deadlock detection

[Andrew Haley]

Boehm, Hans writes:

 > I think that if you want to detect deadlocks on the fly, you really
 > only have to check as you are about to lock a heavyweight lock.
 > And then it suffices to do a depth-first search of only the
 > heavyweight locks reachable from the one we're trying to acquire.

Oh, of course!  Thanks Hans.

That's obvious, but for some reason I didn't think of it -- you only
have to look at the part of the graph you changed.

 > 1) This is typically very cheap, since you only pay on persistent
 > 1) contention, and there probably aren't very many heavyweight
 > 1) locks.
 > 
 > 2) There are probably weird cases for which it's expensive.  It
 > 2) probably shouldn't be the default.
 > 
 > It seems to me that the result is still quite limited though, since
 > it only detects deadlocks involving exclusively locks, as opposed
 > to wait/notify or more obscure dependencies (IO, wait loops on
 > volatiles, etc.)  My guess is that this covers around half the half
 > the Java deadlocks I've seen, though my experience may be atypical.

That's a useful insight.

Andrew.