IB/mlx5: Fix post send fence logic

If the caller specified IB_SEND_FENCE in the send flags of the work
request and no previous work request stated that the successive one
should be fenced, the work request would be executed without a fence.
This could result in RDMA read or atomic operations failure due to a MR
being invalidated. Fix this by adding the mlx5 enumeration for fencing
RDMA/atomic operations and fix the logic to apply this.

Fixes: e126ba97dba9 ('mlx5: Add driver for Mellanox Connect-IB adapters')
Signed-off-by: Eli Cohen <eli@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>

diff --git a/drivers/infiniband/hw/mlx5/qp.c b/drivers/infiniband/hw/mlx5/qp.c
index ce434228a5ead2a90a6375a78203a7591b5f5ae1..ce0a7ab35a227c569deae7f6b322b4c84eb5dcec 100644 (file)
--- a/drivers/infiniband/hw/mlx5/qp.c
+++ b/drivers/infiniband/hw/mlx5/qp.c
@@ -3332,10 +3332,11 @@ static u8 get_fence(u8 fence, struct ib_send_wr *wr)
                        return MLX5_FENCE_MODE_SMALL_AND_FENCE;
                else
                        return fence;
-
-       } else {
-               return 0;
+       } else if (unlikely(wr->send_flags & IB_SEND_FENCE)) {
+               return MLX5_FENCE_MODE_FENCE;
        }
+
+       return 0;
 }
 
 static int begin_wqe(struct mlx5_ib_qp *qp, void **seg,

diff --git a/include/linux/mlx5/qp.h b/include/linux/mlx5/qp.h
index e4e29882fdfd4f15e3663608094115df07bf421d..630f66a186b78b5ccb106610fbc8bf1beea0fc80 100644 (file)
--- a/include/linux/mlx5/qp.h
+++ b/include/linux/mlx5/qp.h
@@ -172,6 +172,7 @@ enum {
 enum {
        MLX5_FENCE_MODE_NONE                    = 0 << 5,
        MLX5_FENCE_MODE_INITIATOR_SMALL         = 1 << 5,
+       MLX5_FENCE_MODE_FENCE                   = 2 << 5,
        MLX5_FENCE_MODE_STRONG_ORDERING         = 3 << 5,
        MLX5_FENCE_MODE_SMALL_AND_FENCE         = 4 << 5,
 };