Patch/feature proposal: Provide access to the runStmt sandbox ThreadID

Edsko de Vries edskodevries at
Wed Jun 5 15:01:09 CEST 2013

Hi all,

This proposal is related to,
which Simon Marlow closed through commit

The problem, in a nutshell, is "how do we terminate a code snippet started
with runStmt"? Before Simon's patch the only way was to disable ghc's
sandboxing, so that the snippet would run in the same thread as the thread
that called runStmt in the first place, and then send an asynchronous
exception to that thread. This is the approach we used to take. It's a
little tricky to get right (have to make sure that these exceptions are
thrown only at the right times), but we thought we had it working okay.

Until, that is, we realized we had a very nasty problem: snippets were
being unexpected interrupted. To debug this, we introduced a
CustomUserInterruption data type to serve as the exception that we were
throwing. This had two purposes: first, we would be use that if we saw a
CustomUserInterrupt that it could only have come from one particular throw,
and second, the CustomUserInterrupt carried an integer which was
incremented on every throw so that we never threw the same exception twice.

What we realized is that snippets were being interrupted by *old*
exceptions; that is, exceptions that we had thrown to *previous* snippets
(and had been caught too). This should obviously never happen. Ian managed
to reproduce this behaviour in a completely different setting ( and we think
that something similar (unsafePerformIO related) must be happening inside

Moreover, a further conjecture is that this only happens when runStmt is
still compiling the snippet to be executed (as opposed to the snippet
actually executing) -- i.e., that the exception gets stuck in the bowels of
ghc somewhere. We don't have any hard evidence for this, other than the
problem has not appeared again with the proposed patch (but that might be
luck, as it depends on timing).

The patch as we currently have it is against 7.4.2, so pre Simon's change
for the sandbox behaviour -- but I don't think that Simon's changes affect
the above problem. The core of our patch is

-sandboxIO :: DynFlags -> MVar Status -> IO [HValue] -> IO Status
-sandboxIO dflags statusMVar thing =
+sandboxIO :: DynFlags -> MVar Status -> MVar (Maybe ThreadId) -> IO
[HValue] -> IO Status
+sandboxIO dflags statusMVar tidMVar thing =
    mask $ \restore -> -- fork starts blocked
-     let runIt = liftM Complete $ try (restore $ rethrow dflags thing)
+     let thing' = gbracket (myThreadId >>= putMVar tidMVar . Just)
+                           (\() -> modifyMVar_ tidMVar (\_ -> return
+                           (\() -> thing)
+         runIt  = liftM Complete $ try (restore $ rethrow dflags thing')
      in if dopt Opt_GhciSandbox dflags
         then do tid <- forkIO $ do res <- runIt
                                    putMVar statusMVar res -- empty: can't

That is, sandboxIO takes an additional MVar (Maybe ThreadId):

   1. Initially this MVar should be empty. The MVar gets initialized to
   Just the Thread ID of the sandbox when the user code starts running. This
   means that if an secondary thread attempts to read the MVar (in order to
   kill the snippet), that secondary thread will block until the user code
   starts running -- it will not interrupt ghc when compiling the snippet.
   2. When the user code exists the MVar is updated to be Nothing. This
   means that if the auxiliary thread reads the MVar and finds Nothing it
   knows that the snippet has already terminated.
   3. When the auxiliary thread finds Just a thread ID (it must use
   withMVar rather than readMVar to avoid a race condition) it can safely
   throw the asynchronous exception.

The remainder of the patch just propagates these changes up so that runStmt
takes this MVar as an argument, too. Probably when integrating the patch
into ghc it would be better to leave runStmt along and provide a runStmt'
that takes the additional argument.

Again, any and all feedback on the above would be appreciated.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <>

More information about the ghc-devs mailing list