On Mon, 2005-02-28 at 07:32 -0800, 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. I don't think that I explained that correctly. The dm code clones the bio and sets the bi_idx to point to where the IO should start. Then when __blk_queue_bounce gets called, it only fills in its bio starting at bi_idx (it uses bio_for_each_segment) and since it calls bio_alloc instead of bio_clone, any pages it doesn't fill in are NULL. I suppose we could call bio_clone instead of bio_alloc in __blk_queue_bounce and fill in the whole bio. Mark. > > Mark, > > > > -- Mark Haverkamp <markh@xxxxxxxx>