Hi, for some concurrency abstractions I decided to use deadlock detection by means of exceptions which didn't work out as expected. Below is some simplifed source code showing the problem: My assumption was since GHC should be able to detect that the child still has a reference to mv_should_work -- although in an exception handler -- the RTS would never decide to kill the other thread.
Unluckily, this assumption turned out to be false. Did I overestimate GHC's capabilities or is this a bug?
It's not a bug, but it's quite tricky to explain. The effect you're seeing is this: GHC 5.04 is a bit more general in dealing with deadlocks, in that it doesn't automatically assume that the main thread is alive (GHC 5.02 did make that assumption). In your example, at the point of the deadlock *both* threads are blocked on unreachable MVars, so both threads get an exception. The RTS has no way to know that if it woke up the child thread first, then this would also unblock the main thread. I've cut-n-pasted the trace below (you can get this by compiling the RTS with -DDEBUG and running the program with +RTS -D1; arguably this is so useful that we ought to take it out from under -DDEBUG). Cheers, Simon DEBUG (-D1): scheduler scheduler: created thread 1, stack size = f4 words scheduler: == scheduler: waiting for thread (1) scheduler: == scheduler: waiting for thread (1) scheduler: == scheduler: new main thread (1) scheduler: all threads: thread 1 @ 0x500c0000 is not blocked scheduler: -->> Running TSO 1 (0x500c0000) ThreadEnterGHC ... scheduler: created thread 2, stack size = f4 words scheduler: --<< thread 1 (0x500c0000; ThreadRunGHC) stopped, yielding scheduler: all threads: thread 2 @ 0x500c0400 is not blocked thread 1 @ 0x500c0000 is not blocked scheduler: -->> Running TSO 2 (0x500c0400) ThreadEnterGHC ... scheduler: --<< thread 2 (0x500c0400) stopped: is blocked on an MVar scheduler: all threads: thread 2 @ 0x500c0400 is blocked on an MVar thread 1 @ 0x500c0000 is not blocked scheduler: -->> Running TSO 1 (0x500c0000) ThreadRunGHC ... scheduler: --<< thread 1 (0x500c0000; ThreadRunGHC) stopped, yielding scheduler: all threads: thread 2 @ 0x500c0400 is blocked on an MVar thread 1 @ 0x500c0000 is not blocked scheduler: -->> Running TSO 1 (0x500c0000) ThreadRunGHC ... scheduler: --<< thread 1 (0x500c0000; ThreadRunGHC) stopped, yielding scheduler: all threads: thread 2 @ 0x500c0400 is blocked on an MVar thread 1 @ 0x500c0000 is not blocked scheduler: -->> Running TSO 1 (0x500c0000) ThreadRunGHC ... scheduler: --<< thread 1 (0x500c0000) stopped: is blocked on an MVar scheduler: all threads: thread 2 @ 0x500c0400 is blocked on an MVar thread 1 @ 0x500c0000 is blocked on an MVar scheduler: deadlocked, forcing major GC... scheduler: resurrecting thread 1 scheduler: raising exception in thread 1. scheduler: raising exception in thread 1. scheduler: resurrecting thread 2 scheduler: raising exception in thread 2. scheduler: raising exception in thread 2. scheduler: -->> Running TSO 2 (0x500bf000) ThreadEnterGHC ... scheduler: --++ thread 2 (0x500bf000) finished scheduler: all threads: thread 2 @ 0x500bf000 has completed thread 1 @ 0x500bf400 is not blocked scheduler: -->> Running TSO 1 (0x500bf400) ThreadEnterGHC ... main caught: thread blocked indefinitely scheduler: --++ thread 1 (0x500bf400) finished scheduler: all threads: thread 2 @ 0x500bf000 has completed thread 1 @ 0x500bf400 has completed == scheduler: main thread (1) finished scheduler: created thread 3, stack size = f4 words scheduler: == scheduler: waiting for thread (3) scheduler: == scheduler: waiting for thread (3) scheduler: == scheduler: new main thread (3) scheduler: all threads: thread 3 @ 0x500bd028 is not blocked thread 2 @ 0x500bf000 has completed thread 1 @ 0x500bf400 has completed scheduler: -->> Running TSO 3 (0x500bd028) ThreadEnterGHC ... scheduler: --++ thread 3 (0x500bd028) finished scheduler: all threads: thread 3 @ 0x500bd028 has completed thread 2 @ 0x500bf000 has completed thread 1 @ 0x500bf400 has completed == scheduler: main thread (3) finished