I bumped into a bug in the new buffering GiST build code. I did this:
create table gisttest (t text);
insert into gisttest select
a||'fooooooooooooooooooooooooooooooooooooooooooooooo' from
generate_series(1,10000000) a;
create index i_gisttest on gisttest using gist (t collate "C") WITH
(fillfactor=10);
After a while, this segfaults.
I debugged this, and traced this into a bug in the
gistRelocateBuildBuffersOnSplit() function. It splits a node buffer into
two (or more) node buffers, when the corresponding index page is split.
It first makes a copy of the old GISTNodeBuffer struct, and then
repurposes the struct to hold the new buffer for the new leftmost page
of the split. The temporary copy of the old page is only needed while
the function moves all the tuples from the old buffer into the new
buffers, after that it can be discarded. The temporary copy of the
struct is kept in the stack. However, the temporary copy can find its
way into the list of "loaded buffers". After the function exits, and
it's time to unload all the currently loaded buffers, you get a segfault
because the pointer now points to garbage. I think the reason this
doesn't always crash is that the copy in the stack usually still happens
to be valid enough that gistUnloadNodeBuffer() just skips it.
I'll commit the attached fix for that, marking the temporary copy
explicitly, so that we can avoid leaking it into any long-lived data
structures.
After fixing that, however, I'm now getting another error, much later in
the build process:
ERROR: failed to re-find parent for block 123002
STATEMENT: create index i_gisttest on gisttest using gist (t collate
"C") WITH (fillfactor=10);
I'll continue debugging that, but it seems to be another, unrelated, bug.
--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com
