The future of the page cache

True. This is not unlike the 'struct buffer_head' used when the fs blocksize is smaller than the page size. This is a little bit of work to do that, but it is no out-of-proportion with the sort of work already done for similar goals.

> We also need to maintain consistent metadata across all pages for a block (i.e. locking, flags, etc.).

I don't see that.
When a page is read, there is no need to also read other pages in the same filesystem block. So different pages in the one block do not all need to be up-to-date for example.
When writing to a block we would need to pre-read all pages which were not being over-written, but that is not intrinsically different to current code.

Certainly all the pages in a block need to be dirtied at the same time. Maybe there would be some awkwardness there...

Thanks.