Use upgradable RwLocks to prevent work duplication

[michaelsproul]

Currently we obtain a read lock while hashing to see if the hash is already computed. If it isn't then we drop the read lock, compute the hash and then re-obtain a write lock to store the computed hash. This leads to a race condition in which work can be duplicated if two threads try to read and compute the hash for a node at the same time.

I think using an upgradable RwLock allows us to prevent work being duplicated, using this algorithm:

1. Obtain an upgradable read lock
2. If the hash is already computed (!= 0), return it
3. Upgrade the lock to a write lock
4. Check again whether the hash is already computed, if so, return it
5. Compute the hash
6. Write the hash using the write lock, release

Step (4) is slightly counter-intuitive but is necessary to prevent this interleaving:

A1, A2, B1, B2, B3, B5, B6, A3, A5, A6

i.e. the case where thread B computes the hash while A is waiting for the write lock at step (3).