I discovered when researching the issue of index-only scans and Hot
Standby that there's a bug (for which I'm responsible) in
lazy_scan_heap[1]. Actually, the code has been like this for a long
time, but I needed to change it when I did the crash-safe visibility
map work, and I didn't. The problem is that lazy_scan_heap() releases
the lock on the buffer it's whacking around before it sets the
visibility map bit. Thus, it's possible for the page-level bit to be
cleared by some other backend before the visibility map bit gets set.
In previous releases that was arguably OK, since the worst thing that
could happen is a postponement of vacuuming on that page until the
next anti-wraparound cycle; but now that we have index-only scans, it
can cause queries to return wrong answers.
Attached is a patch to fix the problem, which rearranges things so
that we set the bit in the visibility map before releasing the buffer
lock. Similar work has already been done for inserts, updates, and
deletes, which in previous releases would *clear* the visibility map
bit after releasing the page lock, and no longer done. But the vacuum
case, which *sets* the bit, was overlooked. Our convention in those
related cases is that it's acceptable to hold the buffer lock while
setting the visibility map bit and issuing the WAL log record, but
there must be no possibility of an I/O to read in the visibility map
page while the buffer lock is held. This turned out to be pretty
simple because, as it turns out, lazy_scan_heap() is almost always
holding a pin on the correct page anyway, so I only had to tweak
things slightly to guarantee it. As a pleasant side effect, the new
code is actually quite a bit simpler and more readable than the old
code, at least IMHO.
While I was at it, I made a couple of minor, related changes which I
believe to be improvements. First, I arranged things so that, when we
pause the first vacuum pass to do an index vac cycle, we release any
buffer pin we're holding on the visibility map page. The fact that we
haven't done that in the past is probably harmless, but I think
there's something to be said for not holding onto buffer pins for long
periods of time when it isn't really necessary. Second, I adjusted
things so that we print a warning if the visibility map bit is set and
the page-level bit is clear, since that should never happen in 9.2+.
It is similar to the existing warning which catches the case where a
page is marked all-visible despite containing dead tuples.
Comments?
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
http://archives.postgresql.org/pgsql-hackers/2012-04/msg00682.php
