On Nov 17, 2006, at 1:12 PM, Jeremy Geras wrote:
- The problem seems to be in TransactionMap::add(..) at the point where it does mMap.erase(i). The iterator appears to be invalid. I have no idea why this might be the case, but I noticed something kinda strange: in TransactionMap::erase(..) there’s a comment: “don't delete it here, the TransactionState deletes itself and removes itself from the map”. However in TransactionMap::add(..) it actually *does* delete the TransactionState (delete i->second). This is totally a guess, but seemed strange to me.
You are correct, this code is broken. If, for some reason, we try to add a TransactionState with the tid of an existing TransactionState, we end up invalidating our iterator. (We end up calling erase() once in ~TransactionState(), and again immediately after.) However, I cannot see how we could possibly end up in this code-path in the first place. We either need to fix this code-path, or put an assert in its place (I prefer this; this code-path is invalid in my opinion) Anyone have a different opinion?
Best regards,
Byron Campen