On Thu, Apr 06, 2023 at 08:47:48AM +0200, Stefano Brivio wrote:
> On Thu, 6 Apr 2023 12:31:55 +1000
> David Gibson <david@gibson.dropbear.id.au> wrote:
> 
> > On Wed, Apr 05, 2023 at 01:58:00PM +0200, Stefano Brivio wrote:
> >
> > [...]
> >
> > > I didn't spot this earlier, but... does it really make sense to wait in
> > > cmd_pid(), also on ENOENT, rather than making 'hold' return only once
> > > the socket is ready?  
> > 
> > So, this is a consequence of the fact that the holder doesn't move
> > into the background itself - it just sits in the foreground until
> > terminated.  That means that the typical usecase puts it into the
> > background from the shell with &, which in turn means that when we
> > reach the next shell command the socket may not be ready - or not even
> > created.
> > 
> > One of the things I had in mind for a hypothetical "nstool unshare"
> > would be to avoid this and have it background itself once the socket
> > is ready.
> 
> Ah, sure, it makes sense now.
> 
> > > I don't think it would be outrageous to have
> > > 'nstool pid' failing if the holding process doesn't exist.
> > > 
> > > Admittely, I'm biased by the few hundreds of times I needed to
> > > 'killall -9 nsholder' in the past months. :)  
> > 
> > So... I agree that's irritating, I've done it a similar number of
> > times.  However, I don't think that's really related to the question
> > above - in my experience it's always been the holder process that's
> > hung around, not something waiting on a holder.
> 
> Yes, same here, but it's something I file under the same category (I
> don't remember why nsholder would hang, you probably explained at some
> point...).

I believe the main reason is because of the holders which are PID 1
within their pid namespaces.  That means that if you interrupt the
tests, the SIGINT or SIGHUP they'll get from tmux etc. shutting down
won't be sufficient to kill them.

> 
> > > >  	rc = getsockopt(fd, SOL_SOCKET, SO_PEERCRED,
> > > >  			&peercred, &optlen);
> > > >  	if (rc < 0)
> > > > -		die("getsockopet(SO_PEERCRED): %s\n", strerror(errno));
> > > > +		die("getsockopet(SO_PEERCRED) %s: %s\n",
> > > > +		    sockpath, strerror(errno));
> > > >  
> > > >  	close(fd);
> > > >  
> > > >  	printf("%d\n", peercred.pid);
> > > >  }
> > > >  
> > > > -static void stop(int fd, const struct sockaddr_un *addr)
> > > > +static void cmd_stop(int argc, char *argv[])
> > > >  {
> > > > -	int rc;
> > > > +	const char *sockpath = argv[1];
> > > > +	int fd, rc;
> > > >  	char buf = 'Q';
> > > >  
> > > > -	rc = connect(fd, (struct sockaddr *)addr, sizeof(*addr));
> > > > -	if (rc < 0)
> > > > -		die("connect(): %s\n", strerror(errno));
> > > > +	if (argc != 2)
> > > > +		usage();
> > > > +
> > > > +	fd = connect_ctl(sockpath, false);
> > > >  
> > > >  	rc = write(fd, &buf, sizeof(buf));  
> > > 
> > > Unrelated: a compound literal would make this more readable.  
> > 
> > Uh.. I don't see where a compound literal would even go here.
> 
> I meant:
> 
> 	rc = write(fd, &(char){ 'Q' }, 1);
> 
> ...so that one doesn't need to look at 'buf'. nstool is C99 anyway.

Oh, ok.  On the other hand it means not using sizeof() to get the
length, which isn't ideal.

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson