net/mlx5: add new memory region support
[dpdk.git] / drivers / net / mlx5 / mlx5.c
index 68783c3..84fc9d5 100644 (file)
@@ -41,6 +41,7 @@
 #include "mlx5_autoconf.h"
 #include "mlx5_defs.h"
 #include "mlx5_glue.h"
+#include "mlx5_mr.h"
 
 /* Device parameter to enable RX completion queue compression. */
 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
@@ -69,6 +70,9 @@
 /* Device parameter to enable hardware Rx vector. */
 #define MLX5_RX_VEC_EN "rx_vec_en"
 
+/* Allow L3 VXLAN flow creation. */
+#define MLX5_L3_VXLAN_EN "l3_vxlan_en"
+
 /* Activate Netlink support in VF mode. */
 #define MLX5_VF_NL_EN "vf_nl_en"
 
 #define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
 #endif
 
+static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
+
+/* Shared memory between primary and secondary processes. */
+struct mlx5_shared_data *mlx5_shared_data;
+
+/* Spinlock for mlx5_shared_data allocation. */
+static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
+
 /** Driver-specific log messages type. */
 int mlx5_logtype;
 
+/**
+ * Prepare shared data between primary and secondary process.
+ */
+static void
+mlx5_prepare_shared_data(void)
+{
+       const struct rte_memzone *mz;
+
+       rte_spinlock_lock(&mlx5_shared_data_lock);
+       if (mlx5_shared_data == NULL) {
+               if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
+                       /* Allocate shared memory. */
+                       mz = rte_memzone_reserve(MZ_MLX5_PMD_SHARED_DATA,
+                                                sizeof(*mlx5_shared_data),
+                                                SOCKET_ID_ANY, 0);
+               } else {
+                       /* Lookup allocated shared memory. */
+                       mz = rte_memzone_lookup(MZ_MLX5_PMD_SHARED_DATA);
+               }
+               if (mz == NULL)
+                       rte_panic("Cannot allocate mlx5 shared data\n");
+               mlx5_shared_data = mz->addr;
+               /* Initialize shared data. */
+               if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
+                       LIST_INIT(&mlx5_shared_data->mem_event_cb_list);
+                       rte_rwlock_init(&mlx5_shared_data->mem_event_rwlock);
+               }
+       }
+       rte_spinlock_unlock(&mlx5_shared_data_lock);
+}
+
 /**
  * Retrieve integer value from environment variable.
  *
@@ -197,6 +240,8 @@ mlx5_dev_close(struct rte_eth_dev *dev)
                priv->txqs_n = 0;
                priv->txqs = NULL;
        }
+       mlx5_flow_delete_drop_queue(dev);
+       mlx5_mr_release(dev);
        if (priv->pd != NULL) {
                assert(priv->ctx != NULL);
                claim_zero(mlx5_glue->dealloc_pd(priv->pd));
@@ -241,10 +286,6 @@ mlx5_dev_close(struct rte_eth_dev *dev)
        if (ret)
                DRV_LOG(WARNING, "port %u some flows still remain",
                        dev->data->port_id);
-       ret = mlx5_mr_verify(dev);
-       if (ret)
-               DRV_LOG(WARNING, "port %u some memory region still remain",
-                       dev->data->port_id);
        memset(priv, 0, sizeof(*priv));
 }
 
