int sync_cgroup(pid_t pid, CGroupUnified unified_requested) {
_cleanup_free_ char *cgroup = NULL;
char tree[] = "/tmp/unifiedXXXXXX", pid_string[DECIMAL_STR_MAX(pid) + 1];
bool undo_mount = false;
const char *fn;
int unified, r;
unified = cg_unified(SYSTEMD_CGROUP_CONTROLLER);
if (unified < 0)
return log_error_errno(unified, "Failed to determine whether the unified hierarchy is used: %m");
if ((unified > 0) == (unified_requested >= CGROUP_UNIFIED_SYSTEMD))
return 0;
/* When the host uses the legacy cgroup setup, but the
* container shall use the unified hierarchy, let's make sure
* we copy the path from the name=systemd hierarchy into the
* unified hierarchy. Similar for the reverse situation. */
r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, pid, &cgroup);
if (r < 0)
return log_error_errno(r, "Failed to get control group of " PID_FMT ": %m", pid);
/* In order to access the unified hierarchy we need to mount it */
if (!mkdtemp(tree))
return log_error_errno(errno, "Failed to generate temporary mount point for unified hierarchy: %m");
if (unified)
r = mount_verbose(LOG_ERR, "cgroup", tree, "cgroup",
MS_NOSUID|MS_NOEXEC|MS_NODEV, "none,name=systemd,xattr");
else
r = mount_verbose(LOG_ERR, "cgroup", tree, "cgroup2",
MS_NOSUID|MS_NOEXEC|MS_NODEV, NULL);
if (r < 0)
goto finish;
undo_mount = true;
fn = strjoina(tree, cgroup, "/cgroup.procs");
(void) mkdir_parents(fn, 0755);
sprintf(pid_string, PID_FMT, pid);
r = write_string_file(fn, pid_string, 0);
if (r < 0)
log_error_errno(r, "Failed to move process: %m");
finish:
if (undo_mount)
(void) umount_verbose(tree);
(void) rmdir(tree);
return r;
}
static int make_volatile(const char *path) {
_cleanup_free_ char *old_usr = NULL;
int r;
assert(path);
r = chase_symlinks("/usr", path, CHASE_PREFIX_ROOT, &old_usr);
if (r < 0)
return log_error_errno(r, "/usr not available in old root: %m");
r = mkdir_p("/run/systemd/volatile-sysroot", 0700);
if (r < 0)
return log_error_errno(r, "Couldn't generate volatile sysroot directory: %m");
r = mount_verbose(LOG_ERR, "tmpfs", "/run/systemd/volatile-sysroot", "tmpfs", MS_STRICTATIME, "mode=755");
if (r < 0)
goto finish_rmdir;
if (mkdir("/run/systemd/volatile-sysroot/usr", 0755) < 0) {
r = log_error_errno(errno, "Failed to create /usr directory: %m");
goto finish_umount;
}
r = mount_verbose(LOG_ERR, old_usr, "/run/systemd/volatile-sysroot/usr", NULL, MS_BIND|MS_REC, NULL);
if (r < 0)
goto finish_umount;
r = bind_remount_recursive("/run/systemd/volatile-sysroot/usr", true, NULL);
if (r < 0) {
log_error_errno(r, "Failed to remount /usr read-only: %m");
goto finish_umount;
}
r = umount_recursive(path, 0);
if (r < 0) {
log_error_errno(r, "Failed to unmount %s: %m", path);
goto finish_umount;
}
if (mount(NULL, "/", NULL, MS_SLAVE|MS_REC, NULL) < 0)
log_warning_errno(errno, "Failed to remount %s MS_SLAVE|MS_REC, ignoring: %m", path);
r = mount_verbose(LOG_ERR, "/run/systemd/volatile-sysroot", path, NULL, MS_MOVE, NULL);
finish_umount:
(void) umount_recursive("/run/systemd/volatile-sysroot", 0);
finish_rmdir:
(void) rmdir("/run/systemd/volatile-sysroot");
return r;
}
static int mount_systemd_cgroup_writable_one(const char *root, const char *own) {
int r;
assert(root);
assert(own);
/* Make our own cgroup a (writable) bind mount */
r = mount_verbose(LOG_ERR, own, own, NULL, MS_BIND, NULL);
if (r < 0)
return r;
/* And then remount the systemd cgroup root read-only */
