[dpdk-dev] [PATCH v2 0/7] vhost: generalize buffer vectors

[dpdk-dev] [PATCH v2 0/7] vhost: generalize buffer vectors

[Maxime Coquelin]

This second revision fixes the arguments swapped in
copy_mbuf_to_desc call.

This series is again preliminray work to ease packed ring
layout integration.

Main changes are using vector buffres also in the dequeue
path, and perform IOVA to HVA translation at vectors fill
time.

I still have to run more benchmarks, but PVP benchmarks does
not show performance changes.

Good thing is that it saves ~140 further lines.

Maxime Coquelin (7):
  vhost: use shadow used ring in dequeue path
  vhost: make gpa to hpa failure an error
  vhost: use buffer vectors in dequeue path
  vhost: translate iovas at vectors fill time
  vhost: improve prefetching in dequeue path
  vhost: prefetch first descriptor in dequeue path
  vhost: improve prefetching in enqueue path

 lib/librte_vhost/vhost.h      |   1 +
 lib/librte_vhost/virtio_net.c | 522 ++++++++++++++++--------------------------
 2 files changed, 192 insertions(+), 331 deletions(-)

-- 
2.14.4

[dpdk-dev] [PATCH v2 1/7] vhost: use shadow used ring in dequeue path

[Maxime Coquelin]

Relax used ring contention by reusing the shadow used
ring feature used by enqueue path.

Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
---
 lib/librte_vhost/virtio_net.c | 45 ++++++++++---------------------------------
 1 file changed, 10 insertions(+), 35 deletions(-)

diff --git a/lib/librte_vhost/virtio_net.c b/lib/librte_vhost/virtio_net.c
index 98ad8e936..7e70a927f 100644
--- a/lib/librte_vhost/virtio_net.c
+++ b/lib/librte_vhost/virtio_net.c
@@ -1019,35 +1019,6 @@ copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 	return error;
 }
 
-static __rte_always_inline void
-update_used_ring(struct virtio_net *dev, struct vhost_virtqueue *vq,
-		 uint32_t used_idx, uint32_t desc_idx)
-{
-	vq->used->ring[used_idx].id  = desc_idx;
-	vq->used->ring[used_idx].len = 0;
-	vhost_log_cache_used_vring(dev, vq,
-			offsetof(struct vring_used, ring[used_idx]),
-			sizeof(vq->used->ring[used_idx]));
-}
-
-static __rte_always_inline void
-update_used_idx(struct virtio_net *dev, struct vhost_virtqueue *vq,
-		uint32_t count)
-{
-	if (unlikely(count == 0))
-		return;
-
-	rte_smp_wmb();
-	rte_smp_rmb();
-
-	vhost_log_cache_sync(dev, vq);
-
-	vq->used->idx += count;
-	vhost_log_used_vring(dev, vq, offsetof(struct vring_used, idx),
-			sizeof(vq->used->idx));
-	vhost_vring_call(dev, vq);
-}
-
 static __rte_always_inline struct zcopy_mbuf *
 get_zmbuf(struct vhost_virtqueue *vq)
 {
@@ -1146,6 +1117,7 @@ rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
 		goto out_access_unlock;
 
 	vq->batch_copy_nb_elems = 0;
+	vq->shadow_used_idx = 0;
 
 	if (dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))
 		vhost_user_iotlb_rd_lock(vq);
@@ -1164,8 +1136,7 @@ rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
 
 			if (mbuf_is_consumed(zmbuf->mbuf)) {
 				used_idx = vq->last_used_idx++ & (vq->size - 1);
-				update_used_ring(dev, vq, used_idx,
-						 zmbuf->desc_idx);
+				update_shadow_used_ring(vq, zmbuf->desc_idx, 0);
 				nr_updated += 1;
 
 				TAILQ_REMOVE(&vq->zmbuf_list, zmbuf, next);
@@ -1176,7 +1147,9 @@ rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
 			}
 		}
 
-		update_used_idx(dev, vq, nr_updated);
+		flush_shadow_used_ring(dev, vq);
+		vhost_vring_call(dev, vq);
+		vq->shadow_used_idx = 0;
 	}
 
