Skip to contentSkip to conversation

net/mlx5: convert meta register to big-endian

PATCH In review Alexander Kozyrev 2 participants 2 messages Jun 16, 2021 – Jun 17, 2021 Network drivers

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 net/mlx5: convert meta register to big-endian

Metadata is stored in the CPU order (little-endian format on x86), while all the packet header fields are stored in the network order. That leads to the wrong results whenever we try to use the metadata value in the modify_field actions: bytes are swapped as a result.

Convert the metadata into the big-endian format before storing it in the Mellanox NIC to achieve consistent behaviour.

---
 drivers/net/mlx5/mlx5_flow_dv.c          | 40 +++++-------------------
 drivers/net/mlx5/mlx5_rx.c               |  5 +--
 drivers/net/mlx5/mlx5_rxtx_vec_altivec.h | 22 +++++++++----
 drivers/net/mlx5/mlx5_rxtx_vec_neon.h    | 30 +++++++++++-------
 drivers/net/mlx5/mlx5_rxtx_vec_sse.h     | 18 ++++++++---
 5 files changed, 59 insertions(+), 56 deletions(-)
drivers/net/mlx5/mlx5_flow_dv.c +8 -32
index dafd37ab93..b36ffde559 100644
@@ -1239,8 +1239,8 @@ flow_dv_convert_action_set_meta
 			 const struct rte_flow_action_set_meta *conf,
 			 struct rte_flow_error *error)
 {
-	uint32_t data = conf->data;
-	uint32_t mask = conf->mask;
+	uint32_t mask = rte_cpu_to_be_32(conf->mask);
+	uint32_t data = rte_cpu_to_be_32(conf->data) & mask;
 	struct rte_flow_item item = {
 		.spec = &data,
 		.mask = &mask,
@@ -1253,25 +1253,14 @@ flow_dv_convert_action_set_meta
 	if (reg < 0)
 		return reg;
 	MLX5_ASSERT(reg != REG_NON);
-	/*
-	 * In datapath code there is no endianness
-	 * coversions for perfromance reasons, all
-	 * pattern conversions are done in rte_flow.
-	 */
 	if (reg == REG_C_0) {
 		struct mlx5_priv *priv = dev->data->dev_private;
 		uint32_t msk_c0 = priv->sh->dv_regc0_mask;
-		uint32_t shl_c0;
+		uint32_t shl_c0 = rte_bsf32(msk_c0);
 
-		MLX5_ASSERT(msk_c0);
-#if RTE_BYTE_ORDER == RTE_BIG_ENDIAN
-		shl_c0 = rte_bsf32(msk_c0);
-#else
-		shl_c0 = sizeof(msk_c0) * CHAR_BIT - rte_fls_u32(msk_c0);
-#endif
-		mask <<= shl_c0;
-		data <<= shl_c0;
-		MLX5_ASSERT(!(~msk_c0 & rte_cpu_to_be_32(mask)));
+		data = rte_cpu_to_be_32(rte_cpu_to_be_32(data) << shl_c0);
+		mask = rte_cpu_to_be_32(mask) & msk_c0;
+		mask = rte_cpu_to_be_32(mask << shl_c0);
 	}
 	reg_c_x[0] = (struct field_modify_info){4, 0, reg_to_field[reg]};
 	/* The routine expects parameters in memory as big-endian ones. */
@@ -9226,27 +9215,14 @@ flow_dv_translate_item_meta(struct rte_eth_dev *dev,
 		if (reg < 0)
 			return;
 		MLX5_ASSERT(reg != REG_NON);
-		/*
-		 * In datapath code there is no endianness
-		 * coversions for perfromance reasons, all
-		 * pattern conversions are done in rte_flow.
-		 */
-		value = rte_cpu_to_be_32(value);
-		mask = rte_cpu_to_be_32(mask);
 		if (reg == REG_C_0) {
 			struct mlx5_priv *priv = dev->data->dev_private;
 			uint32_t msk_c0 = priv->sh->dv_regc0_mask;
 			uint32_t shl_c0 = rte_bsf32(msk_c0);
-#if RTE_BYTE_ORDER == RTE_LITTLE_ENDIAN
-			uint32_t shr_c0 = __builtin_clz(priv->sh->dv_meta_mask);
 
