On timewait_exit cm event we expect to put the last refcount on conn and
release it. Since we already after the connection was closed we
could have tasks in tx_list that won't be processed.
We need to clean them or refcount won't reach 0.
Not cleaning the tasks will also leak any rdma_bufs associated with those
tasks.
Signed-off-by: Roi Dayan <roid@xxxxxxxxxxxx>
---
usr/iscsi/iser.c | 22 +++++++++++++++++++++-
1 file changed, 21 insertions(+), 1 deletion(-)
diff --git a/usr/iscsi/iser.c b/usr/iscsi/iser.c
index ace8da3..f262ed6 100644
--- a/usr/iscsi/iser.c
+++ b/usr/iscsi/iser.c
@@ -1644,7 +1644,27 @@ static void iser_cm_timewait_exit(struct rdma_cm_event *ev)
struct rdma_cm_id *cm_id = ev->id;
struct iser_conn *conn = cm_id->context;
- eprintf("conn:%p cm_id:%p\n", &conn->h, cm_id);
+ eprintf("conn:%p refcnt:%d cm_id:%p\n",
+ &conn->h, conn->h.refcount, cm_id);
+
+ if (conn->h.state == STATE_CLOSE) {
+ /*
+ * Tasks sitting in the conn->tx_list are stuck there after we
+ * close the conn and since each holds a reference on the conn
+ * we need to clean them up explicitly now. Otherwise they will
+ * prevent the conn from being cleaned up (since the refcount
+ * won't reach zero). If the conn doesn't get cleaned up, then
+ * along with leaking the conn itself (and all its resources),
+ * we'll also leak any rdma_buf's associated with the tasks.
+ * Since the rdma_buf pool is associated with the iser_device
+ * and so gets reused when a new iser_conn connection is
+ * established, leaking too many of those bufs will eventually
+ * clear the pool.
+ */
+ iser_ib_clear_tx_list(conn);
+ eprintf("conn:%p refcnt:%d cm_id:%p (after cleanup)\n",
+ &conn->h, conn->h.refcount, cm_id);
+ }
/* Refcount was incremented just before accepting the connection,
typically this is the last decrement and the connection will be
--
1.8.4.3
--
To unsubscribe from this list: send the line "unsubscribe stgt" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
