mlx5: support setting link up or down
[dpdk.git] / drivers / net / mlx5 / mlx5_ethdev.c
index fac685e..463b326 100644 (file)
@@ -47,6 +47,7 @@
 #include <linux/if.h>
 #include <linux/ethtool.h>
 #include <linux/sockios.h>
+#include <fcntl.h>
 
 /* DPDK headers don't like -pedantic. */
 #ifdef PEDANTIC
@@ -56,6 +57,8 @@
 #include <rte_ethdev.h>
 #include <rte_mbuf.h>
 #include <rte_common.h>
+#include <rte_interrupts.h>
+#include <rte_alarm.h>
 #ifdef PEDANTIC
 #pragma GCC diagnostic error "-pedantic"
 #endif
@@ -407,6 +410,9 @@ 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;
+       unsigned int i;
+       unsigned int j;
+       unsigned int reta_idx_n;
 
        priv->rxqs = (void *)dev->data->rx_queues;
        priv->txqs = (void *)dev->data->tx_queues;
@@ -415,11 +421,31 @@ dev_configure(struct rte_eth_dev *dev)
                     (void *)dev, priv->txqs_n, txqs_n);
                priv->txqs_n = txqs_n;
        }
+       if (rxqs_n > priv->ind_table_max_size) {
+               ERROR("cannot handle this many RX queues (%u)", rxqs_n);
+               return EINVAL;
+       }
        if (rxqs_n == priv->rxqs_n)
                return 0;
        INFO("%p: RX queues number update: %u -> %u",
             (void *)dev, priv->rxqs_n, rxqs_n);
        priv->rxqs_n = rxqs_n;
+       /* If the requested number of RX queues is not a power of two, use the
+        * maximum indirection table size for better balancing.
+        * The result is always rounded to the next power of two. */
+       reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
+                                    priv->ind_table_max_size :
+                                    rxqs_n));
+       if (priv_rss_reta_index_resize(priv, reta_idx_n))
+               return ENOMEM;
+       /* When the number of RX queues is not a power of two, the remaining
+        * table entries are padded with reused WQs and hashes are not spread
+        * uniformly. */
+       for (i = 0, j = 0; (i != reta_idx_n); ++i) {
+               (*priv->reta_idx)[i] = j;
+               if (++j == rxqs_n)
+                       j = 0;
+       }
        return 0;
 }
 
@@ -475,8 +501,7 @@ mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
                max = 65535;
        info->max_rx_queues = max;
        info->max_tx_queues = max;
-       /* Last array entry is reserved for broadcast. */
-       info->max_mac_addrs = (RTE_DIM(priv->mac) - 1);
+       info->max_mac_addrs = RTE_DIM(priv->mac);
        info->rx_offload_capa =
                (priv->hw_csum ?
                 (DEV_RX_OFFLOAD_IPV4_CKSUM |
@@ -491,9 +516,34 @@ mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
                 0);
        if (priv_get_ifname(priv, &ifname) == 0)
                info->if_index = if_nametoindex(ifname);
+       /* FIXME: RETA update/query API expects the callee to know the size of
+        * the indirection table, for this PMD the size varies depending on
+        * the number of RX queues, it becomes impossible to find the correct
+        * size if it is not fixed.
+        * The API should be updated to solve this problem. */
+       info->reta_size = priv->ind_table_max_size;
        priv_unlock(priv);
 }
 
