RDMA/qedr: Fix QP state initialization race
authorKalderon, Michal <Michal.Kalderon@cavium.com>
Wed, 21 Mar 2018 12:51:52 +0000 (14:51 +0200)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Wed, 30 May 2018 05:52:18 +0000 (07:52 +0200)
[ Upstream commit caf61b1b8b88ccf1451f7321a176393797e8d292 ]

Once the FW is transitioned to error, FLUSH cqes can be received.
We want the driver to be aware of the fact that QP is already in error.

Without this fix, a user may see false error messages in the dmesg log,
mentioning that a FLUSH cqe was received while QP is not in error state.

Fixes: cecbcddf ("qedr: Add support for QP verbs")
Signed-off-by: Michal Kalderon <Michal.Kalderon@cavium.com>
Signed-off-by: Ariel Elior <Ariel.Elior@cavium.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Sasha Levin <alexander.levin@microsoft.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/infiniband/hw/qedr/verbs.c

index 8bbf64b..7f4cc93 100644 (file)
@@ -1663,14 +1663,15 @@ static void qedr_reset_qp_hwq_info(struct qedr_qp_hwq_info *qph)
 
 static int qedr_update_qp_state(struct qedr_dev *dev,
                                struct qedr_qp *qp,
+                               enum qed_roce_qp_state cur_state,
                                enum qed_roce_qp_state new_state)
 {
        int status = 0;
 
-       if (new_state == qp->state)
+       if (new_state == cur_state)
                return 0;
 
-       switch (qp->state) {
+       switch (cur_state) {
        case QED_ROCE_QP_STATE_RESET:
                switch (new_state) {
                case QED_ROCE_QP_STATE_INIT:
@@ -1774,6 +1775,7 @@ int qedr_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
        struct qedr_dev *dev = get_qedr_dev(&qp->dev->ibdev);
        const struct ib_global_route *grh = rdma_ah_read_grh(&attr->ah_attr);
        enum ib_qp_state old_qp_state, new_qp_state;
+       enum qed_roce_qp_state cur_state;
        int rc = 0;
 
        DP_DEBUG(dev, QEDR_MSG_QP,
@@ -1992,13 +1994,25 @@ int qedr_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
                qp->dest_qp_num = attr->dest_qp_num;
        }
 
+       cur_state = qp->state;
+
+       /* Update the QP state before the actual ramrod to prevent a race with
+        * fast path. Modifying the QP state to error will cause the device to
+        * flush the CQEs and while polling the flushed CQEs will considered as
+        * a potential issue if the QP isn't in error state.
+        */
+       if ((attr_mask & IB_QP_STATE) && qp->qp_type != IB_QPT_GSI &&
+           !udata && qp_params.new_state == QED_ROCE_QP_STATE_ERR)
+               qp->state = QED_ROCE_QP_STATE_ERR;
+
        if (qp->qp_type != IB_QPT_GSI)
                rc = dev->ops->rdma_modify_qp(dev->rdma_ctx,
                                              qp->qed_qp, &qp_params);
 
        if (attr_mask & IB_QP_STATE) {
                if ((qp->qp_type != IB_QPT_GSI) && (!udata))
-                       rc = qedr_update_qp_state(dev, qp, qp_params.new_state);
+                       rc = qedr_update_qp_state(dev, qp, cur_state,
+                                                 qp_params.new_state);
                qp->state = qp_params.new_state;
        }