net/mlx5: add option to allocate memory from system
[dpdk.git] / drivers / net / mlx5 / mlx5.c
index f62ad12..b2a8460 100644 (file)
@@ -38,6 +38,7 @@
 #include <mlx5_glue.h>
 #include <mlx5_devx_cmds.h>
 #include <mlx5_common.h>
+#include <mlx5_common_os.h>
 #include <mlx5_common_mp.h>
 
 #include "mlx5_defs.h"
  */
 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
 
+/*
+ * Device parameter to enable Tx scheduling on timestamps
+ * and specify the packet pacing granularity in nanoseconds.
+ */
+#define MLX5_TX_PP "tx_pp"
+
+/*
+ * Device parameter to specify skew in nanoseconds on Tx datapath,
+ * it represents the time between SQ start WQE processing and
+ * appearing actual packet data on the wire.
+ */
+#define MLX5_TX_SKEW "tx_skew"
+
 /*
  * Device parameter to enable hardware Tx vector.
  * Deprecated, ignored (no vectorized Tx routines anymore).
 /* Enable extensive flow metadata support. */
 #define MLX5_DV_XMETA_EN "dv_xmeta_en"
 
+/* Device parameter to let the user manage the lacp traffic of bonded device */
+#define MLX5_LACP_BY_USER "lacp_by_user"
+
 /* Activate Netlink support in VF mode. */
 #define MLX5_VF_NL_EN "vf_nl_en"
 
 /* Flow memory reclaim mode. */
 #define MLX5_RECLAIM_MEM "reclaim_mem_mode"
 
+/* The default memory allocator used in PMD. */
+#define MLX5_SYS_MEM_EN "sys_mem_en"
+
 static const char *MZ_MLX5_PMD_SHARED_DATA = "mlx5_pmd_shared_data";
 
 /* Shared memory between primary and secondary processes. */
@@ -173,13 +193,12 @@ static rte_spinlock_t mlx5_shared_data_lock = RTE_SPINLOCK_INITIALIZER;
 
 /* Process local data for secondary processes. */
 static struct mlx5_local_data mlx5_local_data;
-/** Driver-specific log messages type. */
-int mlx5_logtype;
 
-static LIST_HEAD(, mlx5_dev_ctx_shared) mlx5_ibv_list = LIST_HEAD_INITIALIZER();
-static pthread_mutex_t mlx5_ibv_list_mutex = PTHREAD_MUTEX_INITIALIZER;
+static LIST_HEAD(, mlx5_dev_ctx_shared) mlx5_dev_ctx_list =
+                                               LIST_HEAD_INITIALIZER();
+static pthread_mutex_t mlx5_dev_ctx_list_mutex = PTHREAD_MUTEX_INITIALIZER;
 
-static struct mlx5_indexed_pool_config mlx5_ipool_cfg[] = {
+static const struct mlx5_indexed_pool_config mlx5_ipool_cfg[] = {
 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
        {
                .size = sizeof(struct mlx5_flow_dv_encap_decap_resource),
@@ -271,7 +290,11 @@ static struct mlx5_indexed_pool_config mlx5_ipool_cfg[] = {
                .type = "mlx5_hrxq_ipool",
        },
        {
-               .size = sizeof(struct mlx5_flow_handle),
+               /*
+                * MLX5_IPOOL_MLX5_FLOW size varies for DV and VERBS flows.
+                * It set in run time according to PCI function configuration.
+                */
+               .size = 0,
                .trunk_size = 64,
                .grow_trunk = 3,
                .grow_shift = 2,
@@ -452,8 +475,13 @@ mlx5_flow_counters_mng_init(struct mlx5_dev_ctx_shared *sh)
        memset(&sh->cmng, 0, sizeof(sh->cmng));
        TAILQ_INIT(&sh->cmng.flow_counters);
        for (i = 0; i < MLX5_CCONT_TYPE_MAX; ++i) {
+               sh->cmng.ccont[i].min_id = MLX5_CNT_BATCH_OFFSET;
+               sh->cmng.ccont[i].max_id = -1;
+               sh->cmng.ccont[i].last_pool_idx = POOL_IDX_INVALID;
                TAILQ_INIT(&sh->cmng.ccont[i].pool_list);
                rte_spinlock_init(&sh->cmng.ccont[i].resize_sl);
+               TAILQ_INIT(&sh->cmng.ccont[i].counters);
+               rte_spinlock_init(&sh->cmng.ccont[i].csl);
        }
 }
 
