Bug in GC's ordering of ForeignPtr finalization?

[Antoine Latter]

On Sun, Aug 28, 2011 at 4:27 PM, Ben Gamari <bgamari.foss at gmail.com> wrote:
> On Tue, 16 Aug 2011 12:32:13 -0400, Ben Gamari <bgamari.foss at gmail.com> wrote:
>> It seems that the notmuch-haskell bindings (version 0.2.2 built against
>> notmuch from git master; passes notmuch-test) aren't dealing with memory
>> management properly. In particular, the attached test code[1] causes
>> talloc to abort.  Unfortunately, while the issue is consistently
>> reproducible, it only occurs with some queries (see source[1]). I have
>> been unable to establish the exact criterion for failure.
>>
>> It seems that the crash is caused by an invalid access to a freed Query
>> object while freeing a Messages object (see Valgrind trace[3]). I've
>> taken a brief look at the bindings themselves but, being only minimally
>> familiar with the FFI, there's nothing obviously wrong (the finalizers
>> passed to newForeignPtr look sane). I was under the impression that
>> talloc was reference counted, so the Query object shouldn't have been
>> freed unless if there was still a Messages object holding a
>> reference. Any idea what might have gone wrong here?  Thanks!
>>
> After looking into this issue in a bit more depth, I'm even more
> confused. In fact, I would not be surprised if I have stumbled into a
> bug in the GC. It seems that the notmuch-haskell bindings follow the
> example of the python bindings in that child objects keep references to
> their parents to prevent the garbage collector from releasing the
> parent, which would in turn cause talloc to free the child objects,
> resulting in odd behavior when the child objects were next accessed. For
> instance, the Query and Messages objects are defined as follows,
>
>    type MessagesPtr = ForeignPtr S__notmuch_messages
>    type MessagePtr = ForeignPtr S__notmuch_message
>    newtype Query = Query (ForeignPtr S__notmuch_query)
>    data MessagesRef = QueryMessages { qmpp :: Query, msp :: MessagesPtr }
>                     | ThreadMessages { tmpp :: Thread, msp :: MessagesPtr }
>                     | MessageMessages { mmspp :: Message, msp :: MessagesPtr }
>    data Message = MessagesMessage { msmpp :: MessagesRef, mp :: MessagePtr }
>                 | Message { mp :: MessagePtr }
>    type Messages = [Message]
>

One problem you might be running in to is that the optimization passes
can notice that a function isn't using all of its arguments, and then
it won't pass them. These even applies if the arguments are bound
together in a record type.

So if you have a record type:

> data QueryResult = QR {qrQueryPtr :: ForeignPtr (), qrResultPointer :: Ptr ()}

and a function:

> processQueryResult :: QueryResult -> IO (...)

If the function doesn't use the 'qrQueryPointer' part of the record,
the compiler may not even pass it in. This might run the finalizer for
the foreign pointer earlier than you expect. If the result pointer is
a part of the query foreign pointer, you're in trouble.

I'm not sure if this is what's happening, but it sounds like it could be.

If this is the case you might want to build some helper functions
using the function 'touchForeignPtr', which does nothing other than
make it look like the foreign pointer is still in use. In my example
it might be something like:

> withQueryResultPtr :: QueryResult -> (Ptr QueryResult -> IO a) -> IO a
> withQueryResultPtr qr k = do
>    x <- k (qrQueryPtr qr)
>    touchForeignPtr (qrResultPointer qr)
>    return x

Antoine