On Mon, Apr 24, 2023 at 02:43:56AM -0700, Christoph Hellwig wrote: > I'm pretty sure DIRECT I/O reads that write into file backed mappings > are out there in the wild. > > So while I wish we had never allowed this, the exercise seems futile and > instead we need to work on supporting this usecase, with the FOLL_PIN > infrastructure being a big step toward that. It's not entirely futile, there's at least one specific use case, which is io_uring which is currently open coding an equivalent check themselves. By introducing this change we prevent them from having to do so and provide a means by which other callers who implicitly need this do not have to do so either. In addition, this change frees up a blocked patch series intending to clean up GUP which should help open the door to further improvements across the system. So I would argue certainly not futile. In addition, I think it's useful to explicitly document that this is a broken case and, through use of the flag, highlight places which are problematic (although perhaps not exhaustively). I know Jason is keen on fixing this at a fundamental level and this flag is ultimately his suggestion, so it certainly doesn't stand in the way of this work moving forward.
