[dpdk-dev] [PATCH 20/24] mlx5: resurrect TX gather support

[Nelio Laranjeiro <nelio.laranjeiro@6wind.com>]

From: Adrien Mazarguil <adrien.mazarguil@6wind.com>

Compared to its previous incarnation, the software limit on the number of
mbuf segments is no more (previously MLX5_PMD_SGE_WR_N, set to 4 by
default) hence no need for linearization code and related buffers that
permanently consumed a non negligible amount of memory to handle oversized
mbufs.

The resulting code is both lighter and faster.

Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>
Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro@6wind.com>
---
 drivers/net/mlx5/mlx5_rxtx.c | 231 +++++++++++++++++++++++++++++++++----------
 drivers/net/mlx5/mlx5_txq.c  |   6 +-
 2 files changed, 182 insertions(+), 55 deletions(-)

diff --git a/drivers/net/mlx5/mlx5_rxtx.c b/drivers/net/mlx5/mlx5_rxtx.c
index 1478b2d..53d2a57 100644
--- a/drivers/net/mlx5/mlx5_rxtx.c
+++ b/drivers/net/mlx5/mlx5_rxtx.c
@@ -301,6 +301,7 @@ mlx5_wqe_write(struct txq *txq, volatile union mlx5_wqe *wqe,
 {
 	wqe->wqe.ctrl.data[0] = htonl((txq->wqe_ci << 8) | MLX5_OPCODE_SEND);
 	wqe->wqe.ctrl.data[1] = htonl((txq->qp_num_8s) | 4);
+	wqe->wqe.ctrl.data[2] = 0;
 	wqe->wqe.ctrl.data[3] = 0;
 	wqe->inl.eseg.rsvd0 = 0;
 	wqe->inl.eseg.rsvd1 = 0;
@@ -346,6 +347,7 @@ mlx5_wqe_write_vlan(struct txq *txq, volatile union mlx5_wqe *wqe,
 
 	wqe->wqe.ctrl.data[0] = htonl((txq->wqe_ci << 8) | MLX5_OPCODE_SEND);
 	wqe->wqe.ctrl.data[1] = htonl((txq->qp_num_8s) | 4);
+	wqe->wqe.ctrl.data[2] = 0;
 	wqe->wqe.ctrl.data[3] = 0;
 	wqe->inl.eseg.rsvd0 = 0;
 	wqe->inl.eseg.rsvd1 = 0;
@@ -423,6 +425,7 @@ mlx5_wqe_write_inline(struct txq *txq, volatile union mlx5_wqe *wqe,
 	assert(size < 64);
 	wqe->inl.ctrl.data[0] = htonl((txq->wqe_ci << 8) | MLX5_OPCODE_SEND);
 	wqe->inl.ctrl.data[1] = htonl(txq->qp_num_8s | size);
+	wqe->inl.ctrl.data[2] = 0;
 	wqe->inl.ctrl.data[3] = 0;
 	wqe->inl.eseg.rsvd0 = 0;
 	wqe->inl.eseg.rsvd1 = 0;
@@ -496,6 +499,7 @@ mlx5_wqe_write_inline_vlan(struct txq *txq, volatile union mlx5_wqe *wqe,
 	assert(size < 64);
 	wqe->inl.ctrl.data[0] = htonl((txq->wqe_ci << 8) | MLX5_OPCODE_SEND);
 	wqe->inl.ctrl.data[1] = htonl(txq->qp_num_8s | size);
+	wqe->inl.ctrl.data[2] = 0;
 	wqe->inl.ctrl.data[3] = 0;
 	wqe->inl.eseg.rsvd0 = 0;
 	wqe->inl.eseg.rsvd1 = 0;
@@ -584,6 +588,7 @@ mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 	uint16_t elts_head = txq->elts_head;
 	const unsigned int elts_n = txq->elts_n;
 	unsigned int i = 0;
+	unsigned int j = 0;
 	unsigned int max;
 	unsigned int comp;
 	volatile union mlx5_wqe *wqe;
@@ -600,21 +605,25 @@ mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 	if (max > elts_n)
 		max -= elts_n;
 	do {
-		struct rte_mbuf *buf;
+		struct rte_mbuf *buf = *(pkts++);
 		unsigned int elts_head_next;
 		uintptr_t addr;
 		uint32_t length;
 		uint32_t lkey;
+		unsigned int segs_n = buf->nb_segs;
+		volatile struct mlx5_wqe_data_seg *dseg;
+		unsigned int ds = sizeof(*wqe) / 16;
 
