Code review comment for lp://staging/~gary/bzr/bug835035

I would tend to say that we should just be more aggressive about unlocking if we're holding the lock and it fails. My big concern is that _check_fallback_repository needs to read data from the repository, and we'd like to hold the read lock so that anything we read stays cached.

However, if you can't see an easy way to do that, this patch would be ok.