On Sun, 24 Sep 2017 10:27:39 -0400,
Al Viro wrote:

BTW, there's something fishy in bio_copy_user_iov().  If the area we'd asked for
had been too large for a single bio, we are going to create a bio and have
 bio_add_pc_page() eventually fill it up to limit.  Then we return into
__blk_rq_map_user_iov(), advance iter and call bio_copy_user_iov() again.
Fine, but... now we might have non-zero iter->iov_offset.  And this
        bmd->is_our_pages = map_data ? 0 : 1;
        memcpy(bmd->iov, iter->iov, sizeof(struct iovec) * iter->nr_segs);
        iov_iter_init(&bmd->iter, iter->type, bmd->iov,
                        iter->nr_segs, iter->count);
does not even look at iter->iov_offset.  As the result, when it gets to
bio_uncopy_user(), we copy the data from each bio into the *beginning* of
the user area, overwriting that from the other bio.

Yeah, something is wrong with bio_copy_user_iov. Our datapath hangs when IO flows through unmodified SG (it forces bio_copy if iov_count is set). I did not look at details, but same IO pattern and memory layout work well with bio_map (module refcount problem).

Anyway, I'd added the obvious fix to #work.iov_iter, reordered it and
force-pushed the result.

I'll give it a try, thanks!
-- 
wbr, Vitaly