ethdev: remove detachable device flag
[dpdk.git] / drivers / net / mlx5 / mlx5.c
index ed77351..e3c9b89 100644 (file)
 #define MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW (1 << 3)
 #endif
 
+#ifndef HAVE_IBV_MLX5_MOD_CQE_128B_COMP
+#define MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP (1 << 4)
+#endif
+
 struct mlx5_args {
        int cqe_comp;
        int txq_inline;
@@ -198,44 +202,23 @@ 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_all(priv);
-       priv_mac_addrs_disable(priv);
-       priv_destroy_hash_rxqs(priv);
-
+       priv_dev_traffic_disable(priv, dev);
        /* Prevent crashes when queues are still in use. */
        dev->rx_pkt_burst = removed_rx_burst;
        dev->tx_pkt_burst = removed_tx_burst;
        if (priv->rxqs != NULL) {
                /* XXX race condition if mlx5_rx_burst() is still running. */
                usleep(1000);
-               for (i = 0; (i != priv->rxqs_n); ++i) {
-                       struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
-                       struct mlx5_rxq_ctrl *rxq_ctrl;
-
-                       if (rxq == NULL)
-                               continue;
-                       rxq_ctrl = container_of(rxq, struct mlx5_rxq_ctrl, rxq);
-                       (*priv->rxqs)[i] = NULL;
-                       mlx5_rxq_cleanup(rxq_ctrl);
-                       rte_free(rxq_ctrl);
-               }
+               for (i = 0; (i != priv->rxqs_n); ++i)
+                       mlx5_priv_rxq_release(priv, i);
                priv->rxqs_n = 0;
                priv->rxqs = NULL;
        }
        if (priv->txqs != NULL) {
                /* XXX race condition if mlx5_tx_burst() is still running. */
                usleep(1000);
-               for (i = 0; (i != priv->txqs_n); ++i) {
-                       struct mlx5_txq_data *txq = (*priv->txqs)[i];
-                       struct mlx5_txq_ctrl *txq_ctrl;
-
-                       if (txq == NULL)
-                               continue;
-                       txq_ctrl = container_of(txq, struct mlx5_txq_ctrl, txq);
-                       (*priv->txqs)[i] = NULL;
-                       mlx5_txq_cleanup(txq_ctrl);
-                       rte_free(txq_ctrl);
-               }
+               for (i = 0; (i != priv->txqs_n); ++i)
+                       mlx5_priv_txq_release(priv, i);
                priv->txqs_n = 0;
                priv->txqs = NULL;
        }
@@ -245,17 +228,29 @@ mlx5_dev_close(struct rte_eth_dev *dev)
                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->rss_conf.rss_key != NULL)
+               rte_free(priv->rss_conf.rss_key);
        if (priv->reta_idx != NULL)
                rte_free(priv->reta_idx);
        priv_socket_uninit(priv);
+       ret = mlx5_priv_hrxq_ibv_verify(priv);
+       if (ret)
+               WARN("%p: some Hash Rx queue still remain", (void *)priv);
+       ret = mlx5_priv_ind_table_ibv_verify(priv);
+       if (ret)
+               WARN("%p: some Indirection table still remain", (void *)priv);
        ret = mlx5_priv_rxq_ibv_verify(priv);
        if (ret)
                WARN("%p: some Verbs Rx queue still remain", (void *)priv);
+       ret = mlx5_priv_rxq_verify(priv);
+       if (ret)
+               WARN("%p: some Rx Queues still remain", (void *)priv);
+       ret = mlx5_priv_txq_ibv_verify(priv);
+       if (ret)
+               WARN("%p: some Verbs Tx queue still remain", (void *)priv);
+       ret = mlx5_priv_txq_verify(priv);
+       if (ret)
+               WARN("%p: some Tx Queues still remain", (void *)priv);
        ret = priv_flow_verify(priv);
        if (ret)
                WARN("%p: some flows still remain", (void *)priv);
@@ -266,7 +261,7 @@ mlx5_dev_close(struct rte_eth_dev *dev)
        memset(priv, 0, sizeof(*priv));
 }
 
