net/mlx5: fix flow engine choice
[dpdk.git] / drivers / net / mlx5 / mlx5.c
index 9a2c711..8abd1b1 100644 (file)
 /* Device parameter to enable multi-packet send WQEs. */
 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
 
+/*
+ * Device parameter to force doorbell register mapping
+ * to non-cahed region eliminating the extra write memory barrier.
+ */
+#define MLX5_TX_DB_NC "tx_db_nc"
+
 /*
  * Device parameter to include 2 dsegs in the title WQEBB.
  * Deprecated, ignored.
@@ -184,6 +190,9 @@ static pthread_mutex_t mlx5_ibv_list_mutex = PTHREAD_MUTEX_INITIALIZER;
 #define MLX5_FLOW_MIN_ID_POOL_SIZE 512
 #define MLX5_ID_GENERATION_ARRAY_FACTOR 16
 
+#define MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE 4096
+#define MLX5_TAGS_HLIST_ARRAY_SIZE 8192
+
 /**
  * Allocate ID pool structure.
  *
@@ -421,6 +430,35 @@ mlx5_get_pdn(struct ibv_pd *pd __rte_unused, uint32_t *pdn __rte_unused)
 }
 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
 
+static int
+mlx5_config_doorbell_mapping_env(const struct mlx5_dev_config *config)
+{
+       char *env;
+       int value;
+
+       assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
+       /* Get environment variable to store. */
+       env = getenv(MLX5_SHUT_UP_BF);
+       value = env ? !!strcmp(env, "0") : MLX5_ARG_UNSET;
+       if (config->dbnc == MLX5_ARG_UNSET)
+               setenv(MLX5_SHUT_UP_BF, MLX5_SHUT_UP_BF_DEFAULT, 1);
+       else
+               setenv(MLX5_SHUT_UP_BF,
+                      config->dbnc == MLX5_TXDB_NCACHED ? "1" : "0", 1);
+       return value;
+}
+
+static void
+mlx5_restore_doorbell_mapping_env(int value)
+{
+       assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
+       /* Restore the original environment variable state. */
+       if (value == MLX5_ARG_UNSET)
+               unsetenv(MLX5_SHUT_UP_BF);
+       else
+               setenv(MLX5_SHUT_UP_BF, value ? "1" : "0", 1);
+}
+
 /**
  * Allocate shared IB device context. If there is multiport device the
  * master and representors will share this context, if there is single
@@ -434,22 +472,26 @@ mlx5_get_pdn(struct ibv_pd *pd __rte_unused, uint32_t *pdn __rte_unused)
  *
  * @param[in] spawn
  *   Pointer to the IB device attributes (name, port, etc).
+ * @param[in] config
+ *   Pointer to device configuration structure.
  *
  * @return
  *   Pointer to mlx5_ibv_shared object on success,
  *   otherwise NULL and rte_errno is set.
  */
 static struct mlx5_ibv_shared *
-mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn)
+mlx5_alloc_shared_ibctx(const struct mlx5_dev_spawn_data *spawn,
+                       const struct mlx5_dev_config *config)
 {
        struct mlx5_ibv_shared *sh;
+       int dbmap_env;
        int err = 0;
        uint32_t i;
 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
        struct mlx5_devx_tis_attr tis_attr = { 0 };
 #endif
 
-assert(spawn);
+       assert(spawn);
        /* Secondary process should not create the shared context. */
        assert(rte_eal_process_type() == RTE_PROC_PRIMARY);
        pthread_mutex_lock(&mlx5_ibv_list_mutex);
@@ -472,18 +514,32 @@ assert(spawn);
                rte_errno  = ENOMEM;
                goto exit;
        }
+       /*
+        * Configure environment variable "MLX5_BF_SHUT_UP"
+        * before the device creation. The rdma_core library
+        * checks the variable at device creation and
+        * stores the result internally.
+        */
+       dbmap_env = mlx5_config_doorbell_mapping_env(config);
        /* Try to open IB device with DV first, then usual Verbs. */
        errno = 0;
        sh->ctx = mlx5_glue->dv_open_device(spawn->ibv_dev);
        if (sh->ctx) {
                sh->devx = 1;
                DRV_LOG(DEBUG, "DevX is supported");
+               /* The device is created, no need for environment. */
+               mlx5_restore_doorbell_mapping_env(dbmap_env);
        } else {
+               /* The environment variable is still configured. */
                sh->ctx = mlx5_glue->open_device(spawn->ibv_dev);
-               if (!sh->ctx) {
-                       err = errno ? errno : ENODEV;
+               err = errno ? errno : ENODEV;
+               /*
+                * The environment variable is not needed anymore,
+                * all device creation attempts are completed.
+                */
+               mlx5_restore_doorbell_mapping_env(dbmap_env);
+               if (!sh->ctx)
                        goto error;
-               }
                DRV_LOG(DEBUG, "DevX is NOT supported");
        }
        err = mlx5_glue->query_device_ex(sh->ctx, NULL, &sh->device_attr);
@@ -659,6 +715,143 @@ exit:
        pthread_mutex_unlock(&mlx5_ibv_list_mutex);
 }
 
