PG Bug reporting form <noreply@postgresql.org> writes:
> The following script:
> ...
> leads to a cassert-enabled server crash with the following messages in the
> log (for the master branch):
Hm, I can get a crash here without valgrind, as long as it's a cassert
build. We're accessing a list that's been thrown away by memory context
cleanup, so this results:
TRAP: FailedAssertion("IsOidList(list)", File: "list.c", Line: 678)
The timing is a bit finicky, but no more so than you report for the
valgrind case.
The difficulty is that the pendingReindexedIndexes list is kept in
some transaction-local context, so it gets flushed during the transaction
abort that is the first step of proc_exit processing. But the static
pointer to it is still set, causing big problems if we do any system
catalog accesses later --- like, say, while dropping the session's
temp tables.
One idea would be to keep the list in TopMemoryContext, but that feels
like a band-aid solution. I think more likely what we ought to do is
stop trying to use a PG_TRY in reindex_relation to drive cleanup,
and instead hook ResetReindexPending into transaction abort processing
honestly.
I wonder how many other uses of PG_TRY have similar issues? It's
not really obvious that this is an unsafe coding pattern.
regards, tom lane