On Mon, Dec 5, 2016 at 3:07 PM, Andreas Seltenreich <seltenreich@gmx.de> wrote:
> on master as of a0ae54d, there's a 1 in 10e6 chance sqlsmith catches
> gather_readnext reading beyond the gatherstate->readers array with
> readers[gatherstate->readnext]. Sample backtrace below.
>
> As readnext is never explicitly initialized, I think what happens is
> that a rescan gets less workers than the initial scan, and the dangling
> readnext points outside the array. I'm no longer seeing these crashes
> when explicitly initializing readnext to 0 like in the attached patch.
Thanks, great catch! Committed and back-patched to 9.6.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company