TimedLock with Stack Traces Strikes Back

code comments edit

By now you probably think I have an unhealthy obsession over the TimedLock struct. Well, you're right. I think it's emblematic of the right way to do things and shows that the right way isn't always the easiest way.

In Ian's last post on the TimedLock, he talked a bit about the performance considerations with my solution to keeping track of stack traces in a multi-threaded situation. To sum, my solution is pokey in certain situations. As always, measure measure measure.

However, Ian mentioned a solution outlined by Marek Malowidzki that avoids creating a stack trace on every lock acquisition. Instead, he only stores the stack trace when a lock timeout occurs, thus avoiding the performance penalty of my implementation. Unfortunately, there's no source code posted for examination.

So I decided to implement Marek's solution based on Ian's write up. As Ian mentioned, it would probably be best if the blocking thread didn't throw an exception, but logged diagnostic information if it detects that some other thread timed out while trying to acquire a lock on the target. I put a very helpful //TODO: right where that should happen since everyone has their own preferred logging framework.

As stated in the write-up, when thread fails to acquire a lock, it adds the object it was trying to lock to a Hashtable as a key with a NULL value. When the blocking thread is about to exit the synchronization block, it checks this hash table and if it finds the object it is locking in there, it will set the value for the Hashtable item as its own stack trace.

Thus if you catch the LockTimeoutException, you can have it try to obtain the stack trace from the Hashtable (supplying a wait since it might not be there immediately). However, there's one potential problem here. If you don't remove that object from the Hashtable after you've looked at it, the next time you lock on that object and then release it, your blocking thread will think an error occured and log some diagnostic information. This isn't too bad since it doesn't cause a LockTimeoutException to be thrown.

One thought I had was to check the Hashtable when I first successfully acquire the lock on a target and remove the target if its already there. However, that's not safe as it's quite possible that another thread failed after the blocking thread acquired the lock but before it examines the Hashtable.

Instead, when you call GetBlockingStackTrace, it retrieves the stack trace from the Hashtable, stores a reference locally, and then removes it from the Hashtable.

In any case, I've posted the source code in my TimedLock repository on GitHub.

DISCLAIMER: The code to keep track of stack traces is designed for debugging only and is turned on by setting the conditional compilation variable #DEBUG = true. I make no guarantees and pity the fool who deploys this version of TimedLock.cs in a production system with #DEBUG = true. Please let me know if you see any problems with this implementation. So far it passes my tiny suite of unit tests.

UPDATE: I accidentally linked to the old TimedLock. The link above has been updated.

UPDATE #2: Eric discovered a subtle bug in my older implementation of TimedLock which he describes here. Nice work in tracking that down!

Comments