In HEAD, the following program sometimes deadlocks (about 1/10 of the time).

import Control.Concurrent
import Control.Monad
import GHC.Conc

main = do
  mainTid <- myThreadId
  labelThread mainTid "main"
  forM_ [0..0] $ \i -> forkIO $ do
    subTid <- myThreadId
    labelThread subTid $ "sub " ++ show i
    forM_ [0..100000000] $ \j -> putStrLn $ "sub " ++ show i ++ ": " ++ show j
  yield
  setNumCapabilities 2

The problem seems to be that there is a race condition between setNumCapabilites and a Task returning from a foreign call. Specifically, a sequence of events like the following can happen:

1. Task 0 makes a foreign call.
2. Task 1 calls setNumCapabilities()
3. The call on task 0 returns; it reads task->cap from the memory.
4. Task 1 moves the Capabilities, invalidating all pointers to them.
5. Task 0 takes over the invalidated Capability.

The attached patch adds an ASSERT to the RTS to demonstrate the problem (it does not fix the problem).