[dpdk-dev] [PATCH v2 12/15] net/avp: packet transmit functions

[Allain Legacy]

Adds support for packet transmit functions so that an application can send
packets to the host application via an AVP device queue.  Both the simple
and scattered functions are supported.

Signed-off-by: Allain Legacy <allain.legacy at windriver.com>
Signed-off-by: Matt Peters <matt.peters at windriver.com>
---
 drivers/net/avp/avp_ethdev.c | 419 +++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 419 insertions(+)

diff --git a/drivers/net/avp/avp_ethdev.c b/drivers/net/avp/avp_ethdev.c
index cd0b0c0..514e27d 100644
--- a/drivers/net/avp/avp_ethdev.c
+++ b/drivers/net/avp/avp_ethdev.c
@@ -92,12 +92,28 @@ static uint16_t avp_recv_scattered_pkts(void *rx_queue,
 static uint16_t avp_recv_pkts(void *rx_queue,
 			      struct rte_mbuf **rx_pkts,
 			      uint16_t nb_pkts);
+
+static uint16_t avp_xmit_scattered_pkts(void *tx_queue,
+					struct rte_mbuf **tx_pkts,
+					uint16_t nb_pkts);
+
+static uint16_t avp_xmit_pkts(void *tx_queue,
+			      struct rte_mbuf **tx_pkts,
+			      uint16_t nb_pkts);
+
 static void avp_dev_rx_queue_release(void *rxq);
 static void avp_dev_tx_queue_release(void *txq);
+
+static void avp_dev_stats_get(struct rte_eth_dev *dev,
+			      struct rte_eth_stats *stats);
+static void avp_dev_stats_reset(struct rte_eth_dev *dev);
+
+
 #define AVP_DEV_TO_PCI(eth_dev) RTE_DEV_TO_PCI((eth_dev)->device)
 
 
 #define RTE_AVP_MAX_RX_BURST 64
+#define RTE_AVP_MAX_TX_BURST 64
 #define RTE_AVP_MAX_MAC_ADDRS 1
 #define RTE_AVP_MIN_RX_BUFSIZE ETHER_MIN_LEN
 
@@ -139,6 +155,8 @@ static uint16_t avp_recv_pkts(void *rx_queue,
 	.dev_configure       = avp_dev_configure,
 	.dev_infos_get       = avp_dev_info_get,
 	.vlan_offload_set    = avp_vlan_offload_set,
+	.stats_get           = avp_dev_stats_get,
+	.stats_reset         = avp_dev_stats_reset,
 	.link_update         = avp_dev_link_update,
 	.rx_queue_setup      = avp_dev_rx_queue_setup,
 	.rx_queue_release    = avp_dev_rx_queue_release,
@@ -946,6 +964,7 @@ struct avp_queue {
 	pci_dev = AVP_DEV_TO_PCI(eth_dev);
 	eth_dev->dev_ops = &avp_eth_dev_ops;
 	eth_dev->rx_pkt_burst = &avp_recv_pkts;
+	eth_dev->tx_pkt_burst = &avp_xmit_pkts;
 
 	if (rte_eal_process_type() != RTE_PROC_PRIMARY) {
 		/*
@@ -959,6 +978,7 @@ struct avp_queue {
 				    "AVP device configured "
 				    "for chained mbufs\n");
 			eth_dev->rx_pkt_burst = avp_recv_scattered_pkts;
+			eth_dev->tx_pkt_burst = avp_xmit_scattered_pkts;
 		}
 		return 0;
 	}
@@ -1109,6 +1129,7 @@ struct avp_queue {
 				    "for chained mbufs\n");
 			eth_dev->data->scattered_rx = 1;
 			eth_dev->rx_pkt_burst = avp_recv_scattered_pkts;
+			eth_dev->tx_pkt_burst = avp_xmit_scattered_pkts;
 		}
 	}
 
