Re: [PATCH 1/2] daemon: use timeout for uninterruptible poll

[Johannes Schindelin <Johannes.Schindelin@gmx.de>]

Hi Kim,

On Sun, 15 Apr 2018, Kim Gybels wrote:

> On (13/04/18 14:36), Johannes Schindelin wrote:
> > > The poll provided in compat/poll.c is not interrupted by receiving
> > > SIGCHLD. Use a timeout for cleaning up dead children in a timely
> > > manner.
> > 
> > Maybe say "When using this poll emulation, use a timeout ..."?
> 
> I will rewrite the commit message when I reroll the patch. Calling the
> poll "uninterruptible" might be wrong as well, although the poll
> doesn't return with EINTR when a child process terminates, it might
> still be interruptible in other ways. On a related note, the handler
> for SIGCHLD is simply not called in Git-for-Windows' daemon.

Right. There is no signal infrastructure on Windows that is an exact
equivalent of what Junio desires.

> > > @@ -1161,8 +1162,13 @@ static int service_loop(struct socketlist *socklist)
> > >  		int i;
> > >  
> > >  		check_dead_children();
> > > -
> > > -		if (poll(pfd, socklist->nr, -1) < 0) {
> > > +#ifdef NO_POLL
> > > +		poll_timeout = live_children ? 100 : -1;
> > > +#endif
> > > +		int ret = poll(pfd, socklist->nr, poll_timeout);
> > > +		if  (ret == 0) {
> > > +			continue;
> > > +		} else if (ret < 0) {
> > 
> > I would find it a bit easier on the eyes if this did not use curlies, and
> > dropped the unnecessary `else` (`continue` will take care of that):
> > 
> > 		if (!ret)
> > 			continue;
> > 		if (ret < 0)
> > 			[...]
> 
> Funny, that's how I would normally write it, if I wasn't so focused on
> trying to follow the coding quidelines. While I'm at it, I will also
> fix that sneaky double space after the if.

:-)

> Is it ok to add the timeout for all platforms using the poll
> emulation, since I only tested for Windows?

From my reading of the patch, it changes only one thing, and only in the
case that the developer asked to build with NO_POLL (which means that the
platform does not have a native poll()): instead of waiting indefinitely,
the poll() call is interrupted in regular intervals to give
reap_dead_children() a chance to clean up.

And that's all it does.

So it is a simply heartbeat for platforms that require it, and that
heartbeat would not even hurt any platform that would *not* require it.

In short: from my point of view, it is fine to add the timeout for all
NO_POLL platforms, even if it was only tested on Windows.

Of course, we *do* know that there is one other user of NO_POLL: the
NonStop platform.

Randall, would you mind testing these two patches on NonStop?

Thanks,
Johannes