On Fri, Apr 21, 2023 at 02:11:45AM +0100, Al Viro wrote: > On Thu, Apr 20, 2023 at 11:36:57PM +0100, Al Viro wrote: > > > Don't get me wrong - I'd love to kill ITER_PIPE off; it's just that tons > > of ->splice_read() duplicating the corresponding ->read_iter() up to the > > point where it would call generic_file_read_iter(), modulo ignoring > > O_DIRECT case and then calling filemap_splice_read() instead... Painful > > and asking for trouble down the road. > > FWIW, broken in this series: > * a bunch of 9p variants (those could probably switch to direct) > * afs > * ceph (very likely) > * ecryptfs (atime propagation) > * dax on various filesystems (probably) > * f2fs > * gfs2 > * nfs > * ocfs2 > * orangefs > * xfs > * zonefs (probably) > * splice from UDP sockets, unless I'm misreading something. > Your sock_splice_read() still falls back to generic_file_splice_read(), > rather than to direct_splice_read() and sockets don't have O_DIRECT > in flags. Neither do they have associated page cache ;-) > > Sure, we could provide a bunch of ->splice_read(), but then we'd have to make > sure that all subsequent changes to matching ->read_iter() get duplicated - > unless they are on O_DIRECT-only paths, that is... OK, how about the following modification of ITER_PIPE? * buffers might extend past the current position * buffers (as now) are either data or zero-copy ones. * advance past the end of data => allocate data-type buffers and fill with zeroes. * copying data to area prior to the end of buffers => each zero-copy-type buffer is converted to a data one; data is copied, but WTF else can you do? If there's a page borrowed from page cache, we can't change its contents anyway. * zero-copy of a page into an area overlapping the existing buffers => treat as data copy for the overlapping part and zero-copy the rest (if any) in normal fashion. * iov_iter_get_pages(): convert everything involved into data-type buffers first, then give caller references to that. Zero-fill new pages (if any). * iov_iter_revert() does *NOT* truncate anything * truncation of stuff past the current position is done by generic_file_splice_read() not only in error case, but on success as well - starting at the end of actually read data, of course. That would make for seekable ITER_PIPE, with zero-copy working for normal cases. I agree that O_DIRECT would probably be better off with ITER_BVEC. Comments?
