net/mlx5: fix representors detection
[dpdk.git] / drivers / net / mlx5 / mlx5.c
index f17bd08..a1c0ad7 100644 (file)
@@ -92,6 +92,9 @@
 /* Activate Netlink support in VF mode. */
 #define MLX5_VF_NL_EN "vf_nl_en"
 
+/* Select port representors to instantiate. */
+#define MLX5_REPRESENTOR "representor"
+
 #ifndef HAVE_IBV_MLX5_MOD_MPW
 #define MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED (1 << 2)
 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
@@ -239,6 +242,7 @@ mlx5_dev_close(struct rte_eth_dev *dev)
        /* In case mlx5_dev_stop() has not been called. */
        mlx5_dev_interrupt_handler_uninstall(dev);
        mlx5_traffic_disable(dev);
+       mlx5_flow_flush(dev, NULL);
        /* Prevent crashes when queues are still in use. */
        dev->rx_pkt_burst = removed_rx_burst;
        dev->tx_pkt_burst = removed_tx_burst;
@@ -258,7 +262,6 @@ mlx5_dev_close(struct rte_eth_dev *dev)
                priv->txqs_n = 0;
                priv->txqs = NULL;
        }
-       mlx5_flow_delete_drop_queue(dev);
        mlx5_mprq_free_mp(dev);
        mlx5_mr_release(dev);
        if (priv->pd != NULL) {
@@ -279,6 +282,8 @@ mlx5_dev_close(struct rte_eth_dev *dev)
                close(priv->nl_socket_route);
        if (priv->nl_socket_rdma >= 0)
                close(priv->nl_socket_rdma);
+       if (priv->mnl_socket)
+               mlx5_nl_flow_socket_destroy(priv->mnl_socket);
        ret = mlx5_hrxq_ibv_verify(dev);
        if (ret)
                DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
@@ -443,6 +448,9 @@ mlx5_args_check(const char *key, const char *val, void *opaque)
        struct mlx5_dev_config *config = opaque;
        unsigned long tmp;
 
+       /* No-op, port representors are processed in mlx5_dev_spawn(). */
+       if (!strcmp(MLX5_REPRESENTOR, key))
+               return 0;
        errno = 0;
        tmp = strtoul(val, NULL, 0);
        if (errno) {
@@ -515,6 +523,7 @@ mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
                MLX5_RX_VEC_EN,
                MLX5_L3_VXLAN_EN,
                MLX5_VF_NL_EN,
+               MLX5_REPRESENTOR,
                NULL,
        };
        struct rte_kvargs *kvlist;
@@ -591,7 +600,7 @@ mlx5_uar_init_primary(struct rte_eth_dev *dev)
        rte_memseg_walk(find_lower_va_bound, &addr);
 
        /* keep distance to hugepages to minimize potential conflicts. */
-       addr = RTE_PTR_SUB(addr, MLX5_UAR_OFFSET + MLX5_UAR_SIZE);
+       addr = RTE_PTR_SUB(addr, (uintptr_t)(MLX5_UAR_OFFSET + MLX5_UAR_SIZE));
        /* anonymous mmap, no real memory consumption. */
        addr = mmap(addr, MLX5_UAR_SIZE,
                    PROT_NONE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
@@ -672,7 +681,9 @@ mlx5_uar_init_secondary(struct rte_eth_dev *dev)
  *
  * @return
  *   A valid Ethernet device object on success, NULL otherwise and rte_errno
- *   is set.
+ *   is set. The following error is defined:
+ *
+ *   EBUSY: device is not supposed to be spawned.
  */
 static struct rte_eth_dev *
 mlx5_dev_spawn(struct rte_device *dpdk_dev,
@@ -709,7 +720,6 @@ mlx5_dev_spawn(struct rte_device *dpdk_dev,
        unsigned int tunnel_en = 0;
        unsigned int mpls_en = 0;
        unsigned int swp = 0;
-       unsigned int verb_priorities = 0;
        unsigned int mprq = 0;
        unsigned int mprq_min_stride_size_n = 0;
        unsigned int mprq_max_stride_size_n = 0;
@@ -723,6 +733,26 @@ mlx5_dev_spawn(struct rte_device *dpdk_dev,
        int own_domain_id = 0;
        unsigned int i;
 