@@ -1586,6 +1607,340 @@ struct avp_queue {
 	return count;
 }
 
+/*
+ * Copy a chained mbuf to a set of host buffers.  This function assumes that
+ * there are sufficient destination buffers to contain the entire source
+ * packet.
+ */
+static inline uint16_t
+avp_dev_copy_to_buffers(struct avp_dev *avp,
+			struct rte_mbuf *mbuf,
+			struct rte_avp_desc **buffers,
+			unsigned int count)
+{
+	struct rte_avp_desc *previous_buf = NULL;
+	struct rte_avp_desc *first_buf = NULL;
+	struct rte_avp_desc *pkt_buf;
+	struct rte_avp_desc *buf;
+	size_t total_length;
+	struct rte_mbuf *m;
+	size_t copy_length;
+	size_t src_offset;
+	char *pkt_data;
+	unsigned int i;
+
+	__rte_mbuf_sanity_check(mbuf, 1);
+
+	m = mbuf;
+	src_offset = 0;
+	total_length = rte_pktmbuf_pkt_len(m);
+	for (i = 0; (i < count) && (m != NULL); i++) {
+		/* fill each destination buffer */
+		buf = buffers[i];
+
+		if (i < count - 1) {
+			/* prefetch next entry while processing this one */
+			pkt_buf = avp_dev_translate_buffer(avp, buffers[i + 1]);
+			rte_prefetch0(pkt_buf);
+		}
+
+		/* Adjust pointers for guest addressing */
+		pkt_buf = avp_dev_translate_buffer(avp, buf);
+		pkt_data = avp_dev_translate_buffer(avp, pkt_buf->data);
+
+		/* setup the buffer chain */
+		if (previous_buf != NULL)
+			previous_buf->next = buf;
+		else
+			first_buf = pkt_buf;
+
+		previous_buf = pkt_buf;
+
+		do {
+			/*
+			 * copy as many source mbuf segments as will fit in the
+			 * destination buffer.
+			 */
+			copy_length = RTE_MIN((avp->host_mbuf_size -
+					       pkt_buf->data_len),
+					      (rte_pktmbuf_data_len(m) -
+					       src_offset));
+			rte_memcpy(RTE_PTR_ADD(pkt_data, pkt_buf->data_len),
+				   RTE_PTR_ADD(rte_pktmbuf_mtod(m, void *),
+					       src_offset),
+				   copy_length);
+			pkt_buf->data_len += copy_length;
+			src_offset += copy_length;
+
+			if (likely(src_offset == rte_pktmbuf_data_len(m))) {
+				/* need a new source buffer */
+				m = m->next;
+				src_offset = 0;
+			}
+
+			if (unlikely(pkt_buf->data_len ==
+				     avp->host_mbuf_size)) {
+				/* need a new destination buffer */
+				break;
+			}
+
+		} while (m != NULL);
+	}
+
+	first_buf->nb_segs = count;
+	first_buf->pkt_len = total_length;
+
+	if (mbuf->ol_flags & PKT_TX_VLAN_PKT) {
+		first_buf->ol_flags |= RTE_AVP_TX_VLAN_PKT;
+		first_buf->vlan_tci = mbuf->vlan_tci;
+	}
+
+	avp_dev_buffer_sanity_check(avp, buffers[0]);
+
+	return total_length;
+}
+
+
+static uint16_t
+avp_xmit_scattered_pkts(void *tx_queue,
+			struct rte_mbuf **tx_pkts,
+			uint16_t nb_pkts)
+{
+	struct rte_avp_desc *avp_bufs[(RTE_AVP_MAX_TX_BURST *
+				       RTE_AVP_MAX_MBUF_SEGMENTS)];
+	struct avp_queue *txq = (struct avp_queue *)tx_queue;
+	struct rte_avp_desc *tx_bufs[RTE_AVP_MAX_TX_BURST];
+	struct avp_dev *avp = txq->avp;
+	struct rte_avp_fifo *alloc_q;
+	struct rte_avp_fifo *tx_q;
+	unsigned int count, avail, n;
+	unsigned int orig_nb_pkts;
+	struct rte_mbuf *m;
+	unsigned int required;
+	unsigned int segments;
+	unsigned int tx_bytes;
+	unsigned int i;
+
+	orig_nb_pkts = nb_pkts;
+	if (unlikely(avp->flags & RTE_AVP_F_DETACHED)) {
+		/* VM live migration in progress */
+		/* TODO ... buffer for X packets then drop? */
+		txq->errors += nb_pkts;
+		return 0;
+	}
+
+	tx_q = avp->tx_q[txq->queue_id];
+	alloc_q = avp->alloc_q[txq->queue_id];
+
+	/* limit the number of transmitted packets to the max burst size */
+	if (unlikely(nb_pkts > RTE_AVP_MAX_TX_BURST))
+		nb_pkts = RTE_AVP_MAX_TX_BURST;
+
+	/* determine how many buffers are available to copy into */
+	avail = avp_fifo_count(alloc_q);
+	if (unlikely(avail > (RTE_AVP_MAX_TX_BURST *
+			      RTE_AVP_MAX_MBUF_SEGMENTS)))
+		avail = RTE_AVP_MAX_TX_BURST * RTE_AVP_MAX_MBUF_SEGMENTS;
+
+	/* determine how many slots are available in the transmit queue */
+	count = avp_fifo_free_count(tx_q);
+
+	/* determine how many packets can be sent */
+	nb_pkts = RTE_MIN(count, nb_pkts);
+
+	/* determine how many packets will fit in the available buffers */
+	count = 0;
+	segments = 0;
+	for (i = 0; i < nb_pkts; i++) {
+		m = tx_pkts[i];
+		if (likely(i < (unsigned int)nb_pkts - 1)) {
+			/* prefetch next entry while processing this one */
+			rte_prefetch0(tx_pkts[i + 1]);
+		}
+		required = (rte_pktmbuf_pkt_len(m) + avp->host_mbuf_size - 1) /
+			avp->host_mbuf_size;
+
+		if (unlikely((required == 0) ||
+			     (required > RTE_AVP_MAX_MBUF_SEGMENTS)))
+			break;
+		else if (unlikely(required + segments > avail))
+			break;
+		segments += required;
+		count++;
+	}
+	nb_pkts = count;
+
+	if (unlikely(nb_pkts == 0)) {
+		/* no available buffers, or no space on the tx queue */
+		txq->errors += orig_nb_pkts;
