net/enic: enable GENEVE with options offload
[dpdk.git] / drivers / net / enic / enic_main.c
index 03c5ef7..6d17d04 100644 (file)
@@ -8,7 +8,6 @@
 #include <sys/stat.h>
 #include <sys/mman.h>
 #include <fcntl.h>
-#include <libgen.h>
 
 #include <rte_pci.h>
 #include <rte_bus_pci.h>
@@ -116,11 +115,18 @@ static void enic_init_soft_stats(struct enic *enic)
        enic_clear_soft_stats(enic);
 }
 
-void enic_dev_stats_clear(struct enic *enic)
+int enic_dev_stats_clear(struct enic *enic)
 {
-       if (vnic_dev_stats_clear(enic->vdev))
+       int ret;
+
+       ret = vnic_dev_stats_clear(enic->vdev);
+       if (ret != 0) {
                dev_err(enic, "Error in clearing stats\n");
+               return ret;
+       }
        enic_clear_soft_stats(enic);
+
+       return 0;
 }
 
 int enic_dev_stats_get(struct enic *enic, struct rte_eth_stats *r_stats)
@@ -418,9 +424,9 @@ enic_free_consistent(void *priv,
        rte_free(mze);
 }
 
-int enic_link_update(struct enic *enic)
+int enic_link_update(struct rte_eth_dev *eth_dev)
 {
-       struct rte_eth_dev *eth_dev = enic->rte_dev;
+       struct enic *enic = pmd_priv(eth_dev);
        struct rte_eth_link link;
 
        memset(&link, 0, sizeof(link));
@@ -439,7 +445,7 @@ enic_intr_handler(void *arg)
 
        vnic_intr_return_all_credits(&enic->intr[ENICPMD_LSC_INTR_OFFSET]);
 
-       enic_link_update(enic);
+       enic_link_update(dev);
        _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
        enic_log_q_error(enic);
 }
@@ -514,27 +520,75 @@ static void enic_prep_wq_for_simple_tx(struct enic *enic, uint16_t queue_idx)
        }
 }
 
-static void pick_rx_handler(struct enic *enic)
+/*
+ * The 'strong' version is in enic_rxtx_vec_avx2.c. This weak version is used
+ * used when that file is not compiled.
+ */
+__rte_weak bool
+enic_use_vector_rx_handler(__rte_unused struct rte_eth_dev *eth_dev)
 {
-       struct rte_eth_dev *eth_dev;
+       return false;
+}
 
-       /* Use the non-scatter, simplified RX handler if possible. */
-       eth_dev = enic->rte_dev;
+void enic_pick_rx_handler(struct rte_eth_dev *eth_dev)
+{
+       struct enic *enic = pmd_priv(eth_dev);
+
+       /*
+        * Preference order:
+        * 1. The vectorized handler if possible and requested.
+        * 2. The non-scatter, simplified handler if scatter Rx is not used.
+        * 3. The default handler as a fallback.
+        */
+       if (enic_use_vector_rx_handler(eth_dev))
+               return;
        if (enic->rq_count > 0 && enic->rq[0].data_queue_enable == 0) {
-               PMD_INIT_LOG(DEBUG, " use the non-scatter Rx handler");
+               ENICPMD_LOG(DEBUG, " use the non-scatter Rx handler");
                eth_dev->rx_pkt_burst = &enic_noscatter_recv_pkts;
        } else {
-               PMD_INIT_LOG(DEBUG, " use the normal Rx handler");
+               ENICPMD_LOG(DEBUG, " use the normal Rx handler");
                eth_dev->rx_pkt_burst = &enic_recv_pkts;
        }
 }
 
