On Thu, Nov 23, 2023 at 11:18 PM Thomas Munro <thomas.munro@gmail.com> wrote:
> Robert pinged me to see if I had any ideas.
Thanks, Thomas.
> The reason it fails on Windows is because there is a special code path
> for WIN32 in the patch's src/bin/pg_combinebackup/copy_file.c, but it
> is incomplete: it returns early without feeding the data into the
> checksum, so all the checksums have the same initial and bogus value.
> I commented that part out so it took the normal path like Unix, and it
> passed.
Yikes, that's embarrassing. Thanks for running it down. There is logic
in the caller to figure out whether we need to recompute the checksum
or can reuse one we already have, but copy_file() doesn't understand
that it should take the slow path if a new checksum computation is
required.
> The reason it fails on Linux 32 bit with -fsanitize is because this
> has uncovered a bug in xlogreader.c, which overflows a 32 bit pointer
> when doing a size test that could easily be changed to non-overflowing
> formulation. AFAICS it is not a live bug because it comes to the
> right conclusion without deferencing the pointer due to other checks,
> but the sanitizer is not wrong to complain about it and I will post a
> patch to fix that in a new thread. With the draft patch I am testing,
> the sanitizer is happy and this passes too.
Thanks so much.
--
Robert Haas
EDB: http://www.enterprisedb.com