On Sun, Dec 4, 2022 at 10:57 AM PG Bug reporting form
<noreply@postgresql.org> wrote:
> We started having random segmentation faults with our postgres 13.4 server,
> running on RHEL 8.7. It was upgraded to 13.9, but the issue persists. The
> database is fairly large, about 250GB on disk.
>
> I got core dump of one of the crashes and it shows SIGSEGV in
> BufFileLoadBuffer. I tried to investigate this a little and it seems the
> reason for it is that ExecHashJoinGetSavedTuple reads {0, 0} as header,
> meaning hashvalue is 0 and tuple length is 0. Line 1277 in nodeHashjoin.c
> subtracts sizeof(uint32) from 0 and passes it as size to BufFileRead(). GDB
> shows size=18446744073627287632 at frame #1 which is not ((uint64_t) -4),
> but -82263984. I think this is caused by BufFileRead which decrements
> parameter 'size' by bytes read, so apparently it has read 82263980 bytes,
> overwriting BufFile struct passed to BufFileLoadBuffer. Its files field now
> contains ascii instead of pointer and file->files[file->curFile]; causes
> SIGSEGV.
>
> Why it has read {0, 0} as saved tuple header, or what could have written
> these zeroes there, I could not find out...
Are you able to reproduce this on demand? Can you get your hands on
the temporary file(s) it's reading? How large is it/are they?
Perhaps we could write a little Python/whatever script to read the
tuples back one at a time until it hits this {0, 0} header to confirm
that it's definitely there, ie the bad header has actually been
written out, which would help narrow down the location of the bug.