On Wed, Jan 06, 2016@04:21:17PM +0000, Keith Busch wrote:

On Wed, Jan 06, 2016@11:43:45PM +0800, Ming Lei wrote:

quoted

Please see the 1st line code of __blk_segment_map_sg(), in which only
one whole bvec is handled, and partial bvec can't be figured out there.

Think of it further, drivers often use bv.bv_len directly in the
iterator, for example:

        bio_for_each_segment(bvec, bio, iter)
                memcpy(page_address(bvec.bv_page) +
                                            bvec.bv_offset,  addr +
offset, bvec.bv_len);

So your patch will break these drivers, won't it?

CC'ing Kent in hopes he will clarify what happens on a split.

The bio_advance() code comments say it's handled:

"
 * This updates bi_sector, bi_size and bi_idx; if the number of bytes to
 * complete doesn't align with a bvec boundary, then bv_len and bv_offset will
 * be updated on the last bvec as well.
"

I admit I'm having a hard time seeing where bv_len and bv_offset updated
in this path. It was obviously handled after 054bdf646e then changed
with 4550dd6c6b.

If I follow correctly, 4550dd6c6b will implicity update the bvec's offset
and length during the split here since bio_iter_iovec resets the bvec's
length and offset:
---
#define __bio_for_each_segment(bvl, bio, iter, start)			\
	for (iter = (start);						\
	     (iter).bi_size &&						\
		((bvl = bio_iter_iovec((bio), (iter))), 1);		\
	     bio_advance_iter((bio), &(iter), (bvl).bv_len))
--

Yes, splitting in the middle of a bvec is perfectly fine. The reason
bio_for_each_segment takes a struct bvec and not a struct bvec * is because it's
computing what bv_len should be (taking the min of bv_len and bi_size, roughly).

See include/linux/bio.h:

bio_for_each_segment()
  bio_iter_iovec()
    bvec_iter_bvec()
      bvec_iter_len()

which does the actual bv_len computation.