RDMA/mlx5: Support real-time timestamp directly from the device

Currently, if the user asks for a real-time timestamp, the device will
return a free-running one, and the timestamp will be translated to
real-time in the user-space.

When the device supports only real-time timestamp and not free-running,
the creation of the QP will fail even though the user needs supported the
real-time one. To prevent this, we will return the real-time timestamp
directly from the device.

Link: https://lore.kernel.org/r/c6cfc8e6f038575c5c2de6505830f7e74e4de80d.1623829775.git.leonro@nvidia.comSigned-off-by: Aharon Landau <aharonl@nvidia.com>
Reviewed-by: Maor Gottlieb <maorg@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>

drivers/infiniband/hw/mlx5/cq.c

@@ -725,7 +725,8 @@ static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata,
 		return -EFAULT;
 
 	if ((ucmd.flags & ~(MLX5_IB_CREATE_CQ_FLAGS_CQE_128B_PAD |
-			    MLX5_IB_CREATE_CQ_FLAGS_UAR_PAGE_INDEX)))
+			    MLX5_IB_CREATE_CQ_FLAGS_UAR_PAGE_INDEX |
+			    MLX5_IB_CREATE_CQ_FLAGS_REAL_TIME_TS)))
 		return -EINVAL;
 
 	if ((ucmd.cqe_size != 64 && ucmd.cqe_size != 128) ||
@@ -826,6 +827,9 @@ static int create_cq_user(struct mlx5_ib_dev *dev, struct ib_udata *udata,
 		cq->private_flags |= MLX5_IB_CQ_PR_FLAGS_CQE_128_PAD;
 	}
 
+	if (ucmd.flags & MLX5_IB_CREATE_CQ_FLAGS_REAL_TIME_TS)
+		cq->private_flags |= MLX5_IB_CQ_PR_FLAGS_REAL_TIME_TS;
+
 	MLX5_SET(create_cq_in, *cqb, uid, context->devx_uid);
 	return 0;
 

drivers/infiniband/hw/mlx5/main.c

@@ -1816,6 +1816,12 @@ static int set_ucontext_resp(struct ib_ucontext *uctx,
 	if (MLX5_CAP_GEN(dev->mdev, ece_support))
 		resp->comp_mask |= MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_ECE;
 
+	if (rt_supported(MLX5_CAP_GEN(dev->mdev, sq_ts_format)) &&
+	    rt_supported(MLX5_CAP_GEN(dev->mdev, rq_ts_format)) &&
+	    rt_supported(MLX5_CAP_ROCE(dev->mdev, qp_ts_format)))
+		resp->comp_mask |=
+			MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_REAL_TIME_TS;
+
 	resp->num_dyn_bfregs = bfregi->num_dyn_bfregs;
 	return 0;
 }

drivers/infiniband/hw/mlx5/mlx5_ib.h

@@ -550,6 +550,7 @@ static inline const struct mlx5_umr_wr *umr_wr(const struct ib_send_wr *wr)
 
 enum mlx5_ib_cq_pr_flags {
 	MLX5_IB_CQ_PR_FLAGS_CQE_128_PAD	= 1 << 0,
+	MLX5_IB_CQ_PR_FLAGS_REAL_TIME_TS = 1 << 1,
 };
 
 struct mlx5_ib_cq {
@@ -1614,4 +1615,10 @@ static inline bool mlx5_ib_lag_should_assign_affinity(struct mlx5_ib_dev *dev)
 		(MLX5_CAP_GEN(dev->mdev, num_lag_ports) > 1 &&
 		 MLX5_CAP_GEN(dev->mdev, lag_tx_port_affinity));
 }
+
+static inline bool rt_supported(int ts_cap)
+{
+	return ts_cap == MLX5_TIMESTAMP_FORMAT_CAP_REAL_TIME ||
+	       ts_cap == MLX5_TIMESTAMP_FORMAT_CAP_FREE_RUNNING_AND_REAL_TIME;
+}
 #endif /* MLX5_IB_H */

