Re: shared memory based stat collector (was: Sharing record typmods between backends)

On 2017-08-14 12:28:39 -0400, Tom Lane wrote:
> Robert Haas <robertmhaas(at)gmail(dot)com> writes:
> > On Mon, Aug 14, 2017 at 12:16 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> >> Just FYI, the only values being reported by buildfarm animals are
> >> "posix", "sysv", and "windows". So while mmap may be a thing,
> >> it's an untested thing.
>
> > I'm pretty sure I dev-tested it before committing anything, but,
> > certainly, having ongoing BF coverage woudn't be a bad thing.
>
> Looking closer, the reason those are the only reported values is
> that those are the only possible results from initdb's
> choose_dsm_implementation(). So the real question here is whether
> "mmap" should be considered to dominate "sysv" if it's available.

No mmap isn't a good option - it's file backed mmap, rather than
anonymous mmap. To my knowledge there's no good portable way to use
anonymous mmap to share memory across processes unless established
before a fork().

> If so, why isn't choose_dsm_implementation() trying it; and if not,
> why are we carrying it?

I think the idea was that there might be platforms that require it, but
...

Greetings,

Andres Freund