core: add DelegateSubgroup= setting

This implements a minimal subset of #24961, but in a lot more
restrictive way: we only allow one level of subcgroup (as that's enough
to address the no-processes in inner cgroups rule), and does not change
anything about threaded cgroup logic or similar, or make any of this new
behaviour mandatory.

All this does is this: all non-control processes we invoke for a unit
we'll invoke in a subgroup by the specified name.

We'll later port all our current services that use cgroup delegation
over to this, i.e. user@.service, systemd-nspawn@.service and
systemd-udevd.service.

src/core/cgroup.c

@@ -293,6 +293,8 @@ void cgroup_context_done(CGroupContext *c) {
         cpu_set_reset(&c->startup_cpuset_cpus);
         cpu_set_reset(&c->cpuset_mems);
         cpu_set_reset(&c->startup_cpuset_mems);
+
+        c->delegate_subgroup = mfree(c->delegate_subgroup);
 }
 
 static int unit_get_kernel_memory_limit(Unit *u, const char *file, uint64_t *ret) {
@@ -570,6 +572,10 @@ void cgroup_context_dump(Unit *u, FILE* f, const char *prefix) {
                 prefix, managed_oom_preference_to_string(c->moom_preference),
                 prefix, cgroup_pressure_watch_to_string(c->memory_pressure_watch));
 
+        if (c->delegate_subgroup)
+                fprintf(f, "%sDelegateSubgroup: %s\n",
+                        prefix, c->delegate_subgroup);
+
         if (c->memory_pressure_threshold_usec != USEC_INFINITY)
                 fprintf(f, "%sMemoryPressureThresholdSec: %s\n",
                         prefix, FORMAT_TIMESPAN(c->memory_pressure_threshold_usec, 1));

src/core/cgroup.h

@@ -133,6 +133,7 @@ struct CGroupContext {
         bool delegate;
         CGroupMask delegate_controllers;
         CGroupMask disable_controllers;
+        char *delegate_subgroup;
 
         /* For unified hierarchy */
         uint64_t cpu_weight;

src/core/dbus-cgroup.c

@@ -435,6 +435,7 @@ const sd_bus_vtable bus_cgroup_vtable[] = {
         SD_BUS_VTABLE_START(0),
         SD_BUS_PROPERTY("Delegate", "b", bus_property_get_bool, offsetof(CGroupContext, delegate), 0),
         SD_BUS_PROPERTY("DelegateControllers", "as", property_get_delegate_controllers, 0, 0),
+        SD_BUS_PROPERTY("DelegateSubgroup", "s", NULL, offsetof(CGroupContext, delegate_subgroup), 0),
         SD_BUS_PROPERTY("CPUAccounting", "b", bus_property_get_bool, offsetof(CGroupContext, cpu_accounting), 0),
         SD_BUS_PROPERTY("CPUWeight", "t", NULL, offsetof(CGroupContext, cpu_weight), 0),
         SD_BUS_PROPERTY("StartupCPUWeight", "t", NULL, offsetof(CGroupContext, startup_cpu_weight), 0),
@@ -536,6 +537,33 @@ static int bus_cgroup_set_transient_property(
 
                 return 1;
 
+        } else if (streq(name, "DelegateSubgroup")) {
+                const char *s;
+
+                if (!UNIT_VTABLE(u)->can_delegate)
+                        return sd_bus_error_set(error, SD_BUS_ERROR_INVALID_ARGS, "Delegation not available for unit type");
+
+                r = sd_bus_message_read(message, "s", &s);
+                if (r < 0)
+                        return r;
+
+                if (!isempty(s) && cg_needs_escape(s))
+                        return sd_bus_error_setf(error, SD_BUS_ERROR_INVALID_ARGS, "Invalid control group name: %s", s);
+
+                if (!UNIT_WRITE_FLAGS_NOOP(flags)) {
+                        if (isempty(s))
+                                c->delegate_subgroup = mfree(c->delegate_subgroup);
+                        else {
+                                r = free_and_strdup_warn(&c->delegate_subgroup, s);
+                                if (r < 0)
+                                        return r;
+                        }
+
+                        unit_write_settingf(u, flags, name, "DelegateSubgroup=%s", s);
+                }
+
+                return 1;
+
         } else if (STR_IN_SET(name, "DelegateControllers", "DisableControllers")) {
                 CGroupMask mask = 0;
 

src/core/execute.c

@@ -4178,8 +4178,12 @@ static int compile_suggested_paths(const ExecContext *c, const ExecParameters *p
         return 0;
 }
 