 	/*
@@ -1233,7 +1206,7 @@ rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
 		desc_indexes[i] = vq->avail->ring[avail_idx];
 
 		if (likely(dev->dequeue_zero_copy == 0))
-			update_used_ring(dev, vq, used_idx, desc_indexes[i]);
+			update_shadow_used_ring(vq, desc_indexes[i], 0);
 	}
 
 	/* Prefetch descriptor index. */
@@ -1326,8 +1299,10 @@ rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
 
 	if (likely(dev->dequeue_zero_copy == 0)) {
 		do_data_copy_dequeue(vq);
-		vq->last_used_idx += i;
-		update_used_idx(dev, vq, i);
+		if (unlikely(i < count))
+			vq->shadow_used_idx = i;
+		flush_shadow_used_ring(dev, vq);
+		vhost_vring_call(dev, vq);
 	}
 
 out:
-- 
2.14.4

[dpdk-dev] [PATCH v2 2/7] vhost: make gpa to hpa failure an error

[Maxime Coquelin]

CVE-2018-1059 fix makes sure gpa contiguous memory is
also contiguous in hva space. Incidentally, it also makes
sure it is contiguous in hpa space.

So we can simplify the code by making gpa contiguous memory
discontiguous in hpa space an error.

Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
---
 lib/librte_vhost/virtio_net.c | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)

diff --git a/lib/librte_vhost/virtio_net.c b/lib/librte_vhost/virtio_net.c
index 7e70a927f..ec4bcc400 100644
--- a/lib/librte_vhost/virtio_net.c
+++ b/lib/librte_vhost/virtio_net.c
@@ -884,13 +884,13 @@ copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 
 		cpy_len = RTE_MIN(desc_chunck_len, mbuf_avail);
 
-		/*
-		 * A desc buf might across two host physical pages that are
-		 * not continuous. In such case (gpa_to_hpa returns 0), data
-		 * 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)))) {
+		if (unlikely(dev->dequeue_zero_copy)) {
+			hpa = gpa_to_hpa(dev,
+					desc_gaddr + desc_offset, cpy_len);
+			if (unlikely(!hpa)) {
+				error = -1;
+				goto out;
+			}
 			cur->data_len = cpy_len;
 			cur->data_off = 0;
 			cur->buf_addr = (void *)(uintptr_t)(desc_addr
-- 
2.14.4

[dpdk-dev] [PATCH v2 3/7] vhost: use buffer vectors in dequeue path

[Maxime Coquelin]

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>
---
 lib/librte_vhost/virtio_net.c | 143 ++++++++++--------------------------------
 1 file changed, 33 insertions(+), 110 deletions(-)

diff --git a/lib/librte_vhost/virtio_net.c b/lib/librte_vhost/virtio_net.c
index ec4bcc400..4816e8003 100644
--- a/lib/librte_vhost/virtio_net.c
+++ b/lib/librte_vhost/virtio_net.c
@@ -750,11 +750,9 @@ 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 mbuf_avail, mbuf_offset;
@@ -764,24 +762,18 @@ copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 	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;
-	}
-
-	desc_chunck_len = desc->len;
-	desc_gaddr = desc->addr;
+	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_RO);
-	if (unlikely(!desc_addr)) {
+	if (unlikely(buf_vec[vec_idx].buf_len < dev->vhost_hlen ||
+				!desc_addr)) {
 		error = -1;
 		goto out;
 	}
@@ -828,16 +820,12 @@ 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 (likely(buf_vec[vec_idx].buf_len == dev->vhost_hlen)) {
+		if (unlikely(++vec_idx >= nr_vec))
 			goto out;
-		}
 
