On Tue, Aug 19, 2025 at 09:49:14AM -0700, Keith Busch wrote: > From: Keith Busch <kbusch@xxxxxxxxxx> > > Previous version: > > https://lore.kernel.org/linux-block/20250805141123.332298-1-kbusch@xxxxxxxx/ > > This series removes the direct io requirement that io vector lengths > align to the logical block size. > > I tested this on a few raw block device types including nvme, > virtio-blk, ahci, and loop. NVMe is the only one I tested with 4k > logical sectors; everything else was 512. > > On each of those, I tested several iomap filesystems: xfs, ext4, and > btrfs. I found it interesting that each behave a little > differently with handling invalid vector alignments: > > - XFS is the most straight forward and reports failures on invalid > vector conditions, same as raw blocks devices. > > - EXT4 falls back to buffered io for writes but not for reads. > > - BTRFS doesn't even try direct io for any unusual alignments; it > chooses buffered io from the start. > > So it has been a little slow going figuring out which results to expect > from various tests, but I think I've got all the corner cases covered. I > can submit the tests cases to blktests and fstests for consideration > separately, too. > > I'm not 100% sure where we're at with the last patch. I think Mike > initially indicated this was okay to remove, but I could swear I read > something saying that might not be the case anymore. I just can't find > the message now. Mike? Hey, Yes, I don't have pointers immediately available but I did mention it and cc'd you. I have found that my work relative to NFS and NFSD does still need to use iov_iter_aligned_bvec -- otherwise misaligned DIO can get issued to the underlying filesystem. I did try to push all the relevant checking down to NFS/NFSD code that assembles their respective bvec into an iov_iter, like you suggested, but came up short after my first attempt. I don't want to speak for the NFS or NFSD miantainers, but I'm personally still OK with the broader iov_iter_is_aligned() interface and even iov_iter_aligned_bvec() going away (and NFS/NFSD carrying their own until I can circle back to hopefully eliminating the need). Either that, or we remove all but iov_iter_aligned_bvec() and export it so that NFS/NFSD can use it, _and_ tweak it so that it offers more coarse-grained length checking, like so: https://lore.kernel.org/linux-nfs/20250708160619.64800-5-snitzer@xxxxxxxxxx/ (this is probably the best intermediate solution actually, though it'd force my NFS and NFSD changes to be dependent on your series landing -- which is probably a perfectly appropriate constraint) Thanks, Mike > > Changes from v2: > > Include vector lengths when validating a split. The length check is > only valid for r/w commands, and skipped for passthrough > DRV_IN/DRV_OUT commands. > > Introduce a prep patch having bio_iov_iter_get_pages() take the > caller's desired length alignment. > > Additional code comments explaing less obvious error conditions. > > Added reviews on the patches that haven't changed. > > Keith Busch (8): > block: check for valid bio while splitting > block: add size alignment to bio_iov_iter_get_pages > block: align the bio after building it > block: simplify direct io validity check > iomap: simplify direct io validity check > block: remove bdev_iter_is_aligned > blk-integrity: use simpler alignment check > iov_iter: remove iov_iter_is_aligned > > block/bio-integrity.c | 4 +- > block/bio.c | 64 ++++++++++++++++++---------- > block/blk-map.c | 2 +- > block/blk-merge.c | 20 +++++++-- > block/fops.c | 13 +++--- > fs/iomap/direct-io.c | 6 +-- > include/linux/bio.h | 13 ++++-- > include/linux/blkdev.h | 20 +++++---- > include/linux/uio.h | 2 - > lib/iov_iter.c | 95 ------------------------------------------ > 10 files changed, 94 insertions(+), 145 deletions(-) > > -- > 2.47.3 >
