net/mlx4: enhance Rx packet type offloads
[dpdk.git] / drivers / net / mlx4 / mlx4.c
index 317d0e6..f9e4f9d 100644 (file)
@@ -44,7 +44,6 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
-#include <unistd.h>
 
 /* Verbs headers do not support -pedantic. */
 #ifdef PEDANTIC
@@ -61,6 +60,7 @@
 #include <rte_ethdev.h>
 #include <rte_ethdev_pci.h>
 #include <rte_ether.h>
+#include <rte_flow.h>
 #include <rte_interrupts.h>
 #include <rte_kvargs.h>
 #include <rte_malloc.h>
@@ -88,8 +88,6 @@ const char *pmd_mlx4_init_params[] = {
 /**
  * DPDK callback for Ethernet device configuration.
  *
- * Prepare the driver for a given number of TX and RX queues.
- *
  * @param dev
  *   Pointer to Ethernet device structure.
  *
@@ -100,28 +98,25 @@ static int
 mlx4_dev_configure(struct rte_eth_dev *dev)
 {
        struct priv *priv = dev->data->dev_private;
-       unsigned int rxqs_n = dev->data->nb_rx_queues;
-       unsigned int txqs_n = dev->data->nb_tx_queues;
-
-       priv->rxqs = (void *)dev->data->rx_queues;
-       priv->txqs = (void *)dev->data->tx_queues;
-       if (txqs_n != priv->txqs_n) {
-               INFO("%p: TX queues number update: %u -> %u",
-                    (void *)dev, priv->txqs_n, txqs_n);
-               priv->txqs_n = txqs_n;
-       }
-       if (rxqs_n != priv->rxqs_n) {
-               INFO("%p: Rx queues number update: %u -> %u",
-                    (void *)dev, priv->rxqs_n, rxqs_n);
-               priv->rxqs_n = rxqs_n;
+       struct rte_flow_error error;
+       int ret;
+
+       /* Prepare internal flow rules. */
+       ret = mlx4_flow_sync(priv, &error);
+       if (ret) {
+               ERROR("cannot set up internal flow rules (code %d, \"%s\"),"
+                     " flow error type %d, cause %p, message: %s",
+                     -ret, strerror(-ret), error.type, error.cause,
+                     error.message ? error.message : "(unspecified)");
        }
-       return 0;
+       return ret;
 }
 
 /**
  * DPDK callback to start the device.
  *
- * Simulate device start by attaching all configured flows.
+ * Simulate device start by initializing common RSS resources and attaching
+ * all configured flows.
  *
  * @param dev
  *   Pointer to Ethernet device structure.
@@ -133,31 +128,40 @@ static int
 mlx4_dev_start(struct rte_eth_dev *dev)
 {
        struct priv *priv = dev->data->dev_private;
+       struct rte_flow_error error;
        int ret;
 
        if (priv->started)
                return 0;
        DEBUG("%p: attaching configured flows to all RX queues", (void *)dev);
        priv->started = 1;
-       ret = mlx4_mac_addr_add(priv);
-       if (ret)
+       ret = mlx4_rss_init(priv);
+       if (ret) {
+               ERROR("%p: cannot initialize RSS resources: %s",
+                     (void *)dev, strerror(-ret));
                goto err;
+       }
        ret = mlx4_intr_install(priv);
        if (ret) {
                ERROR("%p: interrupt handler installation failed",
                     (void *)dev);
                goto err;
        }
-       ret = mlx4_flow_start(priv);
+       ret = mlx4_flow_sync(priv, &error);
        if (ret) {
-               ERROR("%p: flow start failed: %s",
-                     (void *)dev, strerror(ret));
+               ERROR("%p: cannot attach flow rules (code %d, \"%s\"),"
+                     " flow error type %d, cause %p, message: %s",
+                     (void *)dev,
+                     -ret, strerror(-ret), error.type, error.cause,
+                     error.message ? error.message : "(unspecified)");
                goto err;
        }
+       rte_wmb();
+       dev->tx_pkt_burst = mlx4_tx_burst;
+       dev->rx_pkt_burst = mlx4_rx_burst;
        return 0;
 err:
        /* Rollback. */
-       mlx4_mac_addr_del(priv);
        priv->started = 0;
        return ret;
 }
@@ -179,9 +183,12 @@ mlx4_dev_stop(struct rte_eth_dev *dev)
                return;
        DEBUG("%p: detaching flows from all RX queues", (void *)dev);
        priv->started = 0;
-       mlx4_flow_stop(priv);
+       dev->tx_pkt_burst = mlx4_tx_burst_removed;
+       dev->rx_pkt_burst = mlx4_rx_burst_removed;
+       rte_wmb();
+       mlx4_flow_sync(priv, NULL);
        mlx4_intr_uninstall(priv);
-       mlx4_mac_addr_del(priv);
+       mlx4_rss_deinit(priv);
 }
 
 /**
@@ -196,50 +203,19 @@ static void
 mlx4_dev_close(struct rte_eth_dev *dev)
 {
        struct priv *priv = dev->data->dev_private;
-       void *tmp;
        unsigned int i;
 