-		desc_chunck_len = desc->len;
-		desc_gaddr = desc->addr;
+		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,
@@ -848,10 +836,9 @@ copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 		}
 
 		desc_offset = 0;
-		desc_avail  = desc->len;
-		nr_desc    += 1;
+		desc_avail  = buf_vec[vec_idx].buf_len;
 	} else {
-		desc_avail  = desc->len - dev->vhost_hlen;
+		desc_avail  = buf_vec[vec_idx].buf_len - dev->vhost_hlen;
 
 		if (unlikely(desc_chunck_len < dev->vhost_hlen)) {
 			desc_chunck_len = desc_avail;
@@ -906,7 +893,8 @@ copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 			if (likely(cpy_len > MAX_BATCH_LEN ||
 				   vq->batch_copy_nb_elems >= vq->size ||
 				   (hdr && cur == m) ||
-				   desc->len != desc_chunck_len)) {
+				   buf_vec[vec_idx].buf_len !=
+						desc_chunck_len)) {
 				rte_memcpy(rte_pktmbuf_mtod_offset(cur, void *,
 								   mbuf_offset),
 					   (void *)((uintptr_t)(desc_addr +
@@ -933,22 +921,11 @@ copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 
 		/* This desc reaches to its end, get the next one */
 		if (desc_avail == 0) {
-			if ((desc->flags & VRING_DESC_F_NEXT) == 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_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,
@@ -961,7 +938,7 @@ copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 			rte_prefetch0((void *)(uintptr_t)desc_addr);
 
 			desc_offset = 0;
-			desc_avail  = desc->len;
+			desc_avail  = buf_vec[vec_idx].buf_len;
 
 			PRINT_PACKET(dev, (uintptr_t)desc_addr,
 					(uint32_t)desc_chunck_len, 0);
@@ -1085,11 +1062,8 @@ 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 used_idx;
 	uint32_t i = 0;
 	uint16_t free_entries;
-	uint16_t avail_idx;
 
 	dev = get_device(vid);
 	if (!dev)
@@ -1135,7 +1109,6 @@ rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
 			next = TAILQ_NEXT(zmbuf, next);
 
 			if (mbuf_is_consumed(zmbuf->mbuf)) {
-				used_idx = vq->last_used_idx++ & (vq->size - 1);
 				update_shadow_used_ring(vq, zmbuf->desc_idx, 0);
 				nr_updated += 1;
 
@@ -1182,89 +1155,43 @@ rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
 	}
 
 	free_entries = *((volatile uint16_t *)&vq->avail->idx) -
-			vq->last_avail_idx;
+		vq->last_avail_idx;
 	if (free_entries == 0)
 		goto out;
 
 	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);
-	used_idx  = vq->last_used_idx  & (vq->size - 1);
-	rte_prefetch0(&vq->avail->ring[avail_idx]);
-	rte_prefetch0(&vq->used->ring[used_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);
-		used_idx  = (vq->last_used_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) < 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;
 		}
 
@@ -1274,11 +1201,10 @@ 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
@@ -1291,9 +1217,6 @@ 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;
 
-- 
2.14.4

[dpdk-dev] [PATCH v2 4/7] vhost: translate iovas at vectors fill time

[Maxime Coquelin]

This patch aims at simplifying the desc to mbuf and mbuf to desc
copy functions. It performs the iova to hva translations at
vectors fill time.

Doing this, in case desc buffer isn't contiguous in hva space,
it gets split into multiple vectors.

Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
---
 lib/librte_vhost/vhost.h      |   1 +
 lib/librte_vhost/virtio_net.c | 340 ++++++++++++++++++------------------------
 2 files changed, 144 insertions(+), 197 deletions(-)

diff --git a/lib/librte_vhost/vhost.h b/lib/librte_vhost/vhost.h
index 786a74f64..e3b2ed2ff 100644
--- a/lib/librte_vhost/vhost.h
+++ b/lib/librte_vhost/vhost.h
@@ -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;
diff --git a/lib/librte_vhost/virtio_net.c b/lib/librte_vhost/virtio_net.c
index 4816e8003..1ab1edd67 100644
--- a/lib/librte_vhost/virtio_net.c
+++ b/lib/librte_vhost/virtio_net.c
@@ -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;
 
@@ -267,10 +267,31 @@ fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 		}
 
 		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;
+
+			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 +314,8 @@ 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 +337,8 @@ 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_RO) < 0))
 			return -1;
 		len = RTE_MIN(len, size);
 		update_shadow_used_ring(vq, head_idx, len);
@@ -334,21 +357,22 @@ 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 +383,57 @@ 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 +454,12 @@ 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 +467,50 @@ 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 +561,11 @@ 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 +578,8 @@ 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;
 		}
@@ -753,11 +748,10 @@ copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 		  struct buf_vector *buf_vec, uint16_t nr_vec,
 		  struct rte_mbuf *m, struct rte_mempool *mbuf_pool)
 {
-	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;
@@ -766,25 +760,25 @@ copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 	struct batch_copy_elem *batch_copy = vq->batch_copy_elems;
 	int error = 0;
 