+       /* Determine if this port representor is supposed to be spawned. */
+       if (switch_info->representor && dpdk_dev->devargs) {
+               struct rte_eth_devargs eth_da;
+
+               err = rte_eth_devargs_parse(dpdk_dev->devargs->args, &eth_da);
+               if (err) {
+                       rte_errno = -err;
+                       DRV_LOG(ERR, "failed to process device arguments: %s",
+                               strerror(rte_errno));
+                       return NULL;
+               }
+               for (i = 0; i < eth_da.nb_representor_ports; ++i)
+                       if (eth_da.representor_ports[i] ==
+                           (uint16_t)switch_info->port_name)
+                               break;
+               if (i == eth_da.nb_representor_ports) {
+                       rte_errno = EBUSY;
+                       return NULL;
+               }
+       }
        /* Prepare shared data between primary and secondary process. */
        mlx5_prepare_shared_data();
        errno = 0;
@@ -911,6 +941,12 @@ mlx5_dev_spawn(struct rte_device *dpdk_dev,
        priv->device_attr = attr;
        priv->pd = pd;
        priv->mtu = ETHER_MTU;
+#ifndef RTE_ARCH_64
+       /* Initialize UAR access locks for 32bit implementations. */
+       rte_spinlock_init(&priv->uar_lock_cq);
+       for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
+               rte_spinlock_init(&priv->uar_lock[i]);
+#endif
        /* Some internal functions rely on Netlink sockets, open them now. */
        priv->nl_socket_rdma = mlx5_nl_init(0, NETLINK_RDMA);
        priv->nl_socket_route = mlx5_nl_init(RTMGRP_LINK, NETLINK_ROUTE);
@@ -1088,6 +1124,34 @@ mlx5_dev_spawn(struct rte_device *dpdk_dev,
        claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
        if (vf && config.vf_nl_en)
                mlx5_nl_mac_addr_sync(eth_dev);
+       priv->mnl_socket = mlx5_nl_flow_socket_create();
+       if (!priv->mnl_socket) {
+               err = -rte_errno;
+               DRV_LOG(WARNING,
+                       "flow rules relying on switch offloads will not be"
+                       " supported: cannot open libmnl socket: %s",
+                       strerror(rte_errno));
+       } else {
+               struct rte_flow_error error;
+               unsigned int ifindex = mlx5_ifindex(eth_dev);
+
+               if (!ifindex) {
+                       err = -rte_errno;
+                       error.message =
+                               "cannot retrieve network interface index";
+               } else {
+                       err = mlx5_nl_flow_init(priv->mnl_socket, ifindex,
+                                               &error);
+               }
+               if (err) {
+                       DRV_LOG(WARNING,
+                               "flow rules relying on switch offloads will"
+                               " not be supported: %s: %s",
+                               error.message, strerror(rte_errno));
+                       mlx5_nl_flow_socket_destroy(priv->mnl_socket);
+                       priv->mnl_socket = NULL;
+               }
+       }
        TAILQ_INIT(&priv->flows);
        TAILQ_INIT(&priv->ctrl_flows);
        /* Hint libmlx5 to use PMD allocator for data plane resources */
@@ -1110,24 +1174,11 @@ mlx5_dev_spawn(struct rte_device *dpdk_dev,
        mlx5_link_update(eth_dev, 0);
        /* Store device configuration on private structure. */
        priv->config = config;
-       /* Create drop queue. */
-       err = mlx5_flow_create_drop_queue(eth_dev);
-       if (err) {
-               DRV_LOG(ERR, "port %u drop queue allocation failed: %s",
-                       eth_dev->data->port_id, strerror(rte_errno));
-               err = rte_errno;
-               goto error;
-       }
        /* Supported Verbs flow priority number detection. */
-       if (verb_priorities == 0)
-               verb_priorities = mlx5_get_max_verbs_prio(eth_dev);
-       if (verb_priorities < MLX5_VERBS_FLOW_PRIO_8) {
-               DRV_LOG(ERR, "port %u wrong Verbs flow priorities: %u",
-                       eth_dev->data->port_id, verb_priorities);
-               err = ENOTSUP;
+       err = mlx5_flow_discover_priorities(eth_dev);
+       if (err < 0)
                goto error;
-       }
-       priv->config.max_verbs_prio = verb_priorities;
+       priv->config.flow_prio = err;
        /*
         * Once the device is added to the list of memory event
         * callback, its global MR cache table cannot be expanded
@@ -1153,6 +1204,8 @@ error:
                        close(priv->nl_socket_route);
                if (priv->nl_socket_rdma >= 0)
                        close(priv->nl_socket_rdma);
+               if (priv->mnl_socket)
+                       mlx5_nl_flow_socket_destroy(priv->mnl_socket);
                if (own_domain_id)
                        claim_zero(rte_eth_switch_domain_free(priv->domain_id));
                rte_free(priv);
@@ -1277,7 +1330,8 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
         * Netlink calls assuming kernel drivers are recent enough to
         * support them.
         *
-        * In the event of identification failure through Netlink, either:
+        * In the event of identification failure through Netlink, try again
+        * through sysfs, then either:
         *
         * 1. No device matches (n == 0), complain and bail out.
         * 2. A single IB device matches (n == 1) and is not a representor,
@@ -1296,7 +1350,9 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
                if (nl_route < 0 ||
                    !list[i].ifindex ||
                    mlx5_nl_switch_info(nl_route, list[i].ifindex,
-                                       &list[i].info)) {
+                                       &list[i].info) ||
+                   ((!list[i].info.representor && !list[i].info.master) &&
+                    mlx5_sysfs_switch_info(list[i].ifindex, &list[i].info))) {
                        list[i].ifindex = 0;
                        memset(&list[i].info, 0, sizeof(list[i].info));
                        continue;
@@ -1343,8 +1399,12 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
 
                list[i].eth_dev = mlx5_dev_spawn
                        (&pci_dev->device, list[i].ibv_dev, vf, &list[i].info);
-               if (!list[i].eth_dev)
-                       break;
+               if (!list[i].eth_dev) {
+                       if (rte_errno != EBUSY)
+                               break;
+                       /* Device is disabled, ignore it. */
+                       continue;
+               }
                restore = list[i].eth_dev->data->dev_flags;
                rte_eth_copy_pci_info(list[i].eth_dev, pci_dev);
                /* Restore non-PCI flags cleared by the above call. */
@@ -1370,6 +1430,8 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
                ret = -rte_errno;
                /* Roll back. */
                while (i--) {
+                       if (!list[i].eth_dev)
+                               continue;
                        mlx5_dev_close(list[i].eth_dev);
                        if (rte_eal_process_type() == RTE_PROC_PRIMARY)
                                rte_free(list[i].eth_dev->data->dev_private);