[dpdk-dev] [PATCH 1/3] virtio: use indirect ring elements

[Stephen Hemminger <stephen@networkplumber.org>]

The virtio ring in QEMU/KVM is usually limited to 256 entries
and the normal way that virtio driver was queuing mbufs required
nsegs + 1 ring elements. By using the indirect ring element feature
if available, each packet will take only one ring slot even for
multi-segment packets.

Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
---
 drivers/net/virtio/virtio_ethdev.c | 44 +++++++++++++++++-------
 drivers/net/virtio/virtio_rxtx.c   | 69 +++++++++++++++++++++++++++++---------
 drivers/net/virtio/virtqueue.h     | 19 +++++++++++
 3 files changed, 104 insertions(+), 28 deletions(-)

diff --git a/drivers/net/virtio/virtio_ethdev.c b/drivers/net/virtio/virtio_ethdev.c
index caa970c..95c2203 100644
--- a/drivers/net/virtio/virtio_ethdev.c
+++ b/drivers/net/virtio/virtio_ethdev.c
@@ -387,27 +387,47 @@ int virtio_dev_queue_setup(struct rte_eth_dev *dev,
 	vq->virtio_net_hdr_mem = 0;
 
 	if (queue_type == VTNET_TQ) {
+		const struct rte_memzone *hdr_mz;
+		struct virtio_tx_region *txr;
+		unsigned int i;
+
 		/*
 		 * For each xmit packet, allocate a virtio_net_hdr
+		 * and indirect ring elements
 		 */
 		snprintf(vq_name, sizeof(vq_name), "port%d_tvq%d_hdrzone",
-			dev->data->port_id, queue_idx);
-		vq->virtio_net_hdr_mz = rte_memzone_reserve_aligned(vq_name,
-			vq_size * hw->vtnet_hdr_size,
-			socket_id, 0, RTE_CACHE_LINE_SIZE);
-		if (vq->virtio_net_hdr_mz == NULL) {
+			 dev->data->port_id, queue_idx);
+		hdr_mz = rte_memzone_reserve_aligned(vq_name,
+						     vq_size * sizeof(*txr),
+						     socket_id, 0,
+						     RTE_CACHE_LINE_SIZE);
+		if (hdr_mz == NULL) {
 			if (rte_errno == EEXIST)
-				vq->virtio_net_hdr_mz =
-					rte_memzone_lookup(vq_name);
-			if (vq->virtio_net_hdr_mz == NULL) {
+				hdr_mz = rte_memzone_lookup(vq_name);
+			if (hdr_mz == NULL) {
 				rte_free(vq);
 				return -ENOMEM;
 			}
 		}
-		vq->virtio_net_hdr_mem =
-			vq->virtio_net_hdr_mz->phys_addr;
-		memset(vq->virtio_net_hdr_mz->addr, 0,
-			vq_size * hw->vtnet_hdr_size);
+		vq->virtio_net_hdr_mz = hdr_mz;
+		vq->virtio_net_hdr_mem = hdr_mz->phys_addr;
+
+		txr = hdr_mz->addr;
+		memset(txr, 0, vq_size * sizeof(*txr));
+		for (i = 0; i < vq_size; i++) {
+			struct vring_desc *start_dp = txr[i].tx_indir;
+
+			vring_desc_init(start_dp, RTE_DIM(txr[i].tx_indir));
+
+			/* first indirect descriptor is always the tx header */
+			start_dp->addr = vq->virtio_net_hdr_mem
+				+ i * sizeof(*txr)
+				+ offsetof(struct virtio_tx_region, tx_hdr);
+
+			start_dp->len = vq->hw->vtnet_hdr_size;
+			start_dp->flags = VRING_DESC_F_NEXT;
+		}
+
 	} else if (queue_type == VTNET_CQ) {
 		/* Allocate a page for control vq command, data and status */
 		snprintf(vq_name, sizeof(vq_name), "port%d_cvq_hdrzone",
diff --git a/drivers/net/virtio/virtio_rxtx.c b/drivers/net/virtio/virtio_rxtx.c
index e96352c..5fe3eec 100644
--- a/drivers/net/virtio/virtio_rxtx.c
+++ b/drivers/net/virtio/virtio_rxtx.c
@@ -209,14 +209,15 @@ virtqueue_enqueue_recv_refill(struct virtqueue *vq, struct rte_mbuf *cookie)
 }
 
 static int
-virtqueue_enqueue_xmit(struct virtqueue *txvq, struct rte_mbuf *cookie)
+virtqueue_enqueue_xmit(struct virtqueue *txvq, struct rte_mbuf *cookie,
+		       int use_indirect)
 {
 	struct vq_desc_extra *dxp;
 	struct vring_desc *start_dp;
 	uint16_t seg_num = cookie->nb_segs;
-	uint16_t needed = 1 + seg_num;
+	uint16_t needed = use_indirect ? 1 : 1 + seg_num;
 	uint16_t head_idx, idx;
-	size_t head_size = txvq->hw->vtnet_hdr_size;
+	unsigned long offs;
 
 	if (unlikely(txvq->vq_free_cnt == 0))
 		return -ENOSPC;
@@ -232,10 +233,37 @@ virtqueue_enqueue_xmit(struct virtqueue *txvq, struct rte_mbuf *cookie)
 	dxp->ndescs = needed;
 