-	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_RO);
-	if (unlikely(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;
 	}
 
+	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
@@ -792,25 +786,18 @@ 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);
 		}
 	}
@@ -820,68 +807,51 @@ 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(buf_vec[vec_idx].buf_len == dev->vhost_hlen)) {
+	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 = 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_RO);
-		if (unlikely(!desc_addr)) {
-			error = -1;
-			goto out;
-		}
-
-		desc_offset = 0;
-		desc_avail  = buf_vec[vec_idx].buf_len;
+		buf_offset = 0;
+		buf_avail  = buf_len;
 	} else {
-		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_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);
 
 		if (unlikely(dev->dequeue_zero_copy)) {
-			hpa = gpa_to_hpa(dev,
-					desc_gaddr + desc_offset, cpy_len);
+			hpa = gpa_to_hpa(dev, buf_iova + buf_offset, cpy_len);
 			if (unlikely(!hpa)) {
 				error = -1;
 				goto out;
 			}
 			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;
 
 			/*
@@ -892,21 +862,19 @@ 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) ||
-				   buf_vec[vec_idx].buf_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++;
@@ -915,48 +883,25 @@ 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) {
+		/* This buf reaches to its end, get the next one */
+		if (buf_avail == 0) {
 			if (++vec_idx >= nr_vec)
 				break;
 
-			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_RO);
-			if (unlikely(!desc_addr)) {
-				error = -1;
-				goto out;
-			}
-
-			rte_prefetch0((void *)(uintptr_t)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;
 
-			desc_offset = 0;
-			desc_avail  = buf_vec[vec_idx].buf_len;
+			rte_prefetch0((void *)(uintptr_t)buf_addr);
 
-			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);
 		}
 
 		/*
@@ -1175,7 +1120,8 @@ rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
 		if (unlikely(fill_vec_buf(dev, vq,
 						vq->last_avail_idx + i,
 						&nr_vec, buf_vec,
-						&head_idx, &dummy_len) < 0))
+						&head_idx, &dummy_len,
+						VHOST_ACCESS_RW) < 0))
 			break;
 
 		if (likely(dev->dequeue_zero_copy == 0))
-- 
2.14.4

Re: [dpdk-dev] [PATCH v2 4/7] vhost: translate iovas at vectors fill time

[Tiwei Bie]

On Sat, Jun 23, 2018 at 09:11:24AM +0200, Maxime Coquelin wrote:
> This patch aims at simplifying the desc to mbuf and mbuf to desc
> copy functions. It performs the iova to hva translations at
> vectors fill time.
> 
> Doing this, in case desc buffer isn't contiguous in hva space,
> it gets split into multiple vectors.
> 
> Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
> ---
>  lib/librte_vhost/vhost.h      |   1 +
>  lib/librte_vhost/virtio_net.c | 340 ++++++++++++++++++------------------------
>  2 files changed, 144 insertions(+), 197 deletions(-)
> 
> diff --git a/lib/librte_vhost/vhost.h b/lib/librte_vhost/vhost.h
> index 786a74f64..e3b2ed2ff 100644
> --- a/lib/librte_vhost/vhost.h
> +++ b/lib/librte_vhost/vhost.h
> @@ -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;
> diff --git a/lib/librte_vhost/virtio_net.c b/lib/librte_vhost/virtio_net.c
> index 4816e8003..1ab1edd67 100644
> --- a/lib/librte_vhost/virtio_net.c
> +++ b/lib/librte_vhost/virtio_net.c
> @@ -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;
>  
> @@ -267,10 +267,31 @@ fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
>  		}
>  
>  		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) {

We also need to check whether:

vec_id >= BUF_VECTOR_MAX

> +			uint64_t desc_addr;
> +			uint64_t desc_chunck_len = desc_avail;
> +
> +			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 +314,8 @@ 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 +337,8 @@ 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_RO) < 0))

reserve_avail_buf() is called by virtio_dev_rx(),
so the write perm is needed.

