size - from;
+               }
+       } while (nr_left > 0);
+}
+
 uint16_t rte_vhost_poll_enqueue_completed(int vid, uint16_t queue_id,
                struct rte_mbuf **pkts, uint16_t count)
 {
        struct virtio_net *dev = get_device(vid);
        struct vhost_virtqueue *vq;
-       uint16_t n_pkts_cpl = 0, n_pkts_put = 0, n_descs = 0;
+       uint16_t n_pkts_cpl = 0, n_pkts_put = 0, n_descs = 0, n_buffers = 0;
        uint16_t start_idx, pkts_idx, vq_size;
        struct async_inflight_info *pkts_info;
        uint16_t from, i;
@@ -1701,7 +2076,7 @@ uint16_t rte_vhost_poll_enqueue_completed(int vid, uint16_t queue_id,
 
        rte_spinlock_lock(&vq->access_lock);
 
-       pkts_idx = vq->async_pkts_idx & (vq->size - 1);
+       pkts_idx = vq->async_pkts_idx % vq->size;
        pkts_info = vq->async_pkts_info;
        vq_size = vq->size;
        start_idx = virtio_dev_rx_async_get_info_idx(pkts_idx,
@@ -1718,21 +2093,41 @@ uint16_t rte_vhost_poll_enqueue_completed(int vid, uint16_t queue_id,
                goto done;
        }
 
-       for (i = 0; i < n_pkts_put; i++) {
-               from = (start_idx + i) & (vq_size - 1);
-               n_descs += pkts_info[from].descs;
-               pkts[i] = pkts_info[from].mbuf;
+       if (vq_is_packed(dev)) {
+               for (i = 0; i < n_pkts_put; i++) {
+                       from = (start_idx + i) & (vq_size - 1);
+                       n_buffers += pkts_info[from].nr_buffers;
+                       pkts[i] = pkts_info[from].mbuf;
+               }
+       } else {
+               for (i = 0; i < n_pkts_put; i++) {
+                       from = (start_idx + i) & (vq_size - 1);
+                       n_descs += pkts_info[from].descs;
+                       pkts[i] = pkts_info[from].mbuf;
+               }
        }
+
        vq->async_last_pkts_n = n_pkts_cpl - n_pkts_put;
        vq->async_pkts_inflight_n -= n_pkts_put;
 
        if (likely(vq->enabled && vq->access_ok)) {
-               write_back_completed_descs_split(vq, n_descs);
+               if (vq_is_packed(dev)) {
+                       write_back_completed_descs_packed(vq, n_buffers);
 
-               __atomic_add_fetch(&vq->used->idx, n_descs, __ATOMIC_RELEASE);
-               vhost_vring_call_split(dev, vq);
-       } else
-               vq->last_async_desc_idx += n_descs;
+                       vhost_vring_call_packed(dev, vq);
+               } else {
+                       write_back_completed_descs_split(vq, n_descs);
+
+                       __atomic_add_fetch(&vq->used->idx, n_descs,
+                                       __ATOMIC_RELEASE);
+                       vhost_vring_call_split(dev, vq);
+               }
+       } else {
+               if (vq_is_packed(dev))
+                       vq->last_async_buffer_idx_packed += n_buffers;
+               else
+                       vq->last_async_desc_idx_split += n_descs;
+       }
 
 done:
        rte_spinlock_unlock(&vq->access_lock);
@@ -1773,9 +2168,10 @@ virtio_dev_rx_async_submit(struct virtio_net *dev, uint16_t queue_id,
        if (count == 0)
                goto out;
 
-       /* TODO: packed queue not implemented */
        if (vq_is_packed(dev))
-               nb_tx = 0;
+               nb_tx = virtio_dev_rx_async_submit_packed(dev,
+                               vq, queue_id, pkts, count, comp_pkts,
+                               comp_count);
        else
                nb_tx = virtio_dev_rx_async_submit_split(dev,
                                vq, queue_id, pkts, count, comp_pkts,