[DP-103] Race condition in call

[qnikst]

[Imported from JIRA. Reported by Facundo Dominguez @facundominguez) as DP-103 on 2015-02-20 18:27:36]
By Facundo Dominguez at https://cloud-haskell.atlassian.net/browse/DP-103

The implementation of call [1] contains a text like the following:

-- We are guaranteed to receive the reply before the monitor notification
-- (if a reply is sent at all)

As discussed in https://cloud-haskell.atlassian.net/browse/DP-99 this is not true, and I have hit the case in practice where call fails with "call: remote process died: DiedNormal", because the monitor notification arrives before the response.

I don't know how to fix it yet. Maybe using a call to cpDelay after cpSend:

spawnMonitor nid (proc `bindCP` cpSend dict us `bindCP` cpDelay us (cpReturn ()))

and having call send () after unmonitoring.

[1]

distributed-process/src/Control/Distributed/Process/Internal/Spawn.hs

Lines 110 to 132 in 4d3472a

	call dict nid proc = do
	us <- getSelfPid
	(pid, mRef) <- spawnMonitor nid (proc `bindCP` cpSend dict us)
	-- We are guaranteed to receive the reply before the monitor notification
	-- (if a reply is sent at all)
	-- NOTE: This might not be true if we switch to unreliable delivery.
	mResult <- receiveWait
	[ match (return . Right)
	, matchIf (\(ProcessMonitorNotification ref _ _) -> ref == mRef)
	(\(ProcessMonitorNotification _ _ reason) -> return (Left reason))
	]
	case mResult of
	Right a -> do
	-- Wait for the monitor message so that we the mailbox doesn't grow
	receiveWait
	[ matchIf (\(ProcessMonitorNotification ref _ _) -> ref == mRef)
	(\(ProcessMonitorNotification {}) -> return ())
	]
	-- Clean up connection to pid
	reconnect pid
	return a
	Left err ->
	fail $ "call: remote process died: " ++ show err

Fixed by #183

[LaurentRDC]

Fixed by #183