Re: Issue with cancel_before_shmem_exit while searching to remove a particular registered exit callbacks

On Tue, Jul 7, 2020 at 12:55 PM Andres Freund <andres(at)anarazel(dot)de> wrote:
> What are you proposing? For now we could easily enough work around this
> by just making it a on_proc_exit() callback, but that doesn't really
> change the fundamental issue imo.

I think it would be more correct for it to be an on_proc_exit()
callback, because before_shmem_exit() callbacks can and do perform
actions which rely on an awful lot of the system being still in a
working state. RemoveTempRelationsCallback() is a good example: it
thinks it can start and end transactions and make a bunch of catalog
changes. I don't know that any of that could use JIT, but moving the
JIT cleanup to the on_shmem_exit() stage seems better. At that point,
there shouldn't be anybody doing anything that relies on being able to
perform logical changes to the database; we're just shutting down
low-level subsystems at that point, and thus presumably not doing
anything that could possibly need JIT.

But I also agree that what pg_start_backup() was doing before v13 was
wrong; that's why I committed
303640199d0436c5e7acdf50b837a027b5726594. The only reason I didn't
back-patch it is because the consequences are so minor I didn't think
it was worth worrying about. We could, though. I'd be somewhat
inclined to both do that and also change LLVM to use on_proc_exit() in
master, but I don't feel super-strongly about it.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company