RDMA/mlx5: Return ECE DC support

The DC QPs are many-to-one QP types that means that first connection will establish ECE options that coming connections should follow. Due to this property, the ECE code was removed between first [1] and second [2] ECE submissions. This patch returns the dropped code, because ECE is a property of a connection and like any other connection users are needed to manage this data. Allow them to set ECE parameter for DC too and avoid need of having compatibility flag for the DC ECE. [1] https://lore.kernel.org/linux-rdma/20200523132243.817936-1-leon@kernel.org/ [2] https://lore.kernel.org/linux-rdma/20200525174401.71152-1-leon@kernel.org/

Leon Romanovsky committed 6 years ago

Commit a645a891 parent 92cd6672 files · +34 −16Browse files

Unified · Split

2 files

drivers/infiniband/hw/mlx5/qp.c +31 −14
include/linux/mlx5/mlx5_ifc.h +3 −2

drivers/infiniband/hw/mlx5/qp.c+31−14

 	destroy_qp(dev, qp, base, udata);
 }
-static int create_dct(struct ib_pd *pd, struct mlx5_ib_qp *qp,
+static int create_dct(struct mlx5_ib_dev *dev, struct ib_pd *pd,
+		      struct mlx5_ib_qp *qp,
 		      struct mlx5_create_qp_params *params)
 {
 	struct ib_qp_init_attr *attr = params->attr;
 	MLX5_SET(dctc, dctc, cqn, to_mcq(attr->recv_cq)->mcq.cqn);
 	MLX5_SET64(dctc, dctc, dc_access_key, ucmd->access_key);
 	MLX5_SET(dctc, dctc, user_index, uidx);
+	if (MLX5_CAP_GEN(dev->mdev, ece_support))
+		MLX5_SET(dctc, dctc, ece, ucmd->ece_options);
 	if (qp->flags_en & MLX5_QP_FLAG_SCATTER_CQE) {
 		int rcqe_sz = mlx5_ib_get_cqe_size(attr->recv_cq);
 	}
 	if (qp->type == MLX5_IB_QPT_DCT) {
-		err = create_dct(pd, qp, params);
+		err = create_dct(dev, pd, qp, params);
 		goto out;
 	}
 		 */
 		last = sizeof(struct mlx5_ib_create_qp_rss);
 	else
-		/* IB_QPT_RAW_PACKET and IB_QPT_DRIVER don't have ECE data */
+		/* IB_QPT_RAW_PACKET doesn't have ECE data */
 		switch (attr->qp_type) {
-		case IB_QPT_DRIVER:
 		case IB_QPT_RAW_PACKET:
 			last = offsetof(struct mlx5_ib_create_qp, ece_options);
 			break;
  * Other transitions and attributes are illegal
  */
 static int mlx5_ib_modify_dct(struct ib_qp *ibqp, struct ib_qp_attr *attr,
-			      int attr_mask, struct ib_udata *udata)
+			      int attr_mask, struct mlx5_ib_modify_qp *ucmd,
+			      struct ib_udata *udata)
 {
 	struct mlx5_ib_qp *qp = to_mqp(ibqp);
 	struct mlx5_ib_dev *dev = to_mdev(ibqp->device);
 	new_state = attr->qp_state;
 	dctc = MLX5_ADDR_OF(create_dct_in, qp->dct.in, dct_context_entry);
+	if (MLX5_CAP_GEN(dev->mdev, ece_support) && ucmd->ece_options)
+		/*
+		 * DCT doesn't initialize QP till modify command is executed,
+		 * so we need to overwrite previously set ECE field if user
+		 * provided any value except zero, which means not set/not
+		 * valid.
+		 */
+		MLX5_SET(dctc, dctc, ece, ucmd->ece_options);
 	if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
 		u16 set_id;
 		MLX5_SET(dctc, dctc, counter_set_id, set_id);
 	} else if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
 		struct mlx5_ib_modify_qp_resp resp = {};
-		u32 out[MLX5_ST_SZ_DW(create_dct_out)] = {0};
+		u32 out[MLX5_ST_SZ_DW(create_dct_out)] = {};
-		u32 min_resp_len = offsetof(typeof(resp), dctn) +
+		u32 min_resp_len = offsetofend(typeof(resp), dctn);
-				   sizeof(resp.dctn);
 		if (udata->outlen < min_resp_len)
 			return -EINVAL;
 		resp.response_length = min_resp_len;
+		/*
+		 * If we don't have enough space for the ECE options,
+		 * simply indicate it with resp.response_length.
+		 */
+		resp.response_length = (udata->outlen < sizeof(resp)) ?
+					       min_resp_len :
+					       sizeof(resp);
 		required |= IB_QP_MIN_RNR_TIMER | IB_QP_AV | IB_QP_PATH_MTU;
 		if (!is_valid_mask(attr_mask, required, 0))
 			return -EINVAL;
 		if (err)
 			return err;
 		resp.dctn = qp->dct.mdct.mqp.qpn;
+		if (MLX5_CAP_GEN(dev->mdev, ece_support))
+			resp.ece_options = MLX5_GET(create_dct_out, out, ece);
 		err = ib_copy_to_udata(udata, &resp, resp.response_length);
 		if (err) {
 			mlx5_core_destroy_dct(dev, &qp->dct.mdct);
 	qp_type = (unlikely(ibqp->qp_type == MLX5_IB_QPT_HW_GSI)) ? IB_QPT_GSI :
 								    qp->type;
-	if (qp_type == MLX5_IB_QPT_DCT) {
+	if (qp_type == MLX5_IB_QPT_DCT)
-		if (memchr_inv(&ucmd.ece_options, 0, sizeof(ucmd.ece_options)))
+		return mlx5_ib_modify_dct(ibqp, attr, attr_mask, &ucmd, udata);
-			return -EOPNOTSUPP;
-		return mlx5_ib_modify_dct(ibqp, attr, attr_mask, udata);
-	}
 	mutex_lock(&qp->mutex);

include/linux/mlx5/mlx5_ifc.h+3−2

 	u8         ecn[0x2];
 	u8         dscp[0x6];
-	u8         reserved_at_1c0[0x40];
+	u8         reserved_at_1c0[0x20];
+	u8         ece[0x20];
 };
 enum {
 	u8         reserved_at_40[0x8];
 	u8         dctn[0x18];
-	u8         reserved_at_60[0x20];
+	u8         ece[0x20];
 };
 struct mlx5_ifc_create_dct_in_bits {