Re: Our "fallback" atomics implementation doesn't actually work

On 2016-10-06 00:06:33 -0400, Tom Lane wrote:
> Andres Freund <andres@anarazel.de> writes:
> > Hm. After a long battle of head vs. wall I think I see what the problem
> > is.  For the fallback atomics implementation I somehow had assumed that
> > pg_atomic_write_u32() doesn't need to lock, as it's just an unlocked
> > write.  But that's not true, because it has to cause
> > pg_atomic_compare_exchange_u32 to fail.
> 
> Hah ... obvious once you see it.
> 
> > For me the problem often takes a lot longer to reproduce (once only
> > after 40min), could you run with the attached patch, and see whether
> > that fixes things for you?
> 
> For me, with the described test case, HEAD fails within a minute,
> two times out of three or so.  I've not reproduced it after half an
> hour of beating on this patch.  Looks good.

It's not quite there yet, unfortunately. At the moment
pg_atomic_write_u32() is used for local buffers - and we explicitly
don't want that to be locking for temp buffers
(c.f. 6b93fcd149329d4ee7319561b30fc15a573c6307).

Don't really have a great idea about addressing this, besides either
just living with the lock for temp buffers on fallback platforms (which
don't have much of a practical relevance), or introduce
pg_atomic_unlocked_write_u32() or something. Neither seems great.

Regards,

Andres