My nbtree patch [1] needs to call index_getprocinfo() with an
exclusive buffer lock held during a leaf page split. This has an
undetectable self-deadlock (LWLock self-deadlock) risk: a syscache
lookup against pg_proc might have a catcache miss, ending with an
index scan that needs to access the very same buffer. That's not
acceptable.
There is very similar code to this in SP-GiST: there is a
index_getprocinfo() call within doPickSplit(), to get the user-defined
method for a split (SPGIST_PICKSPLIT_PROC). My nbtree patch builds a
new insertion scankey to determine how many attributes we can safely
truncate away in new pivot tuples -- it would be tricky to do this
lookup outside of the split function. I suppose that it's okay to do
this in SP-GiST without special care because there cannot be an
SP-GiST index on a system catalog. I'll need to do something else
about it given that I'm doing this within nbtree, though -- I don't
want to complicate the code that deals with insertion scankeys to make
this work.
Here is a strategy that fixes the problem without complicating matters
for nbtree: It should be safe if I make a point of using a special
comparator (the bitwise one that we already use in other contexts in
the patch) with system catalog indexes. We know that they cannot be of
types that have a varlena header + typstorage != 'p', which ensures
that there are no cases where bitwise equality fails to be a reliable
indicator of opclass equality (e.g. there are no cases like numeric
display scale). We could make sure that this requirement isn't
violated in the future by adding a pg_index test to opr_sanity.sql,
limiting system catalog indexes to opclasses that are known-safe for
the bitwise comparator.
Does that seem sane?
[1] https://commitfest.postgresql.org/21/1787/
--
Peter Geoghegan