+		return 0;
+	}
+
+	PMD_TX_LOG(DEBUG, "Sending %u packets on Tx queue at %p\n",
+		   nb_pkts, tx_q);
+
+	/* retrieve sufficient send buffers */
+	n = avp_fifo_get(alloc_q, (void **)&avp_bufs, segments);
+	if (unlikely(n != segments)) {
+		PMD_TX_LOG(DEBUG, "Failed to allocate buffers "
+			   "n=%u, segments=%u, orig=%u\n",
+			   n, segments, orig_nb_pkts);
+		txq->errors += orig_nb_pkts;
+		return 0;
+	}
+
+	tx_bytes = 0;
+	count = 0;
+	for (i = 0; i < nb_pkts; i++) {
+		/* process each packet to be transmitted */
+		m = tx_pkts[i];
+
+		/* determine how many buffers are required for this packet */
+		required = (rte_pktmbuf_pkt_len(m) + avp->host_mbuf_size - 1) /
+			avp->host_mbuf_size;
+
+		tx_bytes += avp_dev_copy_to_buffers(avp, m,
+						    &avp_bufs[count], required);
+		tx_bufs[i] = avp_bufs[count];
+		count += required;
+
+		/* free the original mbuf */
+		rte_pktmbuf_free(m);
+	}
+
+	txq->packets += nb_pkts;
+	txq->bytes += tx_bytes;
+
+#ifdef RTE_LIBRTE_AVP_DEBUG_BUFFERS
+	for (i = 0; i < nb_pkts; i++)
+		avp_dev_buffer_sanity_check(avp, tx_bufs[i]);
+#endif
+
+	/* send the packets */
+	n = avp_fifo_put(tx_q, (void **)&tx_bufs[0], nb_pkts);
+	if (unlikely(n != orig_nb_pkts))
+		txq->errors += (orig_nb_pkts - n);
+
+	return n;
+}
+
+
+static uint16_t
+avp_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
+{
+	struct avp_queue *txq = (struct avp_queue *)tx_queue;
+	struct rte_avp_desc *avp_bufs[RTE_AVP_MAX_TX_BURST];
+	struct avp_dev *avp = txq->avp;
+	struct rte_avp_desc *pkt_buf;
+	struct rte_avp_fifo *alloc_q;
+	struct rte_avp_fifo *tx_q;
+	unsigned int count, avail, n;
+	struct rte_mbuf *m;
+	unsigned int pkt_len;
+	unsigned int tx_bytes;
+	char *pkt_data;
+	unsigned int i;
+
+	if (unlikely(avp->flags & RTE_AVP_F_DETACHED)) {
+		/* VM live migration in progress */
+		/* TODO ... buffer for X packets then drop?! */
+		txq->errors++;
+		return 0;
+	}
+
+	tx_q = avp->tx_q[txq->queue_id];
+	alloc_q = avp->alloc_q[txq->queue_id];
+
+	/* limit the number of transmitted packets to the max burst size */
+	if (unlikely(nb_pkts > RTE_AVP_MAX_TX_BURST))
+		nb_pkts = RTE_AVP_MAX_TX_BURST;
+
+	/* determine how many buffers are available to copy into */
+	avail = avp_fifo_count(alloc_q);
+
+	/* determine how many slots are available in the transmit queue */
+	count = avp_fifo_free_count(tx_q);
+
+	/* determine how many packets can be sent */
+	count = RTE_MIN(count, avail);
+	count = RTE_MIN(count, nb_pkts);
+
+	if (unlikely(count == 0)) {
+		/* no available buffers, or no space on the tx queue */
+		txq->errors++;
+		return 0;
+	}
+
+	PMD_TX_LOG(DEBUG, "Sending %u packets on Tx queue at %p\n",
+		   count, tx_q);
+
+	/* retrieve sufficient send buffers */
+	n = avp_fifo_get(alloc_q, (void **)&avp_bufs, count);
+	if (unlikely(n != count)) {
+		txq->errors++;
+		return 0;
+	}
+
+	tx_bytes = 0;
+	for (i = 0; i < count; i++) {
+		/* prefetch next entry while processing the current one */
+		if (i < count - 1) {
+			pkt_buf = avp_dev_translate_buffer(avp,
+							   avp_bufs[i + 1]);
+			rte_prefetch0(pkt_buf);
+		}
+
+		/* process each packet to be transmitted */
+		m = tx_pkts[i];
+
+		/* Adjust pointers for guest addressing */
+		pkt_buf = avp_dev_translate_buffer(avp, avp_bufs[i]);
+		pkt_data = avp_dev_translate_buffer(avp, pkt_buf->data);
+		pkt_len = rte_pktmbuf_pkt_len(m);
+
+		if (unlikely((pkt_len > avp->guest_mbuf_size) ||
+					 (pkt_len > avp->host_mbuf_size))) {
+			/*
+			 * application should be using the scattered transmit
+			 * function; send it truncated to avoid the performance
+			 * hit of having to manage returning the already
+			 * allocated buffer to the free list.  This should not
+			 * happen since the application should have set the
+			 * max_rx_pkt_len based on its MTU and it should be
+			 * policing its own packet sizes.
+			 */
+			txq->errors++;
+			pkt_len = RTE_MIN(avp->guest_mbuf_size,
+					  avp->host_mbuf_size);
+		}
+
+		/* copy data out of our mbuf and into the AVP buffer */
+		rte_memcpy(pkt_data, rte_pktmbuf_mtod(m, void *), pkt_len);
+		pkt_buf->pkt_len = pkt_len;
+		pkt_buf->data_len = pkt_len;
+		pkt_buf->nb_segs = 1;
+		pkt_buf->next = NULL;
+
+		if (m->ol_flags & PKT_TX_VLAN_PKT) {
+			pkt_buf->ol_flags |= RTE_AVP_TX_VLAN_PKT;
+			pkt_buf->vlan_tci = m->vlan_tci;
+		}
+
+		tx_bytes += pkt_len;
+
+		/* free the original mbuf */
+		rte_pktmbuf_free(m);
+	}
+
+	txq->packets += count;
+	txq->bytes += tx_bytes;
+
+	/* send the packets */
+	n = avp_fifo_put(tx_q, (void **)&avp_bufs[0], count);
+
+	return n;
+}
+
 static void
 avp_dev_rx_queue_release(void *rx_queue)
 {
@@ -1734,6 +2089,70 @@ struct avp_queue {
 	}
 }
 