+/* Secondary process uses this to set the Tx handler */
+void enic_pick_tx_handler(struct rte_eth_dev *eth_dev)
+{
+       struct enic *enic = pmd_priv(eth_dev);
+
+       if (enic->use_simple_tx_handler) {
+               ENICPMD_LOG(DEBUG, " use the simple tx handler");
+               eth_dev->tx_pkt_burst = &enic_simple_xmit_pkts;
+       } else {
+               ENICPMD_LOG(DEBUG, " use the default tx handler");
+               eth_dev->tx_pkt_burst = &enic_xmit_pkts;
+       }
+}
+
 int enic_enable(struct enic *enic)
 {
        unsigned int index;
        int err;
        struct rte_eth_dev *eth_dev = enic->rte_dev;
        uint64_t simple_tx_offloads;
+       uintptr_t p;
+
+       if (enic->enable_avx2_rx) {
+               struct rte_mbuf mb_def = { .buf_addr = 0 };
+
+               /*
+                * mbuf_initializer contains const-after-init fields of
+                * receive mbufs (i.e. 64 bits of fields from rearm_data).
+                * It is currently used by the vectorized handler.
+                */
+               mb_def.nb_segs = 1;
+               mb_def.data_off = RTE_PKTMBUF_HEADROOM;
+               mb_def.port = enic->port_id;
+               rte_mbuf_refcnt_set(&mb_def, 1);
+               rte_compiler_barrier();
+               p = (uintptr_t)&mb_def.rearm_data;
+               enic->mbuf_initializer = *(uint64_t *)p;
+       }
 
        eth_dev->data->dev_link.link_speed = vnic_dev_port_speed(enic->vdev);
        eth_dev->data->dev_link.link_duplex = ETH_LINK_FULL_DUPLEX;
@@ -584,16 +638,17 @@ int enic_enable(struct enic *enic)
                 DEV_TX_OFFLOAD_TCP_CKSUM);
        if ((eth_dev->data->dev_conf.txmode.offloads &
             ~simple_tx_offloads) == 0) {
-               PMD_INIT_LOG(DEBUG, " use the simple tx handler");
+               ENICPMD_LOG(DEBUG, " use the simple tx handler");
                eth_dev->tx_pkt_burst = &enic_simple_xmit_pkts;
                for (index = 0; index < enic->wq_count; index++)
                        enic_prep_wq_for_simple_tx(enic, index);
+               enic->use_simple_tx_handler = 1;
        } else {
-               PMD_INIT_LOG(DEBUG, " use the default tx handler");
+               ENICPMD_LOG(DEBUG, " use the default tx handler");
                eth_dev->tx_pkt_burst = &enic_xmit_pkts;
        }
 
-       pick_rx_handler(enic);
+       enic_pick_rx_handler(eth_dev);
 
        for (index = 0; index < enic->wq_count; index++)
                enic_start_wq(enic, index);
@@ -683,31 +738,31 @@ void enic_free_rq(void *rxq)
 
 void enic_start_wq(struct enic *enic, uint16_t queue_idx)
 {
-       struct rte_eth_dev *eth_dev = enic->rte_dev;
+       struct rte_eth_dev_data *data = enic->dev_data;
        vnic_wq_enable(&enic->wq[queue_idx]);
-       eth_dev->data->tx_queue_state[queue_idx] = RTE_ETH_QUEUE_STATE_STARTED;
+       data->tx_queue_state[queue_idx] = RTE_ETH_QUEUE_STATE_STARTED;
 }
 
 int enic_stop_wq(struct enic *enic, uint16_t queue_idx)
 {
-       struct rte_eth_dev *eth_dev = enic->rte_dev;
+       struct rte_eth_dev_data *data = enic->dev_data;
        int ret;
 
        ret = vnic_wq_disable(&enic->wq[queue_idx]);
        if (ret)
                return ret;
 
