On Wed, Feb 17, 2016 at 10:03 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
>> Hmm, that's true. I don't think it actually matters all that much,
>> because proclock->tag.myProc->lockGroupLeader == NULL has pretty much
>> the same effect as if proclock->tag.myProc->lockGroupLeader ==
>> proclock->tag.myProc. But not completely. One problem is that we
>> don't currently assume that 8-byte writes are atomic, so somebody
>> might see the group leader field half-set, which would be bad.
>
> Yes, exactly. I'm not certain if there are any real platforms where
> a pointer-sized write wouldn't be atomic (it sure sounds inefficient
> for that to be true), but we have not assumed that to date and I'd
> just as soon not start here.
I guess it's worth pointing out that storing the groupLeader in the
PROCLOCK, as we do currently, avoids this issue entirely. The
deadlock detector can safely follow lockGroupLeader pointers because
it holds all the lock manager partition locks, and proc.c/lock.c don't
need to do so if groupLeader is present in the PROCLOCK. Nor does
that field ever need to be updated, because it's defined to always be
non-NULL: if a process becomes a group leader, the value stored in the
PROCLOCKs does not change.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company