-			value >>= shr_c0;
-			mask >>= shr_c0;
-#endif
-			value <<= shl_c0;
+			mask &= msk_c0;
 			mask <<= shl_c0;
-			MLX5_ASSERT(msk_c0);
-			MLX5_ASSERT(!(~msk_c0 & mask));
+			value <<= shl_c0;
 		}
 		flow_dv_match_meta_reg(matcher, key, reg, value, mask);
 	}
drivers/net/mlx5/mlx5_rx.c +3 -2
index 6cd71a44eb..777a1d6e45 100644
@@ -740,8 +740,9 @@ rxq_cq_to_mbuf(struct mlx5_rxq_data *rxq, struct rte_mbuf *pkt,
 		}
 	}
 	if (rxq->dynf_meta) {
-		uint32_t meta = cqe->flow_table_metadata &
-				rxq->flow_meta_port_mask;
+		uint32_t meta = rte_be_to_cpu_32(cqe->flow_table_metadata >>
+			__builtin_popcount(rxq->flow_meta_port_mask)) &
+			rxq->flow_meta_port_mask;
 
 		if (meta) {
 			pkt->ol_flags |= rxq->flow_meta_mask;
drivers/net/mlx5/mlx5_rxtx_vec_altivec.h +16 -6
index 2d1154b624..648c59e2c2 100644
@@ -1221,23 +1221,33 @@ rxq_cq_process_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq,
 		if (rxq->dynf_meta) {
 			uint64_t flag = rxq->flow_meta_mask;
 			int32_t offs = rxq->flow_meta_offset;
-			uint32_t metadata, mask;
+			uint32_t mask = rxq->flow_meta_port_mask;
+			uint32_t shift =
+				__builtin_popcount(rxq->flow_meta_port_mask);
+			uint32_t metadata;
 
-			mask = rxq->flow_meta_port_mask;
 			/* This code is subject for futher optimization. */
-			metadata = cq[pos].flow_table_metadata & mask;
+			metadata = (rte_be_to_cpu_32
+				(cq[pos].flow_table_metadata) >> shift) &
+				mask;
 			*RTE_MBUF_DYNFIELD(pkts[pos], offs, uint32_t *) =
 								metadata;
 			pkts[pos]->ol_flags |= metadata ? flag : 0ULL;
-			metadata = cq[pos + 1].flow_table_metadata & mask;
+			metadata = (rte_be_to_cpu_32
+				(cq[pos + 1].flow_table_metadata) >> shift) &
+				mask;
 			*RTE_MBUF_DYNFIELD(pkts[pos + 1], offs, uint32_t *) =
 								metadata;
 			pkts[pos + 1]->ol_flags |= metadata ? flag : 0ULL;
-			metadata = cq[pos + 2].flow_table_metadata & mask;
+			metadata = (rte_be_to_cpu_32
+				(cq[pos + 2].flow_table_metadata) >> shift) &
+				mask;
 			*RTE_MBUF_DYNFIELD(pkts[pos + 2], offs, uint32_t *) =
 								metadata;
 			pkts[pos + 2]->ol_flags |= metadata ? flag : 0ULL;
-			metadata = cq[pos + 3].flow_table_metadata & mask;
+			metadata = (rte_be_to_cpu_32
+				(cq[pos + 3].flow_table_metadata) >> shift) &
+				mask;
 			*RTE_MBUF_DYNFIELD(pkts[pos + 3], offs, uint32_t *) =
 								metadata;
 			pkts[pos + 3]->ol_flags |= metadata ? flag : 0ULL;
drivers/net/mlx5/mlx5_rxtx_vec_neon.h +18 -12
index 2234fbe6b2..5c569ee199 100644
@@ -833,23 +833,29 @@ rxq_cq_process_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq,
 			/* This code is subject for futher optimization. */
 			int32_t offs = rxq->flow_meta_offset;
 			uint32_t mask = rxq->flow_meta_port_mask;
+			uint32_t shift =
+				__builtin_popcount(rxq->flow_meta_port_mask);
 
