mlx5: support HW packet padding

Environment variable MLX5_PMD_ENABLE_PADDING enables HW packet padding
in PCI bus transactions.

When packet size is cache aligned and CRC stripping is enabled, 4 fewer
bytes are written to the PCI bus. Enabling padding makes such packets
aligned again.

In cases where PCI bandwidth is the bottleneck, padding can improve
performance by 10%.

This is disabled by default since this can also decrease performance for
unaligned packet sizes.

Signed-off-by: Olga Shern <olgas@mellanox.com>
fix packet padding macro check

Signed-off-by: Adrien Mazarguil <adrien.mazarguil@6wind.com>

diff --git a/doc/guides/nics/mlx5.rst b/doc/guides/nics/mlx5.rst
index 2573bb4..ee96770 100644 (file)
--- a/doc/guides/nics/mlx5.rst
+++ b/doc/guides/nics/mlx5.rst
@@ -151,6 +151,20 @@ Environment variables
   lower performance when there is no backpressure, it is not enabled by
   default.
 
+- ``MLX5_PMD_ENABLE_PADDING``
+
+  Enables HW packet padding in PCI bus transactions.
+
+  When packet size is cache aligned and CRC stripping is enabled, 4 fewer
+  bytes are written to the PCI bus. Enabling padding makes such packets
+  aligned again.
+
+  In cases where PCI bandwidth is the bottleneck, padding can improve
+  performance by 10%.
+
+  This is disabled by default since this can also decrease performance for
+  unaligned packet sizes.
+
 Run-time configuration
 ~~~~~~~~~~~~~~~~~~~~~~
 

diff --git a/doc/guides/rel_notes/release_16_04.rst b/doc/guides/rel_notes/release_16_04.rst
index 37e6407..189b0ee 100644 (file)
--- a/doc/guides/rel_notes/release_16_04.rst
+++ b/doc/guides/rel_notes/release_16_04.rst
@@ -223,6 +223,13 @@ This section should contain new features added in this release. Sample format:
 
   Only available with Mellanox OFED >= 3.2.
 
+* **Added mlx5 optional packet padding by HW.**
+
+  Added an option to make PCI bus transactions rounded to multiple of a
+  cache line size for better alignment.
+
+  Only available with Mellanox OFED >= 3.2.
+
 * **Changed szedata2 type of driver from vdev to pdev.**
 
   Previously szedata2 device had to be added by ``--vdev`` option.

diff --git a/drivers/net/mlx5/Makefile b/drivers/net/mlx5/Makefile
index 0b74991..c1bc427 100644 (file)
--- a/drivers/net/mlx5/Makefile
+++ b/drivers/net/mlx5/Makefile
@@ -136,6 +136,11 @@ mlx5_autoconf.h: $(RTE_SDK)/scripts/auto-config-h.sh
                infiniband/verbs.h \
                enum IBV_EXP_CREATE_WQ_FLAG_SCATTER_FCS \
                $(AUTOCONF_OUTPUT)
+       $Q sh -- '$<' '$@' \
+               HAVE_VERBS_RX_END_PADDING \
+               infiniband/verbs.h \
+               enum IBV_EXP_CREATE_WQ_FLAG_RX_END_PADDING \
+               $(AUTOCONF_OUTPUT)
 
 $(SRCS-$(CONFIG_RTE_LIBRTE_MLX5_PMD):.c=.o): mlx5_autoconf.h
 

diff --git a/drivers/net/mlx5/mlx5.c b/drivers/net/mlx5/mlx5.c
index 02ff362..5dc9f26 100644 (file)
--- a/drivers/net/mlx5/mlx5.c
+++ b/drivers/net/mlx5/mlx5.c
@@ -67,6 +67,25 @@
 #include "mlx5_autoconf.h"
 #include "mlx5_defs.h"
 
+/**
+ * Retrieve integer value from environment variable.
+ *
+ * @param[in] name
+ *   Environment variable name.
+ *
+ * @return
+ *   Integer value, 0 if the variable is not set.
+ */
+int
+mlx5_getenv_int(const char *name)
+{
+       const char *val = getenv(name);