@@ -277,6 +318,7 @@ const struct eth_dev_ops mlx5_dev_ops = {
        .mac_addr_remove = mlx5_mac_addr_remove,
        .mac_addr_add = mlx5_mac_addr_add,
        .mac_addr_set = mlx5_mac_addr_set,
+       .set_mc_addr_list = mlx5_set_mc_addr_list,
        .mtu_set = mlx5_dev_set_mtu,
        .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
        .vlan_offload_set = mlx5_vlan_offload_set,
@@ -329,6 +371,7 @@ const struct eth_dev_ops mlx5_dev_ops_isolate = {
        .mac_addr_remove = mlx5_mac_addr_remove,
        .mac_addr_add = mlx5_mac_addr_add,
        .mac_addr_set = mlx5_mac_addr_set,
+       .set_mc_addr_list = mlx5_set_mc_addr_list,
        .mtu_set = mlx5_dev_set_mtu,
        .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
        .vlan_offload_set = mlx5_vlan_offload_set,
@@ -415,6 +458,8 @@ mlx5_args_check(const char *key, const char *val, void *opaque)
                config->tx_vec_en = !!tmp;
        } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
                config->rx_vec_en = !!tmp;
+       } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
+               config->l3_vxlan_en = !!tmp;
        } else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
                config->vf_nl_en = !!tmp;
        } else {
@@ -448,6 +493,7 @@ mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
                MLX5_TXQ_MAX_INLINE_LEN,
                MLX5_TX_VEC_EN,
                MLX5_RX_VEC_EN,
+               MLX5_L3_VXLAN_EN,
                MLX5_VF_NL_EN,
                NULL,
        };
@@ -619,6 +665,8 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
        unsigned int mps;
        unsigned int cqe_comp;
        unsigned int tunnel_en = 0;
+       unsigned int swp = 0;
+       unsigned int verb_priorities = 0;
        int idx;
        int i;
        struct mlx5dv_context attrs_out = {0};
@@ -626,6 +674,8 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
        struct ibv_counter_set_description cs_desc;
 #endif
 
+       /* Prepare shared data between primary and secondary process. */
+       mlx5_prepare_shared_data();
        assert(pci_drv == &mlx5_driver);
        /* Get mlx5_dev[] index. */
        idx = mlx5_dev_idx(&pci_dev->addr);
@@ -679,21 +729,24 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
                break;
        }
        if (attr_ctx == NULL) {
-               mlx5_glue->free_device_list(list);
                switch (err) {
                case 0:
                        DRV_LOG(ERR,
                                "cannot access device, is mlx5_ib loaded?");
                        err = ENODEV;
-                       goto error;
+                       break;
                case EINVAL:
                        DRV_LOG(ERR,
                                "cannot use device, are drivers up to date?");
-                       goto error;
+                       break;
                }
+               goto error;
        }
        ibv_dev = list[i];
        DRV_LOG(DEBUG, "device opened");
+#ifdef HAVE_IBV_MLX5_MOD_SWP
+       attrs_out.comp_mask |= MLX5DV_CONTEXT_MASK_SWP;
+#endif
        /*
         * Multi-packet send is supported by ConnectX-4 Lx PF as well
         * as all ConnectX-5 devices.
@@ -714,6 +767,11 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
                DRV_LOG(DEBUG, "MPW isn't supported");
                mps = MLX5_MPW_DISABLED;
        }
+#ifdef HAVE_IBV_MLX5_MOD_SWP
+       if (attrs_out.comp_mask & MLX5DV_CONTEXT_MASK_SWP)
+               swp = attrs_out.sw_parsing_caps.sw_parsing_offloads;
+       DRV_LOG(DEBUG, "SWP support: %u", swp);
+#endif
        if (RTE_CACHE_LINE_SIZE == 128 &&
            !(attrs_out.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
                cqe_comp = 0;
@@ -732,8 +790,9 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
        DRV_LOG(WARNING,
                "tunnel offloading disabled due to old OFED/rdma-core version");
 #endif
-       if (mlx5_glue->query_device_ex(attr_ctx, NULL, &device_attr)) {
-               err = errno;
+       err = mlx5_glue->query_device_ex(attr_ctx, NULL, &device_attr);
+       if (err) {
+               DEBUG("ibv_query_device_ex() failed");
                goto error;
        }
        DRV_LOG(INFO, "%u port(s) detected",
@@ -761,6 +820,7 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
                        .txqs_inline = MLX5_ARG_UNSET,
                        .inline_max_packet_sz = MLX5_ARG_UNSET,
                        .vf_nl_en = 1,
+                       .swp = !!swp,
                };
 
