vhost: use buffer vectors in dequeue path
authorMaxime Coquelin <maxime.coquelin@redhat.com>
Fri, 6 Jul 2018 07:04:46 +0000 (09:04 +0200)
committerThomas Monjalon <thomas@monjalon.net>
Tue, 10 Jul 2018 21:13:36 +0000 (23:13 +0200)
To ease packed ring layout integration, this patch makes
the dequeue path to re-use buffer vectors implemented for
enqueue path.

Doing this, copy_desc_to_mbuf() is now ring layout type
agnostic.

Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Acked-by: Tiwei Bie <tiwei.bie@intel.com>
lib/librte_vhost/vhost.h
lib/librte_vhost/virtio_net.c

index 3437b99..79e3117 100644 (file)
@@ -43,6 +43,7 @@
  * from vring to do scatter RX.
  */
 struct buf_vector {
+       uint64_t buf_iova;
        uint64_t buf_addr;
        uint32_t buf_len;
        uint32_t desc_idx;
index 7412673..6339296 100644 (file)
@@ -225,12 +225,12 @@ static __rte_always_inline int
 fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
                         uint32_t avail_idx, uint32_t *vec_idx,
                         struct buf_vector *buf_vec, uint16_t *desc_chain_head,
-                        uint16_t *desc_chain_len)
+                        uint16_t *desc_chain_len, uint8_t perm)
 {
        uint16_t idx = vq->avail->ring[avail_idx & (vq->size - 1)];
        uint32_t vec_id = *vec_idx;
        uint32_t len    = 0;
-       uint64_t dlen;
+       uint64_t dlen, desc_avail, desc_iova;
        struct vring_desc *descs = vq->desc;
        struct vring_desc *idesc = NULL;
 
@@ -261,16 +261,43 @@ fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
        }
 
        while (1) {
-               if (unlikely(vec_id >= BUF_VECTOR_MAX || idx >= vq->size)) {
+               if (unlikely(idx >= vq->size)) {
                        free_ind_table(idesc);
                        return -1;
                }
 
+
                len += descs[idx].len;
-               buf_vec[vec_id].buf_addr = descs[idx].addr;
-               buf_vec[vec_id].buf_len  = descs[idx].len;
-               buf_vec[vec_id].desc_idx = idx;
-               vec_id++;
+               desc_avail = descs[idx].len;
+               desc_iova = descs[idx].addr;
+
+               while (desc_avail) {
+                       uint64_t desc_addr;
+                       uint64_t desc_chunck_len = desc_avail;
+
+                       if (unlikely(vec_id >= BUF_VECTOR_MAX)) {
+                               free_ind_table(idesc);
+                               return -1;
+                       }
+
+                       desc_addr = vhost_iova_to_vva(dev, vq,
+                                       desc_iova,
+                                       &desc_chunck_len,
+                                       perm);
+                       if (unlikely(!desc_addr)) {
+                               free_ind_table(idesc);
+                               return -1;
+                       }
+
+                       buf_vec[vec_id].buf_iova = desc_iova;
+                       buf_vec[vec_id].buf_addr = desc_addr;
+                       buf_vec[vec_id].buf_len  = desc_chunck_len;
+                       buf_vec[vec_id].desc_idx = idx;
+
+                       desc_avail -= desc_chunck_len;
+                       desc_iova += desc_chunck_len;
+                       vec_id++;
+               }
 
                if ((descs[idx].flags & VRING_DESC_F_NEXT) == 0)
                        break;
@@ -293,7 +320,8 @@ fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 static inline int
 reserve_avail_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
                                uint32_t size, struct buf_vector *buf_vec,
-                               uint16_t *num_buffers, uint16_t avail_head)
+                               uint16_t *num_buffers, uint16_t avail_head,
+                               uint16_t *nr_vec)
 {
        uint16_t cur_idx;
        uint32_t vec_idx = 0;
@@ -315,7 +343,8 @@ reserve_avail_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
                        return -1;
 
                if (unlikely(fill_vec_buf(dev, vq, cur_idx, &vec_idx, buf_vec,
-                                               &head_idx, &len) < 0))
+                                               &head_idx, &len,
+                                               VHOST_ACCESS_RW) < 0))
                        return -1;
                len = RTE_MIN(len, size);
                update_shadow_used_ring(vq, head_idx, len);
@@ -334,21 +363,22 @@ reserve_avail_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
                        return -1;
        }
 
