net/mlx5: fix loopback for Direct Verbs queue

In the past, all the queues and other hardware objects were created
through Verbs interface. Currently, most of the objects creation are
migrated to Devx interface by default, including queues. Only when
the DV is disabled by device arg or eswitch is enabled, all or some
of the objects are created through Verbs interface.

When using Devx interface to create queues, the kernel driver
behavior is different from the case using Verbs. The Tx loopback
cannot work properly even if the Tx and Rx queues are configured
with loopback attribute. To fix the support self loopback for Tx, a
Verbs dummy queue pair needs to be created to trigger the kernel to
enable the global loopback capability.

This is only required when TIR is created for Rx and loopback is
needed. Only CQ and QP are needed for this case, no WQ(RQ) needs to
be created.

Bugzilla ID: 645
Fixes: 6deb19e1b2d2 ("net/mlx5: separate Rx queue object creations")
Cc: stable@dpdk.org
Signed-off-by: Bing Zhao <bingz@nvidia.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo@nvidia.com>

diff --git a/drivers/net/mlx5/linux/mlx5_os.c b/drivers/net/mlx5/linux/mlx5_os.c
index ef7ccba5de82de1ee1589b7663c8c555a865a272..534a56a555831d6ba044a1aadf8bbb0b375184e5 100644 (file)
--- a/drivers/net/mlx5/linux/mlx5_os.c
+++ b/drivers/net/mlx5/linux/mlx5_os.c
@@ -1632,7 +1632,10 @@ err_secondary:
                priv->obj_ops.txq_obj_new = mlx5_os_txq_obj_new;
                priv->obj_ops.txq_obj_release = mlx5_os_txq_obj_release;
                mlx5_queue_counter_id_prepare(eth_dev);
-
+               priv->obj_ops.lb_dummy_queue_create =
+                                       mlx5_rxq_ibv_obj_dummy_lb_create;
+               priv->obj_ops.lb_dummy_queue_release =
+                                       mlx5_rxq_ibv_obj_dummy_lb_release;
        } else {
                priv->obj_ops = ibv_obj_ops;
        }

diff --git a/drivers/net/mlx5/linux/mlx5_verbs.c b/drivers/net/mlx5/linux/mlx5_verbs.c
index 0b0759f33fb228b4b98af2fd85094eb664d00c5d..d4fa202ac4be0b33ff6ffc437eafa0d843a75698 100644 (file)
--- a/drivers/net/mlx5/linux/mlx5_verbs.c
+++ b/drivers/net/mlx5/linux/mlx5_verbs.c
@@ -1055,6 +1055,125 @@ error:
        return -rte_errno;
 }
 
