Doing some raid testing threw a bug in the scsi mid layer, because the
segment counts wasn't correct.  Initially I worried that we still had
problems in this area, but it turns out that is due to the raid usage of
bio clones.  Currently you have to hold on to the original bio as well,
since the clone only maintains a pointer to the bio_vec inside the
original bio.  If the original bio is freed first, the clone will have
garbage in its bio->bi_io_vec as soon as that memory is scribbled.

I think the best fix is to maintain flexibility and duplicate the io_vec
inside the clone as well. Attached patch does this.
Signed-off-by: Jens Axboe <axboe@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>