@@ -542,24 +570,29 @@ mlx5_flow_counters_mng_close(struct mlx5_dev_ctx_shared *sh)
  */
 static void
 mlx5_flow_ipool_create(struct mlx5_dev_ctx_shared *sh,
-                      const struct mlx5_dev_config *config __rte_unused)
+                      const struct mlx5_dev_config *config)
 {
        uint8_t i;
+       struct mlx5_indexed_pool_config cfg;
 
-#ifdef HAVE_IBV_FLOW_DV_SUPPORT
-       /*
-        * While DV is supported, user chooses the verbs mode,
-        * the mlx5 flow handle size is different with the
-        * MLX5_FLOW_HANDLE_VERBS_SIZE.
-        */
-       if (!config->dv_flow_en)
-               mlx5_ipool_cfg[MLX5_IPOOL_MLX5_FLOW].size =
-                                       MLX5_FLOW_HANDLE_VERBS_SIZE;
-#endif
        for (i = 0; i < MLX5_IPOOL_MAX; ++i) {
+               cfg = mlx5_ipool_cfg[i];
+               switch (i) {
+               default:
+                       break;
+               /*
+                * Set MLX5_IPOOL_MLX5_FLOW ipool size
+                * according to PCI function flow configuration.
+                */
+               case MLX5_IPOOL_MLX5_FLOW:
+                       cfg.size = config->dv_flow_en ?
+                               sizeof(struct mlx5_flow_handle) :
+                               MLX5_FLOW_HANDLE_VERBS_SIZE;
+                       break;
+               }
                if (config->reclaim_mode)
-                       mlx5_ipool_cfg[i].release_mem_en = 1;
-               sh->ipool[i] = mlx5_ipool_create(&mlx5_ipool_cfg[i]);
+                       cfg.release_mem_en = 1;
+               sh->ipool[i] = mlx5_ipool_create(&cfg);
        }
 }
 
@@ -578,19 +611,124 @@ mlx5_flow_ipool_destroy(struct mlx5_dev_ctx_shared *sh)
                mlx5_ipool_destroy(sh->ipool[i]);
 }
 