+       *nr_vec = vec_idx;
+
        return 0;
 }
 
 static __rte_always_inline int
 copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
                            struct rte_mbuf *m, struct buf_vector *buf_vec,
-                           uint16_t num_buffers)
+                           uint16_t nr_vec, uint16_t num_buffers)
 {
        uint32_t vec_idx = 0;
-       uint64_t desc_addr, desc_gaddr;
        uint32_t mbuf_offset, mbuf_avail;
-       uint32_t desc_offset, desc_avail;
+       uint32_t buf_offset, buf_avail;
+       uint64_t buf_addr, buf_iova, buf_len;
        uint32_t cpy_len;
-       uint64_t desc_chunck_len;
-       uint64_t hdr_addr, hdr_phys_addr;
+       uint64_t hdr_addr;
        struct rte_mbuf *hdr_mbuf;
        struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
        struct virtio_net_hdr_mrg_rxbuf tmp_hdr, *hdr = NULL;
@@ -359,82 +389,57 @@ copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
                goto out;
        }
 
-       desc_chunck_len = buf_vec[vec_idx].buf_len;
-       desc_gaddr = buf_vec[vec_idx].buf_addr;
-       desc_addr = vhost_iova_to_vva(dev, vq,
-                                       desc_gaddr,
-                                       &desc_chunck_len,
-                                       VHOST_ACCESS_RW);
-       if (buf_vec[vec_idx].buf_len < dev->vhost_hlen || !desc_addr) {
+       buf_addr = buf_vec[vec_idx].buf_addr;
+       buf_iova = buf_vec[vec_idx].buf_iova;
+       buf_len = buf_vec[vec_idx].buf_len;
+
+       if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
                error = -1;
                goto out;
        }
 
        hdr_mbuf = m;
-       hdr_addr = desc_addr;
-       if (unlikely(desc_chunck_len < dev->vhost_hlen))
+       hdr_addr = buf_addr;
+       if (unlikely(buf_len < dev->vhost_hlen))
                hdr = &tmp_hdr;
        else
                hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
-       hdr_phys_addr = desc_gaddr;
        rte_prefetch0((void *)(uintptr_t)hdr_addr);
 
        VHOST_LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
                dev->vid, num_buffers);
 
-       desc_avail  = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
-       if (unlikely(desc_chunck_len < dev->vhost_hlen)) {
-               desc_chunck_len = desc_avail;
-               desc_gaddr += dev->vhost_hlen;
-               desc_addr = vhost_iova_to_vva(dev, vq,
-                               desc_gaddr,
-                               &desc_chunck_len,
-                               VHOST_ACCESS_RW);
-               if (unlikely(!desc_addr)) {
-                       error = -1;
-                       goto out;
-               }
-
-               desc_offset = 0;
+       if (unlikely(buf_len < dev->vhost_hlen)) {
+               buf_offset = dev->vhost_hlen - buf_len;
+               vec_idx++;
+               buf_addr = buf_vec[vec_idx].buf_addr;
+               buf_iova = buf_vec[vec_idx].buf_iova;
+               buf_len = buf_vec[vec_idx].buf_len;
+               buf_avail = buf_len - buf_offset;
        } else {
-               desc_offset = dev->vhost_hlen;
-               desc_chunck_len -= dev->vhost_hlen;
+               buf_offset = dev->vhost_hlen;
+               buf_avail = buf_len - dev->vhost_hlen;
        }
 
