Re: Atomics for heap_parallelscan_nextpage()

On 2017-08-16 13:40:09 -0400, Tom Lane wrote:
> I wrote:
> > I can confirm that on dromedary, that regression test case is attempting
> > to create a TOC with a not-well-aligned size: 93268 = 0x16c54 bytes.
>
> ... although, on closer look, it still seems like we have a fundamental
> bit of schizophrenia here, because on this machine
>
> $ grep ALIGN pg_config.h
> #define ALIGNOF_DOUBLE 4
> #define ALIGNOF_INT 4
> #define ALIGNOF_LONG 4
> #define ALIGNOF_LONG_LONG_INT 4
> #define ALIGNOF_SHORT 2
> #define MAXIMUM_ALIGNOF 4
>
> Basically, therefore, ISTM that it is not a good thing that the atomics
> code thinks it can rely on 8-byte-aligned data when the entire rest of
> the system believes that 4-byte alignment is enough for anything.

That's a hardware requirement, we can't do much about it. Several
[micro-]architectures don't support unaligned atomic 8 byte writes.

> I was wondering why the shm_toc code was using BUFFERALIGN and not
> MAXALIGN, and I now suspect that the answer is "it's an entirely
> undocumented kluge to make the atomics code not crash on 32-bit
> machines, so long as nobody puts a pg_atomic_uint64 anywhere except in
> a shm_toc".

I don't think there were any atomics in affected code until earlier
today... And given it didn't work for shm_toc anyway, I'm not quite
following.

> I'm not sure that that's good enough, and I'm damn sure that it
> shouldn't be undocumented.

8 byte alignment would be good enough, so BUFFERALIGN ought to be
sufficient. But it'd be nicer to have a separate more descriptive knob.

Greetings,

Andres Freund