drivers/infiniband/hw/mlx5/qp.c

@@ -1180,8 +1180,16 @@ static bool fr_supported(int ts_cap)
 }
 
 static int get_ts_format(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *cq,
-			 bool fr_sup)
+			 bool fr_sup, bool rt_sup)
 {
+	if (cq->private_flags & MLX5_IB_CQ_PR_FLAGS_REAL_TIME_TS) {
+		if (!rt_sup) {
+			mlx5_ib_dbg(dev,
+				    "Real time TS format is not supported\n");
+			return -EOPNOTSUPP;
+		}
+		return MLX5_TIMESTAMP_FORMAT_REAL_TIME;
+	}
 	if (cq->create_flags & IB_UVERBS_CQ_FLAGS_TIMESTAMP_COMPLETION) {
 		if (!fr_sup) {
 			mlx5_ib_dbg(dev,
@@ -1198,14 +1206,16 @@ static int get_rq_ts_format(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *recv_cq)
 {
 	u8 ts_cap = MLX5_CAP_GEN(dev->mdev, rq_ts_format);
 
-	return get_ts_format(dev, recv_cq, fr_supported(ts_cap));
+	return get_ts_format(dev, recv_cq, fr_supported(ts_cap),
+			     rt_supported(ts_cap));
 }
 
 static int get_sq_ts_format(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *send_cq)
 {
 	u8 ts_cap = MLX5_CAP_GEN(dev->mdev, sq_ts_format);
 
-	return get_ts_format(dev, send_cq, fr_supported(ts_cap));
+	return get_ts_format(dev, send_cq, fr_supported(ts_cap),
+			     rt_supported(ts_cap));
 }
 
 static int get_qp_ts_format(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *send_cq,
@@ -1213,18 +1223,28 @@ static int get_qp_ts_format(struct mlx5_ib_dev *dev, struct mlx5_ib_cq *send_cq,
 {
 	u8 ts_cap = MLX5_CAP_ROCE(dev->mdev, qp_ts_format);
 	bool fr_sup = fr_supported(ts_cap);
+	bool rt_sup = rt_supported(ts_cap);
 	u8 default_ts = fr_sup ? MLX5_TIMESTAMP_FORMAT_FREE_RUNNING :
 				 MLX5_TIMESTAMP_FORMAT_DEFAULT;
 	int send_ts_format =
-		send_cq ? get_ts_format(dev, send_cq, fr_sup) :
+		send_cq ? get_ts_format(dev, send_cq, fr_sup, rt_sup) :
 			  default_ts;
 	int recv_ts_format =
-		recv_cq ? get_ts_format(dev, recv_cq, fr_sup) :
+		recv_cq ? get_ts_format(dev, recv_cq, fr_sup, rt_sup) :
 			  default_ts;
 
 	if (send_ts_format < 0 || recv_ts_format < 0)
 		return -EOPNOTSUPP;
 
+	if (send_ts_format != MLX5_TIMESTAMP_FORMAT_DEFAULT &&
+	    recv_ts_format != MLX5_TIMESTAMP_FORMAT_DEFAULT &&
+	    send_ts_format != recv_ts_format) {
+		mlx5_ib_dbg(
+			dev,
+			"The send ts_format does not match the receive ts_format\n");
+		return -EOPNOTSUPP;
+	}
+
 	return send_ts_format == default_ts ? recv_ts_format : send_ts_format;
 }
 

include/uapi/rdma/mlx5-abi.h

@@ -101,6 +101,7 @@ enum mlx5_ib_alloc_ucontext_resp_mask {
 	MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET = 1UL << 0,
 	MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_DUMP_FILL_MKEY    = 1UL << 1,
 	MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_ECE               = 1UL << 2,
+	MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_REAL_TIME_TS	   = 1UL << 4,
 };
 
 enum mlx5_user_cmds_supp_uhw {
@@ -270,6 +271,7 @@ struct mlx5_ib_query_device_resp {
 enum mlx5_ib_create_cq_flags {
 	MLX5_IB_CREATE_CQ_FLAGS_CQE_128B_PAD	= 1 << 0,
 	MLX5_IB_CREATE_CQ_FLAGS_UAR_PAGE_INDEX  = 1 << 1,
+	MLX5_IB_CREATE_CQ_FLAGS_REAL_TIME_TS	= 1 << 2,
 };
 
 struct mlx5_ib_create_cq {