-
        mbuf_avail  = rte_pktmbuf_data_len(m);
        mbuf_offset = 0;
        while (mbuf_avail != 0 || m->next != NULL) {
-               /* done with current desc buf, get the next one */
-               if (desc_avail == 0) {
+               /* done with current buf, get the next one */
+               if (buf_avail == 0) {
                        vec_idx++;
-                       desc_chunck_len = buf_vec[vec_idx].buf_len;
-                       desc_gaddr = buf_vec[vec_idx].buf_addr;
-                       desc_addr =
-                               vhost_iova_to_vva(dev, vq,
-                                       desc_gaddr,
-                                       &desc_chunck_len,
-                                       VHOST_ACCESS_RW);
-                       if (unlikely(!desc_addr)) {
+                       if (unlikely(vec_idx >= nr_vec)) {
                                error = -1;
                                goto out;
                        }
 
+                       buf_addr = buf_vec[vec_idx].buf_addr;
+                       buf_iova = buf_vec[vec_idx].buf_iova;
+                       buf_len = buf_vec[vec_idx].buf_len;
+
                        /* Prefetch buffer address. */
-                       rte_prefetch0((void *)(uintptr_t)desc_addr);
-                       desc_offset = 0;
-                       desc_avail  = buf_vec[vec_idx].buf_len;
-               } else if (unlikely(desc_chunck_len == 0)) {
-                       desc_chunck_len = desc_avail;
-                       desc_gaddr += desc_offset;
-                       desc_addr = vhost_iova_to_vva(dev, vq,
-                                       desc_gaddr,
-                                       &desc_chunck_len, VHOST_ACCESS_RW);
-                       if (unlikely(!desc_addr)) {
-                               error = -1;
-                               goto out;
-                       }
-                       desc_offset = 0;
+                       rte_prefetch0((void *)(uintptr_t)buf_addr);
+                       buf_offset = 0;
+                       buf_avail  = buf_len;
                }
 
                /* done with current mbuf, get the next one */
@@ -455,18 +460,12 @@ copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
                                uint64_t len;
                                uint64_t remain = dev->vhost_hlen;
                                uint64_t src = (uint64_t)(uintptr_t)hdr, dst;
-                               uint64_t guest_addr = hdr_phys_addr;
+                               uint64_t iova = buf_vec[0].buf_iova;
+                               uint16_t hdr_vec_idx = 0;
 
                                while (remain) {
                                        len = remain;
-                                       dst = vhost_iova_to_vva(dev, vq,
-                                                       guest_addr, &len,
-                                                       VHOST_ACCESS_RW);
-                                       if (unlikely(!dst || !len)) {
-                                               error = -1;
-                                               goto out;
-                                       }
-
+                                       dst = buf_vec[hdr_vec_idx].buf_addr;
                                        rte_memcpy((void *)(uintptr_t)dst,
                                                        (void *)(uintptr_t)src,
                                                        len);
@@ -474,50 +473,50 @@ copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
                                        PRINT_PACKET(dev, (uintptr_t)dst,
                                                        (uint32_t)len, 0);
                                        vhost_log_cache_write(dev, vq,
-                                                       guest_addr, len);
+                                                       iova, len);
 
                                        remain -= len;
-                                       guest_addr += len;
+                                       iova += len;
                                        src += len;
+                                       hdr_vec_idx++;
                                }
                        } else {
                                PRINT_PACKET(dev, (uintptr_t)hdr_addr,
                                                dev->vhost_hlen, 0);
-                               vhost_log_cache_write(dev, vq, hdr_phys_addr,
+                               vhost_log_cache_write(dev, vq,
+                                               buf_vec[0].buf_iova,
                                                dev->vhost_hlen);
                        }
 
