On Fri, Sep 23, 2022 at 05:13:42PM +0100, Al Viro wrote: > You are mixing two issues here - holding references to pages while using > iov_iter instance is obvious; holding them until async IO is complete, even > though struct iov_iter might be long gone by that point is a different > story. But someone needs to hold a refernce until the I/O is completed, because the I/O obviously needs the pages. Yes, we could say the callers holds them and can drop the references right after I/O submission, while the method needs to grab another reference. But that is more complicated and is more costly than just holding the damn reference. > And originating iov_iter instance really can be long-gone by the time > of IO completion - requirement to keep it around would be very hard to > satisfy. I've no objections to requiring the pages in ITER_BVEC to be > preserved at least until the IO completion by means independent of > whatever ->read_iter/->write_iter does to them, but > * that needs to be spelled out very clearly and > * we need to verify that it is, indeed, the case for all existing > iov_iter_bvec callers, preferably with comments next to non-obvious ones > (something that is followed only by the sync IO is obvious) Agreed. > That goes not just for bio - if we make get_pages *NOT* grab references > on ITER_BVEC (and I'm all for it), we need to make sure that those > pages won't be retained after the original protection runs out. Yes.
