Merge pull request #20199 from ddstreet/unit_cgroup_catchup

cgroup: do 'catchup' for unit cgroup inotify watch files

src/core/cgroup.c

@@ -3031,6 +3031,9 @@ static int unit_check_cgroup_events(Unit *u) {
 
         assert(u);
 
+        if (!u->cgroup_path)
+                return 0;
+
         r = cg_get_keyed_attribute_graceful(SYSTEMD_CGROUP_CONTROLLER, u->cgroup_path, "cgroup.events",
                                             STRV_MAKE("populated", "frozen"), values);
         if (r < 0)
@@ -3863,6 +3866,21 @@ void unit_invalidate_cgroup_bpf(Unit *u) {
         }
 }
 
+void unit_cgroup_catchup(Unit *u) {
+        assert(u);
+
+        if (!UNIT_HAS_CGROUP_CONTEXT(u))
+                return;
+
+        /* We dropped the inotify watch during reexec/reload, so we need to
+         * check these as they may have changed.
+         * Note that (currently) the kernel doesn't actually update cgroup
+         * file modification times, so we can't just serialize and then check
+         * the mtime for file(s) we are interested in. */
+        (void) unit_check_cgroup_events(u);
+        unit_add_to_cgroup_oom_queue(u);
+}
+
 bool unit_cgroup_delegate(Unit *u) {
         CGroupContext *c;
 

src/core/cgroup.h

@@ -313,6 +313,8 @@ void manager_invalidate_startup_units(Manager *m);
 const char* cgroup_device_policy_to_string(CGroupDevicePolicy i) _const_;
 CGroupDevicePolicy cgroup_device_policy_from_string(const char *s) _pure_;
 
+void unit_cgroup_catchup(Unit *u);
+
 bool unit_cgroup_delegate(Unit *u);
 
 int compare_job_priority(const void *a, const void *b);

src/core/manager.c

@@ -1437,6 +1437,10 @@ static void manager_clear_jobs_and_units(Manager *m) {
         assert(!m->cleanup_queue);
         assert(!m->gc_unit_queue);
         assert(!m->gc_job_queue);
+        assert(!m->cgroup_realize_queue);
+        assert(!m->cgroup_empty_queue);
+        assert(!m->cgroup_oom_queue);
+        assert(!m->target_deps_queue);
         assert(!m->stop_when_unneeded_queue);
         assert(!m->start_when_upheld_queue);
         assert(!m->stop_when_bound_queue);

src/core/unit.c

@@ -733,6 +733,9 @@ Unit* unit_free(Unit *u) {
         if (u->in_dbus_queue)
                 LIST_REMOVE(dbus_queue, u->manager->dbus_unit_queue, u);
 
+        if (u->in_cleanup_queue)
+                LIST_REMOVE(cleanup_queue, u->manager->cleanup_queue, u);
+
         if (u->in_gc_queue)
                 LIST_REMOVE(gc_queue, u->manager->gc_unit_queue, u);
 
@@ -742,8 +745,8 @@ Unit* unit_free(Unit *u) {
         if (u->in_cgroup_empty_queue)
                 LIST_REMOVE(cgroup_empty_queue, u->manager->cgroup_empty_queue, u);
 
-        if (u->in_cleanup_queue)
+        if (u->in_cgroup_oom_queue)
-                LIST_REMOVE(cleanup_queue, u->manager->cleanup_queue, u);
+                LIST_REMOVE(cgroup_oom_queue, u->manager->cgroup_oom_queue, u);
 
         if (u->in_target_deps_queue)
                 LIST_REMOVE(target_deps_queue, u->manager->target_deps_queue, u);
@@ -3611,6 +3614,8 @@ void unit_catchup(Unit *u) {
 
         if (UNIT_VTABLE(u)->catchup)
                 UNIT_VTABLE(u)->catchup(u);
+
+        unit_cgroup_catchup(u);
 }
 
 static bool fragment_mtime_newer(const char *path, usec_t mtime, bool path_masked) {