>  			return -1;
>  		len = RTE_MIN(len, size);
>  		update_shadow_used_ring(vq, head_idx, len);
> @@ -334,21 +357,22 @@ reserve_avail_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
>  			return -1;
>  	}
>  
> +	*nr_vec = vec_idx;
> +
>  	return 0;
>  }
[...]
> @@ -455,18 +454,12 @@ 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;

There is no need to have two ' ' after '='.

>  					rte_memcpy((void *)(uintptr_t)dst,
>  							(void *)(uintptr_t)src,
>  							len);
[...]
>  
>  		/*
> @@ -1175,7 +1120,8 @@ rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
>  		if (unlikely(fill_vec_buf(dev, vq,
>  						vq->last_avail_idx + i,
>  						&nr_vec, buf_vec,
> -						&head_idx, &dummy_len) < 0))
> +						&head_idx, &dummy_len,
> +						VHOST_ACCESS_RW) < 0))

This is dequeue path, so _RO should be used.

>  			break;
>  
>  		if (likely(dev->dequeue_zero_copy == 0))
> -- 
> 2.14.4
> 

Re: [dpdk-dev] [PATCH v2 4/7] vhost: translate iovas at vectors fill time

[Maxime Coquelin]

On 06/25/2018 04:21 AM, Tiwei Bie wrote:
> On Sat, Jun 23, 2018 at 09:11:24AM +0200, Maxime Coquelin wrote:
>> This patch aims at simplifying the desc to mbuf and mbuf to desc
>> copy functions. It performs the iova to hva translations at
>> vectors fill time.
>>
>> Doing this, in case desc buffer isn't contiguous in hva space,
>> it gets split into multiple vectors.
>>
>> Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
>> ---
>>   lib/librte_vhost/vhost.h      |   1 +
>>   lib/librte_vhost/virtio_net.c | 340 ++++++++++++++++++------------------------
>>   2 files changed, 144 insertions(+), 197 deletions(-)
>>
>> diff --git a/lib/librte_vhost/vhost.h b/lib/librte_vhost/vhost.h
>> index 786a74f64..e3b2ed2ff 100644
>> --- a/lib/librte_vhost/vhost.h
>> +++ b/lib/librte_vhost/vhost.h
>> @@ -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;
>> diff --git a/lib/librte_vhost/virtio_net.c b/lib/librte_vhost/virtio_net.c
>> index 4816e8003..1ab1edd67 100644
>> --- a/lib/librte_vhost/virtio_net.c
>> +++ b/lib/librte_vhost/virtio_net.c
>> @@ -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;
>>   
>> @@ -267,10 +267,31 @@ fill_vec_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
>>   		}
>>   
>>   		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) {
> 
> We also need to check whether:
> 
> vec_id >= BUF_VECTOR_MAX

Right.

>> +			uint64_t desc_addr;
>> +			uint64_t desc_chunck_len = desc_avail;
>> +
>> +			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 +314,8 @@ 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 +337,8 @@ 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_RO) < 0))
> 
> reserve_avail_buf() is called by virtio_dev_rx(),
> so the write perm is needed.
Right.

To avoid having to pass the perms, I wonder if it wouldn't be better to
rely on the descriptors' VRING_DESC_F_WRITE flag.

>>   			return -1;
>>   		len = RTE_MIN(len, size);
>>   		update_shadow_used_ring(vq, head_idx, len);
>> @@ -334,21 +357,22 @@ reserve_avail_buf(struct virtio_net *dev, struct vhost_virtqueue *vq,
>>   			return -1;
>>   	}
>>   
>> +	*nr_vec = vec_idx;
>> +
>>   	return 0;
>>   }
> [...]
>> @@ -455,18 +454,12 @@ 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;
> 
> There is no need to have two ' ' after '='.

Agree.