+
+       if (val == NULL)
+               return 0;
+       return atoi(val);
+}
+
 /**
  * DPDK callback to close the device.
  *
@@ -333,6 +352,9 @@ mlx5_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
 #ifdef HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS
                        IBV_EXP_DEVICE_ATTR_VLAN_OFFLOADS |
 #endif /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
+#ifdef HAVE_VERBS_RX_END_PADDING
+                       IBV_EXP_DEVICE_ATTR_RX_PAD_END_ALIGN |
+#endif /* HAVE_VERBS_RX_END_PADDING */
                        0;
 #endif /* HAVE_EXP_QUERY_DEVICE */
 
@@ -425,6 +447,12 @@ mlx5_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                DEBUG("FCS stripping configuration is %ssupported",
                      (priv->hw_fcs_strip ? "" : "not "));
 
+#ifdef HAVE_VERBS_RX_END_PADDING
+               priv->hw_padding = !!exp_device_attr.rx_pad_end_addr_align;
+#endif /* HAVE_VERBS_RX_END_PADDING */
+               DEBUG("hardware RX end alignment padding is %ssupported",
+                     (priv->hw_padding ? "" : "not "));
+
 #else /* HAVE_EXP_QUERY_DEVICE */
                priv->ind_table_max_size = RSS_INDIRECTION_TABLE_SIZE;
 #endif /* HAVE_EXP_QUERY_DEVICE */

diff --git a/drivers/net/mlx5/mlx5.h b/drivers/net/mlx5/mlx5.h
index 399f885..0a13272 100644 (file)
--- a/drivers/net/mlx5/mlx5.h
+++ b/drivers/net/mlx5/mlx5.h
@@ -104,6 +104,7 @@ struct priv {
        unsigned int hw_csum_l2tun:1; /* Same for L2 tunnels. */
        unsigned int hw_vlan_strip:1; /* VLAN stripping is supported. */
        unsigned int hw_fcs_strip:1; /* FCS stripping is supported. */
+       unsigned int hw_padding:1; /* End alignment padding is supported. */
        unsigned int vf:1; /* This is a VF device. */
        unsigned int pending_alarm:1; /* An alarm is pending. */
        /* RX/TX queues. */
@@ -160,6 +161,10 @@ priv_unlock(struct priv *priv)
        rte_spinlock_unlock(&priv->lock);
 }
 
+/* mlx5.c */
+
+int mlx5_getenv_int(const char *);
+
 /* mlx5_ethdev.c */
 
 struct priv *mlx5_get_priv(struct rte_eth_dev *dev);

diff --git a/drivers/net/mlx5/mlx5_rxq.c b/drivers/net/mlx5/mlx5_rxq.c
index 19a1119..c8af77f 100644 (file)
--- a/drivers/net/mlx5/mlx5_rxq.c
+++ b/drivers/net/mlx5/mlx5_rxq.c
@@ -1282,6 +1282,21 @@ rxq_setup(struct rte_eth_dev *dev, struct rxq *rxq, uint16_t desc,
              tmpl.crc_present << 2);
 #endif /* HAVE_VERBS_FCS */
 
+#ifdef HAVE_VERBS_RX_END_PADDING
+       if (!mlx5_getenv_int("MLX5_PMD_ENABLE_PADDING"))
+               ; /* Nothing else to do. */
+       else if (priv->hw_padding) {
+               INFO("%p: enabling packet padding on queue %p",
+                    (void *)dev, (void *)rxq);
+               attr.wq.flags |= IBV_EXP_CREATE_WQ_FLAG_RX_END_PADDING;
+               attr.wq.comp_mask |= IBV_EXP_CREATE_WQ_FLAGS;
+       } else
+               WARN("%p: packet padding has been requested but is not"
+                    " supported, make sure MLNX_OFED and firmware are"
+                    " up to date",
+                    (void *)dev);
+#endif /* HAVE_VERBS_RX_END_PADDING */
+
        tmpl.wq = ibv_exp_create_wq(priv->ctx, &attr.wq);
        if (tmpl.wq == NULL) {
                ret = (errno ? errno : EINVAL);