(see 'Control.Exception.block').
The newly created thread has an exception handler that discards the
-exceptions 'BlockedOnDeadMVar', 'BlockedIndefinitely', and
+exceptions 'BlockedIndefinitelyOnMVar', 'BlockedIndefinitelyOnSTM', and
'ThreadKilled', and passes all other exceptions to the uncaught
exception handler (see 'setUncaughtExceptionHandler').
-}
Just StackOverflow -> reportStackOverflow
_ -> reportError se
-{- | 'killThread' terminates the given thread (GHC only).
-Any work already done by the thread isn\'t
-lost: the computation is suspended until required by another thread.
-The memory used by the thread will be garbage collected if it isn\'t
-referenced from anywhere. The 'killThread' function is defined in
-terms of 'throwTo':
+{- | 'killThread' raises the 'ThreadKilled' exception in the given
+thread (GHC only).
> killThread tid = throwTo tid ThreadKilled
-Killthread is a no-op if the target thread has already completed.
-}
killThread :: ThreadId -> IO ()
killThread tid = throwTo tid ThreadKilled
can kill each other, it is guaranteed that only one of the threads
will get to kill the other.
+Whatever work the target thread was doing when the exception was
+raised is not lost: the computation is suspended until required by
+another thread.
+
If the target thread is currently making a foreign call, then the
exception will not be raised (and hence 'throwTo' will not return)
until the call has completed. This is the case regardless of whether
a more synchronous design in which 'throwTo' does not return until the exception
is received by the target thread. The trade-off is discussed in Section 9 of the paper.
Like any blocking operation, 'throwTo' is therefore interruptible (see Section 5.3 of
-the paper).
+the paper). Unlike other interruptible operations, however, 'throwTo'
+is /always/ interruptible, even if it does not actually block.
+
+There is no guarantee that the exception will be delivered promptly,
+although the runtime will endeavour to ensure that arbitrary
+delays don't occur. In GHC, an exception can only be raised when a
+thread reaches a /safe point/, where a safe point is where memory
+allocation occurs. Some loops do not perform any memory allocation
+inside the loop and therefore cannot be interrupted by a 'throwTo'.
-There is currently no guarantee that the exception delivered by 'throwTo' will be
-delivered at the first possible opportunity. In particular, a thread may
-unblock and then re-block exceptions (using 'unblock' and 'block') without receiving
-a pending 'throwTo'. This is arguably undesirable behaviour.
+Blocked 'throwTo' is fair: if multiple threads are trying to throw an
+exception to the same target thread, they will succeed in FIFO order.
- -}
+ -}
throwTo :: Exception e => ThreadId -> e -> IO ()
throwTo (ThreadId tid) ex = IO $ \ s ->
case (killThread# tid (toException ex) s) of s1 -> (# s1, () #)
returnSTM :: a -> STM a
returnSTM x = STM (\s -> (# s, x #))
+instance MonadPlus STM where
+ mzero = retry
+ mplus = orElse
+
-- | Unsafely performs IO in the STM monad. Beware: this is a highly
-- dangerous thing to do.
--
prodServiceThread :: IO ()
prodServiceThread = do
- was_set <- readIORef prodding
- writeIORef prodding True
- -- no need for atomicModifyIORef, extra prods are harmless.
- if (not (was_set)) then wakeupIOManager else return ()
+ -- NB. use atomicModifyIORef here, otherwise there are race
+ -- conditions in which prodding is left at True but the server is
+ -- blocked in select().
+ was_set <- atomicModifyIORef prodding $ \b -> (True,b)
+ unless was_set wakeupIOManager
-- Machinery needed to ensure that we only have one copy of certain
-- CAFs in this module even when the base package is present twice, as
-> IO ()
service_loop wakeup readfds writefds ptimeval old_reqs old_delays = do
+ -- reset prodding before we look at the new requests. If a new
+ -- client arrives after this point they will send a wakup which will
+ -- cause the server to loop around again, so we can be sure to not
+ -- miss any requests.
+ --
+ -- NB. it's important to do this in the *first* iteration of
+ -- service_loop, rather than after calling select(), since a client
+ -- may have set prodding to True without sending a wakeup byte down
+ -- the pipe, because the pipe wasn't set up.
+ atomicModifyIORef prodding (\_ -> (False, ()))
+
-- pick up new IO requests
new_reqs <- atomicModifyIORef pendingEvents (\a -> ([],a))
let reqs = new_reqs ++ old_reqs
unless exit $ do
- atomicModifyIORef prodding (\_ -> (False, ()))
-
reqs' <- if wakeup_all then do wakeupAll reqs; return []
else completeRequests reqs readfds writefds []