net/mlx5: enhance SR-IOV detection
[dpdk.git] / drivers / net / mlx5 / mlx5.c
index 52bf4b2..67a541c 100644 (file)
 #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.
  *
@@ -78,7 +97,7 @@
 static void
 mlx5_dev_close(struct rte_eth_dev *dev)
 {
-       struct priv *priv = dev->data->dev_private;
+       struct priv *priv = mlx5_get_priv(dev);
        void *tmp;
        unsigned int i;
 
@@ -88,10 +107,14 @@ mlx5_dev_close(struct rte_eth_dev *dev)
              ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
        /* In case mlx5_dev_stop() has not been called. */
        priv_dev_interrupt_handler_uninstall(priv, dev);
-       priv_special_flow_disable(priv, HASH_RXQ_FLOW_TYPE_ALLMULTI);
-       priv_special_flow_disable(priv, HASH_RXQ_FLOW_TYPE_PROMISC);
+       priv_special_flow_disable_all(priv);
        priv_mac_addrs_disable(priv);
        priv_destroy_hash_rxqs(priv);
+
+       /* Remove flow director elements. */
+       priv_fdir_disable(priv);
+       priv_fdir_delete_filters_list(priv);
+
        /* Prevent crashes when queues are still in use. */
        dev->rx_pkt_burst = removed_rx_burst;
        dev->tx_pkt_burst = removed_tx_burst;
@@ -144,6 +167,8 @@ static const struct eth_dev_ops mlx5_dev_ops = {
        .dev_configure = mlx5_dev_configure,
        .dev_start = mlx5_dev_start,
        .dev_stop = mlx5_dev_stop,
+       .dev_set_link_down = mlx5_set_link_down,
+       .dev_set_link_up = mlx5_set_link_up,
        .dev_close = mlx5_dev_close,
        .promiscuous_enable = mlx5_promiscuous_enable,
        .promiscuous_disable = mlx5_promiscuous_disable,
@@ -153,6 +178,7 @@ static const struct eth_dev_ops mlx5_dev_ops = {
        .stats_get = mlx5_stats_get,
        .stats_reset = mlx5_stats_reset,
        .dev_infos_get = mlx5_dev_infos_get,
+       .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
        .vlan_filter_set = mlx5_vlan_filter_set,
        .rx_queue_setup = mlx5_rx_queue_setup,
        .tx_queue_setup = mlx5_tx_queue_setup,
@@ -164,10 +190,17 @@ static const struct eth_dev_ops mlx5_dev_ops = {
        .mac_addr_add = mlx5_mac_addr_add,
        .mac_addr_set = mlx5_mac_addr_set,
        .mtu_set = mlx5_dev_set_mtu,
+#ifdef HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS
+       .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
+       .vlan_offload_set = mlx5_vlan_offload_set,
+#endif /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
        .reta_update = mlx5_dev_rss_reta_update,
        .reta_query = mlx5_dev_rss_reta_query,
        .rss_hash_update = mlx5_rss_hash_update,
        .rss_hash_conf_get = mlx5_rss_hash_conf_get,
+#ifdef MLX5_FDIR_SUPPORT
+       .filter_ctrl = mlx5_dev_filter_ctrl,
+#endif /* MLX5_FDIR_SUPPORT */
 };
 
 static struct {
@@ -227,7 +260,8 @@ mlx5_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
        int err = 0;
        struct ibv_context *attr_ctx = NULL;
        struct ibv_device_attr device_attr;
-       unsigned int vf;
+       unsigned int sriov;
+       unsigned int mps;
        int idx;
        int i;
 
@@ -269,12 +303,18 @@ mlx5_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                    (pci_dev->addr.devid != pci_addr.devid) ||
                    (pci_dev->addr.function != pci_addr.function))
                        continue;
-               vf = ((pci_dev->id.device_id ==
+               sriov = ((pci_dev->id.device_id ==
                       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF) ||
                      (pci_dev->id.device_id ==
                       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF));
-               INFO("PCI information matches, using device \"%s\" (VF: %s)",
-                    list[i]->name, (vf ? "true" : "false"));
+               /* Multi-packet send is only supported by ConnectX-4 Lx PF. */
+               mps = (pci_dev->id.device_id ==
+                      PCI_DEVICE_ID_MELLANOX_CONNECTX4LX);
+               INFO("PCI information matches, using device \"%s\""
+                    " (SR-IOV: %s, MPS: %s)",
+                    list[i]->name,
+                    sriov ? "true" : "false",
+                    mps ? "true" : "false");
                attr_ctx = ibv_open_device(list[i]);
                err = errno;
                break;
@@ -311,11 +351,19 @@ mlx5_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                struct ibv_exp_device_attr exp_device_attr;
 #endif /* HAVE_EXP_QUERY_DEVICE */
                struct ether_addr mac;
+               uint16_t num_vfs = 0;
 
 #ifdef HAVE_EXP_QUERY_DEVICE
                exp_device_attr.comp_mask =
                        IBV_EXP_DEVICE_ATTR_EXP_CAP_FLAGS |
-                       IBV_EXP_DEVICE_ATTR_RX_HASH;
+                       IBV_EXP_DEVICE_ATTR_RX_HASH |
+#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 */
 
