RDMA/rxe: Recheck the MR in when generating a READ reply

The rping benchmark fails on long runs. The root cause of this failure has
been traced to a failure to compute a nonzero value of mr in rare
situations.

Fix this failure by correctly handling the computation of mr in
read_reply() in rxe_resp.c in the replay flow.

Fixes: 8a1a0be894da ("RDMA/rxe: Replace mr by rkey in responder resources")
Link: https://lore.kernel.org/r/20220418174103.3040-1-rpearsonhpe@gmail.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>

diff --git a/drivers/infiniband/sw/rxe/rxe_resp.c b/drivers/infiniband/sw/rxe/rxe_resp.c
index 1d95fab..9cd0eaf 100644 (file)
--- a/drivers/infiniband/sw/rxe/rxe_resp.c
+++ b/drivers/infiniband/sw/rxe/rxe_resp.c
@@ -745,8 +745,14 @@ static enum resp_states read_reply(struct rxe_qp *qp,
        }
 
        if (res->state == rdatm_res_state_new) {
-               mr = qp->resp.mr;
-               qp->resp.mr = NULL;
+               if (!res->replay) {
+                       mr = qp->resp.mr;
+                       qp->resp.mr = NULL;
+               } else {
+                       mr = rxe_recheck_mr(qp, res->read.rkey);
+                       if (!mr)
+                               return RESPST_ERR_RKEY_VIOLATION;
+               }
 
                if (res->read.resid <= mtu)
                        opcode = IB_OPCODE_RC_RDMA_READ_RESPONSE_ONLY;