Re: Early look: PostgreSQL 15.0

On Wed, Oct 19, 2022 at 01:14:44PM +0100, Stuart Henderson wrote:
> On 2022/10/19 09:47, Stuart Henderson wrote:
> > On 2022/10/19 01:38, Theo Buehler wrote:
> > > > The main part I want feedback on is a new patch that silently treats the
> > > > dynamic_shared_memory_type = posix as dynamic_shared_memory_type = shm.
> > > > Since 15beta1, I had an issue where the PostgreSQL server would stop
> > > > working after a week. I traced that issue to OpenBSD deleting the
> > > > shared memory files in /tmp that shm_open(3) creates, since they were
> > > > more than a week old. Possibly this is only an issue on lightly loaded
> > > > servers, but unless daily(8) skips deleting these files automatically, I
> > > > consider the approach the patch uses necessary for reliable operation.
> > >
> > > I think a patch that extends the list before -prune in daily(8) with
> > > './*.shm' would be perfectly acceptable. I'd ok it once it's tested.
> >
> > Since the decision to put these "files" in /tmp is OpenBSD's
> > (see libc/gen/shm_open.c) it should definitely be handling them
> > automatically in the system scripts.

Exactly.

> It needs to be after the -prune, as that acts only to trim the
> tree descended for the conditons after the "-o".

Ah, right. I knew it was too late for me to cook a diff :)

> OK?

Yes, this works here.

ok tb



>
>
> Index: daily
> ===================================================================
> RCS file: /cvs/src/etc/daily,v
> retrieving revision 1.95
> diff -u -p -r1.95 daily
> --- daily 20 Oct 2020 22:42:29 -0000 1.95
> +++ daily 19 Oct 2022 12:10:56 -0000
> @@ -49,8 +49,8 @@ if [ -d /tmp -a ! -L /tmp ]; then
> cd /tmp && {
> find -x . \
> \( -path './ssh-*' -o -path ./.X11-unix -o -path ./.ICE-unix \
> - -o -path './tmux-*' \) \
> - -prune -o -type f -atime +7 -delete 2>/dev/null
> + -o -path './tmux-*' \) -prune -o \
> + -type f -and ! -path './*.shm' -atime +7 -delete 2>/dev/null
> find -x . -type d -mtime +1 ! -path ./vi.recover ! -path ./.X11-unix \
> ! -path ./.ICE-unix ! -name . \
> -delete >/dev/null 2>&1; }
>
>
> > (I'm not entirely convinced /tmp is a great place to store them
> > but I don't really have a better idea. On most Linux they use
> > a separate /dev/shm ramdisk but I don't think adding a new fstab
> > entry for a libc requirement is going to work out for us!)
>
> ...also, changing the path would require auditing ports for unveil.
>