                        hdr_addr = 0;
                }
 
-               cpy_len = RTE_MIN(desc_chunck_len, mbuf_avail);
+               cpy_len = RTE_MIN(buf_len, mbuf_avail);
 
                if (likely(cpy_len > MAX_BATCH_LEN ||
                                        vq->batch_copy_nb_elems >= vq->size)) {
-                       rte_memcpy((void *)((uintptr_t)(desc_addr +
-                                                       desc_offset)),
+                       rte_memcpy((void *)((uintptr_t)(buf_addr + buf_offset)),
                                rte_pktmbuf_mtod_offset(m, void *, mbuf_offset),
                                cpy_len);
-                       vhost_log_cache_write(dev, vq, desc_gaddr + desc_offset,
+                       vhost_log_cache_write(dev, vq, buf_iova + buf_offset,
                                        cpy_len);
-                       PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
+                       PRINT_PACKET(dev, (uintptr_t)(buf_addr + buf_offset),
                                cpy_len, 0);
                } else {
                        batch_copy[vq->batch_copy_nb_elems].dst =
-                               (void *)((uintptr_t)(desc_addr + desc_offset));
+                               (void *)((uintptr_t)(buf_addr + buf_offset));
                        batch_copy[vq->batch_copy_nb_elems].src =
                                rte_pktmbuf_mtod_offset(m, void *, mbuf_offset);
                        batch_copy[vq->batch_copy_nb_elems].log_addr =
-                               desc_gaddr + desc_offset;
+                               buf_iova + buf_offset;
                        batch_copy[vq->batch_copy_nb_elems].len = cpy_len;
                        vq->batch_copy_nb_elems++;
                }
 
                mbuf_avail  -= cpy_len;
                mbuf_offset += cpy_len;
-               desc_avail  -= cpy_len;
-               desc_offset += cpy_len;
-               desc_chunck_len -= cpy_len;
+               buf_avail  -= cpy_len;
+               buf_offset += cpy_len;
        }
 
 out:
@@ -568,10 +567,11 @@ virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
        avail_head = *((volatile uint16_t *)&vq->avail->idx);
        for (pkt_idx = 0; pkt_idx < count; pkt_idx++) {
                uint32_t pkt_len = pkts[pkt_idx]->pkt_len + dev->vhost_hlen;
+               uint16_t nr_vec = 0;
 
                if (unlikely(reserve_avail_buf(dev, vq,
                                                pkt_len, buf_vec, &num_buffers,
-                                               avail_head) < 0)) {
+                                               avail_head, &nr_vec) < 0)) {
                        VHOST_LOG_DEBUG(VHOST_DATA,
                                "(%d) failed to get enough desc from vring\n",
                                dev->vid);
@@ -584,7 +584,8 @@ virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
                        vq->last_avail_idx + num_buffers);
 
                if (copy_mbuf_to_desc(dev, vq, pkts[pkt_idx],
-                                               buf_vec, num_buffers) < 0) {
+                                               buf_vec, nr_vec,
+                                               num_buffers) < 0) {
                        vq->shadow_used_idx -= num_buffers;
                        break;
                }
@@ -750,49 +751,40 @@ put_zmbuf(struct zcopy_mbuf *zmbuf)
 
 static __rte_always_inline int
 copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
-                 struct vring_desc *descs, uint16_t max_desc,
-                 struct rte_mbuf *m, uint16_t desc_idx,
-                 struct rte_mempool *mbuf_pool)
+                 struct buf_vector *buf_vec, uint16_t nr_vec,
+                 struct rte_mbuf *m, struct rte_mempool *mbuf_pool)
 {
-       struct vring_desc *desc;
-       uint64_t desc_addr, desc_gaddr;
-       uint32_t desc_avail, desc_offset;
+       uint32_t buf_avail, buf_offset;
+       uint64_t buf_addr, buf_iova, buf_len;
        uint32_t mbuf_avail, mbuf_offset;
        uint32_t cpy_len;
-       uint64_t desc_chunck_len;
        struct rte_mbuf *cur = m, *prev = m;
        struct virtio_net_hdr tmp_hdr;
        struct virtio_net_hdr *hdr = NULL;
        /* A counter to avoid desc dead loop chain */
-       uint32_t nr_desc = 1;
+       uint16_t vec_idx = 0;
        struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
        int error = 0;
 
