net/virtio-user: add option to use packed queues
[dpdk.git] / drivers / net / mlx4 / mlx4.c
index 9f8ecd0..7f07b8d 100644 (file)
@@ -98,20 +98,6 @@ mlx4_dev_configure(struct rte_eth_dev *dev)
        if (ret)
                ERROR("%p: interrupt handler installation failed",
                      (void *)dev);
-       /*
-        * Once the device is added to the list of memory event callback, its
-        * global MR cache table cannot be expanded on the fly because of
-        * deadlock. If it overflows, lookup should be done by searching MR list
-        * linearly, which is slow.
-        */
-       if (mlx4_mr_btree_init(&priv->mr.cache, MLX4_MR_BTREE_CACHE_N * 2,
-                              dev->device->numa_node)) {
-               /* rte_errno is already set. */
-               return -rte_errno;
-       }
-       rte_rwlock_write_lock(&mlx4_mem_event_rwlock);
-       LIST_INSERT_HEAD(&mlx4_mem_event_cb_list, priv, mem_event_cb);
-       rte_rwlock_write_unlock(&mlx4_mem_event_rwlock);
 exit:
        return ret;
 }
@@ -589,14 +575,14 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
        ibv_dev = list[i];
        DEBUG("device opened");
        if (mlx4_glue->query_device(attr_ctx, &device_attr)) {
-               rte_errno = ENODEV;
+               err = ENODEV;
                goto error;
        }
        INFO("%u port(s) detected", device_attr.phys_port_cnt);
        conf.ports.present |= (UINT64_C(1) << device_attr.phys_port_cnt) - 1;
        if (mlx4_args(pci_dev->device.devargs, &conf)) {
                ERROR("failed to process device arguments");
-               rte_errno = EINVAL;
+               err = EINVAL;
                goto error;
        }
        /* Use all ports when none are defined */
@@ -604,7 +590,7 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                conf.ports.enabled = conf.ports.present;
        /* Retrieve extended device attributes. */
        if (mlx4_glue->query_device_ex(attr_ctx, NULL, &device_attr_ex)) {
-               rte_errno = ENODEV;
+               err = ENODEV;
                goto error;
        }
        assert(device_attr.max_sge >= MLX4_MAX_SGE);
@@ -623,18 +609,18 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                DEBUG("using port %u", port);
                ctx = mlx4_glue->open_device(ibv_dev);
                if (ctx == NULL) {
-                       rte_errno = ENODEV;
+                       err = ENODEV;
                        goto port_error;
                }
                /* Check port status. */
                err = mlx4_glue->query_port(ctx, port, &port_attr);
                if (err) {
-                       rte_errno = err;
-                       ERROR("port query failed: %s", strerror(rte_errno));
+                       err = ENODEV;
+                       ERROR("port query failed: %s", strerror(err));
                        goto port_error;
                }
                if (port_attr.link_layer != IBV_LINK_LAYER_ETHERNET) {
-                       rte_errno = ENOTSUP;
+                       err = ENOTSUP;
                        ERROR("port %d is not configured in Ethernet mode",
                              port);
                        goto port_error;
@@ -644,15 +630,16 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                              port, mlx4_glue->port_state_str(port_attr.state),
                              port_attr.state);
                /* Make asynchronous FD non-blocking to handle interrupts. */
-               if (mlx4_fd_set_non_blocking(ctx->async_fd) < 0) {
+               err = mlx4_fd_set_non_blocking(ctx->async_fd);
+               if (err) {
                        ERROR("cannot make asynchronous FD non-blocking: %s",
-                             strerror(rte_errno));
+                             strerror(err));
                        goto port_error;
                }
                /* Allocate protection domain. */
                pd = mlx4_glue->alloc_pd(ctx);
                if (pd == NULL) {
-                       rte_errno = ENOMEM;
+                       err = ENOMEM;
                        ERROR("PD allocation failure");
                        goto port_error;
                }
@@ -661,7 +648,7 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                                   sizeof(*priv),
                                   RTE_CACHE_LINE_SIZE);
                if (priv == NULL) {
-                       rte_errno = ENOMEM;
+                       err = ENOMEM;
                        ERROR("priv allocation failure");
                        goto port_error;
                }
@@ -686,14 +673,27 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                                                   &device_attr_ex);
                DEBUG("supported RSS hash fields mask: %016" PRIx64,
                      priv->hw_rss_sup);
+               priv->hw_rss_max_qps =
+                       device_attr_ex.rss_caps.max_rwq_indirection_table_size;
+               DEBUG("MAX RSS queues %d", priv->hw_rss_max_qps);
                priv->hw_fcs_strip = !!(device_attr_ex.raw_packet_caps &
                                        IBV_RAW_PACKET_CAP_SCATTER_FCS);
                DEBUG("FCS stripping toggling is %ssupported",
                      priv->hw_fcs_strip ? "" : "not ");
+               priv->tso =
+                       ((device_attr_ex.tso_caps.max_tso > 0) &&
+                        (device_attr_ex.tso_caps.supported_qpts &
+                         (1 << IBV_QPT_RAW_PACKET)));
+               if (priv->tso)
+                       priv->tso_max_payload_sz =
+                                       device_attr_ex.tso_caps.max_tso;
+               DEBUG("TSO is %ssupported",
+                     priv->tso ? "" : "not ");
                /* Configure the first MAC address by default. */
-               if (mlx4_get_mac(priv, &mac.addr_bytes)) {
+               err = mlx4_get_mac(priv, &mac.addr_bytes);
+               if (err) {
                        ERROR("cannot get MAC address, is mlx4_en loaded?"
