[dpdk-dev] [PATCH 56/56] net/sfc: add callback to send bursts of packets

[Andrew Rybchenko <arybchenko@solarflare.com>]

From: Ivan Malov <ivan.malov@oktetlabs.ru>

Reviewed-by: Andy Moreton <amoreton@solarflare.com>
Signed-off-by: Ivan Malov <ivan.malov@oktetlabs.ru>
Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
---
 drivers/net/sfc/efx/sfc_ethdev.c |   2 +
 drivers/net/sfc/efx/sfc_ev.c     |  22 +++++++-
 drivers/net/sfc/efx/sfc_tweak.h  |   7 +++
 drivers/net/sfc/efx/sfc_tx.c     | 115 +++++++++++++++++++++++++++++++++++++++
 drivers/net/sfc/efx/sfc_tx.h     |  18 ++++++
 5 files changed, 162 insertions(+), 2 deletions(-)

diff --git a/drivers/net/sfc/efx/sfc_ethdev.c b/drivers/net/sfc/efx/sfc_ethdev.c
index 98e5d83..4f94653 100644
--- a/drivers/net/sfc/efx/sfc_ethdev.c
+++ b/drivers/net/sfc/efx/sfc_ethdev.c
@@ -386,6 +386,7 @@ sfc_eth_dev_init(struct rte_eth_dev *dev)
 
 	dev->dev_ops = &sfc_eth_dev_ops;
 	dev->rx_pkt_burst = &sfc_recv_pkts;
+	dev->tx_pkt_burst = &sfc_xmit_pkts;
 
 	sfc_adapter_unlock(sa);
 
@@ -424,6 +425,7 @@ sfc_eth_dev_uninit(struct rte_eth_dev *dev)
 
 	dev->dev_ops = NULL;
 	dev->rx_pkt_burst = NULL;
+	dev->tx_pkt_burst = NULL;
 
 	sfc_kvargs_cleanup(sa);
 
diff --git a/drivers/net/sfc/efx/sfc_ev.c b/drivers/net/sfc/efx/sfc_ev.c
index b3cecbb..8b3f098 100644
--- a/drivers/net/sfc/efx/sfc_ev.c
+++ b/drivers/net/sfc/efx/sfc_ev.c
@@ -142,9 +142,27 @@ static boolean_t
 sfc_ev_tx(void *arg, uint32_t label, uint32_t id)
 {
 	struct sfc_evq *evq = arg;
+	struct sfc_txq *txq;
+	unsigned int stop;
+	unsigned int delta;
 
-	sfc_err(evq->sa, "EVQ %u unexpected Tx event", evq->evq_index);
-	return B_TRUE;
+	txq = evq->txq;
+
+	SFC_ASSERT(txq != NULL);
+	SFC_ASSERT(txq->evq == evq);
+
+	if (unlikely((txq->state & SFC_TXQ_STARTED) == 0))
+		goto done;
+
+	stop = (id + 1) & txq->ptr_mask;
+	id = txq->pending & txq->ptr_mask;
+
+	delta = (stop >= id) ? (stop - id) : (txq->ptr_mask + 1 - id + stop);
+
+	txq->pending += delta;
+
+done:
+	return B_FALSE;
 }
 
 static boolean_t
diff --git a/drivers/net/sfc/efx/sfc_tweak.h b/drivers/net/sfc/efx/sfc_tweak.h
index 24cb9f4..60b618b 100644
--- a/drivers/net/sfc/efx/sfc_tweak.h
+++ b/drivers/net/sfc/efx/sfc_tweak.h
@@ -41,4 +41,11 @@
  */
 #define	SFC_RX_REFILL_BULK	(RTE_CACHE_LINE_SIZE / sizeof(efx_qword_t))
 
+/**
+ * Make the transmit path reap at least one time per a burst;
+ * this improves cache locality because the same mbufs may be used to send
+ * subsequent bursts in certain cases because of well-timed reap
+ */
+#define	SFC_TX_XMIT_PKTS_REAP_AT_LEAST_ONCE	0
+
 #endif /* _SFC_TWEAK_H_ */
diff --git a/drivers/net/sfc/efx/sfc_tx.c b/drivers/net/sfc/efx/sfc_tx.c
index 99d845a..92e81a8 100644
--- a/drivers/net/sfc/efx/sfc_tx.c
+++ b/drivers/net/sfc/efx/sfc_tx.c
@@ -32,6 +32,7 @@
 #include "sfc_log.h"
 #include "sfc_ev.h"
 #include "sfc_tx.h"
+#include "sfc_tweak.h"
 