 		/* Make sure there is enough room to store this packet and
 		 * that one ring entry remains unused. */
-		if (max < 1 + 1)
+		assert(segs_n);
+		if (max < segs_n + 1)
 			break;
-		--max;
+		max -= segs_n;
 		--pkts_n;
-		buf = *(pkts++);
 		elts_head_next = (elts_head + 1) & (elts_n - 1);
 		wqe = &(*txq->wqes)[txq->wqe_ci & (txq->wqe_n - 1)];
+		dseg = &wqe->wqe.dseg;
 		rte_prefetch0(wqe);
 		if (pkts_n)
 			rte_prefetch0(*pkts);
@@ -634,7 +643,6 @@ mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 					    buf->vlan_tci);
 		else
 			mlx5_wqe_write(txq, wqe, addr, length, lkey);
-		wqe->wqe.ctrl.data[2] = 0;
 		/* Should we enable HW CKSUM offload */
 		if (buf->ol_flags &
 		    (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
@@ -643,6 +651,35 @@ mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 				MLX5_ETH_WQE_L4_CSUM;
 		} else
 			wqe->wqe.eseg.cs_flags = 0;
+		while (--segs_n) {
+			/* Spill on next WQE when the current one does not have
+			 * enough room left. Size of WQE must a be a multiple
+			 * of data segment size. */
+			assert(!(sizeof(*wqe) % sizeof(*dseg)));
+			if (!(ds % (sizeof(*wqe) / 16)))
+				dseg = (volatile void *)
+					&(*txq->wqes)[txq->wqe_ci++ &
+						      (txq->wqe_n - 1)];
+			else
+				++dseg;
+			++ds;
+			buf = buf->next;
+			assert(buf);
+			/* Store segment information. */
+			dseg->byte_count = htonl(DATA_LEN(buf));
+			dseg->lkey = txq_mp2mr(txq, txq_mb2mp(buf));
+			dseg->addr = htonll(rte_pktmbuf_mtod(buf, uintptr_t));
+			(*txq->elts)[elts_head_next] = buf;
+			elts_head_next = (elts_head_next + 1) & (elts_n - 1);
+#ifdef MLX5_PMD_SOFT_COUNTERS
+			length += DATA_LEN(buf);
+#endif
+			++j;
+		}
+		/* Update DS field in WQE. */
+		wqe->wqe.ctrl.data[1] &= htonl(0xffffffc0);
+		wqe->wqe.ctrl.data[1] |= htonl(ds & 0x3f);
+		elts_head = elts_head_next;
 #ifdef MLX5_PMD_SOFT_COUNTERS
 		/* Increment sent bytes counter. */
 		txq->stats.obytes += length;
@@ -654,7 +691,7 @@ mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 	if (unlikely(i == 0))
 		return 0;
 	/* Check whether completion threshold has been reached. */
-	comp = txq->elts_comp + i;
+	comp = txq->elts_comp + i + j;
 	if (comp >= MLX5_TX_COMP_THRESH) {
 		/* Request completion on last WQE. */
 		wqe->wqe.ctrl.data[2] = htonl(8);
@@ -693,6 +730,7 @@ mlx5_tx_burst_inline(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 	uint16_t elts_head = txq->elts_head;
 	const unsigned int elts_n = txq->elts_n;
 	unsigned int i = 0;
+	unsigned int j = 0;
 	unsigned int max;
 	unsigned int comp;
 	volatile union mlx5_wqe *wqe;
@@ -710,21 +748,25 @@ mlx5_tx_burst_inline(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 	if (max > elts_n)
 		max -= elts_n;
 	do {
-		struct rte_mbuf *buf;
+		struct rte_mbuf *buf = *(pkts++);
 		unsigned int elts_head_next;
 		uintptr_t addr;
 		uint32_t length;
 		uint32_t lkey;
+		unsigned int segs_n = buf->nb_segs;
+		volatile struct mlx5_wqe_data_seg *dseg;
+		unsigned int ds = sizeof(*wqe) / 16;
 