-       eth_dev->data->tx_queue_state[queue_idx] = RTE_ETH_QUEUE_STATE_STOPPED;
+       data->tx_queue_state[queue_idx] = RTE_ETH_QUEUE_STATE_STOPPED;
        return 0;
 }
 
 void enic_start_rq(struct enic *enic, uint16_t queue_idx)
 {
+       struct rte_eth_dev_data *data = enic->dev_data;
        struct vnic_rq *rq_sop;
        struct vnic_rq *rq_data;
        rq_sop = &enic->rq[enic_rte_rq_idx_to_sop_idx(queue_idx)];
        rq_data = &enic->rq[rq_sop->data_queue_idx];
-       struct rte_eth_dev *eth_dev = enic->rte_dev;
 
        if (rq_data->in_use) {
                vnic_rq_enable(rq_data);
@@ -716,13 +771,13 @@ void enic_start_rq(struct enic *enic, uint16_t queue_idx)
        rte_mb();
        vnic_rq_enable(rq_sop);
        enic_initial_post_rx(enic, rq_sop);
-       eth_dev->data->rx_queue_state[queue_idx] = RTE_ETH_QUEUE_STATE_STARTED;
+       data->rx_queue_state[queue_idx] = RTE_ETH_QUEUE_STATE_STARTED;
 }
 
 int enic_stop_rq(struct enic *enic, uint16_t queue_idx)
 {
+       struct rte_eth_dev_data *data = enic->dev_data;
        int ret1 = 0, ret2 = 0;
-       struct rte_eth_dev *eth_dev = enic->rte_dev;
        struct vnic_rq *rq_sop;
        struct vnic_rq *rq_data;
        rq_sop = &enic->rq[enic_rte_rq_idx_to_sop_idx(queue_idx)];
@@ -738,7 +793,7 @@ int enic_stop_rq(struct enic *enic, uint16_t queue_idx)
        else if (ret1)
                return ret1;
 
-       eth_dev->data->rx_queue_state[queue_idx] = RTE_ETH_QUEUE_STATE_STOPPED;
+       data->rx_queue_state[queue_idx] = RTE_ETH_QUEUE_STATE_STOPPED;
        return 0;
 }
 
@@ -1330,10 +1385,10 @@ int enic_set_vlan_strip(struct enic *enic)
                               enic->rss_enable);
 }
 
-void enic_add_packet_filter(struct enic *enic)
+int enic_add_packet_filter(struct enic *enic)
 {
        /* Args -> directed, multicast, broadcast, promisc, allmulti */
-       vnic_dev_packet_filter(enic->vdev, 1, 1, 1,
+       return vnic_dev_packet_filter(enic->vdev, 1, 1, 1,
                enic->promisc, enic->allmulti);
 }
 
@@ -1344,12 +1399,10 @@ int enic_get_link_status(struct enic *enic)
 
 static void enic_dev_deinit(struct enic *enic)
 {
-       struct rte_eth_dev *eth_dev = enic->rte_dev;
-
        /* stop link status checking */
        vnic_dev_notify_unset(enic->vdev);
 
-       rte_free(eth_dev->data->mac_addrs);
+       /* mac_addrs is freed by rte_eth_dev_release_port() */
        rte_free(enic->cq);
        rte_free(enic->intr);
        rte_free(enic->rq);
@@ -1566,7 +1619,7 @@ int enic_set_mtu(struct enic *enic, uint16_t new_mtu)
 
        /* put back the real receive function */
        rte_mb();
-       pick_rx_handler(enic);
+       enic_pick_rx_handler(eth_dev);
        rte_mb();
 
        /* restart Rx traffic */
@@ -1634,24 +1687,47 @@ static int enic_dev_init(struct enic *enic)
        /* Get the supported filters */
        enic_fdir_info(enic);
 
-       eth_dev->data->mac_addrs = rte_zmalloc("enic_mac_addr", ETH_ALEN
-                                               * ENIC_MAX_MAC_ADDR, 0);
+       eth_dev->data->mac_addrs = rte_zmalloc("enic_mac_addr",
+                                       sizeof(struct rte_ether_addr) *
+                                       ENIC_UNICAST_PERFECT_FILTERS, 0);
        if (!eth_dev->data->mac_addrs) {
                dev_err(enic, "mac addr storage alloc failed, aborting.\n");
                return -1;
        }
-       ether_addr_copy((struct ether_addr *) enic->mac_addr,
+       rte_ether_addr_copy((struct rte_ether_addr *)enic->mac_addr,
                        eth_dev->data->mac_addrs);
 
        vnic_dev_set_reset_flag(enic->vdev, 0);
 
