ethdev: add speed capabilities
[dpdk.git] / drivers / net / mlx5 / mlx5_ethdev.c
index 26b6d73..d7a0eea 100644 (file)
@@ -45,6 +45,9 @@
 #include <sys/socket.h>
 #include <netinet/in.h>
 #include <linux/if.h>
+#include <linux/ethtool.h>
+#include <linux/sockios.h>
+#include <fcntl.h>
 
 /* DPDK headers don't like -pedantic. */
 #ifdef PEDANTIC
@@ -54,6 +57,9 @@
 #include <rte_ethdev.h>
 #include <rte_mbuf.h>
 #include <rte_common.h>
+#include <rte_interrupts.h>
+#include <rte_alarm.h>
+#include <rte_malloc.h>
 #ifdef PEDANTIC
 #pragma GCC diagnostic error "-pedantic"
 #endif
 #include "mlx5_rxtx.h"
 #include "mlx5_utils.h"
 
+/**
+ * Return private structure associated with an Ethernet device.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ *
+ * @return
+ *   Pointer to private structure.
+ */
+struct priv *
+mlx5_get_priv(struct rte_eth_dev *dev)
+{
+       struct mlx5_secondary_data *sd;
+
+       if (!mlx5_is_secondary())
+               return dev->data->dev_private;
+       sd = &mlx5_secondary_data[dev->data->port_id];
+       return sd->data.dev_private;
+}
+
+/**
+ * Check if running as a secondary process.
+ *
+ * @return
+ *   Nonzero if running as a secondary process.
+ */
+inline int
+mlx5_is_secondary(void)
+{
+       return rte_eal_process_type() != RTE_PROC_PRIMARY;
+}
+
 /**
  * Get interface name from private structure.
  *
@@ -392,7 +430,6 @@ priv_set_flags(struct priv *priv, unsigned int keep, unsigned int flags)
  * Ethernet device configuration.
  *
  * Prepare the driver for a given number of TX and RX queues.
- * Allocate parent RSS queue when several RX queues are requested.
  *
  * @param dev
  *   Pointer to Ethernet device structure.
@@ -406,9 +443,11 @@ 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 tmp;
-       int ret;
+       unsigned int i;
+       unsigned int j;
+       unsigned int reta_idx_n;
 
+       priv->rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
        priv->rxqs = (void *)dev->data->rx_queues;
        priv->txqs = (void *)dev->data->tx_queues;
        if (txqs_n != priv->txqs_n) {
@@ -416,51 +455,32 @@ 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);
-       /* If RSS is enabled, disable it first. */
-       if (priv->rss) {
-               unsigned int i;
-
-               /* Only if there are no remaining child RX queues. */
-               for (i = 0; (i != priv->rxqs_n); ++i)
-                       if ((*priv->rxqs)[i] != NULL)
-                               return EINVAL;
-               rxq_cleanup(&priv->rxq_parent);
-               priv->rss = 0;
-               priv->rxqs_n = 0;
-       }
-       if (rxqs_n <= 1) {
-               /* Nothing else to do. */
-               priv->rxqs_n = rxqs_n;
-               return 0;
-       }
-       /* Allocate a new RSS parent queue if supported by hardware. */
-       if (!priv->hw_rss) {
-               ERROR("%p: only a single RX queue can be configured when"
-                     " hardware doesn't support RSS",
-                     (void *)dev);
-               return EINVAL;
-       }
-       /* Fail if hardware doesn't support that many RSS queues. */
-       if (rxqs_n >= priv->max_rss_tbl_sz) {
-               ERROR("%p: only %u RX queues can be configured for RSS",
-                     (void *)dev, priv->max_rss_tbl_sz);
-               return EINVAL;
-       }
-       priv->rss = 1;
-       tmp = priv->rxqs_n;
        priv->rxqs_n = rxqs_n;
-       ret = rxq_setup(dev, &priv->rxq_parent, 0, 0, NULL, NULL);
-       if (!ret)
-               return 0;
-       /* Failure, rollback. */
-       priv->rss = 0;
-       priv->rxqs_n = tmp;
-       assert(ret > 0);
-       return ret;
+       /* 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;
 }
 
 /**
@@ -478,6 +498,9 @@ mlx5_dev_configure(struct rte_eth_dev *dev)
        struct priv *priv = dev->data->dev_private;
        int ret;
 
