On Thu, Nov 5, 2015 at 2:44 PM, Jeff Janes <jeff.janes@gmail.com> wrote:
> The bug theoretically exists in 9.5, but it wasn't until 9.6 (commit
> e95680832854cf300e64c) that free pages were recycled aggressively
> enough that it actually becomes likely to be hit.
In other words: The bug could be in 9.5, but that hasn't been
conclusively demonstrated. Fair?
I'm not an expert on GIN at all; I know far more about B-Tree. But,
commit e95680832854cf300e64c seems a bit odd to me. I don't see any
argument for why it's okay that the recycling of pages can happen
immediately for the pending list, rather than requiring it to happen
at some time later with a safe interlock (some like B-Tree's use of
RecentGlobalXmin). The GIN README discusses a few such issues, but it
wasn't updated by the commit I mentioned, which I would have expected.
OTOH, after all of 10 minutes I can't see what's special about
ginvacuumcleanup() that makes its long established
RecordFreeIndexPage() call fundamentally safer, which if true would be
a surprisingly obvious defect to go unfixed for all these years. This
is what you yourself said about it, I think. I need to look at it
again with fresh eyes, but offhand having no safe interlock for the
well established RecordFreeIndexPage() call for GIN seems like an
implausibly obvious bug.
--
Peter Geoghegan