 			*RTE_MBUF_DYNFIELD(pkts[pos], offs, uint32_t *) =
-				container_of(p0, struct mlx5_cqe,
-					     pkt_info)->flow_table_metadata &
-					     mask;
+				(rte_be_to_cpu_32(container_of
+				(p0, struct mlx5_cqe,
+				pkt_info)->flow_table_metadata) >> shift) &
+				mask;
 			*RTE_MBUF_DYNFIELD(pkts[pos + 1], offs, uint32_t *) =
-				container_of(p1, struct mlx5_cqe,
-					     pkt_info)->flow_table_metadata &
-					     mask;
+				(rte_be_to_cpu_32(container_of
+				(p1, struct mlx5_cqe,
+				pkt_info)->flow_table_metadata) >> shift) &
+				mask;
 			*RTE_MBUF_DYNFIELD(pkts[pos + 2], offs, uint32_t *) =
-				container_of(p2, struct mlx5_cqe,
-					     pkt_info)->flow_table_metadata &
-					     mask;
+				(rte_be_to_cpu_32(container_of
+				(p2, struct mlx5_cqe,
+				pkt_info)->flow_table_metadata) >> shift) &
+				mask;
 			*RTE_MBUF_DYNFIELD(pkts[pos + 3], offs, uint32_t *) =
-				container_of(p3, struct mlx5_cqe,
-					     pkt_info)->flow_table_metadata &
-					     mask;
+				(rte_be_to_cpu_32(container_of
+				(p3, struct mlx5_cqe,
+				pkt_info)->flow_table_metadata) >> shift) &
+				mask;
 			if (*RTE_MBUF_DYNFIELD(pkts[pos], offs, uint32_t *))
 				elts[pos]->ol_flags |= rxq->flow_meta_mask;
 			if (*RTE_MBUF_DYNFIELD(pkts[pos + 1], offs, uint32_t *))
drivers/net/mlx5/mlx5_rxtx_vec_sse.h +14 -4
index c508a7a4f2..661fa7273c 100644
@@ -769,15 +769,25 @@ rxq_cq_process_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq,
 			/* This code is subject for futher optimization. */
 			int32_t offs = rxq->flow_meta_offset;
 			uint32_t mask = rxq->flow_meta_port_mask;
+			uint32_t shift =
+				__builtin_popcount(rxq->flow_meta_port_mask);
 
 			*RTE_MBUF_DYNFIELD(pkts[pos], offs, uint32_t *) =
-				cq[pos].flow_table_metadata & mask;
+				(rte_be_to_cpu_32
+				(cq[pos].flow_table_metadata) >> shift) &
+				mask;
 			*RTE_MBUF_DYNFIELD(pkts[pos + 1], offs, uint32_t *) =
-				cq[pos + p1].flow_table_metadata  & mask;
+				(rte_be_to_cpu_32
+				(cq[pos + p1].flow_table_metadata) >> shift) &
+				mask;
 			*RTE_MBUF_DYNFIELD(pkts[pos + 2], offs, uint32_t *) =
-				cq[pos + p2].flow_table_metadata & mask;
+				(rte_be_to_cpu_32
+				(cq[pos + p2].flow_table_metadata) >> shift) &
+				mask;
 			*RTE_MBUF_DYNFIELD(pkts[pos + 3], offs, uint32_t *) =
-				cq[pos + p3].flow_table_metadata & mask;
+				(rte_be_to_cpu_32
+				(cq[pos + p3].flow_table_metadata) >> shift) &
+				mask;
 			if (*RTE_MBUF_DYNFIELD(pkts[pos], offs, uint32_t *))
 				pkts[pos]->ol_flags |= rxq->flow_meta_mask;
 			if (*RTE_MBUF_DYNFIELD(pkts[pos + 1], offs, uint32_t *))
Signature · 1 line
2.18.2
#1 lore raw

Hi,

Quoted 13 lines

-----Original Message-----
From: Alexander Kozyrev <[email protected]>
Sent: Wednesday, June 16, 2021 5:46 PM
To: [email protected]
Cc: Raslan Darawsheh <[email protected]>; Slava Ovsiienko <[email protected]>; Matan Azrad <[email protected]>
Subject: [PATCH] net/mlx5: convert meta register to big-endian

Metadata is stored in the CPU order (little-endian format on x86), while all the packet header fields are stored in the network order. That leads to the wrong results whenever we try to use the metadata value in the modify_field actions: bytes are swapped as a result.

Quoted 5 lines

Convert the metadata into the big-endian format before storing it in the Mellanox NIC to achieve consistent behaviour.

Patch applied to next-net-mlx,

Kindest regards,
Raslan Darawsheh