Messages in this thread | | | Date | Mon, 13 Jan 2014 20:48:41 -0800 | From | Kent Overstreet <> | Subject | Re: next bio iters break discard? |
| |
On Mon, Jan 13, 2014 at 11:06:33PM -0500, Martin K. Petersen wrote: > >>>>> "Kent" == Kent Overstreet <kmo@daterainc.com> writes: > > Kent, > > Kent> I think for discards we can deal with this easily enough - > Kent> __blk_recalc_rq_segments() will have to special case them - but > Kent> there's a similar (but worse) issue with WRITE_SAME, and looking > Kent> at the code it does attempt to merge WRITE_SAME requests too. > > DISCARD bios have no payload going down the stack. They get a payload > attached in the sd driver and will therefore have a single bvec at > completion time. > > WRITE_SAME bios have a single bvec payload throughout their lifetime. > > For both these types of requests we never attempt to merge the actual > payloads. But the block range worked on may shrink or grow as the bio is > split or merged going down the stack. > > IOW, DISCARD, WRITE SAME and the impending COPY requests do not have a > 1:1 mapping between the block range worked on and the size of any bvecs > attached. Your recent changes must have changed the way we handled that > in the past.
Yeah - but with WRITE_SAME bios, wouldn't we at least have to check that they're writing the same data to merge them?
| |