On 10/23/24 3:15 PM, Uday Shankar wrote: > From: Xinyu Zhang <xizhang@xxxxxxxxxxxxxxx> > > blk_rq_map_user_bvec contains a check bytes + bv->bv_len > nr_iter which > causes unnecessary failures in NVMe passthrough I/O, reproducible as > follows: > > - register a 2 page, page-aligned buffer against a ring > - use that buffer to do a 1 page io_uring NVMe passthrough read > > The second (i = 1) iteration of the loop in blk_rq_map_user_bvec will > then have nr_iter == 1 page, bytes == 1 page, bv->bv_len == 1 page, so > the check bytes + bv->bv_len > nr_iter will succeed, causing the I/O to > fail. This failure is unnecessary, as when the check succeeds, it means > we've checked the entire buffer that will be used by the request - i.e. > blk_rq_map_user_bvec should complete successfully. Therefore, terminate > the loop early and return successfully when the check bytes + bv->bv_len >> nr_iter succeeds. > > While we're at it, also remove the check that all segments in the bvec > are single-page. While this seems to be true for all users of the > function, it doesn't appear to be required anywhere downstream. Yep that looks like an issue. Since this would be nice to backport, please add a Fixes tag as well. -- Jens Axboe
