On 2013-04-04 12:59:36 -0700, Jeff Davis wrote: > Andres, > > Thank you for diagnosing this problem! > > On Thu, 2013-04-04 at 16:53 +0200, Andres Freund wrote: > > I think the route you quickly sketched is more realistic. That would > > remove all knowledge obout XLOG_HINT from generic code hich is a very > > good thing, I spent like 15minutes yesterday wondering whether the early > > return in there might be the cause of the bug... > > I like this approach. It may have some performance impact though, > because there are a couple extra spinlocks taken, and an extra memcpy.
I don't think its really slower. Earlier the code took WalInsertLock everytime, even if we ended up not logging anything. Thats far more epensive than a single spinlock. And the copy should also only be taken in the case we need to log. So I think we end up ahead of the current state. > The code looks good to me except that we should be consistent about the > page hole -- XLogCheckBuffer is calculating it, but then we copy the > entire page. I don't think anything can change the size of the page hole > while we have a shared lock on the buffer, so it seems OK to skip the > page hole during the copy. I don't think it can change either, but I doubt that there's a performance advantage by not copying the hole. I'd guess the simpler code ends up faster. > Another possible approach is to drop the lock on the buffer and > re-acquire it in exclusive mode after we find out we'll need to do > XLogInsert. It means that MarkBufferDirtyHint may end up blocking for > longer, but would avoid the memcpy. I haven't really thought through the > details. That sounds like it would be prone to deadlocks. So I would dislike to go there. Will write up a patch tomorrow. Greetings, Andres Freund -- Andres Freund http://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Training & Services -- Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org) To make changes to your subscription: http://www.postgresql.org/mailpref/pgsql-hackers