From 44d19e4beb486969b25c07f7efdee2e8b8bcf986 Mon Sep 17 00:00:00 2001
From: Lennart Poettering <lennart@poettering.net>
Date: Fri, 16 Nov 2018 20:19:07 +0100
Subject: [PATCH] core: when Delegate=yes is set for a unit, run ExecStartPre=
 and friends in a subcgroup of the unit

Otherwise we might conflict with the "no-processes-in-inner-cgroup" rule
of cgroupsv2. Consider nspawn starting up and initializing its cgroup
hierarchy with "supervisor/" and "payload/" as subcgroup, with itself
moved into the former and the payload into the latter. Now, if an
ExecStartPre= is run right after it cannot be placed in the main cgroup,
because that is now in inner cgroup with populated children.

Hence, let's run these helpers in another sub-cgroup .control/ below it.

This is somewhat ugly since it weakens the clear separation of
ownership, but given that this is an explicit contract, and double opt-in should be acceptable.

Fixes: #10482
(cherry picked from commit 78f93209fc7f61f15b12d7a5f74d712bd020b249)

Resolves: #2215925
---
 src/core/execute.c | 67 +++++++++++++++++++++++++++++++++++++++-------
 src/core/execute.h |  9 ++++---
 src/core/service.c | 13 ++++-----
 3 files changed, 70 insertions(+), 19 deletions(-)

diff --git a/src/core/execute.c b/src/core/execute.c
index b1d8dceb32..7e186c948c 100644
--- a/src/core/execute.c
+++ b/src/core/execute.c
@@ -2834,6 +2834,37 @@ bool exec_context_get_cpu_affinity_from_numa(const ExecContext *c) {
         return c->cpu_affinity_from_numa;
 }
 
+static int exec_parameters_get_cgroup_path(const ExecParameters *params, char **ret) {
+        bool using_subcgroup;
+        char *p;
+
+        assert(params);
+        assert(ret);
+
+        if (!params->cgroup_path)
+                return -EINVAL;
+
+        /* If we are called for a unit where cgroup delegation is on, and the payload created its own populated
+         * subcgroup (which we expect it to do, after all it asked for delegation), then we cannot place the control
+         * processes started after the main unit's process in the unit's main cgroup because it is now an inner one,
+         * and inner cgroups may not contain processes. Hence, if delegation is on, and this is a control process,
+         * let's use ".control" as subcgroup instead. Note that we do so only for ExecStartPost=, ExecReload=,
+         * ExecStop=, ExecStopPost=, i.e. for the commands where the main process is already forked. For ExecStartPre=
+         * 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 = strjoin(params->cgroup_path, "/.control");
+        else
+                p = strdup(params->cgroup_path);
+        if (!p)
+                return -ENOMEM;
+
+        *ret = p;
+        return using_subcgroup;
+}
+
 static int exec_child(
                 Unit *unit,
                 const ExecCommand *command,
@@ -3055,10 +3086,18 @@ static int exec_child(
         }
 
         if (params->cgroup_path) {
-                r = cg_attach_everywhere(params->cgroup_supported, params->cgroup_path, 0, NULL, NULL);
+                _cleanup_free_ char *p = NULL;
+
+                r = exec_parameters_get_cgroup_path(params, &p);
                 if (r < 0) {
                         *exit_status = EXIT_CGROUP;
-                        return log_unit_error_errno(unit, r, "Failed to attach to cgroup %s: %m", params->cgroup_path);
+                        return log_unit_error_errno(unit, r, "Failed to acquire cgroup path: %m");
+                }
+
+                r = cg_attach_everywhere(params->cgroup_supported, p, 0, NULL, NULL);
+                if (r < 0) {
+                        *exit_status = EXIT_CGROUP;
+                        return log_unit_error_errno(unit, r, "Failed to attach to cgroup %s: %m", p);
                 }
         }
 
@@ -3659,6 +3698,7 @@ int exec_spawn(Unit *unit,
                DynamicCreds *dcreds,
                pid_t *ret) {
 
+        _cleanup_free_ char *subcgroup_path = NULL;
         _cleanup_strv_free_ char **files_env = NULL;
         int *fds = NULL;
         size_t n_storage_fds = 0, n_socket_fds = 0;
@@ -3716,6 +3756,17 @@ int exec_spawn(Unit *unit,
                    LOG_UNIT_ID(unit),
                    LOG_UNIT_INVOCATION_ID(unit));
 
+        if (params->cgroup_path) {
+                r = exec_parameters_get_cgroup_path(params, &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);
+                }
+        }
+
         pid = fork();
         if (pid < 0)
                 return log_unit_error_errno(unit, errno, "Failed to fork: %m");
@@ -3754,13 +3805,11 @@ int exec_spawn(Unit *unit,
 
         log_unit_debug(unit, "Forked %s as "PID_FMT, command->path, pid);
 