+/*
+ * Check if dynamic flex parser for eCPRI already exists.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ *
+ * @return
+ *   true on exists, false on not.
+ */
+bool
+mlx5_flex_parser_ecpri_exist(struct rte_eth_dev *dev)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_flex_parser_profiles *prf =
+                               &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
+
+       return !!prf->obj;
+}
+
+/*
+ * Allocation of a flex parser for eCPRI. Once created, this parser related
+ * resources will be held until the device is closed.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ *
+ * @return
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+int
+mlx5_flex_parser_ecpri_alloc(struct rte_eth_dev *dev)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_flex_parser_profiles *prf =
+                               &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
+       struct mlx5_devx_graph_node_attr node = {
+               .modify_field_select = 0,
+       };
+       uint32_t ids[8];
+       int ret;
+
+       if (!priv->config.hca_attr.parse_graph_flex_node) {
+               DRV_LOG(ERR, "Dynamic flex parser is not supported "
+                       "for device %s.", priv->dev_data->name);
+               return -ENOTSUP;
+       }
+       node.header_length_mode = MLX5_GRAPH_NODE_LEN_FIXED;
+       /* 8 bytes now: 4B common header + 4B message body header. */
+       node.header_length_base_value = 0x8;
+       /* After MAC layer: Ether / VLAN. */
+       node.in[0].arc_parse_graph_node = MLX5_GRAPH_ARC_NODE_MAC;
+       /* Type of compared condition should be 0xAEFE in the L2 layer. */
+       node.in[0].compare_condition_value = RTE_ETHER_TYPE_ECPRI;
+       /* Sample #0: type in common header. */
+       node.sample[0].flow_match_sample_en = 1;
+       /* Fixed offset. */
+       node.sample[0].flow_match_sample_offset_mode = 0x0;
+       /* Only the 2nd byte will be used. */
+       node.sample[0].flow_match_sample_field_base_offset = 0x0;
+       /* Sample #1: message payload. */
+       node.sample[1].flow_match_sample_en = 1;
+       /* Fixed offset. */
+       node.sample[1].flow_match_sample_offset_mode = 0x0;
+       /*
+        * Only the first two bytes will be used right now, and its offset will
+        * start after the common header that with the length of a DW(u32).
+        */
+       node.sample[1].flow_match_sample_field_base_offset = sizeof(uint32_t);
+       prf->obj = mlx5_devx_cmd_create_flex_parser(priv->sh->ctx, &node);
+       if (!prf->obj) {
+               DRV_LOG(ERR, "Failed to create flex parser node object.");
+               return (rte_errno == 0) ? -ENODEV : -rte_errno;
+       }
+       prf->num = 2;
+       ret = mlx5_devx_cmd_query_parse_samples(prf->obj, ids, prf->num);
+       if (ret) {
+               DRV_LOG(ERR, "Failed to query sample IDs.");
+               return (rte_errno == 0) ? -ENODEV : -rte_errno;
+       }
+       prf->offset[0] = 0x0;
+       prf->offset[1] = sizeof(uint32_t);
+       prf->ids[0] = ids[0];
+       prf->ids[1] = ids[1];
+       return 0;
+}
+
+/*
+ * Destroy the flex parser node, including the parser itself, input / output
+ * arcs and DW samples. Resources could be reused then.
+ *
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ */
+static void
+mlx5_flex_parser_ecpri_release(struct rte_eth_dev *dev)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_flex_parser_profiles *prf =
+                               &priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
+
+       if (prf->obj)
+               mlx5_devx_cmd_destroy(prf->obj);
+       prf->obj = NULL;
+}
+
 /**
- * Allocate shared IB device context. If there is multiport device the
+ * Allocate shared device context. If there is multiport device the
  * master and representors will share this context, if there is single
- * port dedicated IB device, the context will be used by only given
+ * port dedicated device, the context will be used by only given
  * port due to unification.
  *
- * Routine first searches the context for the specified IB device name,
+ * Routine first searches the context for the specified device name,
  * if found the shared context assumed and reference counter is incremented.
  * If no context found the new one is created and initialized with specified
- * IB device context and parameters.
+ * device context and parameters.
  *
  * @param[in] spawn
- *   Pointer to the IB device attributes (name, port, etc).
+ *   Pointer to the device attributes (name, port, etc).
  * @param[in] config
  *   Pointer to device configuration structure.
  *
@@ -599,8 +737,8 @@ mlx5_flow_ipool_destroy(struct mlx5_dev_ctx_shared *sh)
  *   otherwise NULL and rte_errno is set.
  */
 struct mlx5_dev_ctx_shared *
-mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn,
-                       const struct mlx5_dev_config *config)
+mlx5_alloc_shared_dev_ctx(const struct mlx5_dev_spawn_data *spawn,
+                          const struct mlx5_dev_config *config)
 {
        struct mlx5_dev_ctx_shared *sh;
        int err = 0;
@@ -610,10 +748,11 @@ mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn,
        MLX5_ASSERT(spawn);
        /* Secondary process should not create the shared context. */
        MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
-       pthread_mutex_lock(&mlx5_ibv_list_mutex);
+       pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
        /* Search for IB context by device name. */
-       LIST_FOREACH(sh, &mlx5_ibv_list, next) {
-               if (!strcmp(sh->ibdev_name, spawn->ibv_dev->name)) {
+       LIST_FOREACH(sh, &mlx5_dev_ctx_list, next) {
+               if (!strcmp(sh->ibdev_name,
+                       mlx5_os_get_dev_device_name(spawn->phys_dev))) {
                        sh->refcnt++;
                        goto exit;
                }
@@ -623,7 +762,7 @@ mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn,
        sh = rte_zmalloc("ethdev shared ib context",
                         sizeof(struct mlx5_dev_ctx_shared) +
                         spawn->max_port *
-                        sizeof(struct mlx5_ibv_shared_port),
+                        sizeof(struct mlx5_dev_shared_port),
                         RTE_CACHE_LINE_SIZE);
        if (!sh) {
                DRV_LOG(ERR, "shared context allocation failure");
@@ -678,6 +817,12 @@ mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn,
                        err = ENOMEM;
                        goto error;
                }
+               sh->tx_uar = mlx5_glue->devx_alloc_uar(sh->ctx, 0);
+               if (!sh->tx_uar) {
+                       DRV_LOG(ERR, "Failed to allocate DevX UAR.");
+                       err = ENOMEM;
+                       goto error;
+               }
        }
        sh->flow_id_pool = mlx5_flow_id_pool_alloc
                                        ((1 << HAIRPIN_FLOW_ID_BITS) - 1);
@@ -686,6 +831,12 @@ mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn,
                err = ENOMEM;
                goto error;
        }
+#ifndef RTE_ARCH_64
+       /* Initialize UAR access locks for 32bit implementations. */
+       rte_spinlock_init(&sh->uar_lock_cq);
+       for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
+               rte_spinlock_init(&sh->uar_lock[i]);
+#endif
        /*
         * Once the device is added to the list of memory event
         * callback, its global MR cache table cannot be expanded
@@ -702,7 +853,14 @@ mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn,
                err = rte_errno;
                goto error;
        }
+       mlx5_os_set_reg_mr_cb(&sh->share_cache.reg_mr_cb,
+                             &sh->share_cache.dereg_mr_cb);
        mlx5_os_dev_shared_handler_install(sh);
+       sh->cnt_id_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_DWORD);
+       if (!sh->cnt_id_tbl) {
+               err = rte_errno;
+               goto error;
+       }
        mlx5_flow_aging_init(sh);
        mlx5_flow_counters_mng_init(sh);
        mlx5_flow_ipool_create(sh, config);
@@ -712,13 +870,22 @@ mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn,
                         sh, mem_event_cb);
        rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
        /* Add context to the global device list. */
-       LIST_INSERT_HEAD(&mlx5_ibv_list, sh, next);
+       LIST_INSERT_HEAD(&mlx5_dev_ctx_list, sh, next);
 exit:
-       pthread_mutex_unlock(&mlx5_ibv_list_mutex);
+       pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
        return sh;
 error:
-       pthread_mutex_unlock(&mlx5_ibv_list_mutex);
+       pthread_mutex_destroy(&sh->txpp.mutex);
+       pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
        MLX5_ASSERT(sh);
+       if (sh->cnt_id_tbl) {
+               mlx5_l3t_destroy(sh->cnt_id_tbl);
+               sh->cnt_id_tbl = NULL;
+       }
+       if (sh->tx_uar) {
+               mlx5_glue->devx_free_uar(sh->tx_uar);
+               sh->tx_uar = NULL;
+       }
        if (sh->tis)
                claim_zero(mlx5_devx_cmd_destroy(sh->tis));
        if (sh->td)
@@ -743,14 +910,14 @@ error:
  *   Pointer to mlx5_dev_ctx_shared object to free
  */
 void
-mlx5_free_shared_ibctx(struct mlx5_dev_ctx_shared *sh)
+mlx5_free_shared_dev_ctx(struct mlx5_dev_ctx_shared *sh)
 {
-       pthread_mutex_lock(&mlx5_ibv_list_mutex);
+       pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
 #ifdef RTE_LIBRTE_MLX5_DEBUG
        /* Check the object presence in the list. */
        struct mlx5_dev_ctx_shared *lctx;
 