 /*
  * Maximum number of TX queue flush attempts in case of
@@ -529,3 +530,117 @@ sfc_tx_stop(struct sfc_adapter *sa)
 
 	efx_tx_fini(sa->nic);
 }
+
+uint16_t
+sfc_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
+{
+	struct sfc_txq *txq = (struct sfc_txq *)tx_queue;
+	unsigned int added = txq->added;
+	unsigned int pushed = added;
+	unsigned int pkts_sent = 0;
+	efx_desc_t *pend = &txq->pend_desc[0];
+	const unsigned int hard_max_fill = EFX_TXQ_LIMIT(txq->ptr_mask + 1);
+	const unsigned int soft_max_fill = hard_max_fill -
+					   SFC_TX_MAX_PKT_DESC;
+	unsigned int fill_level = added - txq->completed;
+	boolean_t reap_done;
+	int rc;
+	struct rte_mbuf **pktp;
+
+	if (unlikely((txq->state & SFC_TXQ_RUNNING) == 0))
+		goto done;
+
+	/*
+	 * If insufficient space for a single packet is present,
+	 * we should reap; otherwise, we shouldn't do that all the time
+	 * to avoid latency increase
+	 */
+	reap_done = (fill_level > soft_max_fill);
+
+	if (reap_done) {
+		sfc_tx_reap(txq);
+		/*
+		 * Recalculate fill level since 'txq->completed'
+		 * might have changed on reap
+		 */
+		fill_level = added - txq->completed;
+	}
+
+	for (pkts_sent = 0, pktp = &tx_pkts[0];
+	     (pkts_sent < nb_pkts) && (fill_level <= soft_max_fill);
+	     pkts_sent++, pktp++) {
+		struct rte_mbuf		*m_seg = *pktp;
+		size_t			pkt_len = m_seg->pkt_len;
+		unsigned int		pkt_descs = 0;
+
+		for (; m_seg != NULL; m_seg = m_seg->next) {
+			efsys_dma_addr_t	next_frag;
+			size_t			seg_len;
+
+			seg_len = m_seg->data_len;
+			next_frag = rte_mbuf_data_dma_addr(m_seg);
+
+			do {
+				efsys_dma_addr_t	frag_addr = next_frag;
+				size_t			frag_len;
+
+				next_frag = RTE_ALIGN(frag_addr + 1,
+						      SFC_TX_SEG_BOUNDARY);
+				frag_len = MIN(next_frag - frag_addr, seg_len);
+				seg_len -= frag_len;
+				pkt_len -= frag_len;
+
+				efx_tx_qdesc_dma_create(txq->common,
+							frag_addr, frag_len,
+							(pkt_len == 0),
+							pend++);
+
+				pkt_descs++;
+			} while (seg_len != 0);
+		}
+
+		added += pkt_descs;
+
+		fill_level += pkt_descs;
+		if (unlikely(fill_level > hard_max_fill)) {
+			/*
+			 * Our estimation for maximum number of descriptors
+			 * required to send a packet seems to be wrong.
+			 * Try to reap (if we haven't yet).
+			 */
+			if (!reap_done) {
+				sfc_tx_reap(txq);
+				reap_done = B_TRUE;
+				fill_level = added - txq->completed;
+				if (fill_level > hard_max_fill) {
+					pend -= pkt_descs;
+					break;
+				}
+			} else {
+				pend -= pkt_descs;
+				break;
+			}
+		}
+
+		/* Assign mbuf to the last used desc */
+		txq->sw_ring[(added - 1) & txq->ptr_mask].mbuf = *pktp;
+	}
+
+	if (likely(pkts_sent > 0)) {
+		rc = efx_tx_qdesc_post(txq->common, txq->pend_desc,
+				       pend - &txq->pend_desc[0],
+				       txq->completed, &txq->added);
+		SFC_ASSERT(rc == 0);
+
+		if (likely(pushed != txq->added))
+			efx_tx_qpush(txq->common, txq->added, pushed);
+	}
+
+#if SFC_TX_XMIT_PKTS_REAP_AT_LEAST_ONCE
+	if (!reap_done)
+		sfc_tx_reap(txq);
+#endif
+
+done:
+	return pkts_sent;
+}
diff --git a/drivers/net/sfc/efx/sfc_tx.h b/drivers/net/sfc/efx/sfc_tx.h
index 47970f9..ae13a16 100644
--- a/drivers/net/sfc/efx/sfc_tx.h
+++ b/drivers/net/sfc/efx/sfc_tx.h
@@ -39,6 +39,21 @@
 extern "C" {
 #endif
 
+/**
+ * Estimated maximum number of segments that transmit packet consists of;
+ * it is determined with respect to the expectation of a packet to consist
+ * of a header plus a couple of data segments one of those crossing 4K page;
+ * it is used by transmit path to avoid redundant reaping and, thus,
+ * to avoid increase of latency
+ */
+#define	SFC_TX_MAX_PKT_DESC	4
+
+/**
+ * A segment must not cross 4K boundary
+ * (this is a requirement of NIC TX descriptors)
+ */
+#define	SFC_TX_SEG_BOUNDARY	4096
+
 struct sfc_adapter;
 struct sfc_evq;
 
@@ -100,6 +115,9 @@ void sfc_tx_qstop(struct sfc_adapter *sa, unsigned int sw_index);
 int sfc_tx_start(struct sfc_adapter *sa);
 void sfc_tx_stop(struct sfc_adapter *sa);
 
+uint16_t sfc_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
+		       uint16_t nb_pkts);
+
 #ifdef __cplusplus
 }
 #endif
-- 
2.5.5