 		/* Make sure there is enough room to store this packet and
 		 * that one ring entry remains unused. */
-		if (max < 1 + 1)
+		assert(segs_n);
+		if (max < segs_n + 1)
 			break;
-		--max;
+		max -= segs_n;
 		--pkts_n;
-		buf = *(pkts++);
 		elts_head_next = (elts_head + 1) & (elts_n - 1);
 		wqe = &(*txq->wqes)[txq->wqe_ci & (txq->wqe_n - 1)];
+		dseg = &wqe->wqe.dseg;
 		tx_prefetch_wqe(txq, txq->wqe_ci);
 		tx_prefetch_wqe(txq, txq->wqe_ci + 1);
 		if (pkts_n)
@@ -746,13 +788,14 @@ mlx5_tx_burst_inline(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 		if (pkts_n)
 			rte_prefetch0(rte_pktmbuf_mtod(*pkts,
 						       volatile void *));
-		if (length <= max_inline) {
+		if ((length <= max_inline) && (segs_n == 1)) {
 			if (buf->ol_flags & PKT_TX_VLAN_PKT)
 				mlx5_wqe_write_inline_vlan(txq, wqe,
 							   addr, length,
 							   buf->vlan_tci);
 			else
 				mlx5_wqe_write_inline(txq, wqe, addr, length);
+			goto skip_segs;
 		} else {
 			/* Retrieve Memory Region key for this memory pool. */
 			lkey = txq_mp2mr(txq, txq_mb2mp(buf));
@@ -762,7 +805,35 @@ mlx5_tx_burst_inline(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 			else
 				mlx5_wqe_write(txq, wqe, addr, length, lkey);
 		}
-		wqe->inl.ctrl.data[2] = 0;
+		while (--segs_n) {
+			/* Spill on next WQE when the current one does not have
+			 * enough room left. Size of WQE must a be a multiple
+			 * of data segment size. */
+			assert(!(sizeof(*wqe) % sizeof(*dseg)));
+			if (!(ds % (sizeof(*wqe) / 16)))
+				dseg = (volatile void *)
+					&(*txq->wqes)[txq->wqe_ci++ &
+						      (txq->wqe_n - 1)];
+			else
+				++dseg;
+			++ds;
+			buf = buf->next;
+			assert(buf);
+			/* Store segment information. */
+			dseg->byte_count = htonl(DATA_LEN(buf));
+			dseg->lkey = txq_mp2mr(txq, txq_mb2mp(buf));
+			dseg->addr = htonll(rte_pktmbuf_mtod(buf, uintptr_t));
+			(*txq->elts)[elts_head_next] = buf;
+			elts_head_next = (elts_head_next + 1) & (elts_n - 1);
+#ifdef MLX5_PMD_SOFT_COUNTERS
+			length += DATA_LEN(buf);
+#endif
+			++j;
+		}
+		/* Update DS field in WQE. */
+		wqe->inl.ctrl.data[1] &= htonl(0xffffffc0);
+		wqe->inl.ctrl.data[1] |= htonl(ds & 0x3f);
+	skip_segs:
 		elts_head = elts_head_next;
 #ifdef MLX5_PMD_SOFT_COUNTERS
 		/* Increment sent bytes counter. */
@@ -774,7 +845,7 @@ mlx5_tx_burst_inline(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 	if (unlikely(i == 0))
 		return 0;
 	/* Check whether completion threshold has been reached. */
-	comp = txq->elts_comp + i;
+	comp = txq->elts_comp + i + j;
 	if (comp >= MLX5_TX_COMP_THRESH) {
 		/* Request completion on last WQE. */
 		wqe->inl.ctrl.data[2] = htonl(8);
@@ -878,6 +949,7 @@ mlx5_tx_burst_mpw(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 	uint16_t elts_head = txq->elts_head;
 	const unsigned int elts_n = txq->elts_n;
 	unsigned int i = 0;
+	unsigned int j = 0;
 	unsigned int max;
 	unsigned int comp;
 	struct mlx5_mpw mpw = {
@@ -896,46 +968,67 @@ mlx5_tx_burst_mpw(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 	if (max > elts_n)
 		max -= elts_n;
 	do {
-		struct rte_mbuf *buf;
-		volatile struct mlx5_wqe_data_seg *dseg;
+		struct rte_mbuf *buf = *(pkts++);
 		unsigned int elts_head_next;
-		uintptr_t addr;
 		uint32_t length;
+		unsigned int segs_n = buf->nb_segs;
 		uint32_t cs_flags = 0;
 
