It is pointed that bio_rewind_iter() is one very bad API[1]:
1) bio size may not be restored after rewinding
2) it causes some bogus change, such as 
5151842b9d8732 (block: reset
bi_iter.bi_done after splitting bio)
3) rewinding really makes things complicated wrt. bio splitting
4) unnecessary updating of .bi_done in fast path
[1] https://marc.info/?t=
153549924200005&r=1&w=2
So this patch takes Kent's suggestion to restore one bio into its original
state via saving bio iterator(struct bvec_iter) in bio_integrity_prep(),
given now bio_rewind_iter() is only used by bio integrity code.
Cc: Dmitry Monakhov <dmonakhov@openvz.org>
Cc: Hannes Reinecke <hare@suse.com>
Suggested-by: Kent Overstreet <kent.overstreet@gmail.com>
Acked-by: Kent Overstreet <kent.overstreet@gmail.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
        if (bio_data_dir(bio) == WRITE) {
                bio_integrity_process(bio, &bio->bi_iter,
                                      bi->profile->generate_fn);
+       } else {
+               bip->bio_iter = bio->bi_iter;
        }
        return true;
 
                container_of(work, struct bio_integrity_payload, bip_work);
        struct bio *bio = bip->bip_bio;
        struct blk_integrity *bi = blk_get_integrity(bio->bi_disk);
-       struct bvec_iter iter = bio->bi_iter;
 
        /*
         * At the moment verify is called bio's iterator was advanced
         * during split and completion, we need to rewind iterator to
         * it's original position.
         */
-       if (bio_rewind_iter(bio, &iter, iter.bi_done)) {
-               bio->bi_status = bio_integrity_process(bio, &iter,
-                                                      bi->profile->verify_fn);
-       } else {
-               bio->bi_status = BLK_STS_IOERR;
-       }
-
+       bio->bi_status = bio_integrity_process(bio, &bip->bio_iter,
+                                               bi->profile->verify_fn);
        bio_integrity_free(bio);
        bio_endio(bio);
 }
 
                bio_integrity_trim(split);
 
        bio_advance(bio, split->bi_iter.bi_size);
-       bio->bi_iter.bi_done = 0;
 
        if (bio_flagged(bio, BIO_TRACE_COMPLETION))
                bio_set_flag(split, BIO_TRACE_COMPLETION);
 
 {
        iter->bi_sector += bytes >> 9;
 
-       if (bio_no_advance_iter(bio)) {
+       if (bio_no_advance_iter(bio))
                iter->bi_size -= bytes;
-               iter->bi_done += bytes;
-       } else {
+       else
                bvec_iter_advance(bio->bi_io_vec, iter, bytes);
                /* TODO: It is reasonable to complete bio with error here. */
-       }
-}
-
-static inline bool bio_rewind_iter(struct bio *bio, struct bvec_iter *iter,
-               unsigned int bytes)
-{
-       iter->bi_sector -= bytes >> 9;
-
-       if (bio_no_advance_iter(bio)) {
-               iter->bi_size += bytes;
-               iter->bi_done -= bytes;
-               return true;
-       }
-
-       return bvec_iter_rewind(bio->bi_io_vec, iter, bytes);
 }
 
 #define __bio_for_each_segment(bvl, bio, iter, start)                  \
        unsigned short          bip_max_vcnt;   /* integrity bio_vec slots */
        unsigned short          bip_flags;      /* control flags */
 
+       struct bvec_iter        bio_iter;       /* for rewinding parent bio */
+
        struct work_struct      bip_work;       /* I/O completion */
 
        struct bio_vec          *bip_vec;
 
 
        unsigned int            bi_idx;         /* current index into bvl_vec */
 
-       unsigned int            bi_done;        /* number of bytes completed */
-
        unsigned int            bi_bvec_done;   /* number of bytes completed in
                                                   current bvec */
 };
                bytes -= len;
                iter->bi_size -= len;
                iter->bi_bvec_done += len;
-               iter->bi_done += len;
 
                if (iter->bi_bvec_done == __bvec_iter_bvec(bv, *iter)->bv_len) {
                        iter->bi_bvec_done = 0;