Hey Ludo,

On Wed, Feb 28 2018, Ludovic Courtès wrote:
>> The problem is that shepherd, when run as a user process, can 
>> "lose"
>> services which fork away. Shepherd can still kill them, but a 
>> SIGCHLD
>> won't be delivered if they die, so shepherd can't 
>> restart/disable
>> them. My prime example is emacs, which I run with --daemon. If 
>> I then
>> kill emacs, shepherd will still think that it is running.
>
> There are two issues here, I think.
>
>   1. shepherd cannot lose SIGCHLD: if a process dies immediately 
>   once
>      it’s been spawned, as is the case with “emacs --daemon” or 
>      any
>      other daemon-style program, it should receive SIGCHLD and 
>      process
>      it.

Yeah, that's true, but the problem is that shepherd only processes 
the SIGCHLD if there is a service with its `running` slot set to 
the pid. When emacs forks, the original process may have its 
SIGCHLD handled, but that doesn't affect shepherd's service state 
(as it shouldn't, because it's using #:pid-file to track the 
forked process).

>   2. shepherd currently can’t do much with real daemons.  So 
>   what we do
>      in GuixSD is to either start programs in non-daemon mode, 
>      when
>      that’s an option, or pass #:pid-file to retrieve the forked 
>      process
>      PID.  I think you should do one of these as well.

I am doing that. The problem is that when a service dies (crashes, 
quits, etc.) the `respawn?` option cannot be honoured because 
shepherd is not notified that the process has terminated (because 
it never receives a SIGCHLD for the forked pid). My patch polls 
for the processes we expect, to make up for the lack of 
notification. I would much rather it receive an event/signal to 
notify that the forked process has died, but I don't know how to 
do that in a robust, portable way so I chose to poll instead.

If you look at my test case in tests/respawn-service.sh (which can 
be read in its entirety in the diff attached to my previous email) 
you can see the problem that this patch solves. The test will fail 
without the rest of my patch, but will pass with them (guix build 
container issue notwithstanding).

Carlo