Re: fd.c: flush data problems on osx

On 2016-03-21 14:46:09 +0300, Stas Kelvich wrote:
>
> > On 18 Mar 2016, at 14:45, Stas Kelvich <s(dot)kelvich(at)postgrespro(dot)ru> wrote:
> >>
> >>> One possible solution for that is just fallback to pg_fdatasync in case when offset = nbytes = 0.
> >>
> >> Hm, that's a bit heavyweight. I'd rather do an lseek(SEEK_END) to get
> >> the file size. Could you test that?
> >>
> >
> > It looks like OSX mmap raises EINVAL when length isn’t aligned to pagesize while manual says it can be of arbitrary length, so i aligned it.
> > Also there were call to mmap with PROT_READ | PROT_WRITE, but when called from pre_sync_fname file descriptor is just O_RDONLY, so i changed mmap mode to PROT_READ — seems that PROT_WRITE wasn’t needed anyway.
> >
> > And all of that reduces number of warnings in order of magnitude but there are still some and I don’t yet understand why are they happening.
>
> I’ve spend some more time on this issue and found that remaining
> warnings were caused by mmap-ing directories — that raises EINVAL in
> OSX (probably not only OSX, but I didn’t tried).

> So i’ve skipped mmap for dirs and now restore happens without
> warnings. Also I’ve fixed wrong error check that was in previous
> version of patch.

Hm. I think we should rather just skip calling pg_flush_data in the
directory case, that very likely isn't beneficial on any OS.

I think we still need to fix the mmap() implementation to support the
offset = 0, nbytes = 0 case (via fseek(SEEK_END).

Greetings,

Andres Freund