-       desc = &descs[desc_idx];
-       if (unlikely((desc->len < dev->vhost_hlen)) ||
-                       (desc->flags & VRING_DESC_F_INDIRECT)) {
-               error = -1;
-               goto out;
-       }
+       buf_addr = buf_vec[vec_idx].buf_addr;
+       buf_iova = buf_vec[vec_idx].buf_iova;
+       buf_len = buf_vec[vec_idx].buf_len;
 
-       desc_chunck_len = desc->len;
-       desc_gaddr = desc->addr;
-       desc_addr = vhost_iova_to_vva(dev,
-                                       vq, desc_gaddr,
-                                       &desc_chunck_len,
-                                       VHOST_ACCESS_RO);
-       if (unlikely(!desc_addr)) {
+       if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
                error = -1;
                goto out;
        }
 
+       if (likely(nr_vec > 1))
+               rte_prefetch0((void *)(uintptr_t)buf_vec[1].buf_addr);
+
        if (virtio_net_with_host_offload(dev)) {
-               if (unlikely(desc_chunck_len < sizeof(struct virtio_net_hdr))) {
-                       uint64_t len = desc_chunck_len;
+               if (unlikely(buf_len < sizeof(struct virtio_net_hdr))) {
+                       uint64_t len;
                        uint64_t remain = sizeof(struct virtio_net_hdr);
-                       uint64_t src = desc_addr;
+                       uint64_t src;
                        uint64_t dst = (uint64_t)(uintptr_t)&tmp_hdr;
-                       uint64_t guest_addr = desc_gaddr;
+                       uint16_t hdr_vec_idx = 0;
 
                        /*
                         * No luck, the virtio-net header doesn't fit
@@ -800,25 +792,18 @@ copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
                         */
                        while (remain) {
                                len = remain;
-                               src = vhost_iova_to_vva(dev, vq,
-                                               guest_addr, &len,
-                                               VHOST_ACCESS_RO);
-                               if (unlikely(!src || !len)) {
-                                       error = -1;
-                                       goto out;
-                               }
-
+                               src = buf_vec[hdr_vec_idx].buf_addr;
                                rte_memcpy((void *)(uintptr_t)dst,
                                                   (void *)(uintptr_t)src, len);
 
-                               guest_addr += len;
                                remain -= len;
                                dst += len;
+                               hdr_vec_idx++;
                        }
 
                        hdr = &tmp_hdr;
                } else {
-                       hdr = (struct virtio_net_hdr *)((uintptr_t)desc_addr);
+                       hdr = (struct virtio_net_hdr *)((uintptr_t)buf_addr);
                        rte_prefetch0(hdr);
                }
        }
@@ -828,61 +813,40 @@ copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
         * for Tx: the first for storing the header, and others
         * for storing the data.
         */
-       if (likely((desc->len == dev->vhost_hlen) &&
-                  (desc->flags & VRING_DESC_F_NEXT) != 0)) {
-               desc = &descs[desc->next];
-               if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
-                       error = -1;
+       if (unlikely(buf_len < dev->vhost_hlen)) {
+               buf_offset = dev->vhost_hlen - buf_len;
+               vec_idx++;
+               buf_addr = buf_vec[vec_idx].buf_addr;
+               buf_iova = buf_vec[vec_idx].buf_iova;
+               buf_len = buf_vec[vec_idx].buf_len;
+               buf_avail  = buf_len - buf_offset;
+       } else if (buf_len == dev->vhost_hlen) {
+               if (unlikely(++vec_idx >= nr_vec))
                        goto out;
-               }
+               buf_addr = buf_vec[vec_idx].buf_addr;
+               buf_iova = buf_vec[vec_idx].buf_iova;
+               buf_len = buf_vec[vec_idx].buf_len;
 
