RDMA/rxe: Split qp state for requester and completer
authorBob Pearson <rpearsonhpe@gmail.com>
Wed, 20 Jul 2022 08:56:07 +0000 (04:56 -0400)
committerJason Gunthorpe <jgg@nvidia.com>
Tue, 2 Aug 2022 16:53:36 +0000 (13:53 -0300)
Currently the requester can continue to process send wqes after an local
qp operation error is detected because the setting of the qp state to the
error state is deferred until later. This patch splits the qp state for
the completer and requester into two separate states and sets
qp->req.state = QP_STATE_ERROR as soon as the error is detected before
another wqe can be executed.

Link: https://lore.kernel.org/r/1658307368-1851-4-git-send-email-lizhijian@fujitsu.com
Signed-off-by: Bob Pearson <rpearsonhpe@gmail.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
drivers/infiniband/sw/rxe/rxe_comp.c
drivers/infiniband/sw/rxe/rxe_qp.c
drivers/infiniband/sw/rxe/rxe_req.c
drivers/infiniband/sw/rxe/rxe_verbs.h

index bc53cad..fb0c008 100644 (file)
@@ -567,10 +567,10 @@ int rxe_completer(void *arg)
        if (!rxe_get(qp))
                return -EAGAIN;
 
-       if (!qp->valid || qp->req.state == QP_STATE_ERROR ||
-           qp->req.state == QP_STATE_RESET) {
+       if (!qp->valid || qp->comp.state == QP_STATE_ERROR ||
+           qp->comp.state == QP_STATE_RESET) {
                rxe_drain_resp_pkts(qp, qp->valid &&
-                                   qp->req.state == QP_STATE_ERROR);
+                                   qp->comp.state == QP_STATE_ERROR);
                goto exit;
        }
 
index eef91b8..c6519b9 100644 (file)
@@ -228,6 +228,7 @@ static int rxe_qp_init_req(struct rxe_dev *rxe, struct rxe_qp *qp,
                                               QUEUE_TYPE_FROM_CLIENT);
 
        qp->req.state           = QP_STATE_RESET;
+       qp->comp.state          = QP_STATE_RESET;
        qp->req.opcode          = -1;
        qp->comp.opcode         = -1;
 
@@ -488,6 +489,7 @@ static void rxe_qp_reset(struct rxe_qp *qp)
 
        /* move qp to the reset state */
        qp->req.state = QP_STATE_RESET;
+       qp->comp.state = QP_STATE_RESET;
        qp->resp.state = QP_STATE_RESET;
 
        /* let state machines reset themselves drain work and packet queues
@@ -551,6 +553,7 @@ void rxe_qp_error(struct rxe_qp *qp)
 {
        qp->req.state = QP_STATE_ERROR;
        qp->resp.state = QP_STATE_ERROR;
+       qp->comp.state = QP_STATE_ERROR;
        qp->attr.qp_state = IB_QPS_ERR;
 
        /* drain work and packet queues */
@@ -688,6 +691,7 @@ int rxe_qp_from_attr(struct rxe_qp *qp, struct ib_qp_attr *attr, int mask,
                        pr_debug("qp#%d state -> INIT\n", qp_num(qp));
                        qp->req.state = QP_STATE_INIT;
                        qp->resp.state = QP_STATE_INIT;
+                       qp->comp.state = QP_STATE_INIT;
                        break;
 
                case IB_QPS_RTR:
@@ -698,6 +702,7 @@ int rxe_qp_from_attr(struct rxe_qp *qp, struct ib_qp_attr *attr, int mask,
                case IB_QPS_RTS:
                        pr_debug("qp#%d state -> RTS\n", qp_num(qp));
                        qp->req.state = QP_STATE_READY;
+                       qp->comp.state = QP_STATE_READY;
                        break;
 
                case IB_QPS_SQD:
index f3bd071..f637712 100644 (file)
@@ -816,6 +816,7 @@ err:
        /* update wqe_index for each wqe completion */
        qp->req.wqe_index = queue_next_index(qp->sq.queue, qp->req.wqe_index);
        wqe->state = wqe_state_error;
+       qp->req.state = QP_STATE_ERROR;
        rxe_run_task(&qp->comp.task, 0);
 exit:
        ret = -EAGAIN;
index a24fbe9..96af3e0 100644 (file)
@@ -129,6 +129,7 @@ struct rxe_req_info {
 };
 
 struct rxe_comp_info {
+       enum rxe_qp_state       state;
        u32                     psn;
        int                     opcode;
        int                     timeout;