 		/* Make sure there is enough room to store this packet and
 		 * that one ring entry remains unused. */
-		if (max < 1 + 1)
+		assert(segs_n);
+		if (max < segs_n + 1)
+			break;
+		/* Do not bother with large packets MPW cannot handle. */
+		if (segs_n > MLX5_MPW_DSEG_MAX)
 			break;
-		--max;
+		max -= segs_n;
 		--pkts_n;
-		buf = *(pkts++);
-		elts_head_next = (elts_head + 1) & (elts_n - 1);
 		/* Should we enable HW CKSUM offload */
 		if (buf->ol_flags &
 		    (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM))
 			cs_flags = MLX5_ETH_WQE_L3_CSUM | MLX5_ETH_WQE_L4_CSUM;
-		/* Retrieve buffer information. */
-		addr = rte_pktmbuf_mtod(buf, uintptr_t);
-		length = DATA_LEN(buf);
-		/* Update element. */
-		(*txq->elts)[elts_head] = buf;
+		/* Retrieve packet information. */
+		length = PKT_LEN(buf);
+		assert(length);
 		/* Start new session if packet differs. */
 		if ((mpw.state == MLX5_MPW_STATE_OPENED) &&
 		    ((mpw.len != length) ||
+		     (segs_n != 1) ||
 		     (mpw.wqe->mpw.eseg.cs_flags != cs_flags)))
 			mlx5_mpw_close(txq, &mpw);
 		if (mpw.state == MLX5_MPW_STATE_CLOSED) {
 			mlx5_mpw_new(txq, &mpw, length);
 			mpw.wqe->mpw.eseg.cs_flags = cs_flags;
 		}
-		dseg = mpw.data.dseg[mpw.pkts_n];
-		*dseg = (struct mlx5_wqe_data_seg){
-			.byte_count = htonl(length),
-			.lkey = txq_mp2mr(txq, txq_mb2mp(buf)),
-			.addr = htonll(addr),
-		};
-		++mpw.pkts_n;
+		/* Multi-segment packets must be alone in their MPW. */
+		assert((segs_n == 1) || (mpw.pkts_n == 0));
+#if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
+		length = 0;
+#endif
+		do {
+			volatile struct mlx5_wqe_data_seg *dseg;
+			uintptr_t addr;
+
+			elts_head_next = (elts_head + 1) & (elts_n - 1);
+			assert(buf);
+			(*txq->elts)[elts_head] = buf;
+			dseg = mpw.data.dseg[mpw.pkts_n];
+			addr = rte_pktmbuf_mtod(buf, uintptr_t);
+			*dseg = (struct mlx5_wqe_data_seg){
+				.byte_count = htonl(DATA_LEN(buf)),
+				.lkey = txq_mp2mr(txq, txq_mb2mp(buf)),
+				.addr = htonll(addr),
+			};
+			elts_head = elts_head_next;
+#if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
+			length += DATA_LEN(buf);
+#endif
+			buf = buf->next;
+			++mpw.pkts_n;
+			++j;
+		} while (--segs_n);
+		assert(length == mpw.len);
 		if (mpw.pkts_n == MLX5_MPW_DSEG_MAX)
 			mlx5_mpw_close(txq, &mpw);
 		elts_head = elts_head_next;
@@ -949,7 +1042,8 @@ mlx5_tx_burst_mpw(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 	if (unlikely(i == 0))
 		return 0;
 	/* Check whether completion threshold has been reached. */
-	comp = txq->elts_comp + i;
+	/* "j" includes both packets and segments. */
+	comp = txq->elts_comp + j;
 	if (comp >= MLX5_TX_COMP_THRESH) {
 		volatile union mlx5_wqe *wqe = mpw.wqe;
 
