RDMA/rxe: Fix qp reference counting for atomic ops
authorBob Pearson <rpearsonhpe@gmail.com>
Fri, 4 Jun 2021 23:05:59 +0000 (18:05 -0500)
committerJason Gunthorpe <jgg@nvidia.com>
Wed, 16 Jun 2021 23:20:23 +0000 (20:20 -0300)
Currently the rdma_rxe driver attempts to protect atomic responder
resources by taking a reference to the qp which is only freed when the
resource is recycled for a new read or atomic operation. This means that
in normal circumstances there is almost always an extra qp reference once
an atomic operation has been executed which prevents cleaning up the qp
and associated pd and cqs when the qp is destroyed.

This patch removes the call to rxe_add_ref() in send_atomic_ack() and the
call to rxe_drop_ref() in free_rd_atomic_resource(). If the qp is
destroyed while a peer is retrying an atomic op it will cause the
operation to fail which is acceptable.

Link: https://lore.kernel.org/r/20210604230558.4812-1-rpearsonhpe@gmail.com
Reported-by: Zhu Yanjun <zyjzyj2000@gmail.com>
Fixes: 86af61764151 ("IB/rxe: remove unnecessary skb_clone")
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
drivers/infiniband/sw/rxe/rxe_qp.c
drivers/infiniband/sw/rxe/rxe_resp.c

index a9256862464b85aee09e350dff10aae3684ae1c4..3dd9e9640b943e7f02963fc953dc241a2684b107 100644 (file)
@@ -136,7 +136,6 @@ static void free_rd_atomic_resources(struct rxe_qp *qp)
 void free_rd_atomic_resource(struct rxe_qp *qp, struct resp_res *res)
 {
        if (res->type == RXE_ATOMIC_MASK) {
-               rxe_drop_ref(qp);
                kfree_skb(res->atomic.skb);
        } else if (res->type == RXE_READ_MASK) {
                if (res->read.mr)
index 08f04222dd0d04e71f78d90dd8a5c3729250545f..9c0ce1a4f2eaa2e2b154ddf3b5d6cc26f9bd25df 100644 (file)
@@ -989,8 +989,6 @@ static int send_atomic_ack(struct rxe_qp *qp, struct rxe_pkt_info *pkt,
                goto out;
        }
 
-       rxe_add_ref(qp);
-
        res = &qp->resp.resources[qp->resp.res_head];
        free_rd_atomic_resource(qp, res);
        rxe_advance_resp_resource(qp);