-static const struct eth_dev_ops mlx5_dev_ops = {
+const struct eth_dev_ops mlx5_dev_ops = {
        .dev_configure = mlx5_dev_configure,
        .dev_start = mlx5_dev_start,
        .dev_stop = mlx5_dev_stop,
@@ -309,7 +304,6 @@ static const struct eth_dev_ops mlx5_dev_ops = {
        .rx_queue_intr_disable = mlx5_rx_intr_disable,
 };
 
-
 static const struct eth_dev_ops mlx5_dev_sec_ops = {
        .stats_get = mlx5_stats_get,
        .stats_reset = mlx5_stats_reset,
@@ -321,6 +315,42 @@ static const struct eth_dev_ops mlx5_dev_sec_ops = {
        .tx_descriptor_status = mlx5_tx_descriptor_status,
 };
 
+/* Available operators in flow isolated mode. */
+const struct eth_dev_ops mlx5_dev_ops_isolate = {
+       .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,
+       .link_update = mlx5_link_update,
+       .stats_get = mlx5_stats_get,
+       .stats_reset = mlx5_stats_reset,
+       .xstats_get = mlx5_xstats_get,
+       .xstats_reset = mlx5_xstats_reset,
+       .xstats_get_names = mlx5_xstats_get_names,
+       .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,
+       .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
+       .vlan_offload_set = mlx5_vlan_offload_set,
+       .filter_ctrl = mlx5_dev_filter_ctrl,
+       .rx_descriptor_status = mlx5_rx_descriptor_status,
+       .tx_descriptor_status = mlx5_tx_descriptor_status,
+       .rx_queue_intr_enable = mlx5_rx_intr_enable,
+       .rx_queue_intr_disable = mlx5_rx_intr_disable,
+};
+
 static struct {
        struct rte_pci_addr pci_addr; /* associated PCI address */
        uint32_t ports; /* physical ports bitfield. */
@@ -513,10 +543,14 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
        struct ibv_device_attr_ex device_attr;
        unsigned int sriov;
        unsigned int mps;
+       unsigned int cqe_comp;
        unsigned int tunnel_en = 0;
        int idx;
        int i;
        struct mlx5dv_context attrs_out;
+#ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
+       struct ibv_counter_set_description cs_desc;
+#endif
 
        (void)pci_drv;
        assert(pci_drv == &mlx5_driver);
@@ -605,17 +639,23 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
         * as all ConnectX-5 devices.
         */
        mlx5dv_query_device(attr_ctx, &attrs_out);
-       if (attrs_out.flags & (MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW |
-                              MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED)) {
-               INFO("Enhanced MPW is detected\n");
-               mps = MLX5_MPW_ENHANCED;
-       } else if (attrs_out.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
-               INFO("MPW is detected\n");
-               mps = MLX5_MPW;
+       if (attrs_out.flags & MLX5DV_CONTEXT_FLAGS_MPW_ALLOWED) {
+               if (attrs_out.flags & MLX5DV_CONTEXT_FLAGS_ENHANCED_MPW) {
+                       DEBUG("Enhanced MPW is supported");
+                       mps = MLX5_MPW_ENHANCED;
+               } else {
+                       DEBUG("MPW is supported");
+                       mps = MLX5_MPW;
+               }
        } else {
-               INFO("MPW is disabled\n");
+               DEBUG("MPW isn't supported");
                mps = MLX5_MPW_DISABLED;
        }
+       if (RTE_CACHE_LINE_SIZE == 128 &&
+           !(attrs_out.flags & MLX5DV_CONTEXT_FLAGS_CQE_128B_COMP))
+               cqe_comp = 0;
+       else
+               cqe_comp = 1;
        if (ibv_query_device_ex(attr_ctx, NULL, &device_attr))
                goto error;
        INFO("%u port(s) detected", device_attr.orig_attr.phys_port_cnt);
@@ -631,6 +671,7 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                struct ibv_device_attr_ex device_attr_ex;
                struct ether_addr mac;
                uint16_t num_vfs = 0;
+               struct ibv_device_attr_ex device_attr;
                struct mlx5_args args = {
                        .cqe_comp = MLX5_ARG_UNSET,
                        .txq_inline = MLX5_ARG_UNSET,
@@ -685,6 +726,7 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                        goto port_error;
                }
 
