On 05/24/12 09:39, Kent Overstreet wrote:

On Thu, May 24, 2012 at 09:19:12AM +0900, Jun'ichi Nomura wrote:

quoted

The destructor may also be called from blk_rq_unprep_clone(),
which just puts bio.
So this patch will introduce a memory leak.

Well, keeping around bi_destructor solely for that reason would be
pretty lousy. Can you come up with a better solution?

I don't have good one but here are some ideas:
  a) Do bio_endio() rather than bio_put() in blk_rq_unprep_clone()
     and let bi_end_io reap additional data.
     It looks ugly.
  b) Separate the constructor from blk_rq_prep_clone().
     dm has to do rq_for_each_bio loop again for constructor.
     Possible performance impact.
  c) Open code blk_rq_prep/unprep_clone() in dm.
     It exposes unnecessary block-internals to dm.
  d) Pass destructor function to blk_rq_prep/unprep_clone()
     for them to callback.

Umm, is "d)" better?
-- 
Jun'ichi Nomura, NEC Corporation