return mount_verbose(LOG_ERR, NULL, root, NULL,
MS_BIND|MS_REMOUNT|MS_NOSUID|MS_NOEXEC|MS_NODEV|MS_RDONLY, NULL);
}
static int mount_unified_cgroups(const char *dest) {
const char *p;
int r;
assert(dest);
p = prefix_roota(dest, "/sys/fs/cgroup");
(void) mkdir_p(p, 0755);
r = path_is_mount_point(p, dest, AT_SYMLINK_FOLLOW);
if (r < 0)
return log_error_errno(r, "Failed to determine if %s is mounted already: %m", p);
if (r > 0) {
p = prefix_roota(dest, "/sys/fs/cgroup/cgroup.procs");
if (access(p, F_OK) >= 0)
return 0;
if (errno != ENOENT)
return log_error_errno(errno, "Failed to determine if mount point %s contains the unified cgroup hierarchy: %m", p);
log_error("%s is already mounted but not a unified cgroup hierarchy. Refusing.", p);
return -EINVAL;
}
return mount_verbose(LOG_ERR, "cgroup", p, "cgroup2", MS_NOSUID|MS_NOEXEC|MS_NODEV, NULL);
}
static int mount_legacy_cgroup_hierarchy(
const char *dest,
const char *controller,
const char *hierarchy,
bool read_only) {
const char *to, *fstype, *opts;
int r;
to = strjoina(strempty(dest), "/sys/fs/cgroup/", hierarchy);
r = path_is_mount_point(to, dest, 0);
if (r < 0 && r != -ENOENT)
return log_error_errno(r, "Failed to determine if %s is mounted already: %m", to);
if (r > 0)
return 0;
mkdir_p(to, 0755);
/* The superblock mount options of the mount point need to be
* identical to the hosts', and hence writable... */
if (streq(controller, SYSTEMD_CGROUP_CONTROLLER_HYBRID)) {
fstype = "cgroup2";
opts = NULL;
} else if (streq(controller, SYSTEMD_CGROUP_CONTROLLER_LEGACY)) {
fstype = "cgroup";
opts = "none,name=systemd,xattr";
} else {
fstype = "cgroup";
opts = controller;
}
r = mount_verbose(LOG_ERR, "cgroup", to, fstype, MS_NOSUID|MS_NOEXEC|MS_NODEV, opts);
if (r < 0)
return r;
/* ... hence let's only make the bind mount read-only, not the superblock. */
if (read_only) {
r = mount_verbose(LOG_ERR, NULL, to, NULL,
MS_BIND|MS_REMOUNT|MS_NOSUID|MS_NOEXEC|MS_NODEV|MS_RDONLY, NULL);
if (r < 0)
return r;
}
return 1;
}
static int make_overlay(const char *path) {
_cleanup_free_ char *escaped_path = NULL;
bool tmpfs_mounted = false;
const char *options = NULL;
int r;
assert(path);
r = mkdir_p("/run/systemd/overlay-sysroot", 0700);
if (r < 0)
return log_error_errno(r, "Couldn't create overlay sysroot directory: %m");
r = mount_verbose(LOG_ERR, "tmpfs", "/run/systemd/overlay-sysroot", "tmpfs", MS_STRICTATIME, "mode=755");
if (r < 0)
goto finish;
tmpfs_mounted = true;
if (mkdir("/run/systemd/overlay-sysroot/upper", 0755) < 0) {
r = log_error_errno(errno, "Failed to create /run/systemd/overlay-sysroot/upper: %m");
goto finish;
}
if (mkdir("/run/systemd/overlay-sysroot/work", 0755) < 0) {
r = log_error_errno(errno, "Failed to create /run/systemd/overlay-sysroot/work: %m");
goto finish;
}
escaped_path = shell_escape(path, ",:");
if (!escaped_path) {
r = log_oom();
goto finish;
}
options = strjoina("lowerdir=", escaped_path, ",upperdir=/run/systemd/overlay-sysroot/upper,workdir=/run/systemd/overlay-sysroot/work");
r = mount_verbose(LOG_ERR, "overlay", path, "overlay", 0, options);
finish:
if (tmpfs_mounted)
(void) umount_verbose("/run/systemd/overlay-sysroot");
(void) rmdir("/run/systemd/overlay-sysroot");
return r;
}
int sync_cgroup(pid_t pid, CGroupUnified unified_requested, uid_t uid_shift) {
_cleanup_free_ char *cgroup = NULL;
char tree[] = "/tmp/unifiedXXXXXX", pid_string[DECIMAL_STR_MAX(pid) + 1];
bool undo_mount = false;
const char *fn;
