scope_abandon() called unit_watch_all_pids() to check if there are any pids in
the cgroup, but unit_watch_all_pids() does nothing (and returns an empty set of
pids) when cg_unified_controller(SYSTEMD) returns true. On hybrid or unified,
cg_unified_controller(SYSTEMD) returns 1, so scope_abandon() thinks the scope
is empty, even though it's not, and marks the scope dead prematurely.
Example output after the scope is marked dead with processes still being present:
● session-24.scope - Session 24 of user guest
Loaded: loaded (/run/systemd/transient/session-24.scope; transient; vendor preset: disabled)
Transient: yes
Active: inactive (dead) since Sat 2017-10-14 15:36:22 CEST; 5min ago
Tasks: 1
CGroup: /user.slice/user-1001.slice/session-24.scope
└─17309 sleep infinity
Subsequent calls to stop the scope unit do nothing, because systemd thinks the
scope is already dead.
https://bugzilla.redhat.com/show_bug.cgi?id=
1486859
This is easily reproducible on both F26 and F27:
> ssh guest@localhost
$ pulseaudio & sleep infinity & disown; exit
$ systemctl status $$
> systemctl status session-NN.scope
Tested with unified, hybrid, and legacy layouts, seems to work.
s->was_abandoned = true;
s->controller = mfree(s->controller);
+ scope_set_state(s, SCOPE_ABANDONED);
/* The client is no longer watching the remaining processes,
* so let's step in here, under the assumption that the
unit_tidy_watch_pids(UNIT(s), 0, 0);
unit_watch_all_pids(UNIT(s));
- /* If the PID set is empty now, then let's finish this off */
- if (set_isempty(UNIT(s)->pids))
- scope_notify_cgroup_empty_event(UNIT(s));
- else
- scope_set_state(s, SCOPE_ABANDONED);
-
return 0;
}