-       LIST_FOREACH(lctx, &mlx5_ibv_list, next)
+       LIST_FOREACH(lctx, &mlx5_dev_ctx_list, next)
                if (lctx == sh)
                        break;
        MLX5_ASSERT(lctx);
@@ -780,6 +947,14 @@ mlx5_free_shared_ibctx(struct mlx5_dev_ctx_shared *sh)
        mlx5_flow_counters_mng_close(sh);
        mlx5_flow_ipool_destroy(sh);
        mlx5_os_dev_shared_handler_uninstall(sh);
+       if (sh->cnt_id_tbl) {
+               mlx5_l3t_destroy(sh->cnt_id_tbl);
+               sh->cnt_id_tbl = NULL;
+       }
+       if (sh->tx_uar) {
+               mlx5_glue->devx_free_uar(sh->tx_uar);
+               sh->tx_uar = NULL;
+       }
        if (sh->pd)
                claim_zero(mlx5_glue->dealloc_pd(sh->pd));
        if (sh->tis)
@@ -790,9 +965,10 @@ mlx5_free_shared_ibctx(struct mlx5_dev_ctx_shared *sh)
                claim_zero(mlx5_glue->close_device(sh->ctx));
        if (sh->flow_id_pool)
                mlx5_flow_id_pool_release(sh->flow_id_pool);
+       pthread_mutex_destroy(&sh->txpp.mutex);
        rte_free(sh);
 exit:
-       pthread_mutex_unlock(&mlx5_ibv_list_mutex);
+       pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
 }
 
 /**
@@ -869,7 +1045,7 @@ mlx5_alloc_table_hash_list(struct mlx5_priv *priv)
        snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
        sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE);
        if (!sh->flow_tbls) {
-               DRV_LOG(ERR, "flow tables with hash creation failed.\n");
+               DRV_LOG(ERR, "flow tables with hash creation failed.");
                err = ENOMEM;
                return err;
        }
@@ -1123,6 +1299,8 @@ mlx5_dev_close(struct rte_eth_dev *dev)
        rte_wmb();
        /* Disable datapath on secondary process. */
        mlx5_mp_req_stop_rxtx(dev);
+       /* Free the eCPRI flex parser resource. */
+       mlx5_flex_parser_ecpri_release(dev);
        if (priv->rxqs != NULL) {
                /* XXX race condition if mlx5_rx_burst() is still running. */
                usleep(1000);
@@ -1192,7 +1370,7 @@ mlx5_dev_close(struct rte_eth_dev *dev)
         * mlx5_nl_mac_addr_flush() uses ibdev_path for retrieveing
         * ifindex if Netlink fails.
         */
-       mlx5_free_shared_ibctx(priv->sh);
+       mlx5_free_shared_dev_ctx(priv->sh);
        if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
                unsigned int c = 0;
                uint16_t port_id;
@@ -1221,137 +1399,6 @@ mlx5_dev_close(struct rte_eth_dev *dev)
        dev->data->mac_addrs = NULL;
 }
 
