On Thu, 18 Jun 2009, John Summerfield wrote:

> Jon Peatfield wrote:
>> 
>
>>
>>  BTW the default reboot/shutdown procedures in el5/sl5 don't give user
>>  processes very long to checkpoint themselves, and I *think* that
>>  networking may have been turned off by the time they get signalled.  We 
>
> That's too silly for words. How likely is it that users, somewhere, will have 
> open files on NFS mounts?
>
> 100% where home is on an NFS mount. I don't think any distro would be 
> shutting down networks that soon.

For the case of reboot the directory /etc/rc.d/rc6.d/ contains the 
relevant Kxx* scripts which are run followed by the Sxx* ones.

The set I see on sl53 (and I assume el53) includes K90network, S00killall 
(which despite the name doesn't kill 'all'), and S01reboot - the last of 
these is the only one which appears to signal all *user* processes, e.g.

  ...
  action $"Sending all processes the TERM signal..." /sbin/killall5 -15
  sleep 5
  action $"Sending all processes the KILL signal..."  /sbin/killall5 -9
  ...

So unless / is on NFS the networking is likely to already be down by the 
time that the signals are sent to all user jobs.

Now of course if a process is connected to a tty then that will probably 
get a signal at the point that getty or sshd or X (or whatever) get 
killed, but long-running jobs probably arn't attached to a tty...

I may have managed to miss another place where user jobs get signalled, or 
perhaps I broke something which would do it...

  -- Jon