Michael Paquier <michael@paquier.xyz> writes:
> On Fri, Oct 18, 2019 at 05:55:32PM -0700, Ashwin Agrawal wrote:
>> I am not sure if this causes any potential problems or not, but for
>> consistency of code seems we are missing below. All other places in code
>> where sigsetjmp() exists for top level handling has error_context_stack set
>> to NULL.
> Resetting error_context_stack prevents calling any callbacks which may
> be set. These would not be much useful in this context anyway, and
> visibly that's actually not an issue with the autovacuum code so far
> (I don't recall seeing a custom callback setup in this area, but I may
> have missed something). So fixing it would be a good thing actually,
> on HEAD.
> Any thoughts from others?
This seems like a real and possibly serious bug to me. Backend sigsetjmp
callers *must* clear error_context_stack (or restore it to a previous
value), because if it isn't NULL it's surely pointing at garbage, ie a
local variable that's no longer part of the valid stack.
The issue might be argued to be insignificant because the autovacuum
worker is just going to do proc_exit anyway. But if it encountered
another error during proc_exit, elog.c might try to invoke error
callbacks using garbage callback data.
In short, I think we'd better back-patch too.
regards, tom lane