-static int exec_parameters_get_cgroup_path(const ExecParameters *params, char **ret) {
-        bool using_subcgroup;
+static int exec_parameters_get_cgroup_path(
+                const ExecParameters *params,
+                const CGroupContext *c,
+                char **ret) {
+
+        const char *subgroup = NULL;
         char *p;
 
         assert(params);
@@ -4197,16 +4201,22 @@ static int exec_parameters_get_cgroup_path(const ExecParameters *params, char **
          * this is not necessary, the cgroup is still empty. We distinguish these cases with the EXEC_CONTROL_CGROUP
          * flag, which is only passed for the former statements, not for the latter. */
 
-        using_subcgroup = FLAGS_SET(params->flags, EXEC_CONTROL_CGROUP|EXEC_CGROUP_DELEGATE|EXEC_IS_CONTROL);
-        if (using_subcgroup)
-                p = path_join(params->cgroup_path, ".control");
+        if (FLAGS_SET(params->flags, EXEC_CGROUP_DELEGATE) && (FLAGS_SET(params->flags, EXEC_CONTROL_CGROUP) || c->delegate_subgroup)) {
+                if (FLAGS_SET(params->flags, EXEC_IS_CONTROL))
+                        subgroup = ".control";
+                else
+                        subgroup = c->delegate_subgroup;
+        }
+
+        if (subgroup)
+                p = path_join(params->cgroup_path, subgroup);
         else
                 p = strdup(params->cgroup_path);
         if (!p)
                 return -ENOMEM;
 
         *ret = p;
-        return using_subcgroup;
+        return !!subgroup;
 }
 
 static int exec_context_cpu_affinity_from_numa(const ExecContext *c, CPUSet *ret) {
@@ -4705,7 +4715,7 @@ static int exec_child(
         if (params->cgroup_path) {
                 _cleanup_free_ char *p = NULL;
 
-                r = exec_parameters_get_cgroup_path(params, &p);
+                r = exec_parameters_get_cgroup_path(params, cgroup_context, &p);
                 if (r < 0) {
                         *exit_status = EXIT_CGROUP;
                         return log_unit_error_errno(unit, r, "Failed to acquire cgroup path: %m");
@@ -4880,11 +4890,26 @@ static int exec_child(
                  * touch a single hierarchy too. */
 
                 if (params->flags & EXEC_CGROUP_DELEGATE) {
+                        _cleanup_free_ char *p = NULL;
+
                         r = cg_set_access(SYSTEMD_CGROUP_CONTROLLER, params->cgroup_path, uid, gid);
                         if (r < 0) {
                                 *exit_status = EXIT_CGROUP;
                                 return log_unit_error_errno(unit, r, "Failed to adjust control group access: %m");
                         }
+
+                        r = exec_parameters_get_cgroup_path(params, cgroup_context, &p);
+                        if (r < 0) {
+                                *exit_status = EXIT_CGROUP;
+                                return log_unit_error_errno(unit, r, "Failed to acquire cgroup path: %m");
+                        }
+                        if (r > 0) {
+                                r = cg_set_access(SYSTEMD_CGROUP_CONTROLLER, p, uid, gid);
+                                if (r < 0) {
+                                        *exit_status = EXIT_CGROUP;
+                                        return log_unit_error_errno(unit, r, "Failed to adjust control subgroup access: %m");
+                                }
+                        }
                 }
 
                 if (cgroup_context && cg_unified() > 0 && is_pressure_supported() > 0) {
@@ -5635,13 +5660,13 @@ int exec_spawn(Unit *unit,
         log_command_line(unit, "About to execute", command->path, command->argv);
 