                DEBUG("using port %u (%08" PRIx32 ")", port, test);
@@ -330,6 +378,13 @@ mlx5_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                        ERROR("port query failed: %s", strerror(err));
                        goto port_error;
                }
+
+               if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
+                       ERROR("port %d is not configured in Ethernet mode",
+                             port);
+                       goto port_error;
+               }
+
                if (port_attr.state != IBV_PORT_ACTIVE)
                        DEBUG("port %d is not active: \"%s\" (%d)",
                              port, ibv_port_state_str(port_attr.state),
@@ -386,12 +441,33 @@ mlx5_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                        priv->ind_table_max_size = RSS_INDIRECTION_TABLE_SIZE;
                DEBUG("maximum RX indirection table size is %u",
                      priv->ind_table_max_size);
+#ifdef HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS
+               priv->hw_vlan_strip = !!(exp_device_attr.wq_vlan_offloads_cap &
+                                        IBV_EXP_RECEIVE_WQ_CVLAN_STRIP);
+#endif /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
+               DEBUG("VLAN stripping is %ssupported",
+                     (priv->hw_vlan_strip ? "" : "not "));
+
+#ifdef HAVE_VERBS_FCS
+               priv->hw_fcs_strip = !!(exp_device_attr.exp_device_cap_flags &
+                                       IBV_EXP_DEVICE_SCATTER_FCS);
+#endif /* HAVE_VERBS_FCS */
+               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 */
 
-               priv->vf = vf;
+               priv_get_num_vfs(priv, &num_vfs);
+               priv->sriov = (num_vfs || sriov);
+               priv->mps = mps;
                /* Allocate and register default RSS hash keys. */
                priv->rss_conf = rte_calloc(__func__, hash_rxq_init_n,
                                            sizeof((*priv->rss_conf)[0]), 0);
@@ -416,13 +492,14 @@ mlx5_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                     mac.addr_bytes[0], mac.addr_bytes[1],
                     mac.addr_bytes[2], mac.addr_bytes[3],
                     mac.addr_bytes[4], mac.addr_bytes[5]);
-               /* Register MAC and broadcast addresses. */
+               /* Register MAC address. */
                claim_zero(priv_mac_addr_add(priv, 0,
                                             (const uint8_t (*)[ETHER_ADDR_LEN])
                                             mac.addr_bytes));
-               claim_zero(priv_mac_addr_add(priv, (RTE_DIM(priv->mac) - 1),
-                                            &(const uint8_t [ETHER_ADDR_LEN])
-                                            { "\xff\xff\xff\xff\xff\xff" }));
+               /* Initialize FD filters list. */
+               err = fdir_init_filters_list(priv);
+               if (err)
+                       goto port_error;
 #ifndef NDEBUG
                {
                        char ifname[IF_NAMESIZE];
@@ -452,18 +529,44 @@ mlx5_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                        goto port_error;
                }
 
-               eth_dev->data->dev_private = priv;
-               eth_dev->pci_dev = pci_dev;
+               /* Secondary processes have to use local storage for their
+                * private data as well as a copy of eth_dev->data, but this
+                * pointer must not be modified before burst functions are
+                * actually called. */
+               if (mlx5_is_secondary()) {
+                       struct mlx5_secondary_data *sd =
+                               &mlx5_secondary_data[eth_dev->data->port_id];
+                       sd->primary_priv = eth_dev->data->dev_private;
+                       if (sd->primary_priv == NULL) {
+                               ERROR("no private data for port %u",
+                                               eth_dev->data->port_id);
+                               err = EINVAL;
+                               goto port_error;
+                       }
+                       sd->shared_dev_data = eth_dev->data;
+                       rte_spinlock_init(&sd->lock);
+                       memcpy(sd->data.name, sd->shared_dev_data->name,
+                                  sizeof(sd->data.name));
+                       sd->data.dev_private = priv;
+                       sd->data.rx_mbuf_alloc_failed = 0;
+                       sd->data.mtu = ETHER_MTU;
+                       sd->data.port_id = sd->shared_dev_data->port_id;
+                       sd->data.mac_addrs = priv->mac;
+                       eth_dev->tx_pkt_burst = mlx5_tx_burst_secondary_setup;
+                       eth_dev->rx_pkt_burst = mlx5_rx_burst_secondary_setup;
+               } else {
+                       eth_dev->data->dev_private = priv;
+                       eth_dev->data->rx_mbuf_alloc_failed = 0;
+                       eth_dev->data->mtu = ETHER_MTU;
+                       eth_dev->data->mac_addrs = priv->mac;
+               }
 
+               eth_dev->pci_dev = pci_dev;
                rte_eth_copy_pci_info(eth_dev, pci_dev);
-
                eth_dev->driver = &mlx5_driver;
-               eth_dev->data->rx_mbuf_alloc_failed = 0;
-               eth_dev->data->mtu = ETHER_MTU;
-
                priv->dev = eth_dev;
                eth_dev->dev_ops = &mlx5_dev_ops;
-               eth_dev->data->mac_addrs = priv->mac;
+
                TAILQ_INIT(&eth_dev->link_intr_cbs);
 
                /* Bring Ethernet device up. */
@@ -472,8 +575,10 @@ mlx5_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                continue;
 
 port_error:
-               rte_free(priv->rss_conf);
-               rte_free(priv);
+               if (priv) {
+                       rte_free(priv->rss_conf);
+                       rte_free(priv);
+               }
                if (pd)
                        claim_zero(ibv_dealloc_pd(pd));
                if (ctx)