+               ibv_query_device_ex(ctx, NULL, &device_attr);
                /* Check port status. */
                err = ibv_query_port(ctx, port, &port_attr);
                if (err) {
@@ -732,7 +774,7 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                priv->pd = pd;
                priv->mtu = ETHER_MTU;
                priv->mps = mps; /* Enable MPW by default if supported. */
-               priv->cqe_comp = 1; /* Enable compression by default. */
+               priv->cqe_comp = cqe_comp;
                priv->tunnel_en = tunnel_en;
                /* Enable vector by default if supported. */
                priv->tx_vec_en = 1;
@@ -762,6 +804,13 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                DEBUG("L2 tunnel checksum offloads are %ssupported",
                      (priv->hw_csum_l2tun ? "" : "not "));
 
+#ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
+               priv->counter_set_supported = !!(device_attr.max_counter_sets);
+               ibv_describe_counter_set(ctx, 0, &cs_desc);
+               DEBUG("counter type = %d, num of cs = %ld, attributes = %d",
+                     cs_desc.counter_type, cs_desc.num_of_cs,
+                     cs_desc.attributes);
+#endif
                priv->ind_table_max_size =
                        device_attr_ex.rss_caps.max_rwq_indirection_table_size;
                /* Remove this check once DPDK supports larger/variable
@@ -821,19 +870,10 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                                priv->txq_inline = MLX5_WQE_SIZE_MAX -
                                                   MLX5_WQE_SIZE;
                }
-               /* 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;
+               if (priv->cqe_comp && !cqe_comp) {
+                       WARN("Rx CQE compression isn't supported");
+                       priv->cqe_comp = 0;
                }
-               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?"
@@ -846,10 +886,6 @@ mlx5_pci_probe(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 address. */
-               claim_zero(priv_mac_addr_add(priv, 0,
-                                            (const uint8_t (*)[ETHER_ADDR_LEN])
-                                            mac.addr_bytes));
 #ifndef NDEBUG
                {
                        char ifname[IF_NAMESIZE];
@@ -882,11 +918,13 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                eth_dev->data->mac_addrs = priv->mac;
                eth_dev->device = &pci_dev->device;
                rte_eth_copy_pci_info(eth_dev, pci_dev);
-               eth_dev->data->dev_flags |= RTE_ETH_DEV_DETACHABLE;
                eth_dev->device->driver = &mlx5_driver.driver;
                priv->dev = eth_dev;
                eth_dev->dev_ops = &mlx5_dev_ops;
+               /* Register MAC address. */
+               claim_zero(mlx5_mac_addr_add(eth_dev, &mac, 0, 0));
                TAILQ_INIT(&priv->flows);
+               TAILQ_INIT(&priv->ctrl_flows);
 
                /* Hint libmlx5 to use PMD allocator for data plane resources */
                struct mlx5dv_ctx_allocators alctr = {
@@ -904,10 +942,8 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                continue;
 
 port_error:
-               if (priv) {
-                       rte_free(priv->rss_conf);
+               if (priv)
                        rte_free(priv);
-               }
                if (pd)
                        claim_zero(ibv_dealloc_pd(pd));
                if (ctx)
@@ -1002,6 +1038,9 @@ rte_mlx5_pmd_init(void)
        setenv("RDMAV_HUGEPAGES_SAFE", "1", 1);
        /* Don't map UAR to WC if BlueFlame is not used.*/
        setenv("MLX5_SHUT_UP_BF", "1", 1);
+       /* Match the size of Rx completion entry to the size of a cacheline. */
+       if (RTE_CACHE_LINE_SIZE == 128)
+               setenv("MLX5_CQE_SIZE", "128", 0);
        ibv_fork_init();
        rte_pci_register(&mlx5_driver);
 }