-const struct eth_dev_ops mlx5_dev_ops = {
-       .dev_configure = mlx5_dev_configure,
-       .dev_start = mlx5_dev_start,
-       .dev_stop = mlx5_dev_stop,
-       .dev_set_link_down = mlx5_set_link_down,
-       .dev_set_link_up = mlx5_set_link_up,
-       .dev_close = mlx5_dev_close,
-       .promiscuous_enable = mlx5_promiscuous_enable,
-       .promiscuous_disable = mlx5_promiscuous_disable,
-       .allmulticast_enable = mlx5_allmulticast_enable,
-       .allmulticast_disable = mlx5_allmulticast_disable,
-       .link_update = mlx5_link_update,
-       .stats_get = mlx5_stats_get,
-       .stats_reset = mlx5_stats_reset,
-       .xstats_get = mlx5_xstats_get,
-       .xstats_reset = mlx5_xstats_reset,
-       .xstats_get_names = mlx5_xstats_get_names,
-       .fw_version_get = mlx5_fw_version_get,
-       .dev_infos_get = mlx5_dev_infos_get,
-       .read_clock = mlx5_read_clock,
-       .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
-       .vlan_filter_set = mlx5_vlan_filter_set,
-       .rx_queue_setup = mlx5_rx_queue_setup,
-       .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
-       .tx_queue_setup = mlx5_tx_queue_setup,
-       .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
-       .rx_queue_release = mlx5_rx_queue_release,
-       .tx_queue_release = mlx5_tx_queue_release,
-       .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
-       .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
-       .mac_addr_remove = mlx5_mac_addr_remove,
-       .mac_addr_add = mlx5_mac_addr_add,
-       .mac_addr_set = mlx5_mac_addr_set,
-       .set_mc_addr_list = mlx5_set_mc_addr_list,
-       .mtu_set = mlx5_dev_set_mtu,
-       .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
-       .vlan_offload_set = mlx5_vlan_offload_set,
-       .reta_update = mlx5_dev_rss_reta_update,
-       .reta_query = mlx5_dev_rss_reta_query,
-       .rss_hash_update = mlx5_rss_hash_update,
-       .rss_hash_conf_get = mlx5_rss_hash_conf_get,
-       .filter_ctrl = mlx5_dev_filter_ctrl,
-       .rx_descriptor_status = mlx5_rx_descriptor_status,
-       .tx_descriptor_status = mlx5_tx_descriptor_status,
-       .rxq_info_get = mlx5_rxq_info_get,
-       .txq_info_get = mlx5_txq_info_get,
-       .rx_burst_mode_get = mlx5_rx_burst_mode_get,
-       .tx_burst_mode_get = mlx5_tx_burst_mode_get,
-       .rx_queue_count = mlx5_rx_queue_count,
-       .rx_queue_intr_enable = mlx5_rx_intr_enable,
-       .rx_queue_intr_disable = mlx5_rx_intr_disable,
-       .is_removed = mlx5_is_removed,
-       .udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
-       .get_module_info = mlx5_get_module_info,
-       .get_module_eeprom = mlx5_get_module_eeprom,
-       .hairpin_cap_get = mlx5_hairpin_cap_get,
-       .mtr_ops_get = mlx5_flow_meter_ops_get,
-};
-
-/* Available operations from secondary process. */
-const struct eth_dev_ops mlx5_dev_sec_ops = {
-       .stats_get = mlx5_stats_get,
-       .stats_reset = mlx5_stats_reset,
-       .xstats_get = mlx5_xstats_get,
-       .xstats_reset = mlx5_xstats_reset,
-       .xstats_get_names = mlx5_xstats_get_names,
-       .fw_version_get = mlx5_fw_version_get,
-       .dev_infos_get = mlx5_dev_infos_get,
-       .rx_descriptor_status = mlx5_rx_descriptor_status,
-       .tx_descriptor_status = mlx5_tx_descriptor_status,
-       .rxq_info_get = mlx5_rxq_info_get,
-       .txq_info_get = mlx5_txq_info_get,
-       .rx_burst_mode_get = mlx5_rx_burst_mode_get,
-       .tx_burst_mode_get = mlx5_tx_burst_mode_get,
-       .get_module_info = mlx5_get_module_info,
-       .get_module_eeprom = mlx5_get_module_eeprom,
-};
-
-/* Available operations in flow isolated mode. */
-const struct eth_dev_ops mlx5_dev_ops_isolate = {
-       .dev_configure = mlx5_dev_configure,
-       .dev_start = mlx5_dev_start,
-       .dev_stop = mlx5_dev_stop,
-       .dev_set_link_down = mlx5_set_link_down,
-       .dev_set_link_up = mlx5_set_link_up,
-       .dev_close = mlx5_dev_close,
-       .promiscuous_enable = mlx5_promiscuous_enable,
-       .promiscuous_disable = mlx5_promiscuous_disable,
-       .allmulticast_enable = mlx5_allmulticast_enable,
-       .allmulticast_disable = mlx5_allmulticast_disable,
-       .link_update = mlx5_link_update,
-       .stats_get = mlx5_stats_get,
-       .stats_reset = mlx5_stats_reset,
-       .xstats_get = mlx5_xstats_get,
-       .xstats_reset = mlx5_xstats_reset,
-       .xstats_get_names = mlx5_xstats_get_names,
-       .fw_version_get = mlx5_fw_version_get,
-       .dev_infos_get = mlx5_dev_infos_get,
-       .dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
-       .vlan_filter_set = mlx5_vlan_filter_set,
-       .rx_queue_setup = mlx5_rx_queue_setup,
-       .rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
-       .tx_queue_setup = mlx5_tx_queue_setup,
-       .tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
-       .rx_queue_release = mlx5_rx_queue_release,
-       .tx_queue_release = mlx5_tx_queue_release,
-       .flow_ctrl_get = mlx5_dev_get_flow_ctrl,
-       .flow_ctrl_set = mlx5_dev_set_flow_ctrl,
-       .mac_addr_remove = mlx5_mac_addr_remove,
-       .mac_addr_add = mlx5_mac_addr_add,
-       .mac_addr_set = mlx5_mac_addr_set,
-       .set_mc_addr_list = mlx5_set_mc_addr_list,
-       .mtu_set = mlx5_dev_set_mtu,
-       .vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
-       .vlan_offload_set = mlx5_vlan_offload_set,
-       .filter_ctrl = mlx5_dev_filter_ctrl,
-       .rx_descriptor_status = mlx5_rx_descriptor_status,
-       .tx_descriptor_status = mlx5_tx_descriptor_status,
-       .rxq_info_get = mlx5_rxq_info_get,
-       .txq_info_get = mlx5_txq_info_get,
-       .rx_burst_mode_get = mlx5_rx_burst_mode_get,
-       .tx_burst_mode_get = mlx5_tx_burst_mode_get,
-       .rx_queue_intr_enable = mlx5_rx_intr_enable,
-       .rx_queue_intr_disable = mlx5_rx_intr_disable,
-       .is_removed = mlx5_is_removed,
-       .get_module_info = mlx5_get_module_info,
-       .get_module_eeprom = mlx5_get_module_eeprom,
-       .hairpin_cap_get = mlx5_hairpin_cap_get,
-       .mtr_ops_get = mlx5_flow_meter_ops_get,
-};
-
 /**
  * Verify and store value for device argument.
  *
@@ -1369,18 +1416,26 @@ static int
 mlx5_args_check(const char *key, const char *val, void *opaque)
 {
        struct mlx5_dev_config *config = opaque;
-       unsigned long tmp;
+       unsigned long mod;
+       signed long tmp;
 