@@ -1050,6 +1144,7 @@ mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
 	uint16_t elts_head = txq->elts_head;
 	const unsigned int elts_n = txq->elts_n;
 	unsigned int i = 0;
+	unsigned int j = 0;
 	unsigned int max;
 	unsigned int comp;
 	unsigned int inline_room = txq->max_inline;
@@ -1069,36 +1164,38 @@ mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
 	if (max > elts_n)
 		max -= elts_n;
 	do {
-		struct rte_mbuf *buf;
+		struct rte_mbuf *buf = *(pkts++);
 		unsigned int elts_head_next;
 		uintptr_t addr;
 		uint32_t length;
+		unsigned int segs_n = buf->nb_segs;
 		uint32_t cs_flags = 0;
 
 		/* Make sure there is enough room to store this packet and
 		 * that one ring entry remains unused. */
-		if (max < 1 + 1)
+		assert(segs_n);
+		if (max < segs_n + 1)
 			break;
-		--max;
+		/* Do not bother with large packets MPW cannot handle. */
+		if (segs_n > MLX5_MPW_DSEG_MAX)
+			break;
+		max -= segs_n;
 		--pkts_n;
-		buf = *(pkts++);
-		elts_head_next = (elts_head + 1) & (elts_n - 1);
 		/* Should we enable HW CKSUM offload */
 		if (buf->ol_flags &
 		    (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM))
 			cs_flags = MLX5_ETH_WQE_L3_CSUM | MLX5_ETH_WQE_L4_CSUM;
-		/* Retrieve buffer information. */
-		addr = rte_pktmbuf_mtod(buf, uintptr_t);
-		length = DATA_LEN(buf);
-		/* Update element. */
-		(*txq->elts)[elts_head] = buf;
+		/* Retrieve packet information. */
+		length = PKT_LEN(buf);
 		/* Start new session if packet differs. */
 		if (mpw.state == MLX5_MPW_STATE_OPENED) {
 			if ((mpw.len != length) ||
+			    (segs_n != 1) ||
 			    (mpw.wqe->mpw.eseg.cs_flags != cs_flags))
 				mlx5_mpw_close(txq, &mpw);
 		} else if (mpw.state == MLX5_MPW_INL_STATE_OPENED) {
 			if ((mpw.len != length) ||
+			    (segs_n != 1) ||
 			    (length > inline_room) ||
 			    (mpw.wqe->mpw_inl.eseg.cs_flags != cs_flags)) {
 				mlx5_mpw_inline_close(txq, &mpw);
@@ -1106,7 +1203,8 @@ mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
 			}
 		}
 		if (mpw.state == MLX5_MPW_STATE_CLOSED) {
-			if (length > inline_room) {
+			if ((segs_n != 1) ||
+			    (length > inline_room)) {
 				mlx5_mpw_new(txq, &mpw, length);
 				mpw.wqe->mpw.eseg.cs_flags = cs_flags;
 			} else {
@@ -1114,17 +1212,36 @@ mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
 				mpw.wqe->mpw_inl.eseg.cs_flags = cs_flags;
 			}
 		}
+		/* Multi-segment packets must be alone in their MPW. */
+		assert((segs_n == 1) || (mpw.pkts_n == 0));
 		if (mpw.state == MLX5_MPW_STATE_OPENED) {
-			volatile struct mlx5_wqe_data_seg *dseg;
-
 			assert(inline_room == txq->max_inline);
-			dseg = mpw.data.dseg[mpw.pkts_n];
-			*dseg = (struct mlx5_wqe_data_seg){
-				.byte_count = htonl(length),
-				.lkey = txq_mp2mr(txq, txq_mb2mp(buf)),
-				.addr = htonll(addr),
-			};
-			++mpw.pkts_n;
+#if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
+			length = 0;
+#endif
+			do {
+				volatile struct mlx5_wqe_data_seg *dseg;
+
+				elts_head_next =
+					(elts_head + 1) & (elts_n - 1);
+				assert(buf);
+				(*txq->elts)[elts_head] = buf;
+				dseg = mpw.data.dseg[mpw.pkts_n];
+				addr = rte_pktmbuf_mtod(buf, uintptr_t);
+				*dseg = (struct mlx5_wqe_data_seg){
+					.byte_count = htonl(DATA_LEN(buf)),
+					.lkey = txq_mp2mr(txq, txq_mb2mp(buf)),
+					.addr = htonll(addr),
+				};
+				elts_head = elts_head_next;
+#if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
+				length += DATA_LEN(buf);
+#endif
+				buf = buf->next;
+				++mpw.pkts_n;
+				++j;
+			} while (--segs_n);
+			assert(length == mpw.len);
 			if (mpw.pkts_n == MLX5_MPW_DSEG_MAX)
 				mlx5_mpw_close(txq, &mpw);
 		} else {
@@ -1132,6 +1249,10 @@ mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
 
