On Sun 28-08-22 21:59:49, John Hubbard wrote: > On 8/27/22 17:39, Al Viro wrote: > > On Sun, Aug 28, 2022 at 01:38:57AM +0100, Al Viro wrote: > >> On Sat, Aug 27, 2022 at 04:55:18PM -0700, John Hubbard wrote: > >>> On 8/27/22 15:48, Al Viro wrote: > >>>> On Sat, Aug 27, 2022 at 01:36:06AM -0700, John Hubbard wrote: > >>>>> Convert the NFS Direct IO layer to use pin_user_pages_fast() and > >>>>> unpin_user_page(), instead of get_user_pages_fast() and put_page(). > >>>> > >>>> Again, this stuff can be hit with ITER_BVEC iterators > >>>> > >>>>> - result = iov_iter_get_pages_alloc2(iter, &pagevec, > >>>>> + result = dio_w_iov_iter_pin_pages_alloc(iter, &pagevec, > >>>>> rsize, &pgbase); > >>>> > >>>> and this will break on those. > >>> > >>> If anyone has an example handy, of a user space program that leads > >>> to this situation (O_DIRECT with ITER_BVEC), it would really help > >>> me reach enlightenment a lot quicker in this area. :) > >> > >> Er... splice(2) to O_DIRECT-opened file on e.g. ext4? Or > >> sendfile(2) to the same, for that matter... > > > > s/ext4/nfs/ to hit this particular codepath, obviously. > > OK, I have a solution to this that's pretty easy: > > 1) Get rid of the user_backed_iter(i) check in > dio_w_iov_iter_pin_pages() and dio_w_iov_iter_pin_pages_alloc(), and > > 2) At the call sites, match up the unpin calls appropriately. > > ...and apply a similar fix for the fuse conversion patch. > > However, the core block/bio conversion in patch 4 still does depend upon > a key assumption, which I got from a 2019 email discussion with > Christoph Hellwig and others here [1], which says: > > "All pages released by bio_release_pages should come from > get_get_user_pages...". > > I really hope that still holds true. Otherwise this whole thing is in > trouble. > > [1] https://lore.kernel.org/kvm/20190724053053.GA18330@xxxxxxxxxxxxx/ Well as far as I've checked that discussion, Christoph was aware of pipe pages etc. (i.e., bvecs) entering direct IO code. But he had some patches [2] which enabled GUP to work for bvecs as well (using the kernel mapping under the hood AFAICT from a quick glance at the series). I suppose we could also handle this in __iov_iter_get_pages_alloc() by grabbing pin reference instead of plain get_page() for the case of bvec iter. That way we should have only pinned pages in bio_release_pages() even for the bvec case. [2] http://git.infradead.org/users/hch/misc.git/shortlog/refs/heads/gup-bvec Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
