When a Haskell function exported with foreign export is called from outside Haskell a new Task may be allocated for it. There will be one Task allocated for each OS thread calling into Haskell (this happens in rts_lock()). This Task is, by design, never deallocated. This works fine with a small number of threads, but if the external caller generates a lot of threads that die the Haskell RTS will just get more and more Task objects that are never freed. The Task also contains a condition variable and a mutex, which incurs a HANDLE leak on Windows.

I don't see an elegant fix to this, but the attached code remembers the OS thread for "incall" Task, and every so often (every 100 Tasks created) if will scan the Tasks looking for ones where the OS thread is dead, and remove these. The fix only works on Windows, I'm afraid.