Re: Sockets in Shepherd, and Emacs, oh my!

[Liliana Marie Prikler]

Hi,

Am Sonntag, dem 10.04.2022 um 22:40 +0200 schrieb Ludovic Courtès:
> Hallo!
> 
> Liliana Marie Prikler <liliana.prikler@gmail.com> skribis:
> 
> > given that Shepherd 0.9.0 adds support for systemd-style socket
> > activation and Emacs supports that (if linked against systemd!) I
> > wanted to try it out.  Here's my "generic" recipe for constructing
> > an Emacs service:
> > 
> > (define* (emacs #:key (name 'server)
> >                 (uid (getuid))
> >                 (gid (getgid)))
> >   (make <service>
> >     #:provides (list
> >                 (symbol-append 'emacs@ name))
> >     #:requires '()
> >     #:start
> >     (make-systemd-constructor
> >      (list "emacs" "--fg-daemon")
> >      (list (endpoint
> >             (make-socket-address AF_UNIX
> >                                  (format #f "/run/user/~d/emacs/~s"
> >                                          uid name))
> >             #:socket-owner uid
> >             #:socket-group gid)))
> >     #:stop (make-systemd-destructor)))
Addendum: This only worked, because I had the directory already
created.  For the directory to be created with the correct permissions,
you need to add #:socket-directory-permissions #o700.

> Fun!  Good to see this put to good use.  :-)
> 
> > There appear to be some bugs, though.  If shepherd launches a
> > process via systemd-constructor and it dies before claiming the
> > socket, it will become a zombie.
> 
> What do you mean by “claiming the socket”?  Do you have a log file
> showing this, or at (better yet) a small reproducer?  You can use
> ‘tests/systemd.sh’ as a starting point.
Emacs without systemd support simply exits, because Shepherd owns the
socket.  I think a rather simple command that exits immediately or
after sleep 5 ought to do the trick.

> > Further, if the process dies after having claimed the socket,
> > shepherd will not reset service's entry points and running value. 
> > Both should be fixable in my opinion.
> 
> The running value is changed as soon as the child process is started
> by ‘make-systemd-constructor’, AFAICS.  But maybe I’m overlooking
> something.
Sure, but in neither case appears the dying process to be handled
correctly.  If shepherd doesn't check whether the service takes over,
it might even be the same bug, but I didn't check if the correctly
spawned emacs becomes a zombie.  At the very least I'd argue it'd be
cleaned up during manual service restart, which is needed to get your
socket back.

> 
Cheers