+       if (mlx5_is_secondary())
+               return -E_RTE_SECONDARY;
+
        priv_lock(priv);
        ret = dev_configure(dev);
        assert(ret >= 0);
@@ -496,7 +519,7 @@ mlx5_dev_configure(struct rte_eth_dev *dev)
 void
 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
 {
-       struct priv *priv = dev->data->dev_private;
+       struct priv *priv = mlx5_get_priv(dev);
        unsigned int max;
        char ifname[IF_NAMESIZE];
 
@@ -515,25 +538,126 @@ 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 |
                  DEV_RX_OFFLOAD_UDP_CKSUM |
                  DEV_RX_OFFLOAD_TCP_CKSUM) :
                 0);
-       info->tx_offload_capa =
-               (priv->hw_csum ?
-                (DEV_TX_OFFLOAD_IPV4_CKSUM |
-                 DEV_TX_OFFLOAD_UDP_CKSUM |
-                 DEV_TX_OFFLOAD_TCP_CKSUM) :
-                0);
+       info->tx_offload_capa = DEV_TX_OFFLOAD_VLAN_INSERT;
+       if (priv->hw_csum)
+               info->tx_offload_capa |=
+                       (DEV_TX_OFFLOAD_IPV4_CKSUM |
+                        DEV_TX_OFFLOAD_UDP_CKSUM |
+                        DEV_TX_OFFLOAD_TCP_CKSUM);
        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;
+       info->speed_capa =
+                       ETH_LINK_SPEED_1G |
+                       ETH_LINK_SPEED_10G |
+                       ETH_LINK_SPEED_20G |
+                       ETH_LINK_SPEED_25G |
+                       ETH_LINK_SPEED_40G |
+                       ETH_LINK_SPEED_50G |
+                       ETH_LINK_SPEED_56G;
        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).
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param wait_to_complete
+ *   Wait for request completion (ignored).
+ */
+static int
+mlx5_link_update_unlocked(struct rte_eth_dev *dev, int wait_to_complete)
+{
+       struct priv *priv = mlx5_get_priv(dev);
+       struct ethtool_cmd edata = {
+               .cmd = ETHTOOL_GSET
+       };
+       struct ifreq ifr;
+       struct rte_eth_link dev_link;
+       int link_speed = 0;
+
+       (void)wait_to_complete;
+       if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
+               WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
+               return -1;
+       }
+       memset(&dev_link, 0, sizeof(dev_link));
+       dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
+                               (ifr.ifr_flags & IFF_RUNNING));
+       ifr.ifr_data = &edata;
+       if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
+               WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
+                    strerror(errno));
+               return -1;
+       }
+       link_speed = ethtool_cmd_speed(&edata);
+       if (link_speed == -1)
+               dev_link.link_speed = 0;
+       else
+               dev_link.link_speed = link_speed;
+       dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
+                               ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
+       if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
+               /* Link status changed. */
+               dev->data->dev_link = dev_link;
+               return 0;
+       }
+       /* Link status is still the same. */
+       return -1;
+}
+
+/**
+ * DPDK callback to retrieve physical link information.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param wait_to_complete
+ *   Wait for request completion (ignored).
+ */
+int
+mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
+{
+       struct priv *priv = mlx5_get_priv(dev);
+       int ret;
+
+       priv_lock(priv);
+       ret = mlx5_link_update_unlocked(dev, wait_to_complete);
+       priv_unlock(priv);
+       return ret;
+}
+
 /**
  * DPDK callback to change the MTU.
  *
@@ -560,6 +684,9 @@ mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
        uint16_t (*rx_func)(void *, struct rte_mbuf **, uint16_t) =
                mlx5_rx_burst;
 
+       if (mlx5_is_secondary())
+               return -E_RTE_SECONDARY;
+
        priv_lock(priv);
        /* Set kernel interface MTU first. */
        if (priv_set_mtu(priv, mtu)) {
@@ -600,16 +727,6 @@ mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
                                rx_func = mlx5_rx_burst_sp;
                        break;
                }
-               /* Reenable non-RSS queue attributes. No need to check
-                * for errors at this stage. */
-               if (!priv->rss) {
-                       if (priv->started)
-                               rxq_mac_addrs_add(rxq);
-                       if (priv->started && priv->promisc_req)
-                               rxq_promiscuous_enable(rxq);
-                       if (priv->started && priv->allmulti_req)
-                               rxq_allmulticast_enable(rxq);
-               }
                /* Scattered burst function takes priority. */
                if (rxq->sp)
                        rx_func = mlx5_rx_burst_sp;
@@ -623,6 +740,111 @@ out:
        return -ret;
 }
 