-        /* We add the new process to the cgroup both in the child (so
-         * that we can be sure that no user code is ever executed
-         * outside of the cgroup) and in the parent (so that we can be
-         * sure that when we kill the cgroup the process will be
-         * killed too). */
-        if (params->cgroup_path)
-                (void) cg_attach(SYSTEMD_CGROUP_CONTROLLER, params->cgroup_path, pid);
+        /* We add the new process to the cgroup both in the child (so that we can be sure that no user code is ever
+         * executed outside of the cgroup) and in the parent (so that we can be sure that when we kill the cgroup the
+         * process will be killed too). */
+        if (subcgroup_path)
+                (void) cg_attach(SYSTEMD_CGROUP_CONTROLLER, subcgroup_path, pid);
 
         exec_status_start(&command->exec_status, pid);
 
diff --git a/src/core/execute.h b/src/core/execute.h
index 62c6229621..f5bc180ece 100644
--- a/src/core/execute.h
+++ b/src/core/execute.h
@@ -289,12 +289,13 @@ typedef enum ExecFlags {
         EXEC_CHOWN_DIRECTORIES = 1 << 5, /* chown() the runtime/state/cache/log directories to the user we run as, under all conditions */
         EXEC_NSS_BYPASS_BUS    = 1 << 6, /* Set the SYSTEMD_NSS_BYPASS_BUS environment variable, to disable nss-systemd for dbus */
         EXEC_CGROUP_DELEGATE   = 1 << 7,
+        EXEC_IS_CONTROL        = 1 << 8,
+        EXEC_CONTROL_CGROUP    = 1 << 9, /* Place the process not in the indicated cgroup but in a subcgroup '/.control', but only EXEC_CGROUP_DELEGATE and EXEC_IS_CONTROL is set, too */
 
         /* The following are not used by execute.c, but by consumers internally */
-        EXEC_PASS_FDS          = 1 << 8,
-        EXEC_IS_CONTROL        = 1 << 9,
-        EXEC_SETENV_RESULT     = 1 << 10,
-        EXEC_SET_WATCHDOG      = 1 << 11,
+        EXEC_PASS_FDS          = 1 << 10,
+        EXEC_SETENV_RESULT     = 1 << 11,
+        EXEC_SET_WATCHDOG      = 1 << 12,
 } ExecFlags;
 
 struct ExecParameters {
diff --git a/src/core/service.c b/src/core/service.c
index e05d0e0514..0423f2c73e 100644
--- a/src/core/service.c
+++ b/src/core/service.c
@@ -1429,7 +1429,7 @@ static int service_spawn(
         assert(c);
         assert(_pid);
 
-        r = unit_prepare_exec(UNIT(s));
+        r = unit_prepare_exec(UNIT(s)); /* This realizes the cgroup, among other things */
         if (r < 0)
                 return r;
 
@@ -1798,7 +1798,7 @@ static void service_enter_stop_post(Service *s, ServiceResult f) {
                 r = service_spawn(s,
                                   s->control_command,
                                   s->timeout_stop_usec,
-                                  EXEC_APPLY_SANDBOXING|EXEC_APPLY_CHROOT|EXEC_APPLY_TTY_STDIN|EXEC_IS_CONTROL|EXEC_SETENV_RESULT,
+                                  EXEC_APPLY_SANDBOXING|EXEC_APPLY_CHROOT|EXEC_APPLY_TTY_STDIN|EXEC_IS_CONTROL|EXEC_SETENV_RESULT|EXEC_CONTROL_CGROUP,
                                   &s->control_pid);
                 if (r < 0)
                         goto fail;
@@ -1913,7 +1913,7 @@ static void service_enter_stop(Service *s, ServiceResult f) {
                 r = service_spawn(s,
                                   s->control_command,
                                   s->timeout_stop_usec,
-                                  EXEC_APPLY_SANDBOXING|EXEC_APPLY_CHROOT|EXEC_IS_CONTROL|EXEC_SETENV_RESULT,
+                                  EXEC_APPLY_SANDBOXING|EXEC_APPLY_CHROOT|EXEC_IS_CONTROL|EXEC_SETENV_RESULT|EXEC_CONTROL_CGROUP,
                                   &s->control_pid);
                 if (r < 0)
                         goto fail;
@@ -1991,7 +1991,7 @@ static void service_enter_start_post(Service *s) {
                 r = service_spawn(s,
                                   s->control_command,
                                   s->timeout_start_usec,
-                                  EXEC_APPLY_SANDBOXING|EXEC_APPLY_CHROOT|EXEC_IS_CONTROL,
+                                  EXEC_APPLY_SANDBOXING|EXEC_APPLY_CHROOT|EXEC_IS_CONTROL|EXEC_CONTROL_CGROUP,
                                   &s->control_pid);
                 if (r < 0)
                         goto fail;
@@ -2269,7 +2269,7 @@ static void service_enter_reload(Service *s) {
                 r = service_spawn(s,
                                   s->control_command,
                                   s->timeout_start_usec,
-                                  EXEC_APPLY_SANDBOXING|EXEC_APPLY_CHROOT|EXEC_IS_CONTROL,
+                                  EXEC_APPLY_SANDBOXING|EXEC_APPLY_CHROOT|EXEC_IS_CONTROL|EXEC_CONTROL_CGROUP,
                                   &s->control_pid);
                 if (r < 0)
                         goto fail;
@@ -2309,7 +2309,8 @@ static void service_run_next_control(Service *s) {
                           timeout,
                           EXEC_APPLY_SANDBOXING|EXEC_APPLY_CHROOT|EXEC_IS_CONTROL|
                           (IN_SET(s->control_command_id, SERVICE_EXEC_CONDITION, SERVICE_EXEC_START_PRE, SERVICE_EXEC_STOP_POST) ? EXEC_APPLY_TTY_STDIN : 0)|
-                          (IN_SET(s->control_command_id, SERVICE_EXEC_STOP, SERVICE_EXEC_STOP_POST) ? EXEC_SETENV_RESULT : 0),
+                          (IN_SET(s->control_command_id, SERVICE_EXEC_STOP, SERVICE_EXEC_STOP_POST) ? EXEC_SETENV_RESULT : 0)|
+                          (IN_SET(s->control_command_id, SERVICE_EXEC_START_POST, SERVICE_EXEC_RELOAD, SERVICE_EXEC_STOP, SERVICE_EXEC_STOP_POST) ? EXEC_CONTROL_CGROUP : 0),
                           &s->control_pid);
         if (r < 0)
                 goto fail;