On Mon, Dec 5, 2011 at 4:09 PM, Marcelo Tosatti <mtosa...@redhat.com> wrote: > On Wed, Nov 23, 2011 at 11:47:53AM +0000, Stefan Hajnoczi wrote: > On earlier discussion, you replied to me: > > " >>> req = tracked_request_add(bs, sector_num, nb_sectors, false); >> >> The tracked request should include cluster round info? > > When checking A and B for overlap, only one of them needs to be > rounded in order for overlap detection to be correct. Therefore we > can store the original request [start, length) in tracked_requests and > only round the new request. > " > > The problem AFAICS is this: > > - Store a non-cluster-aligned request in the tracked request list. > - Wait on that non-cluster-aligned request > (wait_for_overlapping_requests). > - Submit cluster-aligned request for COR request. > > So, the tracked request list does not properly reflect the in-flight > COR requests. Which can result in: > > 1) guest reads sector 10. > 2) <sector_num=10,nb_sectors=2> added to tracked request list. > 3) COR code submits read for <sector_num=10,nb_sectors=2+cluster_align>
It will also round down to sector_num=0 when cluster size is 128 sectors (e.g. qcow2 and qed). > 4) unrelated guest operation writes to sector 13, nb_sectors=1. That is > allowed to proceed without waiting because tracked request list does not > reflect what COR in-flight requests. The tracked request list has <sector_num=10, nb_sectors=2> and the candidate write request is <sector_num=13, nb_sectors=1>. In wait_for_overlapping_requests() we round the candidate request to <sector_num=0, nb_sectors=cluster_size>. This rounded request does overlap <sector_num=10, sectors=2> so it will need to wait. Therefore CoR and write will not execute at the same time. Does this make more sense? Stefan
