net/mlx5: enhance SR-IOV detection
[dpdk.git] / drivers / net / mlx5 / mlx5.c
index ddd74d0..67a541c 100644 (file)
 #include "mlx5_utils.h"
 #include "mlx5_rxtx.h"
 #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.
@@ -77,7 +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;
 
@@ -85,6 +105,16 @@ mlx5_dev_close(struct rte_eth_dev *dev)
        DEBUG("%p: closing device \"%s\"",
              (void *)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_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;
@@ -116,14 +146,19 @@ mlx5_dev_close(struct rte_eth_dev *dev)
                priv->txqs_n = 0;
                priv->txqs = NULL;
        }
-       if (priv->rss)
-               rxq_cleanup(&priv->rxq_parent);
        if (priv->pd != NULL) {
                assert(priv->ctx != NULL);
                claim_zero(ibv_dealloc_pd(priv->pd));
                claim_zero(ibv_close_device(priv->ctx));
        } else
                assert(priv->ctx == NULL);
+       if (priv->rss_conf != NULL) {
+               for (i = 0; (i != hash_rxq_init_n); ++i)
+                       rte_free((*priv->rss_conf)[i]);
+               rte_free(priv->rss_conf);
+       }
+       if (priv->reta_idx != NULL)
+               rte_free(priv->reta_idx);
        priv_unlock(priv);
        memset(priv, 0, sizeof(*priv));
 }
@@ -132,15 +167,40 @@ 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,
+       .allmulticast_enable = mlx5_allmulticast_enable,
+       .allmulticast_disable = mlx5_allmulticast_disable,
+       .link_update = mlx5_link_update,
+       .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,
        .rx_queue_release = mlx5_rx_queue_release,
        .tx_queue_release = mlx5_tx_queue_release,
+       .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
+       .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
        .mac_addr_remove = mlx5_mac_addr_remove,
        .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 {
@@ -200,7 +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;
 
@@ -242,12 +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;
@@ -284,12 +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;
-#ifdef RSS_SUPPORT
-               exp_device_attr.comp_mask |= IBV_EXP_DEVICE_ATTR_RSS_TBL_SZ;
-#endif /* RSS_SUPPORT */
+               exp_device_attr.comp_mask =
+                       IBV_EXP_DEVICE_ATTR_EXP_CAP_FLAGS |
+                       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);
@@ -304,6 +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),
@@ -339,32 +420,6 @@ mlx5_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                        ERROR("ibv_exp_query_device() failed");
                        goto port_error;
                }
-#ifdef RSS_SUPPORT
-               if ((exp_device_attr.exp_device_cap_flags &
-                    IBV_EXP_DEVICE_QPG) &&
-                   (exp_device_attr.exp_device_cap_flags &
-                    IBV_EXP_DEVICE_UD_RSS) &&
-                   (exp_device_attr.comp_mask &
-                    IBV_EXP_DEVICE_ATTR_RSS_TBL_SZ) &&
-                   (exp_device_attr.max_rss_tbl_sz > 0)) {
-                       priv->hw_qpg = 1;
-                       priv->hw_rss = 1;
-                       priv->max_rss_tbl_sz = exp_device_attr.max_rss_tbl_sz;
-               } else {
-                       priv->hw_qpg = 0;
-                       priv->hw_rss = 0;
-                       priv->max_rss_tbl_sz = 0;
-               }
-               priv->hw_tss = !!(exp_device_attr.exp_device_cap_flags &
-                                 IBV_EXP_DEVICE_UD_TSS);
-               DEBUG("device flags: %s%s%s",
-                     (priv->hw_qpg ? "IBV_DEVICE_QPG " : ""),
-                     (priv->hw_tss ? "IBV_DEVICE_TSS " : ""),
-                     (priv->hw_rss ? "IBV_DEVICE_RSS " : ""));
-               if (priv->hw_rss)
-                       DEBUG("maximum RSS indirection table size: %u",
-                             exp_device_attr.max_rss_tbl_sz);
-#endif /* RSS_SUPPORT */
 
                priv->hw_csum =
                        ((exp_device_attr.exp_device_cap_flags &
@@ -379,9 +434,53 @@ 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;
+               /* Remove this check once DPDK supports larger/variable
+                * indirection tables. */
+               if (priv->ind_table_max_size > (unsigned int)RSS_INDIRECTION_TABLE_SIZE)
+                       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);
+               if (priv->rss_conf == NULL) {
+                       err = ENOMEM;
+                       goto port_error;
+               }
+               err = rss_hash_rss_conf_new_key(priv,
+                                               rss_hash_default_key,
+                                               rss_hash_default_key_len,
+                                               ETH_RSS_PROTO_MASK);
+               if (err)
+                       goto port_error;
                /* Configure the first MAC address by default. */
                if (priv_get_mac(priv, &mac.addr_bytes)) {
                        ERROR("cannot get MAC address, is mlx5_en loaded?"
@@ -393,13 +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];
@@ -429,15 +529,45 @@ mlx5_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                        goto port_error;
                }
 
-               eth_dev->data->dev_private = priv;
+               /* 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. */
                DEBUG("forcing Ethernet interface up");
@@ -445,7 +575,10 @@ mlx5_pci_devinit(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                continue;
 
 port_error:
-               rte_free(priv);
+               if (priv) {
+                       rte_free(priv->rss_conf);
+                       rte_free(priv);
+               }
                if (pd)
                        claim_zero(ibv_dealloc_pd(pd));
                if (ctx)
@@ -510,6 +643,7 @@ static struct eth_driver mlx5_driver = {
                .name = MLX5_DRIVER_NAME,
                .id_table = mlx5_pci_id_map,
                .devinit = mlx5_pci_devinit,
+               .drv_flags = RTE_PCI_DRV_INTR_LSC,
        },
        .dev_private_size = sizeof(struct priv)
 };