        LIST_INIT(&enic->flows);
-       rte_spinlock_init(&enic->flows_lock);
 
        /* set up link status checking */
        vnic_dev_notify_set(enic->vdev, -1); /* No Intr for notify */
 
+       /*
+        * When Geneve with options offload is available, always disable it
+        * first as it can interfere with user flow rules.
+        */
+       if (enic->geneve_opt_avail &&
+           vnic_dev_overlay_offload_ctrl(enic->vdev,
+                       OVERLAY_FEATURE_GENEVE,
+                       OVERLAY_OFFLOAD_DISABLE)) {
+               dev_err(enic, "failed to disable geneve+option\n");
+       }
        enic->overlay_offload = false;
+       if (enic->disable_overlay && enic->vxlan) {
+               /*
+                * Explicitly disable overlay offload as the setting is
+                * sticky, and resetting vNIC does not disable it.
+                */
+               if (vnic_dev_overlay_offload_ctrl(enic->vdev,
+                                                 OVERLAY_FEATURE_VXLAN,
+                                                 OVERLAY_OFFLOAD_DISABLE)) {
+                       dev_err(enic, "failed to disable overlay offload\n");
+               } else {
+                       dev_info(enic, "Overlay offload is disabled\n");
+               }
+       }
        if (!enic->disable_overlay && enic->vxlan &&
            /* 'VXLAN feature' enables VXLAN, NVGRE, and GENEVE. */
            vnic_dev_overlay_offload_ctrl(enic->vdev,
@@ -1661,16 +1737,33 @@ static int enic_dev_init(struct enic *enic)
                        DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM |
                        DEV_TX_OFFLOAD_GENEVE_TNL_TSO |
                        DEV_TX_OFFLOAD_VXLAN_TNL_TSO;
-               /*
-                * Do not add PKT_TX_OUTER_{IPV4,IPV6} as they are not
-                * 'offload' flags (i.e. not part of PKT_TX_OFFLOAD_MASK).
-                */
                enic->tx_offload_mask |=
+                       PKT_TX_OUTER_IPV6 |
+                       PKT_TX_OUTER_IPV4 |
                        PKT_TX_OUTER_IP_CKSUM |
                        PKT_TX_TUNNEL_MASK;
                enic->overlay_offload = true;
-               enic->vxlan_port = ENIC_DEFAULT_VXLAN_PORT;
                dev_info(enic, "Overlay offload is enabled\n");
+       }
+       /* Geneve with options offload requires overlay offload */
+       if (enic->overlay_offload && enic->geneve_opt_avail &&
+           enic->geneve_opt_request) {
+               if (vnic_dev_overlay_offload_ctrl(enic->vdev,
+                               OVERLAY_FEATURE_GENEVE,
+                               OVERLAY_OFFLOAD_ENABLE)) {
+                       dev_err(enic, "failed to enable geneve+option\n");
+               } else {
+                       enic->geneve_opt_enabled = 1;
+                       dev_info(enic, "Geneve with options is enabled\n");
+               }
+       }
+       /*
+        * Reset the vxlan port if HW vxlan parsing is available. It
+        * is always enabled regardless of overlay offload
+        * enable/disable.
+        */
+       if (enic->vxlan) {
+               enic->vxlan_port = ENIC_DEFAULT_VXLAN_PORT;
                /*
                 * Reset the vxlan port to the default, as the NIC firmware
                 * does not reset it automatically and keeps the old setting.
@@ -1692,7 +1785,7 @@ int enic_probe(struct enic *enic)
        struct rte_pci_device *pdev = enic->pdev;
        int err = -1;
 
-       dev_debug(enic, " Initializing ENIC PMD\n");
+       dev_debug(enic, "Initializing ENIC PMD\n");
 
        /* if this is a secondary process the hardware is already initialized */
        if (rte_eal_process_type() != RTE_PROC_PRIMARY)