Peter Eisentraut <peter.eisentr...@2ndquadrant.com> writes:
> On 2019-11-04 16:01, Tom Lane wrote:
>> Now that I've actually looked at the patched code, there's a far
>> more severe problem with it. Namely, that use of PG_FINALLY
>> means that the "finally" segment is run without having popped
>> the error context stack, which means that any error thrown
>> within that segment will sigsetjmp right back to the top,
>> resulting in an infinite loop. (Well, not infinite, because
>> it'll crash out once the error nesting depth limit is hit.)
>> We *must* pop the stack before running recovery code.
> I can confirm that that indeed happens. :(
> Here is a patch to fix it.
This seems all right from here. Since PG_RE_THROW() is guaranteed
noreturn, I personally wouldn't have bothered with an "else" after it,
but that's just stylistic.
regards, tom lane