+/**
+ * Destroy table hash list and all the root entries per domain.
+ *
+ * @param[in] priv
+ *   Pointer to the private device data structure.
+ */
+static void
+mlx5_free_table_hash_list(struct mlx5_priv *priv)
+{
+       struct mlx5_ibv_shared *sh = priv->sh;
+       struct mlx5_flow_tbl_data_entry *tbl_data;
+       union mlx5_flow_tbl_key table_key = {
+               {
+                       .table_id = 0,
+                       .reserved = 0,
+                       .domain = 0,
+                       .direction = 0,
+               }
+       };
+       struct mlx5_hlist_entry *pos;
+
+       if (!sh->flow_tbls)
+               return;
+       pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
+       if (pos) {
+               tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
+                                       entry);
+               assert(tbl_data);
+               mlx5_hlist_remove(sh->flow_tbls, pos);
+               rte_free(tbl_data);
+       }
+       table_key.direction = 1;
+       pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
+       if (pos) {
+               tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
+                                       entry);
+               assert(tbl_data);
+               mlx5_hlist_remove(sh->flow_tbls, pos);
+               rte_free(tbl_data);
+       }
+       table_key.direction = 0;
+       table_key.domain = 1;
+       pos = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64);
+       if (pos) {
+               tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
+                                       entry);
+               assert(tbl_data);
+               mlx5_hlist_remove(sh->flow_tbls, pos);
+               rte_free(tbl_data);
+       }
+       mlx5_hlist_destroy(sh->flow_tbls, NULL, NULL);
+}
+
+/**
+ * Initialize flow table hash list and create the root tables entry
+ * for each domain.
+ *
+ * @param[in] priv
+ *   Pointer to the private device data structure.
+ *
+ * @return
+ *   Zero on success, positive error code otherwise.
+ */
+static int
+mlx5_alloc_table_hash_list(struct mlx5_priv *priv)
+{
+       struct mlx5_ibv_shared *sh = priv->sh;
+       char s[MLX5_HLIST_NAMESIZE];
+       int err = 0;
+
+       assert(sh);
+       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");
+               err = ENOMEM;
+               return err;
+       }
+#ifndef HAVE_MLX5DV_DR
+       /*
+        * In case we have not DR support, the zero tables should be created
+        * because DV expect to see them even if they cannot be created by
+        * RDMA-CORE.
+        */
+       union mlx5_flow_tbl_key table_key = {
+               {
+                       .table_id = 0,
+                       .reserved = 0,
+                       .domain = 0,
+                       .direction = 0,
+               }
+       };
+       struct mlx5_flow_tbl_data_entry *tbl_data = rte_zmalloc(NULL,
+                                                         sizeof(*tbl_data), 0);
+
+       if (!tbl_data) {
+               err = ENOMEM;
+               goto error;
+       }
+       tbl_data->entry.key = table_key.v64;
+       err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
+       if (err)
+               goto error;
+       rte_atomic32_init(&tbl_data->tbl.refcnt);
+       rte_atomic32_inc(&tbl_data->tbl.refcnt);
+       table_key.direction = 1;
+       tbl_data = rte_zmalloc(NULL, sizeof(*tbl_data), 0);
+       if (!tbl_data) {
+               err = ENOMEM;
+               goto error;
+       }
+       tbl_data->entry.key = table_key.v64;
+       err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
+       if (err)
+               goto error;
+       rte_atomic32_init(&tbl_data->tbl.refcnt);
+       rte_atomic32_inc(&tbl_data->tbl.refcnt);
+       table_key.direction = 0;
+       table_key.domain = 1;
+       tbl_data = rte_zmalloc(NULL, sizeof(*tbl_data), 0);
+       if (!tbl_data) {
+               err = ENOMEM;
+               goto error;
+       }
+       tbl_data->entry.key = table_key.v64;
+       err = mlx5_hlist_insert(sh->flow_tbls, &tbl_data->entry);
+       if (err)
+               goto error;
+       rte_atomic32_init(&tbl_data->tbl.refcnt);
+       rte_atomic32_inc(&tbl_data->tbl.refcnt);
+       return err;
+error:
+       mlx5_free_table_hash_list(priv);
+#endif /* HAVE_MLX5DV_DR */
+       return err;
+}
+
 /**
  * Initialize DR related data within private structure.
  * Routine checks the reference counter and does actual
@@ -673,12 +866,23 @@ exit:
 static int
 mlx5_alloc_shared_dr(struct mlx5_priv *priv)
 {
-#ifdef HAVE_MLX5DV_DR
        struct mlx5_ibv_shared *sh = priv->sh;
-       int err = 0;
+       char s[MLX5_HLIST_NAMESIZE];
+       int err = mlx5_alloc_table_hash_list(priv);
+
+       if (err)
+               return err;
+       /* Create tags hash list table. */
+       snprintf(s, sizeof(s), "%s_tags", sh->ibdev_name);
+       sh->tag_table = mlx5_hlist_create(s, MLX5_TAGS_HLIST_ARRAY_SIZE);
+       if (!sh->tag_table) {
+               DRV_LOG(ERR, "tags with hash creation failed.\n");
+               err = ENOMEM;
+               goto error;
+       }
+#ifdef HAVE_MLX5DV_DR
        void *domain;
 