-               desc_chunck_len = desc->len;
-               desc_gaddr = desc->addr;
-               desc_addr = vhost_iova_to_vva(dev,
-                                                       vq, desc_gaddr,
-                                                       &desc_chunck_len,
-                                                       VHOST_ACCESS_RO);
-               if (unlikely(!desc_addr)) {
-                       error = -1;
-                       goto out;
-               }
-
-               desc_offset = 0;
-               desc_avail  = desc->len;
-               nr_desc    += 1;
+               buf_offset = 0;
+               buf_avail = buf_len;
        } else {
-               desc_avail  = desc->len - dev->vhost_hlen;
-
-               if (unlikely(desc_chunck_len < dev->vhost_hlen)) {
-                       desc_chunck_len = desc_avail;
-                       desc_gaddr += dev->vhost_hlen;
-                       desc_addr = vhost_iova_to_vva(dev,
-                                       vq, desc_gaddr,
-                                       &desc_chunck_len,
-                                       VHOST_ACCESS_RO);
-                       if (unlikely(!desc_addr)) {
-                               error = -1;
-                               goto out;
-                       }
-
-                       desc_offset = 0;
-               } else {
-                       desc_offset = dev->vhost_hlen;
-                       desc_chunck_len -= dev->vhost_hlen;
-               }
+               buf_offset = dev->vhost_hlen;
+               buf_avail = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
        }
 
-       rte_prefetch0((void *)(uintptr_t)(desc_addr + desc_offset));
+       rte_prefetch0((void *)(uintptr_t)
+                       (buf_addr + buf_offset));
 
-       PRINT_PACKET(dev, (uintptr_t)(desc_addr + desc_offset),
-                       (uint32_t)desc_chunck_len, 0);
+       PRINT_PACKET(dev,
+                       (uintptr_t)(buf_addr + buf_offset),
+                       (uint32_t)buf_avail, 0);
 
        mbuf_offset = 0;
        mbuf_avail  = m->buf_len - RTE_PKTMBUF_HEADROOM;
        while (1) {
                uint64_t hpa;
 
-               cpy_len = RTE_MIN(desc_chunck_len, mbuf_avail);
+               cpy_len = RTE_MIN(buf_avail, mbuf_avail);
 
                /*
                 * A desc buf might across two host physical pages that are
@@ -890,11 +854,11 @@ copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
                 * will be copied even though zero copy is enabled.
                 */
                if (unlikely(dev->dequeue_zero_copy && (hpa = gpa_to_hpa(dev,
-                                       desc_gaddr + desc_offset, cpy_len)))) {
+                                       buf_iova + buf_offset, cpy_len)))) {
                        cur->data_len = cpy_len;
                        cur->data_off = 0;
-                       cur->buf_addr = (void *)(uintptr_t)(desc_addr
-                               + desc_offset);
+                       cur->buf_addr =
+                               (void *)(uintptr_t)(buf_addr + buf_offset);
                        cur->buf_iova = hpa;
 
                        /*
@@ -905,20 +869,19 @@ copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
                } else {
                        if (likely(cpy_len > MAX_BATCH_LEN ||
                                   vq->batch_copy_nb_elems >= vq->size ||
-                                  (hdr && cur == m) ||
-                                  desc->len != desc_chunck_len)) {
+                                  (hdr && cur == m))) {
                                rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
                                                                   mbuf_offset),
-                                          (void *)((uintptr_t)(desc_addr +
-                                                               desc_offset)),
+                                          (void *)((uintptr_t)(buf_addr +
+                                                          buf_offset)),
                                           cpy_len);
                        } else {
                                batch_copy[vq->batch_copy_nb_elems].dst =
                                        rte_pktmbuf_mtod_offset(cur, void *,
                                                                mbuf_offset);
                                batch_copy[vq->batch_copy_nb_elems].src =
-                                       (void *)((uintptr_t)(desc_addr +
-                                                            desc_offset));
+                                       (void *)((uintptr_t)(buf_addr +
+                                                               buf_offset));
                                batch_copy[vq->batch_copy_nb_elems].len =
                                        cpy_len;
                                vq->batch_copy_nb_elems++;
@@ -927,59 +890,25 @@ copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 
                mbuf_avail  -= cpy_len;
                mbuf_offset += cpy_len;
-               desc_avail  -= cpy_len;
-               desc_chunck_len -= cpy_len;
-               desc_offset += cpy_len;
+               buf_avail -= cpy_len;
+               buf_offset += cpy_len;
 