-                             " (rte_errno: %s)", strerror(rte_errno));
+                             " (error: %s)", strerror(err));
                        goto port_error;
                }
                INFO("port %u MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
@@ -726,15 +726,14 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                        eth_dev = rte_eth_dev_allocate(name);
                }
                if (eth_dev == NULL) {
+                       err = ENOMEM;
                        ERROR("can not allocate rte ethdev");
-                       rte_errno = ENOMEM;
                        goto port_error;
                }
                eth_dev->data->dev_private = priv;
                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 = &mlx4_driver.driver;
                /* Initialize local interrupt handle for current port. */
                priv->intr_handle = (struct rte_intr_handle){
                        .fd = -1,
@@ -761,20 +760,40 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                /* Update link status once if waiting for LSC. */
                if (eth_dev->data->dev_flags & RTE_ETH_DEV_INTR_LSC)
                        mlx4_link_update(eth_dev, 0);
+               /*
+                * Once the device is added to the list of memory event
+                * callback, its global MR cache table cannot be expanded
+                * on the fly because of deadlock. If it overflows, lookup
+                * should be done by searching MR list linearly, which is slow.
+                */
+               err = mlx4_mr_btree_init(&priv->mr.cache,
+                                        MLX4_MR_BTREE_CACHE_N * 2,
+                                        eth_dev->device->numa_node);
+               if (err) {
+                       /* rte_errno is already set. */
+                       goto port_error;
+               }
+               /* Add device to memory callback list. */
+               rte_rwlock_write_lock(&mlx4_mem_event_rwlock);
+               LIST_INSERT_HEAD(&mlx4_mem_event_cb_list, priv, mem_event_cb);
+               rte_rwlock_write_unlock(&mlx4_mem_event_rwlock);
                rte_eth_dev_probing_finish(eth_dev);
                continue;
 port_error:
                rte_free(priv);
+               if (eth_dev != NULL)
+                       eth_dev->data->dev_private = NULL;
                if (pd)
                        claim_zero(mlx4_glue->dealloc_pd(pd));
                if (ctx)
                        claim_zero(mlx4_glue->close_device(ctx));
-               if (eth_dev)
+               if (eth_dev != NULL) {
+                       /* mac_addrs must not be freed because part of dev_private */
+                       eth_dev->data->mac_addrs = NULL;
                        rte_eth_dev_release_port(eth_dev);
+               }
                break;
        }
-       if (i == device_attr.phys_port_cnt)
-               return 0;
        /*
         * XXX if something went wrong in the loop above, there is a resource
         * leak (ctx, pd, priv, dpdk ethdev) but we can do nothing about it as
@@ -786,8 +805,9 @@ error:
                claim_zero(mlx4_glue->close_device(attr_ctx));
        if (list)
                mlx4_glue->free_device_list(list);
-       assert(rte_errno >= 0);
-       return -rte_errno;
+       if (err)
+               rte_errno = err;
+       return -err;
 }
 
 static const struct rte_pci_id mlx4_pci_id_map[] = {
@@ -951,9 +971,7 @@ glue_error:
 /**
  * Driver initialization routine.
  */
-RTE_INIT(rte_mlx4_pmd_init);
-static void
-rte_mlx4_pmd_init(void)
+RTE_INIT(rte_mlx4_pmd_init)
 {
        /*
         * MLX4_DEVICE_FATAL_CLEANUP tells ibv_destroy functions we