mlx5: DR, Force QP drain on table creation

[ Upstream commit 6b8e9bc ]

When a table is created we first write the data to the STE anchor that
represents the table and then create/register the table using a FW
CREATE_FLOW_TABLE command. Since we write data using an RC QP a
theoretical race can happen between FW and SW, in case the FW will allow
to access the FT anchor ICM address prior to getting a CQE on the anchor
write. To resolve this a force drain will be done on the QP to assure
the anchor is written successfully.

Fixes: 94300b3 ("mlx5: Expose steering table functionality")
Signed-off-by: Alex Vesker <valex@nvidia.com>
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Signed-off-by: Nicolas Morey-Chaisemartin <nmoreychaisemartin@suse.com>

providers/mlx5/dr_table.c

@@ -199,6 +199,10 @@ struct mlx5dv_dr_table *mlx5dv_dr_table_create(struct mlx5dv_dr_domain *dmn,
 		if (ret)
 			goto free_tbl;
 
+		ret = dr_send_ring_force_drain(dmn);
+		if (ret)
+			goto uninit_tbl;
+
 		ret = dr_table_create_devx_tbl(tbl);
 		if (ret)
 			goto uninit_tbl;