Generic semaphores

> It looks like a parent/child locking rule would be sufficient for most
> trees.

Only if you never need to hold more than two locks at a time.  In general that isn't good
enough.  (It certainly isn't good enough for the device tree!)  Furthermore lockdep has a
limit of 8 subclasses; thus you won't be able to acquire the locks for all the nodes along a
branch if the branch is too long.

In addition, trees can have other, more complicated, access patterns which lockdep can't even
come close to handling.  For instance, the rule about always acquiring locks going down a
branch can be generalized as follows:

    Whenever you hold a lock on a node A, you must not acquire
    a lock on node B unless you already hold the lock for the
    closest common ancestor of A and B.

This rule allows you to acquire locks down a branch, but it allows other patterns as well.  If
all threads follow the rule then deadlock can never occur (exercize!).  Clearly this is far
beyond lockdep's ability to express.