[dpdk-dev] [PATCH 2/4] event/octeontx2: improve single flow performance

[Harman Kalra <hkalra@marvell.com>]

From: Pavan Nikhilesh <pbhagavatula@marvell.com>

Improve single flow performance by moving the point of coherence
to the end of transmit sequence.

Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
---
 drivers/event/octeontx2/otx2_worker.h | 35 +++++++++++++++++----------
 drivers/net/octeontx2/otx2_tx.h       | 18 ++++++++++++++
 2 files changed, 40 insertions(+), 13 deletions(-)

diff --git a/drivers/event/octeontx2/otx2_worker.h b/drivers/event/octeontx2/otx2_worker.h
index 1bf8afedf..32d611458 100644
--- a/drivers/event/octeontx2/otx2_worker.h
+++ b/drivers/event/octeontx2/otx2_worker.h
@@ -247,15 +247,6 @@ otx2_ssogws_head_wait(struct otx2_ssogws *ws)
 #endif
 }
 
-static __rte_always_inline void
-otx2_ssogws_order(struct otx2_ssogws *ws, const uint8_t wait_flag)
-{
-	if (wait_flag)
-		otx2_ssogws_head_wait(ws);
-
-	rte_cio_wmb();
-}
-
 static __rte_always_inline const struct otx2_eth_txq *
 otx2_ssogws_xtract_meta(struct rte_mbuf *m,
 			const uint64_t txq_data[][RTE_MAX_QUEUES_PER_PORT])
@@ -287,10 +278,9 @@ otx2_ssogws_event_tx(struct otx2_ssogws *ws, struct rte_event ev[],
 		return otx2_sec_event_tx(ws, ev, m, txq, flags);
 	}
 
-	rte_prefetch_non_temporal(&txq_data[m->port][0]);
 	/* Perform header writes before barrier for TSO */
 	otx2_nix_xmit_prepare_tso(m, flags);
-	otx2_ssogws_order(ws, !ev->sched_type);
+	rte_cio_wmb();
 	txq = otx2_ssogws_xtract_meta(m, txq_data);
 	otx2_ssogws_prepare_pkt(txq, m, cmd, flags);
 
@@ -298,12 +288,31 @@ otx2_ssogws_event_tx(struct otx2_ssogws *ws, struct rte_event ev[],
 		const uint16_t segdw = otx2_nix_prepare_mseg(m, cmd, flags);
 		otx2_nix_xmit_prepare_tstamp(cmd, &txq->cmd[0],
 					     m->ol_flags, segdw, flags);
-		otx2_nix_xmit_mseg_one(cmd, txq->lmt_addr, txq->io_addr, segdw);
+		if (!ev->sched_type) {
+			otx2_nix_xmit_mseg_prep_lmt(cmd, txq->lmt_addr, segdw);
+			otx2_ssogws_head_wait(ws);
+			if (otx2_nix_xmit_submit_lmt(txq->io_addr) == 0)
+				otx2_nix_xmit_mseg_one(cmd, txq->lmt_addr,
+						       txq->io_addr, segdw);
+		} else {
+			otx2_nix_xmit_mseg_one(cmd, txq->lmt_addr, txq->io_addr,
+					       segdw);
+		}
 	} else {
 		/* Passing no of segdw as 4: HDR + EXT + SG + SMEM */
 		otx2_nix_xmit_prepare_tstamp(cmd, &txq->cmd[0],
 					     m->ol_flags, 4, flags);
-		otx2_nix_xmit_one(cmd, txq->lmt_addr, txq->io_addr, flags);
+
+		if (!ev->sched_type) {
+			otx2_nix_xmit_prep_lmt(cmd, txq->lmt_addr, flags);
+			otx2_ssogws_head_wait(ws);
+			if (otx2_nix_xmit_submit_lmt(txq->io_addr) == 0)
+				otx2_nix_xmit_one(cmd, txq->lmt_addr,
+						  txq->io_addr, flags);
+		} else {
+			otx2_nix_xmit_one(cmd, txq->lmt_addr, txq->io_addr,
+					  flags);
+		}
 	}
 
 	otx2_write64(0, ws->swtag_flush_op);
diff --git a/drivers/net/octeontx2/otx2_tx.h b/drivers/net/octeontx2/otx2_tx.h
index 3c4317092..caf170fd1 100644
--- a/drivers/net/octeontx2/otx2_tx.h
+++ b/drivers/net/octeontx2/otx2_tx.h
@@ -383,6 +383,18 @@ otx2_nix_xmit_one(uint64_t *cmd, void *lmt_addr,
 	} while (lmt_status == 0);
 }
 
+static __rte_always_inline void
+otx2_nix_xmit_prep_lmt(uint64_t *cmd, void *lmt_addr, const uint32_t flags)
+{
+	otx2_lmt_mov(lmt_addr, cmd, otx2_nix_tx_ext_subs(flags));
+}
+
+static __rte_always_inline uint64_t
+otx2_nix_xmit_submit_lmt(const rte_iova_t io_addr)
+{
+	return otx2_lmt_submit(io_addr);
+}
+
 static __rte_always_inline uint16_t
 otx2_nix_prepare_mseg(struct rte_mbuf *m, uint64_t *cmd, const uint16_t flags)
 {
@@ -453,6 +465,12 @@ otx2_nix_prepare_mseg(struct rte_mbuf *m, uint64_t *cmd, const uint16_t flags)
 	return segdw;
 }
 
+static __rte_always_inline void
+otx2_nix_xmit_mseg_prep_lmt(uint64_t *cmd, void *lmt_addr, uint16_t segdw)
+{
+	otx2_lmt_mov_seg(lmt_addr, (const void *)cmd, segdw);
+}
+
 static __rte_always_inline void
 otx2_nix_xmit_mseg_one(uint64_t *cmd, void *lmt_addr,
 		       rte_iova_t io_addr, uint16_t segdw)
-- 
2.18.0