+/*
+ * Create the dummy QP with minimal resources for loopback.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ *
+ * @return
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+int
+mlx5_rxq_ibv_obj_dummy_lb_create(struct rte_eth_dev *dev)
+{
+#if defined(HAVE_IBV_DEVICE_TUNNEL_SUPPORT) && defined(HAVE_IBV_FLOW_DV_SUPPORT)
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_dev_ctx_shared *sh = priv->sh;
+       struct ibv_context *ctx = sh->ctx;
+       struct mlx5dv_qp_init_attr qp_init_attr = {0};
+       struct {
+               struct ibv_cq_init_attr_ex ibv;
+               struct mlx5dv_cq_init_attr mlx5;
+       } cq_attr = {{0}};
+
+       if (dev->data->dev_conf.lpbk_mode) {
+               /* Allow packet sent from NIC loop back w/o source MAC check. */
+               qp_init_attr.comp_mask |=
+                               MLX5DV_QP_INIT_ATTR_MASK_QP_CREATE_FLAGS;
+               qp_init_attr.create_flags |=
+                               MLX5DV_QP_CREATE_TIR_ALLOW_SELF_LOOPBACK_UC;
+       } else {
+               return 0;
+       }
+       /* Only need to check refcnt, 0 after "sh" is allocated. */
+       if (!!(__atomic_fetch_add(&sh->self_lb.refcnt, 1, __ATOMIC_RELAXED))) {
+               MLX5_ASSERT(sh->self_lb.ibv_cq && sh->self_lb.qp);
+               priv->lb_used = 1;
+               return 0;
+       }
+       cq_attr.ibv = (struct ibv_cq_init_attr_ex){
+               .cqe = 1,
+               .channel = NULL,
+               .comp_mask = 0,
+       };
+       cq_attr.mlx5 = (struct mlx5dv_cq_init_attr){
+               .comp_mask = 0,
+       };
+       /* Only CQ is needed, no WQ(RQ) is required in this case. */
+       sh->self_lb.ibv_cq = mlx5_glue->cq_ex_to_cq(mlx5_glue->dv_create_cq(ctx,
+                                                       &cq_attr.ibv,
+                                                       &cq_attr.mlx5));
+       if (!sh->self_lb.ibv_cq) {
+               DRV_LOG(ERR, "Port %u cannot allocate CQ for loopback.",
+                       dev->data->port_id);
+               rte_errno = errno;
+               goto error;
+       }
+       sh->self_lb.qp = mlx5_glue->dv_create_qp(ctx,
+                               &(struct ibv_qp_init_attr_ex){
+                                       .qp_type = IBV_QPT_RAW_PACKET,
+                                       .comp_mask = IBV_QP_INIT_ATTR_PD,
+                                       .pd = sh->pd,
+                                       .send_cq = sh->self_lb.ibv_cq,
+                                       .recv_cq = sh->self_lb.ibv_cq,
+                                       .cap.max_recv_wr = 1,
+                               },
+                               &qp_init_attr);
+       if (!sh->self_lb.qp) {
+               DRV_LOG(DEBUG, "Port %u cannot allocate QP for loopback.",
+                       dev->data->port_id);
+               rte_errno = errno;
+               goto error;
+       }
+       priv->lb_used = 1;
+       return 0;
+error:
+       if (sh->self_lb.ibv_cq) {
+               claim_zero(mlx5_glue->destroy_cq(sh->self_lb.ibv_cq));
+               sh->self_lb.ibv_cq = NULL;
+       }
+       (void)__atomic_sub_fetch(&sh->self_lb.refcnt, 1, __ATOMIC_RELAXED);
+       return -rte_errno;
+#else
+       RTE_SET_USED(dev);
+       return 0;
+#endif
+}
+
+/*
+ * Release the dummy queue resources for loopback.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ */
+void
+mlx5_rxq_ibv_obj_dummy_lb_release(struct rte_eth_dev *dev)
+{
+#if defined(HAVE_IBV_DEVICE_TUNNEL_SUPPORT) && defined(HAVE_IBV_FLOW_DV_SUPPORT)
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_dev_ctx_shared *sh = priv->sh;
+
+       if (!priv->lb_used)
+               return;
+       MLX5_ASSERT(__atomic_load_n(&sh->self_lb.refcnt, __ATOMIC_RELAXED));
+       if (!(__atomic_sub_fetch(&sh->self_lb.refcnt, 1, __ATOMIC_RELAXED))) {
+               if (sh->self_lb.qp) {
+                       claim_zero(mlx5_glue->destroy_qp(sh->self_lb.qp));
+                       sh->self_lb.qp = NULL;
+               }
+               if (sh->self_lb.ibv_cq) {
+                       claim_zero(mlx5_glue->destroy_cq(sh->self_lb.ibv_cq));
+                       sh->self_lb.ibv_cq = NULL;
+               }
+       }
+       priv->lb_used = 0;
+#else
+       RTE_SET_USED(dev);
+       return;
+#endif
+}
+
 /**
  * Release an Tx verbs queue object.
  *
@@ -1084,4 +1203,6 @@ struct mlx5_obj_ops ibv_obj_ops = {
        .txq_obj_new = mlx5_txq_ibv_obj_new,
        .txq_obj_modify = mlx5_ibv_modify_qp,
        .txq_obj_release = mlx5_txq_ibv_obj_release,
+       .lb_dummy_queue_create = NULL,
+       .lb_dummy_queue_release = NULL,
 };

diff --git a/drivers/net/mlx5/linux/mlx5_verbs.h b/drivers/net/mlx5/linux/mlx5_verbs.h
index 76a79bf4f474fa99c02ad30502ee3cdb06f87960..f7e8e2fe982fc34d6c841345bd19bce826b8256f 100644 (file)
--- a/drivers/net/mlx5/linux/mlx5_verbs.h
+++ b/drivers/net/mlx5/linux/mlx5_verbs.h
@@ -9,6 +9,8 @@
 
 int mlx5_txq_ibv_obj_new(struct rte_eth_dev *dev, uint16_t idx);
 void mlx5_txq_ibv_obj_release(struct mlx5_txq_obj *txq_obj);
+int mlx5_rxq_ibv_obj_dummy_lb_create(struct rte_eth_dev *dev);
+void mlx5_rxq_ibv_obj_dummy_lb_release(struct rte_eth_dev *dev);
 
 /* Verbs ops struct */
 extern const struct mlx5_mr_ops mlx5_mr_verbs_ops;

diff --git a/drivers/net/mlx5/mlx5.h b/drivers/net/mlx5/mlx5.h
index b8a29dd36935726d32f9300fa893d00728602095..32b2817bf2868fcd93c4403feefb0702aae6aa19 100644 (file)
--- a/drivers/net/mlx5/mlx5.h
+++ b/drivers/net/mlx5/mlx5.h
@@ -287,6 +287,13 @@ struct mlx5_drop {
        struct mlx5_rxq_obj *rxq; /* Rx queue object. */
 };
 