-               /* This desc reaches to its end, get the next one */
-               if (desc_avail == 0) {
-                       if ((desc->flags & VRING_DESC_F_NEXT) == 0)
+               /* This buf reaches to its end, get the next one */
+               if (buf_avail == 0) {
+                       if (++vec_idx >= nr_vec)
                                break;
 
-                       if (unlikely(desc->next >= max_desc ||
-                                    ++nr_desc > max_desc)) {
-                               error = -1;
-                               goto out;
-                       }
-                       desc = &descs[desc->next];
-                       if (unlikely(desc->flags & VRING_DESC_F_INDIRECT)) {
-                               error = -1;
-                               goto out;
-                       }
-
-                       desc_chunck_len = desc->len;
-                       desc_gaddr = desc->addr;
-                       desc_addr = vhost_iova_to_vva(dev,
-                                                       vq, desc_gaddr,
-                                                       &desc_chunck_len,
-                                                       VHOST_ACCESS_RO);
-                       if (unlikely(!desc_addr)) {
-                               error = -1;
-                               goto out;
-                       }
+                       buf_addr = buf_vec[vec_idx].buf_addr;
+                       buf_iova = buf_vec[vec_idx].buf_iova;
+                       buf_len = buf_vec[vec_idx].buf_len;
 
-                       rte_prefetch0((void *)(uintptr_t)desc_addr);
+                       rte_prefetch0((void *)(uintptr_t)buf_addr);
 
-                       desc_offset = 0;
-                       desc_avail  = desc->len;
-
-                       PRINT_PACKET(dev, (uintptr_t)desc_addr,
-                                       (uint32_t)desc_chunck_len, 0);
-               } else if (unlikely(desc_chunck_len == 0)) {
-                       desc_chunck_len = desc_avail;
-                       desc_gaddr += desc_offset;
-                       desc_addr = vhost_iova_to_vva(dev, vq,
-                                       desc_gaddr,
-                                       &desc_chunck_len,
-                                       VHOST_ACCESS_RO);
-                       if (unlikely(!desc_addr)) {
-                               error = -1;
-                               goto out;
-                       }
-                       desc_offset = 0;
+                       buf_offset = 0;
+                       buf_avail  = buf_len;
 
-                       PRINT_PACKET(dev, (uintptr_t)desc_addr,
-                                       (uint32_t)desc_chunck_len, 0);
+                       PRINT_PACKET(dev, (uintptr_t)buf_addr,
+                                       (uint32_t)buf_avail, 0);
                }
 
                /*
@@ -1085,10 +1014,8 @@ rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
        struct virtio_net *dev;
        struct rte_mbuf *rarp_mbuf = NULL;
        struct vhost_virtqueue *vq;
-       uint32_t desc_indexes[MAX_PKT_BURST];
        uint32_t i = 0;
        uint16_t free_entries;
-       uint16_t avail_idx;
 
        dev = get_device(vid);
        if (!dev)
@@ -1186,80 +1113,38 @@ rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
 
        VHOST_LOG_DEBUG(VHOST_DATA, "(%d) %s\n", dev->vid, __func__);
 
-       /* Prefetch available and used ring */
-       avail_idx = vq->last_avail_idx & (vq->size - 1);
-       rte_prefetch0(&vq->avail->ring[avail_idx]);
-
        count = RTE_MIN(count, MAX_PKT_BURST);
        count = RTE_MIN(count, free_entries);
        VHOST_LOG_DEBUG(VHOST_DATA, "(%d) about to dequeue %u buffers\n",
                        dev->vid, count);
 