-       if (priv == NULL)
-               return;
        DEBUG("%p: closing device \"%s\"",
              (void *)dev,
              ((priv->ctx != NULL) ? priv->ctx->device->name : ""));
-       mlx4_mac_addr_del(priv);
-       /*
-        * Prevent crashes when queues are still in use. This is unfortunately
-        * still required for DPDK 1.3 because some programs (such as testpmd)
-        * never release them before closing the device.
-        */
        dev->rx_pkt_burst = mlx4_rx_burst_removed;
        dev->tx_pkt_burst = mlx4_tx_burst_removed;
-       if (priv->rxqs != NULL) {
-               /* XXX race condition if mlx4_rx_burst() is still running. */
-               usleep(1000);
-               for (i = 0; (i != priv->rxqs_n); ++i) {
-                       tmp = (*priv->rxqs)[i];
-                       if (tmp == NULL)
-                               continue;
-                       (*priv->rxqs)[i] = NULL;
-                       mlx4_rxq_cleanup(tmp);
-                       rte_free(tmp);
-               }
-               priv->rxqs_n = 0;
-               priv->rxqs = NULL;
-       }
-       if (priv->txqs != NULL) {
-               /* XXX race condition if mlx4_tx_burst() is still running. */
-               usleep(1000);
-               for (i = 0; (i != priv->txqs_n); ++i) {
-                       tmp = (*priv->txqs)[i];
-                       if (tmp == NULL)
-                               continue;
-                       (*priv->txqs)[i] = NULL;
-                       mlx4_txq_cleanup(tmp);
-                       rte_free(tmp);
-               }
-               priv->txqs_n = 0;
-               priv->txqs = NULL;
-       }
+       rte_wmb();
+       mlx4_flow_clean(priv);
+       for (i = 0; i != dev->data->nb_rx_queues; ++i)
+               mlx4_rx_queue_release(dev->data->rx_queues[i]);
+       for (i = 0; i != dev->data->nb_tx_queues; ++i)
+               mlx4_tx_queue_release(dev->data->tx_queues[i]);
        if (priv->pd != NULL) {
                assert(priv->ctx != NULL);
                claim_zero(ibv_dealloc_pd(priv->pd));
@@ -258,9 +234,18 @@ static const struct eth_dev_ops mlx4_dev_ops = {
        .dev_set_link_up = mlx4_dev_set_link_up,
        .dev_close = mlx4_dev_close,
        .link_update = mlx4_link_update,
+       .promiscuous_enable = mlx4_promiscuous_enable,
+       .promiscuous_disable = mlx4_promiscuous_disable,
+       .allmulticast_enable = mlx4_allmulticast_enable,
+       .allmulticast_disable = mlx4_allmulticast_disable,
+       .mac_addr_remove = mlx4_mac_addr_remove,
+       .mac_addr_add = mlx4_mac_addr_add,
+       .mac_addr_set = mlx4_mac_addr_set,
        .stats_get = mlx4_stats_get,
        .stats_reset = mlx4_stats_reset,
        .dev_infos_get = mlx4_dev_infos_get,
+       .dev_supported_ptypes_get = mlx4_dev_supported_ptypes_get,
+       .vlan_filter_set = mlx4_vlan_filter_set,
        .rx_queue_setup = mlx4_rx_queue_setup,
        .tx_queue_setup = mlx4_tx_queue_setup,
        .rx_queue_release = mlx4_rx_queue_release,
@@ -577,6 +562,17 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                priv->pd = pd;
                priv->mtu = ETHER_MTU;
                priv->vf = vf;
+               priv->hw_csum = !!(device_attr.device_cap_flags &
+                                  IBV_DEVICE_RAW_IP_CSUM);
+               DEBUG("checksum offloading is %ssupported",
+                     (priv->hw_csum ? "" : "not "));
+               /* Only ConnectX-3 Pro supports tunneling. */
+               priv->hw_csum_l2tun =
+                       priv->hw_csum &&
+                       (device_attr.vendor_part_id ==
+                        PCI_DEVICE_ID_MELLANOX_CONNECTX3PRO);
+               DEBUG("L2 tunnel checksum offloads are %ssupported",
+                     (priv->hw_csum_l2tun ? "" : "not "));
                /* Configure the first MAC address by default. */
                if (mlx4_get_mac(priv, &mac.addr_bytes)) {
                        ERROR("cannot get MAC address, is mlx4_en loaded?"
@@ -589,9 +585,7 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                     mac.addr_bytes[2], mac.addr_bytes[3],
                     mac.addr_bytes[4], mac.addr_bytes[5]);
                /* Register MAC address. */
-               priv->mac = mac;
-               if (mlx4_mac_addr_add(priv))
-                       goto port_error;
+               priv->mac[0] = mac;
 #ifndef NDEBUG
                {
                        char ifname[IF_NAMESIZE];
@@ -620,7 +614,7 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                        goto port_error;
                }
                eth_dev->data->dev_private = priv;
-               eth_dev->data->mac_addrs = &priv->mac;
+               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;
@@ -644,7 +638,6 @@ mlx4_pci_probe(struct rte_pci_driver *pci_drv, struct rte_pci_device *pci_dev)
                eth_dev->intr_handle = &priv->intr_handle;
                priv->dev = eth_dev;
                eth_dev->dev_ops = &mlx4_dev_ops;
-               eth_dev->data->dev_flags |= RTE_ETH_DEV_DETACHABLE;
                /* Bring Ethernet device up. */
                DEBUG("forcing Ethernet interface up");
                mlx4_dev_set_link_up(priv->dev);