On Tue, Sep 27, 2022 at 9:44 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Right, the missing piece is the intentional clobber. Thomas indicated
> he'd made such a test locally, but I think it needs full support with
> the same options that mcxt.c has (CLOBBER_FREED_MEMORY and so on
> --- although unmapping the memory is equally good for that, if we
> always do it).
I'd also like to use Valgrind here. That's how I noticed this issue,
in fact. I wrote a very rough patch (too rough to even post as an
FYI).
I don't think that CLOBBER_FREED_MEMORY would necessarily have
detected the problems with
PARALLEL_KEY_BUFFER_USAGE/PARALLEL_KEY_WAL_USAGE (or the analogous
problems in nbtsort.c and vacuumparallel.c). The big advantage of
custom Valgrind instrumentation that marks newly allocated memory
undefined (not inaccessible) is the provenance stuff. Valgrind will
track the provenance of the uninitialized memory reliably, even as the
values are copied around. Any control flow that relies on the value
will make Valgrind complain.
--
Peter Geoghegan