-       assert(sh);
        if (sh->dv_refcnt) {
                /* Shared DV/DR structures is already initialized. */
                sh->dv_refcnt++;
@@ -717,12 +921,12 @@ mlx5_alloc_shared_dr(struct mlx5_priv *priv)
        }
 #endif
        sh->pop_vlan_action = mlx5_glue->dr_create_flow_action_pop_vlan();
+#endif /* HAVE_MLX5DV_DR */
        sh->dv_refcnt++;
        priv->dr_shared = 1;
        return 0;
-
 error:
-       /* Rollback the created objects. */
+       /* Rollback the created objects. */
        if (sh->rx_domain) {
                mlx5_glue->dr_destroy_domain(sh->rx_domain);
                sh->rx_domain = NULL;
@@ -743,11 +947,13 @@ error:
                mlx5_glue->destroy_flow_action(sh->pop_vlan_action);
                sh->pop_vlan_action = NULL;
        }
+       if (sh->tag_table) {
+               /* tags should be destroyed with flow before. */
+               mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
+               sh->tag_table = NULL;
+       }
+       mlx5_free_table_hash_list(priv);
        return err;
-#else
-       (void)priv;
-       return 0;
-#endif
 }
 
 /**
@@ -759,7 +965,6 @@ error:
 static void
 mlx5_free_shared_dr(struct mlx5_priv *priv)
 {
-#ifdef HAVE_MLX5DV_DR
        struct mlx5_ibv_shared *sh;
 
        if (!priv->dr_shared)
@@ -767,6 +972,7 @@ mlx5_free_shared_dr(struct mlx5_priv *priv)
        priv->dr_shared = 0;
        sh = priv->sh;
        assert(sh);
+#ifdef HAVE_MLX5DV_DR
        assert(sh->dv_refcnt);
        if (sh->dv_refcnt && --sh->dv_refcnt)
                return;
@@ -793,9 +999,13 @@ mlx5_free_shared_dr(struct mlx5_priv *priv)
                sh->pop_vlan_action = NULL;
        }
        pthread_mutex_destroy(&sh->dv_mutex);
-#else
-       (void)priv;
-#endif
+#endif /* HAVE_MLX5DV_DR */
+       if (sh->tag_table) {
+               /* tags should be destroyed with flow before. */
+               mlx5_hlist_destroy(sh->tag_table, NULL, NULL);
+               sh->tag_table = NULL;
+       }
+       mlx5_free_table_hash_list(priv);
 }
 
