In commit 0fc94a5ba I wrote:
+ * ... It's okay to update the [ session-wide ]
+ * hash table with the new tree because all plpgsql functions within a
+ * given transaction share the same simple_eval_estate.
Um. Well, that's true for actual functions, but plpgsql DO blocks use
their own private simple_eval_estate. That means that after a DO block
runs, the cast_hash contains dangling pointers to expression eval state
trees, which a subsequent plpgsql execution in the same transaction will
think are still valid. Ooops. (See bug #13571.)
The simplest fix for this would be to give up on the idea that DO blocks
use private simple_eval_estates, and make them use the shared one.
However, that would result in intra-transaction memory bloat for
transactions executing large numbers of DO blocks; see commit c7b849a89,
which installed that arrangement to begin with. Since that change was
based on a user complaint, this answer doesn't seem appetizing.
Or we could try to use the shared simple_eval_estate for CAST expressions
even within DO blocks, but I'm afraid that would break things in subtle
ways. We need to do actual execution in the block's own eval_estate,
or we will have problems with leakage of pass-by-reference cast results
because exec_eval_cleanup() won't know to clean them up. It's possible
that we could get away with putting the expression state tree into the
shared simple_eval_estate's per-query memory and then executing it with
the block's private simple_eval_estate, but I'm afraid there are probably
places in execQual and/or C functions that suppose that the expression
state tree is in the estate's per-query memory. (That is, I doubt that
we're totally consistent about whether we use fcinfo->flinfo->fn_mcxt or
econtext->ecxt_per_query_memory for long-lived data, in which case an
arrangement like this could lead to dangling pointers.)
Or we could change things so that DO blocks use private cast_hash
hashtables along with their private simple_eval_estates. This would
give up some efficiency (since a DO block would then always need to do
its own cast lookups) but it would be a simple and reliable fix.
I'm kind of inclined to go with the last choice, but I wonder if anyone
wants to argue differently, or sees another feasible solution.
regards, tom lane
