One question, doesn't the kernel already have to support restricting writing out blocks that are larger than the page size "as a whole"? The hardware, block or the fs layer may require this for regular (non-atomic) operation anyway, depending on their respective block sizes, right? So I don't quite understand why writeback is fundamentally incompatible with atomic writes (for block sizes larger than page size - I'm assuming if the blocks are at most page size there is no problem anyway?).

As an idea, one could also consider an API that allows increasing these restrictions on a per-file level.