        /* No-op, port representors are processed in mlx5_dev_spawn(). */
        if (!strcmp(MLX5_REPRESENTOR, key))
                return 0;
        errno = 0;
-       tmp = strtoul(val, NULL, 0);
+       tmp = strtol(val, NULL, 0);
        if (errno) {
                rte_errno = errno;
                DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
                return -rte_errno;
        }
+       if (tmp < 0 && strcmp(MLX5_TX_PP, key) && strcmp(MLX5_TX_SKEW, key)) {
+               /* Negative values are acceptable for some keys only. */
+               rte_errno = EINVAL;
+               DRV_LOG(WARNING, "%s: invalid negative value \"%s\"", key, val);
+               return -rte_errno;
+       }
+       mod = tmp >= 0 ? tmp : -tmp;
        if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
                config->cqe_comp = !!tmp;
        } else if (strcmp(MLX5_RXQ_CQE_PAD_EN, key) == 0) {
@@ -1431,6 +1486,15 @@ mlx5_args_check(const char *key, const char *val, void *opaque)
                config->txq_inline_mpw = tmp;
        } else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
                DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
+       } else if (strcmp(MLX5_TX_PP, key) == 0) {
+               if (!mod) {
+                       DRV_LOG(ERR, "Zero Tx packet pacing parameter");
+                       rte_errno = EINVAL;
+                       return -rte_errno;
+               }
+               config->tx_pp = tmp;
+       } else if (strcmp(MLX5_TX_SKEW, key) == 0) {
+               config->tx_skew = tmp;
        } else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
                config->rx_vec_en = !!tmp;
        } else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
@@ -1451,6 +1515,8 @@ mlx5_args_check(const char *key, const char *val, void *opaque)
                        return -rte_errno;
                }
                config->dv_xmeta_en = tmp;
+       } else if (strcmp(MLX5_LACP_BY_USER, key) == 0) {
+               config->lacp_by_user = !!tmp;
        } else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
                config->mr_ext_memseg_en = !!tmp;
        } else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
@@ -1470,6 +1536,8 @@ mlx5_args_check(const char *key, const char *val, void *opaque)
                        return -rte_errno;
                }
                config->reclaim_mode = tmp;
+       } else if (strcmp(MLX5_SYS_MEM_EN, key) == 0) {
+               config->sys_mem_en = !!tmp;
        } else {
                DRV_LOG(WARNING, "%s: unknown parameter", key);
                rte_errno = EINVAL;
@@ -1511,6 +1579,8 @@ mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
                MLX5_TXQ_MPW_HDR_DSEG_EN,
                MLX5_TXQ_MAX_INLINE_LEN,
                MLX5_TX_DB_NC,
+               MLX5_TX_PP,
+               MLX5_TX_SKEW,
                MLX5_TX_VEC_EN,
                MLX5_RX_VEC_EN,
                MLX5_L3_VXLAN_EN,
@@ -1518,6 +1588,7 @@ mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
                MLX5_DV_ESW_EN,
                MLX5_DV_FLOW_EN,
                MLX5_DV_XMETA_EN,
+               MLX5_LACP_BY_USER,
                MLX5_MR_EXT_MEMSEG_EN,
                MLX5_REPRESENTOR,
                MLX5_MAX_DUMP_FILES_NUM,
@@ -1525,6 +1596,7 @@ mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
                MLX5_CLASS_ARG_NAME,
                MLX5_HP_BUF_SIZE,
                MLX5_RECLAIM_MEM,
+               MLX5_SYS_MEM_EN,
                NULL,
        };
        struct rte_kvargs *kvlist;
@@ -1783,132 +1855,13 @@ mlx5_set_metadata_mask(struct rte_eth_dev *dev)
        DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
 }
 