         if (params->cgroup_path) {
-                r = exec_parameters_get_cgroup_path(params, &subcgroup_path);
+                r = exec_parameters_get_cgroup_path(params, cgroup_context, &subcgroup_path);
                 if (r < 0)
                         return log_unit_error_errno(unit, r, "Failed to acquire subcgroup path: %m");
                 if (r > 0) { /* We are using a child cgroup */
                         r = cg_create(SYSTEMD_CGROUP_CONTROLLER, subcgroup_path);
                         if (r < 0)
-                                return log_unit_error_errno(unit, r, "Failed to create control group '%s': %m", subcgroup_path);
+                                return log_unit_error_errno(unit, r, "Failed to create subcgroup '%s': %m", subcgroup_path);
 
                         /* Normally we would not propagate the xattrs to children but since we created this
                          * sub-cgroup internally we should do it. */

src/core/load-fragment-gperf.gperf.in

@@ -237,6 +237,7 @@
 {{type}}.TasksAccounting,                  config_parse_bool,                           0,                                  offsetof({{type}}, cgroup_context.tasks_accounting)
 {{type}}.TasksMax,                         config_parse_tasks_max,                      0,                                  offsetof({{type}}, cgroup_context.tasks_max)
 {{type}}.Delegate,                         config_parse_delegate,                       0,                                  offsetof({{type}}, cgroup_context)
+{{type}}.DelegateSubgroup,                 config_parse_delegate_subgroup ,             0,                                  offsetof({{type}}, cgroup_context)
 {{type}}.DisableControllers,               config_parse_disable_controllers,            0,                                  offsetof({{type}}, cgroup_context)
 {{type}}.IPAccounting,                     config_parse_bool,                           0,                                  offsetof({{type}}, cgroup_context.ip_accounting)
 {{type}}.IPAddressAllow,                   config_parse_in_addr_prefixes,               AF_UNSPEC,                          offsetof({{type}}, cgroup_context.ip_address_allow)

src/core/load-fragment.c

@@ -4029,6 +4029,42 @@ int config_parse_delegate(
         return 0;
 }
 
+int config_parse_delegate_subgroup(
+                const char *unit,
+                const char *filename,
+                unsigned line,
+                const char *section,
+                unsigned section_line,
+                const char *lvalue,
+                int ltype,
+                const char *rvalue,
+                void *data,
+                void *userdata) {
+
+        CGroupContext *c = ASSERT_PTR(data);
+        UnitType t;
+
+        t = unit_name_to_type(unit);
+        assert(t >= 0);
+
+        if (!unit_vtable[t]->can_delegate) {
+                log_syntax(unit, LOG_WARNING, filename, line, 0, "DelegateSubgroup= setting not supported for this unit type, ignoring.");
+                return 0;
+        }
+
+        if (isempty(rvalue)) {
+                c->delegate_subgroup = mfree(c->delegate_subgroup);
+                return 0;
+        }
+
+        if (cg_needs_escape(rvalue)) { /* Insist that specified names don't need escaping */
+                log_syntax(unit, LOG_WARNING, filename, line, 0, "Invalid control group name, ignoring: %s", rvalue);
+                return 0;
+        }
+
+        return free_and_strdup_warn(&c->delegate_subgroup, rvalue);
+}
+
 int config_parse_managed_oom_mode(
                 const char *unit,
                 const char *filename,

src/core/load-fragment.h

@@ -83,6 +83,7 @@ CONFIG_PARSER_PROTOTYPE(config_parse_cpu_shares);
 CONFIG_PARSER_PROTOTYPE(config_parse_memory_limit);
 CONFIG_PARSER_PROTOTYPE(config_parse_tasks_max);
 CONFIG_PARSER_PROTOTYPE(config_parse_delegate);
+CONFIG_PARSER_PROTOTYPE(config_parse_delegate_subgroup);
 CONFIG_PARSER_PROTOTYPE(config_parse_managed_oom_mode);
 CONFIG_PARSER_PROTOTYPE(config_parse_managed_oom_mem_pressure_limit);
 CONFIG_PARSER_PROTOTYPE(config_parse_managed_oom_preference);