>>   					rte_memcpy((void *)(uintptr_t)dst,
>>   							(void *)(uintptr_t)src,
>>   							len);
> [...]
>>   
>>   		/*
>> @@ -1175,7 +1120,8 @@ rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
>>   		if (unlikely(fill_vec_buf(dev, vq,
>>   						vq->last_avail_idx + i,
>>   						&nr_vec, buf_vec,
>> -						&head_idx, &dummy_len) < 0))
>> +						&head_idx, &dummy_len,
>> +						VHOST_ACCESS_RW) < 0))
> 
> This is dequeue path, so _RO should be used.

Right.

>>   			break;
>>   
>>   		if (likely(dev->dequeue_zero_copy == 0))
>> -- 
>> 2.14.4
>>

Thanks the review,
Maxime

Re: [dpdk-dev] [PATCH v2 4/7] vhost: translate iovas at vectors fill time

[Tiwei Bie]

On Mon, Jun 25, 2018 at 09:19:34AM +0200, Maxime Coquelin wrote:
> On 06/25/2018 04:21 AM, Tiwei Bie wrote:
> > On Sat, Jun 23, 2018 at 09:11:24AM +0200, Maxime Coquelin wrote:
[...]
> > > @@ -293,7 +314,8 @@ 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 +337,8 @@ 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_RO) < 0))
> > 
> > reserve_avail_buf() is called by virtio_dev_rx(),
> > so the write perm is needed.
> Right.
> 
> To avoid having to pass the perms, I wonder if it wouldn't be better to
> rely on the descriptors' VRING_DESC_F_WRITE flag.
> 

Currently, DPDK vhost net doesn't check this flag,
so it could cause problems in some cases. If we
want to rely on this flag, I think we still need
to pass something similar to tell fill_vec_buf()
whether the bufs will be written or read, so the
flag can be checked.

Best regards,
Tiwei Bie

Re: [dpdk-dev] [PATCH v2 4/7] vhost: translate iovas at vectors fill time

[Maxime Coquelin]

On 06/25/2018 12:31 PM, Tiwei Bie wrote:
> On Mon, Jun 25, 2018 at 09:19:34AM +0200, Maxime Coquelin wrote:
>> On 06/25/2018 04:21 AM, Tiwei Bie wrote:
>>> On Sat, Jun 23, 2018 at 09:11:24AM +0200, Maxime Coquelin wrote:
> [...]
>>>> @@ -293,7 +314,8 @@ 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 +337,8 @@ 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_RO) < 0))
>>>
>>> reserve_avail_buf() is called by virtio_dev_rx(),
>>> so the write perm is needed.
>> Right.
>>
>> To avoid having to pass the perms, I wonder if it wouldn't be better to
>> rely on the descriptors' VRING_DESC_F_WRITE flag.
>>
> 
> Currently, DPDK vhost net doesn't check this flag,
> so it could cause problems in some cases. If we
> want to rely on this flag, I think we still need
> to pass something similar to tell fill_vec_buf()
> whether the bufs will be written or read, so the
> flag can be checked.

Right, let's keep the perm parameter for now.

Thanks,
Maxime

> Best regards,
> Tiwei Bie
> 

[dpdk-dev] [PATCH v2 5/7] vhost: improve prefetching in dequeue path

[Maxime Coquelin]

This is an optimization to prefetch next buffer while the
current one is being processed.

Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
---
 lib/librte_vhost/virtio_net.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/lib/librte_vhost/virtio_net.c b/lib/librte_vhost/virtio_net.c
index 1ab1edd67..0980efe53 100644
--- a/lib/librte_vhost/virtio_net.c
+++ b/lib/librte_vhost/virtio_net.c
@@ -895,7 +895,13 @@ copy_desc_to_mbuf(struct virtio_net *dev, struct vhost_virtqueue *vq,
 			buf_iova = buf_vec[vec_idx].buf_iova;
 			buf_len = buf_vec[vec_idx].buf_len;
 
-			rte_prefetch0((void *)(uintptr_t)buf_addr);
+			/*
+			 * Prefecth desc n + 1 buffer while
+			 * desc n buffer is processed.
+			 */
+			if (vec_idx + 1 < nr_vec)
+				rte_prefetch0((void *)(uintptr_t)
+						buf_vec[vec_idx + 1].buf_addr);
 
 			buf_offset = 0;
 			buf_avail  = buf_len;
@@ -1127,6 +1133,8 @@ rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
 		if (likely(dev->dequeue_zero_copy == 0))
 			update_shadow_used_ring(vq, head_idx, 0);
 
