nandi/memwatch · df1c928

⑂ df1c928

oomd drop-in: arm user@.service, not user.slice

oomd kills a descendant of what it monitors, choosing among that
cgroup's children. Armed on user.slice the only candidate is
user-$UID.slice -- the entire login session, compositor included -- so
the backstop's kill action was "log the user out" rather than "kill the
hog". Move it to user@.service, where the candidates are app.slice,
session.slice and background.slice.

This is the same trap as the session-3.scope victim that eligible_cgroup()
exists to block; it just arrived via a different mechanism.

Applied and verified: oomctl shows user@1000.service monitored at 60%
over oomd's 20s window, with app.slice and descendants as candidates.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

nandithebull committed 2026-09-20T16:45:34-07:00 Browse files

df1c928 parent: 937fc8b

modified optional/10-user-pressure.conf +10 -3 Viewed

@@ -1,12 +1,19 @@
1		-# Drop-in for user.slice: /etc/systemd/system/user.slice.d/10-user-pressure.conf
	1	+# Drop-in for user@.service: /etc/systemd/system/user@.service.d/10-user-pressure.conf
2	2	#
3	3	# Out of the box user.slice has ManagedOOMMemoryPressure=auto, which inherits
4	4	# from the root slice, which is off. The result is that systemd-oomd watches
5	5	# system.slice and completely ignores every application you actually run.
6	6	#
7		-# This arms it at 60% pressure sustained over oomd's default 30s window. It is
	7	+# This arms it at 60% pressure sustained over oomd's default 20s window. It is
8	8	# deliberately slower and dumber than memwatchd -- it is the backstop for the
9	9	# case where memwatchd is dead, wedged, or wrong. It kills without asking.
10		-[Slice]
	10	+#
	11	+# Note the placement. oomd kills a descendant of what it monitors, picking
	12	+# among that cgroup's children. Armed on user.slice, the only candidate is
	13	+# user-$UID.slice -- the whole login session, compositor included, so firing
	14	+# means a logout. Armed here on user@.service the candidates are app.slice,
	15	+# session.slice and background.slice, so it takes out applications and leaves
	16	+# the session standing. Same reasoning as eligible_cgroup() in memwatchd.
	17	+[Service]
11	18	ManagedOOMMemoryPressure=kill
12	19	ManagedOOMMemoryPressureLimit=60%

--# Drop-in for user.slice: /etc/systemd/system/user.slice.d/10-user-pressure.conf
++# Drop-in for user@.service: /etc/systemd/system/user@.service.d/10-user-pressure.conf
  #
  # Out of the box user.slice has ManagedOOMMemoryPressure=auto, which inherits
  # from the root slice, which is off. The result is that systemd-oomd watches
  # system.slice and completely ignores every application you actually run.
  #
--# This arms it at 60% pressure sustained over oomd's default 30s window. It is
++# This arms it at 60% pressure sustained over oomd's default 20s window. It is
  # deliberately slower and dumber than memwatchd -- it is the backstop for the
  # case where memwatchd is dead, wedged, or wrong. It kills without asking.
--[Slice]
++#
++# Note the placement. oomd kills a *descendant* of what it monitors, picking
++# among that cgroup's children. Armed on user.slice, the only candidate is
++# user-$UID.slice -- the whole login session, compositor included, so firing
++# means a logout. Armed here on user@.service the candidates are app.slice,
++# session.slice and background.slice, so it takes out applications and leaves
++# the session standing. Same reasoning as eligible_cgroup() in memwatchd.
++[Service]
  ManagedOOMMemoryPressure=kill
  ManagedOOMMemoryPressureLimit=60%