RDMA/hns: Fix incorrect post-send with direct wqe of wr-list

Currently, direct wqe is not supported for wr-list. RoCE driver excludes
direct wqe for wr-list by judging whether the number of wr is 1.

For a wr-list where the second wr is a length-error atomic wr, the
post-send driver handles the first wr and adds 1 to the wr number counter
firstly. While handling the second wr, the driver finds out a length error
and terminates the wr handle process, remaining the counter at 1. This
causes the driver mistakenly judges there is only 1 wr and thus enters
the direct wqe process, carrying the current length-error atomic wqe.

This patch fixes the error by adding a judgement whether the current wr
is a bad wr. If so, use the normal doorbell process but not direct wqe
despite the wr number is 1.

Fixes: 01584a5edcc4 ("RDMA/hns: Add support of direct wqe")
Signed-off-by: Junxian Huang <huangjunxian6@hisilicon.com>
Link: https://lore.kernel.org/r/20230804012711.808069-3-huangjunxian6@hisilicon.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>

diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
index 30451ce..97ff7f7 100644 (file)
--- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
+++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
@@ -750,7 +750,8 @@ out:
                qp->sq.head += nreq;
                qp->next_sge = sge_idx;
 
-               if (nreq == 1 && (qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE))
+               if (nreq == 1 && !ret &&
+                   (qp->en_flags & HNS_ROCE_QP_CAP_DIRECT_WQE))
                        write_dwqe(hr_dev, qp, wqe);
                else
                        update_sq_db(hr_dev, qp);