[dpdk-dev] [PATCH 06/17] mlx5: adapt indirection table size depending on RX queues number

Adrien Mazarguil adrien.mazarguil at 6wind.com
Mon Oct 5 19:54:41 CEST 2015


From: Nelio Laranjeiro <nelio.laranjeiro at 6wind.com>

Use the maximum size of the indirection table when the number of requested
RX queues is not a power of two, this help to improve RSS balancing.

A message informs users that balancing is not optimal in such cases.

Signed-off-by: Nelio Laranjeiro <nelio.laranjeiro at 6wind.com>
Signed-off-by: Adrien Mazarguil <adrien.mazarguil at 6wind.com>
---
 drivers/net/mlx5/mlx5.c      | 10 +++++++++-
 drivers/net/mlx5/mlx5.h      |  1 +
 drivers/net/mlx5/mlx5_defs.h |  3 +++
 drivers/net/mlx5/mlx5_rxq.c  | 21 ++++++++++++++-------
 4 files changed, 27 insertions(+), 8 deletions(-)

diff --git a/drivers/net/mlx5/mlx5.c b/drivers/net/mlx5/mlx5.c
index a316989..167e14b 100644
--- a/drivers/net/mlx5/mlx5.c
+++ b/drivers/net/mlx5/mlx5.c
@@ -301,7 +301,9 @@ mlx5_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
 		struct ether_addr mac;
 
 #ifdef HAVE_EXP_QUERY_DEVICE
-		exp_device_attr.comp_mask = IBV_EXP_DEVICE_ATTR_EXP_CAP_FLAGS;
+		exp_device_attr.comp_mask =
+			IBV_EXP_DEVICE_ATTR_EXP_CAP_FLAGS |
+			IBV_EXP_DEVICE_ATTR_RX_HASH;
 #endif /* HAVE_EXP_QUERY_DEVICE */
 
 		DEBUG("using port %u (%08" PRIx32 ")", port, test);
@@ -365,6 +367,12 @@ mlx5_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
 		DEBUG("L2 tunnel checksum offloads are %ssupported",
 		      (priv->hw_csum_l2tun ? "" : "not "));
 
+		priv->ind_table_max_size = exp_device_attr.rx_hash_caps.max_rwq_indirection_table_size;
+		DEBUG("maximum RX indirection table size is %u",
+		      priv->ind_table_max_size);
+
+#else /* HAVE_EXP_QUERY_DEVICE */
+		priv->ind_table_max_size = RSS_INDIRECTION_TABLE_SIZE;
 #endif /* HAVE_EXP_QUERY_DEVICE */
 
 		priv->vf = vf;
diff --git a/drivers/net/mlx5/mlx5.h b/drivers/net/mlx5/mlx5.h
index 08900f5..b099dac 100644
--- a/drivers/net/mlx5/mlx5.h
+++ b/drivers/net/mlx5/mlx5.h
@@ -114,6 +114,7 @@ struct priv {
 	/* Indirection tables referencing all RX WQs. */
 	struct ibv_exp_rwq_ind_table *(*ind_tables)[];
 	unsigned int ind_tables_n; /* Number of indirection tables. */
+	unsigned int ind_table_max_size; /* Maximum indirection table size. */
 	/* Hash RX QPs feeding the indirection table. */
 	struct hash_rxq (*hash_rxqs)[];
 	unsigned int hash_rxqs_n; /* Hash RX QPs array size. */
diff --git a/drivers/net/mlx5/mlx5_defs.h b/drivers/net/mlx5/mlx5_defs.h
index 79de609..e697764 100644
--- a/drivers/net/mlx5/mlx5_defs.h
+++ b/drivers/net/mlx5/mlx5_defs.h
@@ -53,6 +53,9 @@
 /* Request send completion once in every 64 sends, might be less. */
 #define MLX5_PMD_TX_PER_COMP_REQ 64
 
+/* RSS Indirection table size. */
+#define RSS_INDIRECTION_TABLE_SIZE 128
+
 /* Maximum number of Scatter/Gather Elements per Work Request. */
 #ifndef MLX5_PMD_SGE_WR_N
 #define MLX5_PMD_SGE_WR_N 4
diff --git a/drivers/net/mlx5/mlx5_rxq.c b/drivers/net/mlx5/mlx5_rxq.c
index b5084f8..606367c 100644
--- a/drivers/net/mlx5/mlx5_rxq.c
+++ b/drivers/net/mlx5/mlx5_rxq.c
@@ -224,7 +224,13 @@ priv_make_ind_table_init(struct priv *priv,
 int
 priv_create_hash_rxqs(struct priv *priv)
 {
-	unsigned int wqs_n = (1 << log2above(priv->rxqs_n));
+	/* If the requested number of WQs is not a power of two, use the
+	 * maximum indirection table size for better balancing.
+	 * The result is always rounded to the next power of two. */
+	unsigned int wqs_n =
+		(1 << log2above((priv->rxqs_n & (priv->rxqs_n - 1)) ?
+				priv->ind_table_max_size :
+				priv->rxqs_n));
 	struct ibv_exp_wq *wqs[wqs_n];
 	struct ind_table_init ind_table_init[IND_TABLE_INIT_N];
 	unsigned int ind_tables_n =
@@ -251,16 +257,17 @@ priv_create_hash_rxqs(struct priv *priv)
 		      " indirection table cannot be created");
 		return EINVAL;
 	}
-	if (wqs_n < priv->rxqs_n) {
+	if ((wqs_n < priv->rxqs_n) || (wqs_n > priv->ind_table_max_size)) {
 		ERROR("cannot handle this many RX queues (%u)", priv->rxqs_n);
 		err = ERANGE;
 		goto error;
 	}
-	if (wqs_n != priv->rxqs_n)
-		WARN("%u RX queues are configured, consider rounding this"
-		     " number to the next power of two (%u) for optimal"
-		     " performance",
-		     priv->rxqs_n, wqs_n);
+	if (wqs_n != priv->rxqs_n) {
+		INFO("%u RX queues are configured, consider rounding this"
+		     " number to the next power of two for better balancing",
+		     priv->rxqs_n);
+		DEBUG("indirection table extended to assume %u WQs", wqs_n);
+	}
 	/* When the number of RX queues is not a power of two, the remaining
 	 * table entries are padded with reused WQs and hashes are not spread
 	 * uniformly. */
-- 
2.1.0



More information about the dev mailing list