+const uint32_t *
+mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
+{
+       static const uint32_t ptypes[] = {
+               /* refers to rxq_cq_to_pkt_type() */
+               RTE_PTYPE_L3_IPV4,
+               RTE_PTYPE_L3_IPV6,
+               RTE_PTYPE_INNER_L3_IPV4,
+               RTE_PTYPE_INNER_L3_IPV6,
+               RTE_PTYPE_UNKNOWN
+
+       };
+
+       if (dev->rx_pkt_burst == mlx5_rx_burst ||
+           dev->rx_pkt_burst == mlx5_rx_burst_sp)
+               return ptypes;
+       return NULL;
+}
+
 /**
  * DPDK callback to retrieve physical link information (unlocked version).
  *
@@ -790,3 +840,235 @@ mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
        fclose(file);
        return 0;
 }
+
+/**
+ * Link status handler.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param dev
+ *   Pointer to the rte_eth_dev structure.
+ *
+ * @return
+ *   Nonzero if the callback process can be called immediately.
+ */
+static int
+priv_dev_link_status_handler(struct priv *priv, struct rte_eth_dev *dev)
+{
+       struct ibv_async_event event;
+       int port_change = 0;
+       int ret = 0;
+
+       /* Read all message and acknowledge them. */
+       for (;;) {
+               if (ibv_get_async_event(priv->ctx, &event))
+                       break;
+
+               if (event.event_type == IBV_EVENT_PORT_ACTIVE ||
+                   event.event_type == IBV_EVENT_PORT_ERR)
+                       port_change = 1;
+               else
+                       DEBUG("event type %d on port %d not handled",
+                             event.event_type, event.element.port_num);
+               ibv_ack_async_event(&event);
+       }
+
+       if (port_change ^ priv->pending_alarm) {
+               struct rte_eth_link *link = &dev->data->dev_link;
+
+               priv->pending_alarm = 0;
+               mlx5_link_update_unlocked(dev, 0);
+               if (((link->link_speed == 0) && link->link_status) ||
+                   ((link->link_speed != 0) && !link->link_status)) {
+                       /* Inconsistent status, check again later. */
+                       priv->pending_alarm = 1;
+                       rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
+                                         mlx5_dev_link_status_handler,
+                                         dev);
+               } else
+                       ret = 1;
+       }
+       return ret;
+}
+
+/**
+ * Handle delayed link status event.
+ *
+ * @param arg
+ *   Registered argument.
+ */
+void
+mlx5_dev_link_status_handler(void *arg)
+{
+       struct rte_eth_dev *dev = arg;
+       struct priv *priv = dev->data->dev_private;
+       int ret;
+
+       priv_lock(priv);
+       assert(priv->pending_alarm == 1);
+       ret = priv_dev_link_status_handler(priv, dev);
+       priv_unlock(priv);
+       if (ret)
+               _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC);
+}
+
+/**
+ * Handle interrupts from the NIC.
+ *
+ * @param[in] intr_handle
+ *   Interrupt handler.
+ * @param cb_arg
+ *   Callback argument.
+ */
+void
+mlx5_dev_interrupt_handler(struct rte_intr_handle *intr_handle, void *cb_arg)
+{
+       struct rte_eth_dev *dev = cb_arg;
+       struct priv *priv = dev->data->dev_private;
+       int ret;
+
+       (void)intr_handle;
+       priv_lock(priv);
+       ret = priv_dev_link_status_handler(priv, dev);
+       priv_unlock(priv);
+       if (ret)
+               _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC);
+}
+
+/**
+ * Uninstall interrupt handler.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param dev
+ *   Pointer to the rte_eth_dev structure.
+ */
+void
+priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev)
+{
+       if (!dev->data->dev_conf.intr_conf.lsc)
+               return;
+       rte_intr_callback_unregister(&priv->intr_handle,
+                                    mlx5_dev_interrupt_handler,
+                                    dev);
+       if (priv->pending_alarm)
+               rte_eal_alarm_cancel(mlx5_dev_link_status_handler, dev);
+       priv->pending_alarm = 0;
+       priv->intr_handle.fd = 0;
+       priv->intr_handle.type = 0;
+}
+
+/**
+ * Install interrupt handler.
+ *
+ * @param priv
+ *   Pointer to private structure.
+ * @param dev
+ *   Pointer to the rte_eth_dev structure.
+ */
+void
+priv_dev_interrupt_handler_install(struct priv *priv, struct rte_eth_dev *dev)
+{
+       int rc, flags;
+
+       if (!dev->data->dev_conf.intr_conf.lsc)
+               return;
+       assert(priv->ctx->async_fd > 0);
+       flags = fcntl(priv->ctx->async_fd, F_GETFL);
+       rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
+       if (rc < 0) {
+               INFO("failed to change file descriptor async event queue");
+               dev->data->dev_conf.intr_conf.lsc = 0;
+       } else {
+               priv->intr_handle.fd = priv->ctx->async_fd;
+               priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
+               rte_intr_callback_register(&priv->intr_handle,
+                                          mlx5_dev_interrupt_handler,
+                                          dev);
+       }
+}
+
+/**
+ * Change the link state (UP / DOWN).
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param up
+ *   Nonzero for link up, otherwise link down.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+static int
+priv_set_link(struct priv *priv, int up)
+{
+       struct rte_eth_dev *dev = priv->dev;
+       int err;
+       unsigned int i;
+
+       if (up) {
+               err = priv_set_flags(priv, ~IFF_UP, IFF_UP);
+               if (err)
+                       return err;
+               for (i = 0; i < priv->rxqs_n; i++)
+                       if ((*priv->rxqs)[i]->sp)
+                               break;
+               /* Check if an sp queue exists.
+                * Note: Some old frames might be received.
+                */
+               if (i == priv->rxqs_n)
+                       dev->rx_pkt_burst = mlx5_rx_burst;
+               else
+                       dev->rx_pkt_burst = mlx5_rx_burst_sp;
+               dev->tx_pkt_burst = mlx5_tx_burst;
+       } else {
+               err = priv_set_flags(priv, ~IFF_UP, ~IFF_UP);
+               if (err)
+                       return err;
+               dev->rx_pkt_burst = removed_rx_burst;
+               dev->tx_pkt_burst = removed_tx_burst;
+       }
+       return 0;
+}
+
+/**
+ * DPDK callback to bring the link DOWN.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+int
+mlx5_set_link_down(struct rte_eth_dev *dev)
+{
+       struct priv *priv = dev->data->dev_private;
+       int err;
+
+       priv_lock(priv);
+       err = priv_set_link(priv, 0);
+       priv_unlock(priv);
+       return err;
+}
+
+/**
+ * DPDK callback to bring the link UP.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ *
+ * @return
+ *   0 on success, errno value on failure.
+ */
+int
+mlx5_set_link_up(struct rte_eth_dev *dev)
+{
+       struct priv *priv = dev->data->dev_private;
+       int err;
+
+       priv_lock(priv);
+       err = priv_set_link(priv, 1);
+       priv_unlock(priv);
+       return err;
+}