+/**
+ * DPDK callback to get flow control status.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param[out] fc_conf
+ *   Flow control output buffer.
+ *
+ * @return
+ *   0 on success, negative errno value on failure.
+ */
+int
+mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
+{
+       struct priv *priv = dev->data->dev_private;
+       struct ifreq ifr;
+       struct ethtool_pauseparam ethpause = {
+               .cmd = ETHTOOL_GPAUSEPARAM
+       };
+       int ret;
+
+       if (mlx5_is_secondary())
+               return -E_RTE_SECONDARY;
+
+       ifr.ifr_data = &ethpause;
+       priv_lock(priv);
+       if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
+               ret = errno;
+               WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
+                    " failed: %s",
+                    strerror(ret));
+               goto out;
+       }
+
+       fc_conf->autoneg = ethpause.autoneg;
+       if (ethpause.rx_pause && ethpause.tx_pause)
+               fc_conf->mode = RTE_FC_FULL;
+       else if (ethpause.rx_pause)
+               fc_conf->mode = RTE_FC_RX_PAUSE;
+       else if (ethpause.tx_pause)
+               fc_conf->mode = RTE_FC_TX_PAUSE;
+       else
+               fc_conf->mode = RTE_FC_NONE;
+       ret = 0;
+
+out:
+       priv_unlock(priv);
+       assert(ret >= 0);
+       return -ret;
+}
+
+/**
+ * DPDK callback to modify flow control parameters.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param[in] fc_conf
+ *   Flow control parameters.
+ *
+ * @return
+ *   0 on success, negative errno value on failure.
+ */
+int
+mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
+{
+       struct priv *priv = dev->data->dev_private;
+       struct ifreq ifr;
+       struct ethtool_pauseparam ethpause = {
+               .cmd = ETHTOOL_SPAUSEPARAM
+       };
+       int ret;
+
+       if (mlx5_is_secondary())
+               return -E_RTE_SECONDARY;
+
+       ifr.ifr_data = &ethpause;
+       ethpause.autoneg = fc_conf->autoneg;
+       if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
+           (fc_conf->mode & RTE_FC_RX_PAUSE))
+               ethpause.rx_pause = 1;
+       else
+               ethpause.rx_pause = 0;
+
+       if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
+           (fc_conf->mode & RTE_FC_TX_PAUSE))
+               ethpause.tx_pause = 1;
+       else
+               ethpause.tx_pause = 0;
+
+       priv_lock(priv);
+       if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
+               ret = errno;
+               WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
+                    " failed: %s",
+                    strerror(ret));
+               goto out;
+       }
+       ret = 0;
+
+out:
+       priv_unlock(priv);
+       assert(ret >= 0);
+       return -ret;
+}
+
 /**
  * Get PCI information from struct ibv_device.
  *
@@ -672,3 +894,386 @@ 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;
+}
+
+/**
+ * Configure secondary process queues from a private data pointer (primary
+ * or secondary) and update burst callbacks. Can take place only once.
+ *
+ * All queues must have been previously created by the primary process to
+ * avoid undefined behavior.
+ *
+ * @param priv
+ *   Private data pointer from either primary or secondary process.
+ *
+ * @return
+ *   Private data pointer from secondary process, NULL in case of error.
+ */
+struct priv *
+mlx5_secondary_data_setup(struct priv *priv)
+{
+       unsigned int port_id = 0;
+       struct mlx5_secondary_data *sd;
+       void **tx_queues;
+       void **rx_queues;
+       unsigned int nb_tx_queues;
+       unsigned int nb_rx_queues;
+       unsigned int i;
+
+       /* priv must be valid at this point. */
+       assert(priv != NULL);
+       /* priv->dev must also be valid but may point to local memory from
+        * another process, possibly with the same address and must not
+        * be dereferenced yet. */
+       assert(priv->dev != NULL);
+       /* Determine port ID by finding out where priv comes from. */
+       while (1) {
+               sd = &mlx5_secondary_data[port_id];
+               rte_spinlock_lock(&sd->lock);
+               /* Primary process? */
+               if (sd->primary_priv == priv)
+                       break;
+               /* Secondary process? */
+               if (sd->data.dev_private == priv)
+                       break;
+               rte_spinlock_unlock(&sd->lock);
+               if (++port_id == RTE_DIM(mlx5_secondary_data))
+                       port_id = 0;
