Hi,
Browsing LWN, I ran across this comment:
http://lwn.net/Articles/336039/
The author makes a bunch of unsubstantiated claims about STM, namely
that all implementations use locking under the hood, and that STM can
live- and deadlock. I've seen a lot of similar sentiments in other
places as well (comp.arch springs to mind).
Now, I'm no expert on STM, but I was pretty sure these are incorrect,
and I certainly had the impression that Haskell's STM guarantees some
progress - which it couldn't in a deadlock situation.
MVars can be simulated with STM, and MVars can semantically get in a deadlock situation, so STM can also deadlock.
Admittedly, if you're using STM to simulate MVars, you're doing it wrong.
Luke