-       /* Retrieve all of the head indexes first to avoid caching issues. */
-       for (i = 0; i < count; i++) {
-               avail_idx = (vq->last_avail_idx + i) & (vq->size - 1);
-               desc_indexes[i] = vq->avail->ring[avail_idx];
-
-               if (likely(dev->dequeue_zero_copy == 0))
-                       update_shadow_used_ring(vq, desc_indexes[i], 0);
-       }
-
-       /* Prefetch descriptor index. */
-       rte_prefetch0(&vq->desc[desc_indexes[0]]);
        for (i = 0; i < count; i++) {
-               struct vring_desc *desc, *idesc = NULL;
-               uint16_t sz, idx;
-               uint64_t dlen;
+               struct buf_vector buf_vec[BUF_VECTOR_MAX];
+               uint16_t head_idx, dummy_len;
+               uint32_t nr_vec = 0;
                int err;
 
-               if (likely(i + 1 < count))
-                       rte_prefetch0(&vq->desc[desc_indexes[i + 1]]);
-
-               if (vq->desc[desc_indexes[i]].flags & VRING_DESC_F_INDIRECT) {
-                       dlen = vq->desc[desc_indexes[i]].len;
-                       desc = (struct vring_desc *)(uintptr_t)
-                               vhost_iova_to_vva(dev, vq,
-                                               vq->desc[desc_indexes[i]].addr,
-                                               &dlen,
-                                               VHOST_ACCESS_RO);
-                       if (unlikely(!desc))
-                               break;
-
-                       if (unlikely(dlen < vq->desc[desc_indexes[i]].len)) {
-                               /*
-                                * The indirect desc table is not contiguous
-                                * in process VA space, we have to copy it.
-                                */
-                               idesc = alloc_copy_ind_table(dev, vq,
-                                               &vq->desc[desc_indexes[i]]);
-                               if (unlikely(!idesc))
-                                       break;
-
-                               desc = idesc;
-                       }
+               if (unlikely(fill_vec_buf(dev, vq,
+                                               vq->last_avail_idx + i,
+                                               &nr_vec, buf_vec,
+                                               &head_idx, &dummy_len,
+                                               VHOST_ACCESS_RO) < 0))
+                       break;
 
-                       rte_prefetch0(desc);
-                       sz = vq->desc[desc_indexes[i]].len / sizeof(*desc);
-                       idx = 0;
-               } else {
-                       desc = vq->desc;
-                       sz = vq->size;
-                       idx = desc_indexes[i];
-               }
+               if (likely(dev->dequeue_zero_copy == 0))
+                       update_shadow_used_ring(vq, head_idx, 0);
 
                pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
                if (unlikely(pkts[i] == NULL)) {
                        RTE_LOG(ERR, VHOST_DATA,
                                "Failed to allocate memory for mbuf.\n");
-                       free_ind_table(idesc);
                        break;
                }
 
-               err = copy_desc_to_mbuf(dev, vq, desc, sz, pkts[i], idx,
-                                       mbuf_pool);
+               err = copy_desc_to_mbuf(dev, vq, buf_vec, nr_vec, pkts[i],
+                               mbuf_pool);
                if (unlikely(err)) {
                        rte_pktmbuf_free(pkts[i]);
-                       free_ind_table(idesc);
                        break;
                }
 
@@ -1269,11 +1154,10 @@ rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
                        zmbuf = get_zmbuf(vq);
                        if (!zmbuf) {
                                rte_pktmbuf_free(pkts[i]);
-                               free_ind_table(idesc);
                                break;
                        }
                        zmbuf->mbuf = pkts[i];
-                       zmbuf->desc_idx = desc_indexes[i];
+                       zmbuf->desc_idx = head_idx;
 
                        /*
                         * Pin lock the mbuf; we will check later to see
@@ -1286,9 +1170,6 @@ rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
                        vq->nr_zmbuf += 1;
                        TAILQ_INSERT_TAIL(&vq->zmbuf_list, zmbuf, next);
                }
-
-               if (unlikely(!!idesc))
-                       free_ind_table(idesc);
        }
        vq->last_avail_idx += i;