Merge tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma

Pull rdma fixes from Jason Gunthorpe:
 "It has been fairly silent lately on our -rc front. Big queue of
  patches on the mailing list going to for-next though.

  Bug fixes:
   - qedr driver bugfixes causing application hangs, wrong uapi errnos,
     and a race condition
   - three syzkaller found bugfixes in the ucma uapi

  Regression fixes for things introduced in 4.16:
   - Crash on error introduced in mlx5 UMR flow
   - Crash on module unload/etc introduced by bad interaction of
     restrack and mlx5 patches this cycle
   - Typo in a two line syzkaller bugfix causing a bad regression
   - Coverity report of nonsense code in hns driver"

* tag 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rdma/rdma:
  RDMA/ucma: Introduce safer rdma_addr_size() variants
  RDMA/hns: ensure for-loop actually iterates and free's buffers
  RDMA/ucma: Check that device exists prior to accessing it
  RDMA/ucma: Check that device is connected prior to access it
  RDMA/rdma_cm: Fix use after free race with process_one_req
  RDMA/qedr: Fix QP state initialization race
  RDMA/qedr: Fix rc initialization on CNQ allocation failure
  RDMA/qedr: fix QP's ack timeout configuration
  RDMA/ucma: Correct option size check using optlen
  RDMA/restrack: Move restrack_clean to be symmetrical to restrack_init
  IB/mlx5: Don't clean uninitialized UMR resources

1 files changed, 32 insertions, 13 deletions

diff --git a/drivers/infiniband/hw/qedr/verbs.c b/drivers/infiniband/hw/qedr/verbs.c
index 875b17272d65..419a158e8fca 100644
--- a/drivers/infiniband/hw/qedr/verbs.c
+++ b/drivers/infiniband/hw/qedr/verbs.c
@@ -1841,14 +1841,15 @@ 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:
@@ -1955,6 +1956,7 @@ 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,
@@ -2086,18 +2088,23 @@ int qedr_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 		SET_FIELD(qp_params.modify_flags,
 			  QED_ROCE_MODIFY_QP_VALID_ACK_TIMEOUT, 1);
 
-		qp_params.ack_timeout = attr->timeout;
-		if (attr->timeout) {
-			u32 temp;
-
-			temp = 4096 * (1UL << attr->timeout) / 1000 / 1000;
-			/* FW requires [msec] */
-			qp_params.ack_timeout = temp;
-		} else {
-			/* Infinite */
+		/* The received timeout value is an exponent used like this:
+		 *    "12.7.34 LOCAL ACK TIMEOUT
+		 *    Value representing the transport (ACK) timeout for use by
+		 *    the remote, expressed as: 4.096 * 2^timeout [usec]"
+		 * The FW expects timeout in msec so we need to divide the usec
+		 * result by 1000. We'll approximate 1000~2^10, and 4.096 ~ 2^2,
+		 * so we get: 2^2 * 2^timeout / 2^10 = 2^(timeout - 8).
+		 * The value of zero means infinite so we use a 'max_t' to make
+		 * sure that sub 1 msec values will be configured as 1 msec.
+		 */
+		if (attr->timeout)
+			qp_params.ack_timeout =
+					1 << max_t(int, attr->timeout - 8, 0);
+		else
 			qp_params.ack_timeout = 0;
-		}
 	}
+
 	if (attr_mask & IB_QP_RETRY_CNT) {
 		SET_FIELD(qp_params.modify_flags,
 			  QED_ROCE_MODIFY_QP_VALID_RETRY_CNT, 1);
@@ -2170,13 +2177,25 @@ 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;
 	}