-/**
- * Allocate page of door-bells and register it using DevX API.
- *
- * @param [in] dev
- *   Pointer to Ethernet device.
- *
- * @return
- *   Pointer to new page on success, NULL otherwise.
- */
-static struct mlx5_devx_dbr_page *
-mlx5_alloc_dbr_page(struct rte_eth_dev *dev)
-{
-       struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_devx_dbr_page *page;
-
-       /* Allocate space for door-bell page and management data. */
-       page = rte_calloc_socket(__func__, 1, sizeof(struct mlx5_devx_dbr_page),
-                                RTE_CACHE_LINE_SIZE, dev->device->numa_node);
-       if (!page) {
-               DRV_LOG(ERR, "port %u cannot allocate dbr page",
-                       dev->data->port_id);
-               return NULL;
-       }
-       /* Register allocated memory. */
-       page->umem = mlx5_glue->devx_umem_reg(priv->sh->ctx, page->dbrs,
-                                             MLX5_DBR_PAGE_SIZE, 0);
-       if (!page->umem) {
-               DRV_LOG(ERR, "port %u cannot umem reg dbr page",
-                       dev->data->port_id);
-               rte_free(page);
-               return NULL;
-       }
-       return page;
-}
-
-/**
- * Find the next available door-bell, allocate new page if needed.
- *
- * @param [in] dev
- *   Pointer to Ethernet device.
- * @param [out] dbr_page
- *   Door-bell page containing the page data.
- *
- * @return
- *   Door-bell address offset on success, a negative error value otherwise.
- */
-int64_t
-mlx5_get_dbr(struct rte_eth_dev *dev, struct mlx5_devx_dbr_page **dbr_page)
-{
-       struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_devx_dbr_page *page = NULL;
-       uint32_t i, j;
-
-       LIST_FOREACH(page, &priv->dbrpgs, next)
-               if (page->dbr_count < MLX5_DBR_PER_PAGE)
-                       break;
-       if (!page) { /* No page with free door-bell exists. */
-               page = mlx5_alloc_dbr_page(dev);
-               if (!page) /* Failed to allocate new page. */
-                       return (-1);
-               LIST_INSERT_HEAD(&priv->dbrpgs, page, next);
-       }
-       /* Loop to find bitmap part with clear bit. */
-       for (i = 0;
-            i < MLX5_DBR_BITMAP_SIZE && page->dbr_bitmap[i] == UINT64_MAX;
-            i++)
-               ; /* Empty. */
-       /* Find the first clear bit. */
-       MLX5_ASSERT(i < MLX5_DBR_BITMAP_SIZE);
-       j = rte_bsf64(~page->dbr_bitmap[i]);
-       page->dbr_bitmap[i] |= (UINT64_C(1) << j);
-       page->dbr_count++;
-       *dbr_page = page;
-       return (((i * 64) + j) * sizeof(uint64_t));
-}
-
-/**
- * Release a door-bell record.
- *
- * @param [in] dev
- *   Pointer to Ethernet device.
- * @param [in] umem_id
- *   UMEM ID of page containing the door-bell record to release.
- * @param [in] offset
- *   Offset of door-bell record in page.
- *
- * @return
- *   0 on success, a negative error value otherwise.
- */
-int32_t
-mlx5_release_dbr(struct rte_eth_dev *dev, uint32_t umem_id, uint64_t offset)
-{
-       struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_devx_dbr_page *page = NULL;
-       int ret = 0;
-
-       LIST_FOREACH(page, &priv->dbrpgs, next)
-               /* Find the page this address belongs to. */
-               if (mlx5_os_get_umem_id(page->umem) == umem_id)
-                       break;
-       if (!page)
-               return -EINVAL;
-       page->dbr_count--;
-       if (!page->dbr_count) {
-               /* Page not used, free it and remove from list. */
-               LIST_REMOVE(page, next);
-               if (page->umem)
-                       ret = -mlx5_glue->devx_umem_dereg(page->umem);
-               rte_free(page);
-       } else {
-               /* Mark in bitmap that this door-bell is not in use. */
-               offset /= MLX5_DBR_SIZE;
-               int i = offset / 64;
-               int j = offset % 64;
-
-               page->dbr_bitmap[i] &= ~(UINT64_C(1) << j);
-       }
-       return ret;
-}
-
 int
 rte_pmd_mlx5_get_dyn_flag_names(char *names[], unsigned int n)
 {
        static const char *const dynf_names[] = {
                RTE_PMD_MLX5_FINE_GRANULARITY_INLINE,
-               RTE_MBUF_DYNFLAG_METADATA_NAME
+               RTE_MBUF_DYNFLAG_METADATA_NAME,
+               RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME
        };
        unsigned int i;
 