                len = snprintf(name, sizeof(name), PCI_PRI_FMT,
@@ -780,16 +840,22 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
                        eth_dev->device = &pci_dev->device;
                        eth_dev->dev_ops = &mlx5_dev_sec_ops;
                        err = mlx5_uar_init_secondary(eth_dev);
-                       if (err)
+                       if (err) {
+                               err = rte_errno;
                                goto error;
+                       }
                        /* Receive command fd from primary process */
                        err = mlx5_socket_connect(eth_dev);
-                       if (err)
+                       if (err < 0) {
+                               err = rte_errno;
                                goto error;
+                       }
                        /* Remap UAR for Tx queues. */
                        err = mlx5_tx_uar_remap(eth_dev, err);
-                       if (err)
+                       if (err) {
+                               err = rte_errno;
                                goto error;
+                       }
                        /*
                         * Ethdev pointer is still required as input since
                         * the primary device is not accessible from the
@@ -853,11 +919,12 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
                if (err) {
                        DRV_LOG(ERR, "failed to process device arguments: %s",
                                strerror(err));
+                       err = rte_errno;
                        goto port_error;
                }
-               if (mlx5_glue->query_device_ex(ctx, NULL, &device_attr_ex)) {
+               err = mlx5_glue->query_device_ex(ctx, NULL, &device_attr_ex);
+               if (err) {
                        DRV_LOG(ERR, "ibv_query_device_ex() failed");
-                       err = errno;
                        goto port_error;
                }
                config.hw_csum = !!(device_attr_ex.device_cap_flags_ex &
@@ -926,14 +993,16 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
                        goto port_error;
                }
                eth_dev->data->dev_private = priv;
-               priv->dev = eth_dev;
+               priv->dev_data = eth_dev->data;
                eth_dev->data->mac_addrs = priv->mac;
                eth_dev->device = &pci_dev->device;
                rte_eth_copy_pci_info(eth_dev, pci_dev);
                eth_dev->device->driver = &mlx5_driver.driver;
                err = mlx5_uar_init_primary(eth_dev);
-               if (err)
+               if (err) {
+                       err = rte_errno;
                        goto port_error;
+               }
                /* Configure the first MAC address by default. */
                if (mlx5_get_mac(eth_dev, &mac.addr_bytes)) {
                        DRV_LOG(ERR,
@@ -963,8 +1032,10 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
 #endif
                /* Get actual MTU if possible. */
                err = mlx5_get_mtu(eth_dev, &priv->mtu);
-               if (err)
+               if (err) {
+                       err = rte_errno;
                        goto port_error;
+               }
                DRV_LOG(DEBUG, "port %u MTU is %u", eth_dev->data->port_id,
                        priv->mtu);
                /*
@@ -1006,6 +1077,23 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
                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 port_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);
+                       goto port_error;
+               }
+               priv->config.max_verbs_prio = verb_priorities;
                continue;
 port_error:
                if (priv)
@@ -1014,6 +1102,8 @@ port_error:
                        claim_zero(mlx5_glue->dealloc_pd(pd));
                if (ctx)
                        claim_zero(mlx5_glue->close_device(ctx));
+               if (eth_dev && rte_eal_process_type() == RTE_PROC_PRIMARY)
+                       rte_eth_dev_release_port(eth_dev);
                break;
        }
        /*
@@ -1225,8 +1315,10 @@ RTE_INIT(rte_mlx5_pmd_init);
 static void
 rte_mlx5_pmd_init(void)
 {
-       /* Build the static table for ptype conversion. */
+       /* Build the static tables for Verbs conversion. */
        mlx5_set_ptype_table();
+       mlx5_set_cksum_table();
+       mlx5_set_swp_types_table();
        /*
         * RDMAV_HUGEPAGES_SAFE tells ibv_fork_init() we intend to use
         * huge pages. Calling ibv_fork_init() during init allows
@@ -1259,6 +1351,8 @@ rte_mlx5_pmd_init(void)
        }
        mlx5_glue->fork_init();
        rte_pci_register(&mlx5_driver);
+       rte_mem_event_callback_register("MLX5_MEM_EVENT_CB",
+                                       mlx5_mr_mem_event_cb, NULL);
 }
 
 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);