On Wed, Feb 6, 2019 at 12:17 PM David Zafman <dzafman@xxxxxxxxxx> wrote: > > > Hey guys, > > Regarding http://tracker.ceph.com/issues/38034 > > A MBackfillReserve::RELEASE following by a MBackfillReserve::REQUEST > from the primary are processed out of order at the backfill target. > > This can be triggered by a preemption where the same PG immediately > attempts to backfill again. What ensures that backfill node processes > the messages from the primary in order? Is this a result of changes > in the messenger that were being tested? It's a lot more likely that there's something in the OSD that pauses the RELEASE but lets the REQUEST through than that the messenger is messing up the dispatch order of these two messages. I'd look at the processing steps for those and what conditions cause them to bail out and re-queue. -Greg
