Re: better atomics - v0.6

On 2014-09-24 21:19:06 +0300, Heikki Linnakangas wrote:
> On 09/24/2014 07:57 PM, Andres Freund wrote:
> >On 2014-09-24 12:44:09 -0400, Tom Lane wrote:
> >>Andres Freund <andres(at)2ndquadrant(dot)com> writes:
> >>>On 2014-09-24 18:55:51 +0300, Heikki Linnakangas wrote:
> >>>>There doesn't seem to be any hardware implementations of that in the patch.
> >>>>Is there any architecture that has an instruction or compiler intrinsic for
> >>>>that?
> >>
> >>>You can implement it rather efficiently on ll/sc architectures. But I
> >>>don't really think it matters. I prefer add_until (I've seen it named
> >>>saturated add before as well) to live in the atomics code, rather than
> >>>reimplement it in atomics employing code. I guess you see that
> >>>differently?
> >>
> >>I think the question is more like "what in the world happened to confining
> >>ourselves to a small set of atomics".
> >
> >I fail to see why the existance of a wrapper around compare-exchange
> >(which is one of the primitives we'd agreed upon) runs counter to
> >the agreement that we'll only rely on a limited number of atomics on the
> >hardware level?
>
> It might be a useful function, but if there's no hardware implementation for
> it, it doesn't belong in atomics.h. We don't want to turn it into a general
> library of useful little functions.

Uh. It belongs there because it *atomically* does a saturated add?
That's precisely what atomics.h is about, so I don't see why it'd belong
anywhere else.

> >>I doubt either that this exists
> >>natively anywhere, or ethat it's so useful that we should expect platforms
> >>to have efficient implementations.
>
> Googling around, ARM seems to have a QADD instruction that does that. But
> AFAICS it's not an atomic instruction that you could use for synchronization
> purposes, just a regular instruction.

On ARM - and many other LL/SC architectures - you can implement it
atomically using LL/SC. In pseudocode:

while (true)
{
load_linked(mem, reg);
if (reg + add < limit)
reg = reg + add;
else
reg = limit;
if (store_conditional(mem, reg))
break
}

That's how pretty much *all* atomic instructions work on such architectures.

> >It's useful for my work to get rid of most LockBufHdr() calls (to
> >manipulate usagecount locklessly). That's why I added it. We can delay
> >it till that patch is ready, but I don't really see the benefit.
>
> Yeah, please leave it out for now, we can argue about it later. Even if we
> want it in the future, it would be just dead, untested code now.

Ok, will remove it for now.

I won't repost a version with it removed, as removing a function as the
only doesn't seem to warrant reposting it.

Greetings,

Andres Freund

--
Andres Freund http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Training & Services