Are negative dentries the right abstraction in the first place?

I'd be somewhat concerned about potential impact on reading positive dentries, especially given that should be the common case. The cost of a bloom filter lookup should be small. but without looking at the code I assume the reading of dentries is already very optimised.

The other problem I see is how to delete a dentrie, as you can't remove something from a bloom filter without rebuilding it entirely. While the probabilistic nature of a bloom filter means you would need to handle that case anyway (its the maybe), a common use case for a negative dentrie would be for a small number of entries queried very frequently, which would lose the performance benefit if deleted. But perhaps these frequently queried files wouldn't have existed in the first place.