+static void
+avp_dev_stats_get(struct rte_eth_dev *eth_dev, struct rte_eth_stats *stats)
+{
+	struct avp_dev *avp =
+		RTE_AVP_DEV_PRIVATE_TO_HW(eth_dev->data->dev_private);
+	unsigned int i;
+
+	memset(stats, 0, sizeof(*stats));
+	for (i = 0; i < avp->num_rx_queues; i++) {
+		struct avp_queue *rxq = avp->dev_data->rx_queues[i];
+
+		if (rxq) {
+			stats->ipackets += rxq->packets;
+			stats->ibytes += rxq->bytes;
+			stats->ierrors += rxq->errors;
+
+			stats->q_ipackets[i] += rxq->packets;
+			stats->q_ibytes[i] += rxq->bytes;
+			stats->q_errors[i] += rxq->errors;
+		}
+	}
+
+	for (i = 0; i < avp->num_tx_queues; i++) {
+		struct avp_queue *txq = avp->dev_data->tx_queues[i];
+
+		if (txq) {
+			stats->opackets += txq->packets;
+			stats->obytes += txq->bytes;
+			stats->oerrors += txq->errors;
+
+			stats->q_opackets[i] += txq->packets;
+			stats->q_obytes[i] += txq->bytes;
+			stats->q_errors[i] += txq->errors;
+		}
+	}
+}
+
+static void
+avp_dev_stats_reset(struct rte_eth_dev *eth_dev)
+{
+	struct avp_dev *avp =
+		RTE_AVP_DEV_PRIVATE_TO_HW(eth_dev->data->dev_private);
+	unsigned int i;
+
+	for (i = 0; i < avp->num_rx_queues; i++) {
+		struct avp_queue *rxq = avp->dev_data->rx_queues[i];
+
+		if (rxq) {
+			rxq->bytes = 0;
+			rxq->packets = 0;
+			rxq->errors = 0;
+		}
+	}
+
+	for (i = 0; i < avp->num_tx_queues; i++) {
+		struct avp_queue *txq = avp->dev_data->tx_queues[i];
+
+		if (txq) {
+			txq->bytes = 0;
+			txq->packets = 0;
+			txq->errors = 0;
+		}
+	}
+}
 
 RTE_PMD_REGISTER_PCI(rte_avp, rte_avp_pmd.pci_drv);
 RTE_PMD_REGISTER_PCI_TABLE(rte_avp, pci_id_avp_map);
-- 
1.8.3.1