Re: Weird failure with latches in curculio on v15

Hi,

On 2023-02-01 12:08:24 -0500, Robert Haas wrote:
> On Wed, Feb 1, 2023 at 11:58 AM Andres Freund <andres(at)anarazel(dot)de> wrote:
> > > 9a740f81e clearly made things a lot worse, but it wasn't great
> > > before. Can we see a way forward to removing the problem entirely?
> >
> > Yea, I think we can - we should stop relying on system(). If we instead
> > run the command properly as a subprocess, we don't need to do bad things
> > in the signal handler anymore.
>
> I like the idea of not relying on system(). In most respects, doing
> fork() + exec() ourselves seems superior. We can control where the
> output goes, what we do while waiting, etc. But system() runs the
> command through the shell, so that for example you don't have to
> invent your own way of splitting a string into words to be passed to
> exec[whatever](). I've never understood how you're supposed to get
> that behavior other than by calling system().

We could just exec the shell in the forked process, using -c to invoke
the command. That should give us pretty much the same efficiency as
system(), with a lot more control.

I think we already do that somewhere. <dig>. Ah, yes, spawn_process() in
pg_regress.c. I suspect we couldn't use exec for restore_command etc,
as I think it's not uncommon to use && in the command.

Perhaps we should abstract the relevant pieces of spawn_process() that
into something more general? The OS specifics are sufficiently
complicated that I don't think it'd be good to have multiple copies.

It's too bad that we have the history of passing things to shell,
otherwise we could define a common argument handling of the GUC and just
execve ourselves, but that ship has sailed.

Greetings,

Andres Freund