[dpdk-dev] [PATCH 3/4] virtio: use indirect ring elements

[Stephen Hemminger]

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 at networkplumber.org>
---
 drivers/net/virtio/virtio_ethdev.c | 11 +++++---
 drivers/net/virtio/virtio_ethdev.h |  3 ++-
 drivers/net/virtio/virtio_rxtx.c   | 51 ++++++++++++++++++++++++++++++--------
 drivers/net/virtio/virtqueue.h     |  8 ++++++
 4 files changed, 57 insertions(+), 16 deletions(-)

diff --git a/drivers/net/virtio/virtio_ethdev.c b/drivers/net/virtio/virtio_ethdev.c
index 465d3cd..bcfb87b 100644
--- a/drivers/net/virtio/virtio_ethdev.c
+++ b/drivers/net/virtio/virtio_ethdev.c
@@ -359,12 +359,15 @@ int virtio_dev_queue_setup(struct rte_eth_dev *dev,
 	if (queue_type == VTNET_TQ) {
 		/*
 		 * 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);
+			 dev->data->port_id, queue_idx);
+
+		vq->virtio_net_hdr_mz =
+			rte_memzone_reserve_aligned(vq_name,
+						    vq_size * sizeof(struct virtio_tx_region),
+						    socket_id, 0, RTE_CACHE_LINE_SIZE);
 		if (vq->virtio_net_hdr_mz == NULL) {
 			if (rte_errno == EEXIST)
 				vq->virtio_net_hdr_mz =
diff --git a/drivers/net/virtio/virtio_ethdev.h b/drivers/net/virtio/virtio_ethdev.h
index 9026d42..07a9265 100644
--- a/drivers/net/virtio/virtio_ethdev.h
+++ b/drivers/net/virtio/virtio_ethdev.h
@@ -64,7 +64,8 @@
 	 1u << VIRTIO_NET_F_CTRL_VQ	  |	\
 	 1u << VIRTIO_NET_F_CTRL_RX	  |	\
 	 1u << VIRTIO_NET_F_CTRL_VLAN	  |	\
-	 1u << VIRTIO_NET_F_MRG_RXBUF)
+	 1u << VIRTIO_NET_F_MRG_RXBUF     |	\
+	 1u << VIRTIO_RING_F_INDIRECT_DESC)
 
 /*
  * CQ function prototype
diff --git a/drivers/net/virtio/virtio_rxtx.c b/drivers/net/virtio/virtio_rxtx.c
index dbe6665..8979695 100644
--- a/drivers/net/virtio/virtio_rxtx.c
+++ b/drivers/net/virtio/virtio_rxtx.c
@@ -199,14 +199,15 @@ 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;
-	uint16_t head_size = txvq->hw->vtnet_hdr_size;
+	unsigned long offs;
 
 	if (unlikely(txvq->vq_free_cnt == 0))
 		return -ENOSPC;
@@ -220,11 +221,26 @@ virtqueue_enqueue_xmit(struct virtqueue *txvq, struct rte_mbuf *cookie)
 	dxp = &txvq->vq_descx[idx];
 	dxp->cookie = (void *)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 = (uint32_t)head_size;
+
+	if (use_indirect) {
+		offs = offsetof(struct virtio_tx_region, tx_indir)
+			+ idx * sizeof(struct virtio_tx_region);
+
+		start_dp[idx].addr = txvq->virtio_net_hdr_mem + offs;
+		start_dp[idx].len = sizeof(struct vring_desc);
+		start_dp[idx].flags = VRING_DESC_F_INDIRECT;
+
+		start_dp = (struct vring_desc *)
+			((char *)txvq->virtio_net_hdr_mz->addr + offs);
+		idx = 0;
+	}
+
+	offs = offsetof(struct virtio_tx_region, tx_hdr)
+		+ idx * sizeof(struct virtio_tx_region);
+
+	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--) {
@@ -236,7 +252,12 @@ 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;
@@ -760,7 +781,15 @@ 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];
-		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 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 (need > 0) {
@@ -769,7 +798,7 @@ 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");
@@ -787,7 +816,7 @@ 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 7789411..a9410b4 100644
--- a/drivers/net/virtio/virtqueue.h
+++ b/drivers/net/virtio/virtqueue.h
@@ -237,6 +237,14 @@ 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)));
+};
+
 /**
  * Tell the backend not to interrupt us.
  */
-- 
2.1.4