On Tue, Jun 18, 2019 at 08:26:14PM +0300, Leon Romanovsky wrote:

+/**
+ * rdma_counter_bind_qp_auto - Check and bind the QP to a counter base on
+ *   the auto-mode rule
+ */
+int rdma_counter_bind_qp_auto(struct ib_qp *qp, u8 port)
+{
+	struct rdma_port_counter *port_counter;
+	struct ib_device *dev = qp->device;
+	struct rdma_counter *counter;
+	int ret;
+
+	if (!rdma_is_port_valid(dev, port))
+		return -EINVAL;
+
+	port_counter = &dev->port_data[port].port_counter;
+	if (port_counter->mode.mode != RDMA_COUNTER_MODE_AUTO)
+		return 0;
+
+	counter = rdma_get_counter_auto_mode(qp, port);
+	if (counter) {
+		ret = __rdma_counter_bind_qp(counter, qp);
+		if (ret) {
+			rdma_restrack_put(&counter->res);
+			return ret;
+		}
+		kref_get(&counter->kref);

The counter is left in the xarray while the kref is zero, this
kref_get is wrong..

Using two kref like things at the same time is a bad idea, the
'rdma_get_counter_auto_mode' should return the kref held, not the
restrack get. The restrack_del doesn't happen as long as the kref is
positive, so we don't need the retrack thing here..

+	} else {
+		counter = rdma_counter_alloc(dev, port, RDMA_COUNTER_MODE_AUTO);
+		if (!counter)
+			return -ENOMEM;
+
+		auto_mode_init_counter(counter, qp, port_counter->mode.mask);
+
+		ret = __rdma_counter_bind_qp(counter, qp);
+		if (ret)
+			goto err_bind;
+
+		rdma_counter_res_add(counter, qp);
+		if (!rdma_restrack_get(&counter->res)) {
+			ret = -EINVAL;
+			goto err_get;
+		}

and this shouldn't be needed as the kref is inited to 1 by the
rdma_counter_alloc..

+	}
+
+	return 0;
+
+err_get:
+	 __rdma_counter_unbind_qp(qp);
+	__rdma_counter_dealloc(counter);
+err_bind:
+	rdma_counter_free(counter);
+	return ret;
+}

And then all this error unwind and all the twisty __ functions should
just be a single kref_put and the release should handle everything.

Jason