Unless I'm mistaken, here the "mask" call inside bracket already makes sure you don't receive asynchronous exceptions unless you call a function that is interruptible (i.e. goes back into the runtime system). The hClose example you give doesn't fall in this category, as something inside the RTS needs to call "allowInterrupt" (or otherwise unmask exceptions) in order for async exceptions to be delivered. The "readMVar" example you give *does* have this issue (because putMVar does an implicit allowInterrupt) but in recent GHC readMVar has been redefined as a primop.

The danger of deadlock is *not* minimal here, doing what you suggest will transform many valid programs (i.e. if you block on a "takeMVar" in the cleanup action) into ones that have unkillable orphan threads.

On Wed, Sep 3, 2014 at 1:56 PM, Eyal Lotem <eyal.lotem@gmail.com> wrote:

I'd like to propose a change in the behavior of Control.Exception to help guarantee cleanups are not accidentally lost.

For example, bracket is defined as:
bracket before after thing =
  mask $ \restore -> do
    a <- before
    r <- restore (thing a) `onException` after a
    _ <- after a
    return r
This definition has a serious problem: "after a" (in either the exception handling case, or the ordinary case) can include interruptible actions which abort the cleanup.

This means bracket does not in fact guarantee the cleanup occurs.

For example:

readMVar = bracket takeMVar putMVar -- If async exception occurs during putMVar, MVar is broken!

withFile .. = bracket (openFile ..) hClose -- Async exception during hClose leaks the file handle!

Interruptible actions during "before" are fine (as long as "before" handles them properly). Interruptible actions during "after" are virtually always a bug -- at best leaking a resource, and at worst breaking the program's invariants.

I propose changing all the cleanup handlers to run under uninterruptibleMask, specifically:

bracket, bracketOnError, bracket_, catch, catchJust, finally, handle, handleJust, onException

should all simply wrap their exception/cancellation handler with uninterruptibleMask.

The danger of a deadlock is minimal when compared with the virtually guaranteed buggy incorrect handling of async exceptions during cleanup.

--
Eyal
_______________________________________________
Libraries mailing list
Libraries@haskell.org
http://www.haskell.org/mailman/listinfo/libraries

--
Gregory Collins <greg@gregorycollins.net>