bug#58485: [shepherd] Restarting guix-publish fails

[Lars-Dominik Braun <lars@6xq.net>]

Hi,

it seems that `herd restart guix-publish` stopped working after the
introduction of socket activation into shepherd. This is a problem,
because I restart guix-publish automatically after unattended-upgrades. It
fails with the following error for me:

---snip---
Backtrace:
           7 (primitive-load "/gnu/store/7xrg2sbb529ki6hv99n27svg0fi?")
In ice-9/boot-9.scm:
    724:2  6 (call-with-prompt ("prompt") #<procedure 7f8173184940 ?> ?)
  1752:10  5 (with-exception-handler _ _ #:unwind? _ # _)
In ice-9/eval.scm:
    619:8  4 (_ #(#(#<directory (guile-user) 7f817318ac80>)))
In ice-9/boot-9.scm:
   260:13  3 (for-each #<procedure restart-service (name)> _)
In gnu/services/herd.scm:
    168:4  2 (invoke-action guix-publish restart () #<procedure 7f81?>)
    176:7  1 (failure)
In ice-9/boot-9.scm:
  1685:16  0 (raise-exception _ #:continuable? _)

ice-9/boot-9.scm:1685:16: In procedure raise-exception:
ERROR:
  1. &action-exception-error:
      service: guix-publish
      action: start
      key: system-error
      args: ("bind" "~A" ("Address already in use") (98))
---snap---

Note that due to the socket activation you must visit the URL at least
once to start up the guix-publish process. Otherwise a restart will
work fine. It also works fine the second time I invoke `herd restart
guix-publish`, because `guix-publish` is dead by that time.

Looking at an strace shepherd is indeed trying to kill `guix-publish`
and re-bind to the same address:

---snip---
1     read(23, "(shepherd-command (version 0) (action restart) (service guix-publish) (arguments ()) (directory \"/root\"))", 1024) = 105
1     getpgid(18096)                    = 18096
1     getpgid(0)                        = 0
1     kill(-18096, SIGTERM)             = 0
1     newfstatat(AT_FDCWD, "/etc/localtime", {st_mode=S_IFREG|0444, st_size=2298, ...}, 0) = 0
1     write(17, "shepherd[1]: Service guix-publish has been stopped.\n", 52) = 52
1     socket(AF_INET, SOCK_STREAM|SOCK_CLOEXEC|SOCK_NONBLOCK, IPPROTO_IP) = 36
1     setsockopt(36, SOL_SOCKET, SO_REUSEADDR, [1], 4) = 0
1     bind(36, {sa_family=AF_INET, sin_port=htons(8082), sin_addr=inet_addr("127.0.0.1")}, 16) = -1 EADDRINUSE (Address already in use)
1     write(23, "(reply (version 0) (result #f) (error (error (version 0) action-exception start guix-publish system-error (\"bind\" \"~A\" (\"Address already in use\") (98)))) (messages (\"Service guix-publish has been stopped.\")))", 208) = 208
1     close(23)
---snap---

The obvious explanation would be that stopping does not wait for the
process to actually exit. make-kill-destructor does not waitpid it seems
and 'running is set unconditionally to #f after 'stop has finished.

Cheers,
Lars