On Mon, Feb 28 2005, Mark Haverkamp wrote: > On Sat, 2005-02-26 at 13:39 +0100, Jens Axboe wrote: > > On Fri, Feb 25 2005, Linus Torvalds wrote: > > > > > > > > > On Fri, 25 Feb 2005, Andrew Morton wrote: > > > > > > > > It seems very weird for dm to be shoving NULL page*'s into the middle of a > > > > bio's bvec array, so your fix might end up being a workaround pending a > > > > closer look at what's going on in there. > > > > > > Yes. I don't see how this patch can be anything but bandaid to hide the > > > real bug. Where do these "non-page" bvec's originate? > > > > Yep that's the fishy part, there should not be NULL pages in the middle > > (or empty bios, for that matter) submitted for io. > > > > Mark, what was the bug that triggered you to write this patch? > > It happened when some pages of IO from a dm device were bounced. It > looks to me when bio's are cloned in the dm code to split it for > physical devices that only the pointers to pages that apply to that > device are copied and th bi_idx is adjusted to point to the start, > leaving some NULL pointers at the start of the bio_vec. This should fix it. Signed-off-by: Jens Axboe <axboe@xxxxxxx> ===== mm/highmem.c 1.55 vs edited ===== --- 1.55/mm/highmem.c 2005-01-08 06:44:13 +01:00 +++ edited/mm/highmem.c 2005-02-28 16:50:59 +01:00 @@ -425,7 +425,7 @@ * at least one page was bounced, fill in possible non-highmem * pages */ - bio_for_each_segment(from, *bio_orig, i) { + __bio_for_each_segment(from, *bio_orig, i) { to = bio_iovec_idx(bio, i); if (!to->bv_page) { to->bv_page = from->bv_page; -- Jens Axboe