May 25, 2008
I posted this as a page by accident – so here it is as a blog entry and I’ll delete the page some day.
My previous post discussed how inet_ntoa uses a static buffer which can cause a race condition. Unlike in ‘C’, this is particularly likely to cause a race in Haskell programs due to the quick, easy, and cheap threads using forkIO that (potentially) share a single OS thread. Two bright spots were that inet_ntoa was marked as IO and that the result is usually unimportant.
Another FFI binding, nano-md5, has a similar race condition but is much more series (not marked as IO and the result is a digest).
An even-handed note: iirc, nano-md5 remains hackage mostly as an FFI example – not that this is advertised in the nano-md5 description. “Real” users are told to look at hsOpenSSL and hopenssl – a cursory glance at the code suggests they don’t have this bug. Also, the other bindings don’t require O(n) space – so they are certainly worth switching to.
The nano-md5 line:
digest <- c_md5 ptr (fromIntegral n) nullPtr
is the culprit. It uses ‘nullPtr’ and according to the OpenSSL manual “If md is NULL, the digest is placed in a static array”.
Test code that confirms the bug can be found here – this will run three hash operations in parallel and eventually one result will be the correct first bits with ending bits from one of the other digests. The developer has already fixed the issue for versions 0.1.2+. I’ll wrap this post up with a request for library developers to please work to avoid use of static buffers – they have no place in this forkIO happy playland I call Haskell.