+/* Loopback dummy queue resources required due to Verbs API. */
+struct mlx5_lb_ctx {
+       struct ibv_qp *qp; /* QP object. */
+       void *ibv_cq; /* Completion queue. */
+       uint16_t refcnt; /* Reference count for representors. */
+};
+
 #define MLX5_COUNTERS_PER_POOL 512
 #define MLX5_MAX_PENDING_QUERIES 4
 #define MLX5_CNT_CONTAINER_RESIZE 64
@@ -1128,6 +1135,7 @@ struct mlx5_dev_ctx_shared {
        /* Meter management structure. */
        struct mlx5_aso_ct_pools_mng *ct_mng;
        /* Management data for ASO connection tracking. */
+       struct mlx5_lb_ctx self_lb; /* QP to enable self loopback for Devx. */
        struct mlx5_dev_shared_port port[]; /* per device port data array. */
 };
 
@@ -1287,6 +1295,8 @@ struct mlx5_obj_ops {
        int (*txq_obj_modify)(struct mlx5_txq_obj *obj,
                              enum mlx5_txq_modify_type type, uint8_t dev_port);
        void (*txq_obj_release)(struct mlx5_txq_obj *txq_obj);
+       int (*lb_dummy_queue_create)(struct rte_eth_dev *dev);
+       void (*lb_dummy_queue_release)(struct rte_eth_dev *dev);
 };
 
 #define MLX5_RSS_HASH_FIELDS_LEN RTE_DIM(mlx5_rss_hash_fields)
@@ -1316,6 +1326,7 @@ struct mlx5_priv {
        unsigned int sampler_en:1; /* Whether support sampler. */
        unsigned int mtr_en:1; /* Whether support meter. */
        unsigned int mtr_reg_share:1; /* Whether support meter REG_C share. */
+       unsigned int lb_used:1; /* Loopback queue is referred to. */
        uint16_t domain_id; /* Switch domain identifier. */
        uint16_t vport_id; /* Associated VF vport index (if any). */
        uint32_t vport_meta_tag; /* Used for vport index match ove VF LAG. */

diff --git a/drivers/net/mlx5/mlx5_devx.c b/drivers/net/mlx5/mlx5_devx.c
index 531a81d7fa578875421571ec4413303c755056ab..78b88f99b4bed315a679e5cd187d70ba8f734f8d 100644 (file)
--- a/drivers/net/mlx5/mlx5_devx.c
+++ b/drivers/net/mlx5/mlx5_devx.c
@@ -1188,4 +1188,6 @@ struct mlx5_obj_ops devx_obj_ops = {
        .txq_obj_new = mlx5_txq_devx_obj_new,
        .txq_obj_modify = mlx5_devx_modify_sq,
        .txq_obj_release = mlx5_txq_devx_obj_release,
+       .lb_dummy_queue_create = NULL,
+       .lb_dummy_queue_release = NULL,
 };

diff --git a/drivers/net/mlx5/mlx5_trigger.c b/drivers/net/mlx5/mlx5_trigger.c
index 879d3171e9c72c6b28f05dc9ed386aeb076405f6..ae7fcca229f84d1aab8e761494bc188f0ee7e7e4 100644 (file)
--- a/drivers/net/mlx5/mlx5_trigger.c
+++ b/drivers/net/mlx5/mlx5_trigger.c
@@ -1068,6 +1068,12 @@ mlx5_dev_start(struct rte_eth_dev *dev)
                        dev->data->port_id, strerror(rte_errno));
                goto error;
        }
+       if ((priv->config.devx && priv->config.dv_flow_en &&
+           priv->config.dest_tir) && priv->obj_ops.lb_dummy_queue_create) {
+               ret = priv->obj_ops.lb_dummy_queue_create(dev);
+               if (ret)
+                       goto error;
+       }
        ret = mlx5_txq_start(dev);
        if (ret) {
                DRV_LOG(ERR, "port %u Tx queue allocation failed: %s",
@@ -1148,6 +1154,8 @@ error:
        mlx5_traffic_disable(dev);
        mlx5_txq_stop(dev);
        mlx5_rxq_stop(dev);
+       if (priv->obj_ops.lb_dummy_queue_release)
+               priv->obj_ops.lb_dummy_queue_release(dev);
        mlx5_txpp_stop(dev); /* Stop last. */
        rte_errno = ret; /* Restore rte_errno. */
        return -rte_errno;
@@ -1186,6 +1194,8 @@ mlx5_dev_stop(struct rte_eth_dev *dev)
        priv->sh->port[priv->dev_port - 1].devx_ih_port_id = RTE_MAX_ETHPORTS;
        mlx5_txq_stop(dev);
        mlx5_rxq_stop(dev);
+       if (priv->obj_ops.lb_dummy_queue_release)
+               priv->obj_ops.lb_dummy_queue_release(dev);
        mlx5_txpp_stop(dev);
 
        return 0;