 	start_dp = txvq->vq_ring.desc;
-	start_dp[idx].addr =
-		txvq->virtio_net_hdr_mem + idx * head_size;
-	start_dp[idx].len = head_size;
-	start_dp[idx].flags = VRING_DESC_F_NEXT;
+
+	if (use_indirect) {
+		/* setup tx ring slot to point to indirect
+		 * descriptor list stored in reserved region.
+		 *
+		 * the first slot in indirect ring is already preset
+		 * to point to the header in reserved region
+		 */
+		struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
+
+		offs = idx * sizeof(struct virtio_tx_region)
+			+ offsetof(struct virtio_tx_region, tx_indir);
+
+		start_dp[idx].addr  = txvq->virtio_net_hdr_mem + offs;
+		start_dp[idx].len   = (seg_num + 1) * sizeof(struct vring_desc);
+		start_dp[idx].flags = VRING_DESC_F_INDIRECT;
+
+		/* loop below will fill in rest of the indirect elements */
+		start_dp = txr[idx].tx_indir;
+		idx = 0;
+	} else {
+		/* setup first tx ring slot to point to header
+		 * stored in reserved region.
+		 */
+		offs = idx * sizeof(struct virtio_tx_region)
+			+ offsetof(struct virtio_tx_region, tx_hdr);
+
+		start_dp[idx].addr  = txvq->virtio_net_hdr_mem + offs;
+		start_dp[idx].len   = txvq->hw->vtnet_hdr_size;
+		start_dp[idx].flags = VRING_DESC_F_NEXT;
+	}
 
 	for (; ((seg_num > 0) && (cookie != NULL)); seg_num--) {
 		idx = start_dp[idx].next;
@@ -246,7 +274,12 @@ virtqueue_enqueue_xmit(struct virtqueue *txvq, struct rte_mbuf *cookie)
 	}
 
 	start_dp[idx].flags &= ~VRING_DESC_F_NEXT;
-	idx = start_dp[idx].next;
+
+	if (use_indirect)
+		idx = txvq->vq_ring.desc[head_idx].next;
+	else
+		idx = start_dp[idx].next;
+
 	txvq->vq_desc_head_idx = idx;
 	if (txvq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
 		txvq->vq_desc_tail_idx = idx;
@@ -289,10 +322,7 @@ virtio_dev_vring_start(struct virtqueue *vq, int queue_type)
 	vq->vq_free_cnt = vq->vq_nentries;
 	memset(vq->vq_descx, 0, sizeof(struct vq_desc_extra) * vq->vq_nentries);
 
-	/* Chain all the descriptors in the ring with an END */
-	for (i = 0; i < size - 1; i++)
-		vr->desc[i].next = (uint16_t)(i + 1);
-	vr->desc[i].next = VQ_RING_DESC_CHAIN_END;
+	vring_desc_init(vr->desc, size);
 
 	/*
 	 * Disable device(host) interrupting guest
@@ -848,8 +878,15 @@ virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 
 	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
 		struct rte_mbuf *txm = tx_pkts[nb_tx];
-		/* Need one more descriptor for virtio header. */
-		int need = txm->nb_segs - txvq->vq_free_cnt + 1;
+		int use_indirect, slots, need;
+
+		use_indirect = vtpci_with_feature(txvq->hw,
+						  VIRTIO_RING_F_INDIRECT_DESC)
+			&& (txm->nb_segs < VIRTIO_MAX_TX_INDIRECT);
+
+		/* How many main ring entries are needed to this Tx? */
+		slots = use_indirect ? 1 : 1 + txm->nb_segs;
+		need = slots - txvq->vq_free_cnt;
 
 		/* Positive value indicates it need free vring descriptors */
 		if (unlikely(need > 0)) {
@@ -858,7 +895,7 @@ virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 			need = RTE_MIN(need, (int)nb_used);
 
 			virtio_xmit_cleanup(txvq, need);
-			need = txm->nb_segs - txvq->vq_free_cnt + 1;
+			need = slots - txvq->vq_free_cnt;
 			if (unlikely(need > 0)) {
 				PMD_TX_LOG(ERR,
 					   "No free tx descriptors to transmit");
@@ -876,7 +913,7 @@ virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 		}
 
 		/* Enqueue Packet buffers */
-		error = virtqueue_enqueue_xmit(txvq, txm);
+		error = virtqueue_enqueue_xmit(txvq, txm, use_indirect);
 		if (unlikely(error)) {
 			if (error == ENOSPC)
 				PMD_TX_LOG(ERR, "virtqueue_enqueue Free count = 0");
diff --git a/drivers/net/virtio/virtqueue.h b/drivers/net/virtio/virtqueue.h
index 68e0b4b..4e9239e 100644
--- a/drivers/net/virtio/virtqueue.h
+++ b/drivers/net/virtio/virtqueue.h
@@ -243,6 +243,25 @@ struct virtio_net_hdr_mrg_rxbuf {
 	uint16_t num_buffers; /**< Number of merged rx buffers */
 };
 
+/* Region reserved to allow for transmit header and indirect ring */
+#define VIRTIO_MAX_TX_INDIRECT 8
+struct virtio_tx_region {
+	struct virtio_net_hdr_mrg_rxbuf tx_hdr;
+	struct vring_desc tx_indir[VIRTIO_MAX_TX_INDIRECT]
+			   __attribute__((__aligned__(16)));
+};
+
+/* Chain all the descriptors in the ring with an END */
+static inline void
+vring_desc_init(struct vring_desc *dp, uint16_t n)
+{
+	uint16_t i;
+
+	for (i = 0; i < n - 1; i++)
+		dp[i].next = (uint16_t)(i + 1);
+	dp[i].next = VQ_RING_DESC_CHAIN_END;
+}
+
 /**
  * Tell the backend not to interrupt us.
  */
-- 
2.1.4