[systemd-devel] Concurrent login / daemon-reload produces abandoned sessions

[Michal Koutný]

Hello Nicolas.

On Wed, Jul 07, 2021 at 02:12:51PM +0200, Nicolas Bock <nicolas.bock at canonical.com> wrote:
> Using systemd-248.3-1ubuntu1 on Ubuntu Impish the following
> script produces multiple abandoned sessions:
> 
> 	$ for i in {1..100}; do sleep 0.2; ssh localhost sudo systemctl daemon-reload & ssh localhost sleep 1 & done
> 	$ sleep 2
> 	$ jobs -p | xargs --verbose --no-run-if-empty kill -KILL
> 	$ systemctl | grep abandoned
> 	  session-174.scope                    loaded active     abandoned Session 174 of user ubuntu
> 	  session-175.scope                    loaded active     abandoned Session 175 of user ubuntu
> 	  session-176.scope                    loaded active     abandoned Session 176 of user ubuntu
> 	  session-25.scope                     loaded active     abandoned Session 25 of user ubuntu
> 
> I would like to debug this behavior further and understand
> why this is happening but don't know where to look next.

It might be a bit challenging :)

> Is there any information in particular I should look at?

I assume you use hybrid or unified cgroup setup and that the abandoned
scopes are empty (no processes in their cgroups), correct?

My hypothesis is following

// race between scope abandonement, emptiness notification -> abandon comes first
manager_reload
  manager_clear_jobs_and_units
    unit_release_cgroup
      inotify_rm_watch(u->manager->cgroup_inotify_fd, u->cgroup_control_inotify_wd)
[...]
// last process terminates somewhere here but we're not watching emptiness yet
scope_coldplug()
  // scope should be checked for emptiness here

I _think_ this could be fixed with the patch
--- a/src/core/scope.c
+++ b/src/core/scope.c
@@ -243,8 +243,8 @@ static int scope_coldplug(Unit *u) {
                                 if (r < 0 && r != -EEXIST)
                                         return r;
                         }
-                } else
-                        (void) unit_enqueue_rewatch_pids(u);
+                }
+                (void) unit_enqueue_rewatch_pids(u);
         }

         bus_scope_track_controller(s);

Can you file a Github issue to track this (and possibly try if this
works for you)?

Thanks,
Michal

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: Digital signature
URL: <https://lists.freedesktop.org/archives/systemd-devel/attachments/20210708/1e0d526c/attachment.sig>