+       }
+       /* Switch to secondary private structure. If private data has already
+        * been updated by another thread, there is nothing else to do. */
+       priv = sd->data.dev_private;
+       if (priv->dev->data == &sd->data)
+               goto end;
+       /* Sanity checks. Secondary private structure is supposed to point
+        * to local eth_dev, itself still pointing to the shared device data
+        * structure allocated by the primary process. */
+       assert(sd->shared_dev_data != &sd->data);
+       assert(sd->data.nb_tx_queues == 0);
+       assert(sd->data.tx_queues == NULL);
+       assert(sd->data.nb_rx_queues == 0);
+       assert(sd->data.rx_queues == NULL);
+       assert(priv != sd->primary_priv);
+       assert(priv->dev->data == sd->shared_dev_data);
+       assert(priv->txqs_n == 0);
+       assert(priv->txqs == NULL);
+       assert(priv->rxqs_n == 0);
+       assert(priv->rxqs == NULL);
+       nb_tx_queues = sd->shared_dev_data->nb_tx_queues;
+       nb_rx_queues = sd->shared_dev_data->nb_rx_queues;
+       /* Allocate local storage for queues. */
+       tx_queues = rte_zmalloc("secondary ethdev->tx_queues",
+                               sizeof(sd->data.tx_queues[0]) * nb_tx_queues,
+                               RTE_CACHE_LINE_SIZE);
+       rx_queues = rte_zmalloc("secondary ethdev->rx_queues",
+                               sizeof(sd->data.rx_queues[0]) * nb_rx_queues,
+                               RTE_CACHE_LINE_SIZE);
+       if (tx_queues == NULL || rx_queues == NULL)
+               goto error;
+       /* Lock to prevent control operations during setup. */
+       priv_lock(priv);
+       /* TX queues. */
+       for (i = 0; i != nb_tx_queues; ++i) {
+               struct txq *primary_txq = (*sd->primary_priv->txqs)[i];
+               struct txq *txq;
+
+               if (primary_txq == NULL)
+                       continue;
+               txq = rte_calloc_socket("TXQ", 1, sizeof(*txq), 0,
+                                       primary_txq->socket);
+               if (txq != NULL) {
+                       if (txq_setup(priv->dev,
+                                     txq,
+                                     primary_txq->elts_n * MLX5_PMD_SGE_WR_N,
+                                     primary_txq->socket,
+                                     NULL) == 0) {
+                               txq->stats.idx = primary_txq->stats.idx;
+                               tx_queues[i] = txq;
+                               continue;
+                       }
+                       rte_free(txq);
+               }
+               while (i) {
+                       txq = tx_queues[--i];
+                       txq_cleanup(txq);
+                       rte_free(txq);
+               }
+               goto error;
+       }
+       /* RX queues. */
+       for (i = 0; i != nb_rx_queues; ++i) {
+               struct rxq *primary_rxq = (*sd->primary_priv->rxqs)[i];
+
+               if (primary_rxq == NULL)
+                       continue;
+               /* Not supported yet. */
+               rx_queues[i] = NULL;
+       }
+       /* Update everything. */
+       priv->txqs = (void *)tx_queues;
+       priv->txqs_n = nb_tx_queues;
+       priv->rxqs = (void *)rx_queues;
+       priv->rxqs_n = nb_rx_queues;
+       sd->data.rx_queues = rx_queues;
+       sd->data.tx_queues = tx_queues;
+       sd->data.nb_rx_queues = nb_rx_queues;
+       sd->data.nb_tx_queues = nb_tx_queues;
+       sd->data.dev_link = sd->shared_dev_data->dev_link;
+       sd->data.mtu = sd->shared_dev_data->mtu;
+       memcpy(sd->data.rx_queue_state, sd->shared_dev_data->rx_queue_state,
+              sizeof(sd->data.rx_queue_state));
+       memcpy(sd->data.tx_queue_state, sd->shared_dev_data->tx_queue_state,
+              sizeof(sd->data.tx_queue_state));
+       sd->data.dev_flags = sd->shared_dev_data->dev_flags;
+       /* Use local data from now on. */
+       rte_mb();
+       priv->dev->data = &sd->data;
+       rte_mb();
+       priv->dev->tx_pkt_burst = mlx5_tx_burst;
+       priv->dev->rx_pkt_burst = removed_rx_burst;
+       priv_unlock(priv);
+end:
+       /* More sanity checks. */
+       assert(priv->dev->tx_pkt_burst == mlx5_tx_burst);
+       assert(priv->dev->rx_pkt_burst == removed_rx_burst);
+       assert(priv->dev->data == &sd->data);
+       rte_spinlock_unlock(&sd->lock);
+       return priv;
+error:
+       priv_unlock(priv);
+       rte_free(tx_queues);
+       rte_free(rx_queues);
+       rte_spinlock_unlock(&sd->lock);
+       return NULL;
+}