+		rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
+
 		pkts[i] = rte_pktmbuf_alloc(mbuf_pool);
 		if (unlikely(pkts[i] == NULL)) {
 			RTE_LOG(ERR, VHOST_DATA,
-- 
2.14.4

[dpdk-dev] [PATCH v2 6/7] vhost: prefetch first descriptor in dequeue path

[Maxime Coquelin]

Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
---
 lib/librte_vhost/virtio_net.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/lib/librte_vhost/virtio_net.c b/lib/librte_vhost/virtio_net.c
index 0980efe53..115bb885d 100644
--- a/lib/librte_vhost/virtio_net.c
+++ b/lib/librte_vhost/virtio_net.c
@@ -1076,6 +1076,8 @@ rte_vhost_dequeue_burst(int vid, uint16_t queue_id,
 		vq->shadow_used_idx = 0;
 	}
 
+	rte_prefetch0(&vq->avail->ring[vq->last_avail_idx & (vq->size - 1)]);
+
 	/*
 	 * Construct a RARP broadcast packet, and inject it to the "pkts"
 	 * array, to looks like that guest actually send such packet.
-- 
2.14.4

[dpdk-dev] [PATCH v2 7/7] vhost: improve prefetching in enqueue path

[Maxime Coquelin]

This is an optimization to prefetch next buffer while the
current one is being processed.

Signed-off-by: Maxime Coquelin <maxime.coquelin@redhat.com>
---
 lib/librte_vhost/virtio_net.c | 12 +++++++++---
 1 file changed, 9 insertions(+), 3 deletions(-)

diff --git a/lib/librte_vhost/virtio_net.c b/lib/librte_vhost/virtio_net.c
index 115bb885d..2c13a59a2 100644
--- a/lib/librte_vhost/virtio_net.c
+++ b/lib/librte_vhost/virtio_net.c
@@ -387,6 +387,9 @@ copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 	buf_iova = buf_vec[vec_idx].buf_iova;
 	buf_len = buf_vec[vec_idx].buf_len;
 
+	if (nr_vec > 1)
+		rte_prefetch0((void *)(uintptr_t)buf_vec[1].buf_addr);
+
 	if (unlikely(buf_len < dev->vhost_hlen && nr_vec <= 1)) {
 		error = -1;
 		goto out;
@@ -398,7 +401,6 @@ copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 		hdr = &tmp_hdr;
 	else
 		hdr = (struct virtio_net_hdr_mrg_rxbuf *)(uintptr_t)hdr_addr;
-	rte_prefetch0((void *)(uintptr_t)hdr_addr);
 
 	VHOST_LOG_DEBUG(VHOST_DATA, "(%d) RX: num merge buffers %d\n",
 		dev->vid, num_buffers);
@@ -430,8 +432,10 @@ copy_mbuf_to_desc(struct virtio_net *dev, struct vhost_virtqueue *vq,
 			buf_iova = buf_vec[vec_idx].buf_iova;
 			buf_len = buf_vec[vec_idx].buf_len;
 
-			/* Prefetch buffer address. */
-			rte_prefetch0((void *)(uintptr_t)buf_addr);
+			/* Prefetch next buffer address. */
+			if (vec_idx + 1 < nr_vec)
+				rte_prefetch0((void *)(uintptr_t)
+						buf_vec[vec_idx + 1].buf_addr);
 			buf_offset = 0;
 			buf_avail  = buf_len;
 		}
@@ -573,6 +577,8 @@ virtio_dev_rx(struct virtio_net *dev, uint16_t queue_id,
 			break;
 		}
 
+		rte_prefetch0((void *)(uintptr_t)buf_vec[0].buf_addr);
+
 		VHOST_LOG_DEBUG(VHOST_DATA, "(%d) current index %d | end index %d\n",
 			dev->vid, vq->last_avail_idx,
 			vq->last_avail_idx + num_buffers);
-- 
2.14.4