 			assert(mpw.state == MLX5_MPW_INL_STATE_OPENED);
 			assert(length <= inline_room);
+			assert(length == DATA_LEN(buf));
+			elts_head_next = (elts_head + 1) & (elts_n - 1);
+			addr = rte_pktmbuf_mtod(buf, uintptr_t);
+			(*txq->elts)[elts_head] = buf;
 			/* Maximum number of bytes before wrapping. */
 			max = ((uintptr_t)&(*txq->wqes)[txq->wqe_n] -
 			       (uintptr_t)mpw.data.raw);
@@ -1156,6 +1277,7 @@ mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
 				mpw.data.raw =
 					(volatile void *)&(*txq->wqes)[0];
 			++mpw.pkts_n;
+			++j;
 			if (mpw.pkts_n == MLX5_MPW_DSEG_MAX) {
 				mlx5_mpw_inline_close(txq, &mpw);
 				inline_room = txq->max_inline;
@@ -1174,7 +1296,8 @@ mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
 	if (unlikely(i == 0))
 		return 0;
 	/* Check whether completion threshold has been reached. */
-	comp = txq->elts_comp + i;
+	/* "j" includes both packets and segments. */
+	comp = txq->elts_comp + j;
 	if (comp >= MLX5_TX_COMP_THRESH) {
 		volatile union mlx5_wqe *wqe = mpw.wqe;
 
diff --git a/drivers/net/mlx5/mlx5_txq.c b/drivers/net/mlx5/mlx5_txq.c
index d013230..38900c1 100644
--- a/drivers/net/mlx5/mlx5_txq.c
+++ b/drivers/net/mlx5/mlx5_txq.c
@@ -311,7 +311,11 @@ txq_ctrl_setup(struct rte_eth_dev *dev, struct txq_ctrl *txq_ctrl,
 			.max_send_wr = ((priv->device_attr.max_qp_wr < desc) ?
 					priv->device_attr.max_qp_wr :
 					desc),
-			/* Max number of scatter/gather elements in a WR. */
+			/* Max number of scatter/gather elements in a WR,
+			 * must be 1 to prevent libmlx5 from trying to affect
+			 * too much memory. TX gather is not impacted by the
+			 * priv->device_attr.max_sge limit and will still work
+			 * properly. */
 			.max_send_sge = 1,
 		},
 		.qp_type = IBV_QPT_RAW_PACKET,
-- 
2.1.4