Robert Haas <robertmhaas@gmail.com> writes:
> What if we change the hash table to have RelFileNode as the key and an
> array of MAX_FORKNUM bitmapsets as the value? Then when you get a
> "forget" request, you can just zap all the sets to empty.
Hm ... the only argument I can really make against that is that there'll
be no way to move such a table into shared memory; but there's probably
little hope of that anyway, given points made upthread. The bitmapset
manipulations are a bit tricky but solvable, and I agree there's
something to be said for not tying this stuff so closely to the
mechanism for relfilenode recycling.
regards, tom lane