 /**
@@ -1300,6 +1510,16 @@ mlx5_args_check(const char *key, const char *val, void *opaque)
                DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
        } else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
                config->mps = !!tmp;
+       } else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
+               if (tmp != MLX5_TXDB_CACHED &&
+                   tmp != MLX5_TXDB_NCACHED &&
+                   tmp != MLX5_TXDB_HEURISTIC) {
+                       DRV_LOG(ERR, "invalid Tx doorbell "
+                                    "mapping parameter");
+                       rte_errno = EINVAL;
+                       return -rte_errno;
+               }
+               config->dbnc = tmp;
        } else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
                DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
        } else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
@@ -1322,8 +1542,8 @@ mlx5_args_check(const char *key, const char *val, void *opaque)
                if (tmp != MLX5_XMETA_MODE_LEGACY &&
                    tmp != MLX5_XMETA_MODE_META16 &&
                    tmp != MLX5_XMETA_MODE_META32) {
-                       DRV_LOG(WARNING, "invalid extensive "
-                                        "metadata parameter");
+                       DRV_LOG(ERR, "invalid extensive "
+                                    "metadata parameter");
                        rte_errno = EINVAL;
                        return -rte_errno;
                }
@@ -1373,6 +1593,7 @@ mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
                MLX5_TXQ_MPW_EN,
                MLX5_TXQ_MPW_HDR_DSEG_EN,
                MLX5_TXQ_MAX_INLINE_LEN,
+               MLX5_TX_DB_NC,
                MLX5_TX_VEC_EN,
                MLX5_RX_VEC_EN,
                MLX5_L3_VXLAN_EN,
@@ -1938,7 +2159,20 @@ mlx5_dev_spawn(struct rte_device *dpdk_dev,
                eth_dev->tx_pkt_burst = mlx5_select_tx_function(eth_dev);
                return eth_dev;
        }
-       sh = mlx5_alloc_shared_ibctx(spawn);
+       /*
+        * Some parameters ("tx_db_nc" in particularly) are needed in
+        * advance to create dv/verbs device context. We proceed the
+        * devargs here to get ones, and later proceed devargs again
+        * to override some hardware settings.
+        */
+       err = mlx5_args(&config, dpdk_dev->devargs);
+       if (err) {
+               err = rte_errno;
+               DRV_LOG(ERR, "failed to process device arguments: %s",
+                       strerror(rte_errno));
+               goto error;
+       }
+       sh = mlx5_alloc_shared_ibctx(spawn, &config);
        if (!sh)
                return NULL;
        config.devx = sh->devx;
@@ -2123,7 +2357,8 @@ mlx5_dev_spawn(struct rte_device *dpdk_dev,
                        err = ENOTSUP;
                        goto error;
                }
-       } else if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
+       }
+       if (devx_port.comp_mask & MLX5DV_DEVX_PORT_VPORT) {
                priv->vport_id = devx_port.vport_num;
        } else if (spawn->pf_bond >= 0) {
                DRV_LOG(ERR, "can't deduce vport index for port %d"
@@ -2180,13 +2415,8 @@ mlx5_dev_spawn(struct rte_device *dpdk_dev,
                }
                own_domain_id = 1;
        }
-       err = mlx5_args(&config, dpdk_dev->devargs);
-       if (err) {
-               err = rte_errno;
-               DRV_LOG(ERR, "failed to process device arguments: %s",
-                       strerror(rte_errno));
-               goto error;
-       }
+       /* Override some values set by hardware configuration. */
+       mlx5_args(&config, dpdk_dev->devargs);
        err = mlx5_dev_check_sibling_config(priv, &config);
        if (err)
                goto error;
@@ -2198,7 +2428,7 @@ mlx5_dev_spawn(struct rte_device *dpdk_dev,
        !defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
        DRV_LOG(DEBUG, "counters are not supported");
 #endif
-#ifndef HAVE_IBV_FLOW_DV_SUPPORT
+#if !defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_MLX5DV_DR)
        if (config.dv_flow_en) {
                DRV_LOG(WARNING, "DV flow is not supported");
                config.dv_flow_en = 0;
@@ -3031,6 +3261,7 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
        dev_config = (struct mlx5_dev_config){
                .hw_padding = 0,
                .mps = MLX5_ARG_UNSET,
+               .dbnc = MLX5_ARG_UNSET,
                .rx_vec_en = 1,
                .txq_inline_max = MLX5_ARG_UNSET,
                .txq_inline_min = MLX5_ARG_UNSET,
@@ -3045,6 +3276,7 @@ mlx5_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
                        .min_rxqs_num = MLX5_MPRQ_MIN_RXQS,
                },
                .dv_esw_en = 1,
+               .dv_flow_en = 1,
        };
        /* Device specific configuration. */
        switch (pci_dev->id.device_id) {