@@ -2100,6 +2053,10 @@ static const struct rte_pci_id mlx5_pci_id_map[] = {
                RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
                                PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
        },
+       {
+               RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
+                               PCI_DEVICE_ID_MELLANOX_CONNECTX6LX)
+       },
        {
                .vendor_id = 0
        }
@@ -2114,20 +2071,17 @@ struct rte_pci_driver mlx5_driver = {
        .remove = mlx5_pci_remove,
        .dma_map = mlx5_dma_map,
        .dma_unmap = mlx5_dma_unmap,
-       .drv_flags = RTE_PCI_DRV_INTR_LSC | RTE_PCI_DRV_INTR_RMV |
-                    RTE_PCI_DRV_PROBE_AGAIN,
+       .drv_flags = PCI_DRV_FLAGS,
 };
 
+/* Initialize driver log type. */
+RTE_LOG_REGISTER(mlx5_logtype, pmd.net.mlx5, NOTICE)
+
 /**
  * Driver initialization routine.
  */
 RTE_INIT(rte_mlx5_pmd_init)
 {
-       /* Initialize driver log type. */
-       mlx5_logtype = rte_log_register("pmd.net.mlx5");
-       if (mlx5_logtype >= 0)
-               rte_log_set_level(mlx5_logtype, RTE_LOG_NOTICE);
-
        /* Build the static tables for Verbs conversion. */
        mlx5_set_ptype_table();
        mlx5_set_cksum_table();