Process hierarchy on a DC?

[Simo Sorce]

On Fri, Apr 28, 2017 at 01:11:30PM +0200, Stefan Metzmacher via samba-technical wrote:
> Am 28.04.2017 um 12:06 schrieb Ralph Böhme:
> > On Thu, Apr 27, 2017 at 02:39:21PM -0700, Jeremy Allison wrote:
> >> Much though I hate to revert, I think this might be a good case
> >> for it temporarily.
> > 
> > two patchsets attached:
> > 
> > 1) revert the use of tfork in samba_runcmd_send(). This restores the process
> > hierarchy and stopping samba by sending a single SIGTERM
> 
> Please add a large explaination to the commit message, basically a summary
> of this mail thread.
> 
> > If this process hierarchy is a problem then I guess we need two functions in the
> > future:
> > 
> > - apply the revert and rename samba_runcmd_send (and the callers) to something
> >   like samba_runchld_send()
> > 
> > - re-add a samba_runcmd_send() that uses tfork
> > 
> > samba_runchld_send() would be used for starting long lived processes where you
> > care about the process hierarchy, preserving the problematic behaviour that you
> > have to correctly handle SIGCHLD yourself. So just the existing behaviour.
> 
> I'd name it samba_start_child_send().
> 
> Here I'd still like to deliver the exit status of the process via a pipe,
> maybe just doing a double fork, basically skipping the first child that
> directly exists.
> The main process would still call waitpid() after getting the exit status
> from the pipe, to avoid zombies, but the result of waitpid() will be
> ignored.

Sorry for the dumb question, but why are you doing double forks and escaping
the original process group, and then try to undo all this stuff by sending
around new process groups to kill ?

Why not just do a simple straight fork and retain the process group and
hierarchy ?

> > samba_runcmd_send() would be used for starting short lived processes or when you
> > just don't care about the process hierarchy and you wan't to be sure that the
> > command you run doesn't generate a SIGCHLD
> 
> And samba_run_command_send() which would be what we currently have in
> master.
> 
> Please note that some current callers of samba_runcmd_send() would want
> the samba_run_command_send() and some samba_start_child_send().
> 
> metze
>