Thread: Nicely exiting PG_TRY and PG_CATCH
Hi hackers,
I've found some odd lines in plpython-related code. These look to me like a potential source of problems, but maybe I'm not fully aware of some nuances.
Usually it's not a good idea to exit PG_TRY() block via return statement. Otherwise it would leave PG_exception_stack global variable in a wrong state and next ereport() will jump to some junk address. But here it is a straight return in plpy_exec.c:
PG_TRY();
{
args = PyList_New(proc->nargs);
if (!args)
return NULL;
{
args = PyList_New(proc->nargs);
if (!args)
return NULL;
...
Two more cases could be found further in the same file:
Isn't it a problem?
Another suspicious case is PG_CATCH block in jsonb_plpython.c:
PG_CATCH();
{
ereport(ERROR,
(errcode(ERRCODE_DATATYPE_MISMATCH),
errmsg("could not convert value \"%s\" to jsonb", str)));
}
{
ereport(ERROR,
(errcode(ERRCODE_DATATYPE_MISMATCH),
errmsg("could not convert value \"%s\" to jsonb", str)));
}
The problem is that leaving PG_CATCH() without PG_RE_THROW(), ReThrowError() or FlushErrorState() will consume an errordata stack slot, while the stack size is only 5. Do this five times and we'll have a PANIC on the next ereport().
As it's stated in elog.c (comment about the error stack depth at the beginning of FlushErrorState()), "The only case where it would be more than one deep is if we serviced an error that interrupted construction of another message."
Looks to me that FlushErrorState() should be inserted before this ereport. Shouldn't it?
--
best regards,
Mikhail A. Gribkov
Mikhail A. Gribkov
Mikhail Gribkov <youzhick@gmail.com> writes:
> Usually it's not a good idea to exit PG_TRY() block via return statement.
> Otherwise it would leave PG_exception_stack global variable in a wrong
> state and next ereport() will jump to some junk address.
Yeah, you can't return or goto out of the PG_TRY part.
> Another suspicious case is PG_CATCH block in jsonb_plpython.c:
This should be OK. The PG_CATCH and PG_FINALLY macros are set up so that
we've fully restored that state *before* we execute any of the
error-handling code. It would be basically impossible to have a guarantee
that CATCH blocks never throw errors; they'd be so restricted as to be
near useless, like signal handlers.
regards, tom lane
> Yeah, you can't return or goto out of the PG_TRY part.
So this is a problem if the check would ever work.
(Sorry for such a delayed answer.)
Then we need to fix it. Attached is a minimal patch, which changes nothing except for correct PG_TRY exiting.
Isn't it better this way?
CCing to Peter Eisentraut, whose patch originally introduced these returns.
Peter, will such a patch work somehow against your initial idea?
--
On Sat, Oct 8, 2022 at 12:19 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:
Mikhail Gribkov <youzhick@gmail.com> writes:
> Usually it's not a good idea to exit PG_TRY() block via return statement.
> Otherwise it would leave PG_exception_stack global variable in a wrong
> state and next ereport() will jump to some junk address.
Yeah, you can't return or goto out of the PG_TRY part.
> Another suspicious case is PG_CATCH block in jsonb_plpython.c:
This should be OK. The PG_CATCH and PG_FINALLY macros are set up so that
we've fully restored that state *before* we execute any of the
error-handling code. It would be basically impossible to have a guarantee
that CATCH blocks never throw errors; they'd be so restricted as to be
near useless, like signal handlers.
regards, tom lane