[ofa-general] Re: having problems with the latest for-2.6.27 tree

[Roland Dreier]

It looks like an existing bug that now bites harder after the conversion
to mutex... on the normal exit path, iw_conn_req_handler never did
cma_enable_remove(conn_id) and now doesn't unlock the mutex.

So I think the fix is:

diff --git a/drivers/infiniband/core/cma.c b/drivers/infiniband/core/cma.c
index ead2b33..ae11d5c 100644
--- a/drivers/infiniband/core/cma.c
+++ b/drivers/infiniband/core/cma.c
@@ -1342,8 +1342,11 @@ static int iw_conn_req_handler(struct iw_cm_id *cm_id,
 		cma_exch(conn_id, CMA_DESTROYING);
 		mutex_unlock(&conn_id->handler_mutex);
 		rdma_destroy_id(&conn_id->id);
+		goto out;
 	}
 
+	mutex_unlock(&conn_id->handler_mutex);
+
 out:
 	if (dev)
 		dev_put(dev);

I've rolled this and the mutex_lock_nested() fixes into the patch in my
tree and pushed it out... will test rping shortly.

Would have been nice for the original author of a patch changing locking
to test with lockdep (IB needed the mutex_lock_nested treatment too),
but oh well.

 - R.