It's a pity that flush() is defined like this. Ideally, if mmap claims to mimick ordinary file objects, flush() should be a no-op() and there should be a separate sync() method.

On the other hand, your (Charles-François's) patch is already much better than the statu quo.
If nobody objects, I think it should be committed to 3.2. Whether or not we should be backport it to the stable branches is a bit more delicate, since it /could/ break badly written applications...

On a sidenote, the mmap object has received a *lot* less attention during the years than the other IO primitives (especially file objects in 3.x). It should probably only be used for specialized cases.