int r, unified_controller;
unified_controller = cg_unified_controller(SYSTEMD_CGROUP_CONTROLLER);
if (unified_controller < 0)
return log_error_errno(unified_controller, "Failed to determine whether the systemd hierarchy is unified: %m");
if ((unified_controller > 0) == (unified_requested >= CGROUP_UNIFIED_SYSTEMD))
return 0;
/* When the host uses the legacy cgroup setup, but the
* container shall use the unified hierarchy, let's make sure
* we copy the path from the name=systemd hierarchy into the
* unified hierarchy. Similar for the reverse situation. */
r = cg_pid_get_path(SYSTEMD_CGROUP_CONTROLLER, pid, &cgroup);
if (r < 0)
return log_error_errno(r, "Failed to get control group of " PID_FMT ": %m", pid);
/* In order to access the unified hierarchy we need to mount it */
if (!mkdtemp(tree))
return log_error_errno(errno, "Failed to generate temporary mount point for unified hierarchy: %m");
if (unified_controller > 0)
r = mount_verbose(LOG_ERR, "cgroup", tree, "cgroup",
MS_NOSUID|MS_NOEXEC|MS_NODEV, "none,name=systemd,xattr");
else
r = mount_verbose(LOG_ERR, "cgroup", tree, "cgroup2",
MS_NOSUID|MS_NOEXEC|MS_NODEV, NULL);
if (r < 0)
goto finish;
undo_mount = true;
/* If nspawn dies abruptly the cgroup hierarchy created below
* its unit isn't cleaned up. So, let's remove it
* https://github.com/systemd/systemd/pull/4223#issuecomment-252519810 */
fn = strjoina(tree, cgroup);
(void) rm_rf(fn, REMOVE_ROOT|REMOVE_ONLY_DIRECTORIES);
fn = strjoina(tree, cgroup, "/cgroup.procs");
(void) mkdir_parents(fn, 0755);
sprintf(pid_string, PID_FMT, pid);
r = write_string_file(fn, pid_string, 0);
if (r < 0) {
log_error_errno(r, "Failed to move process: %m");
goto finish;
}
fn = strjoina(tree, cgroup);
r = chown_cgroup_path(fn, uid_shift);
if (r < 0)
log_error_errno(r, "Failed to chown() cgroup %s: %m", fn);
finish:
if (undo_mount)
(void) umount_verbose(tree);
(void) rmdir(tree);
return r;
}
/* Mount legacy cgroup hierarchy when cgroup namespaces are unsupported. */
static int mount_legacy_cgns_unsupported(
const char *dest,
CGroupUnified unified_requested,
bool userns,
uid_t uid_shift,
uid_t uid_range,
const char *selinux_apifs_context) {
_cleanup_set_free_free_ Set *controllers = NULL;
const char *cgroup_root;
int r;
cgroup_root = prefix_roota(dest, "/sys/fs/cgroup");
(void) mkdir_p(cgroup_root, 0755);
/* Mount a tmpfs to /sys/fs/cgroup if it's not mounted there yet. */
r = path_is_mount_point(cgroup_root, dest, AT_SYMLINK_FOLLOW);
if (r < 0)
return log_error_errno(r, "Failed to determine if /sys/fs/cgroup is already mounted: %m");
if (r == 0) {
_cleanup_free_ char *options = NULL;
r = tmpfs_patch_options("mode=755", uid_shift == 0 ? UID_INVALID : uid_shift, selinux_apifs_context, &options);
if (r < 0)
return log_oom();
r = mount_verbose(LOG_ERR, "tmpfs", cgroup_root, "tmpfs",
MS_NOSUID|MS_NOEXEC|MS_NODEV|MS_STRICTATIME, options);
if (r < 0)
return r;
}
r = cg_all_unified();
if (r < 0)
return r;
if (r > 0)
goto skip_controllers;
r = cg_kernel_controllers(&controllers);
if (r < 0)
return log_error_errno(r, "Failed to determine cgroup controllers: %m");
for (;;) {
_cleanup_free_ char *controller = NULL, *origin = NULL, *combined = NULL;
controller = set_steal_first(controllers);
if (!controller)
break;
origin = prefix_root("/sys/fs/cgroup/", controller);
if (!origin)
return log_oom();
r = readlink_malloc(origin, &combined);
if (r == -EINVAL) {
/* Not a symbolic link, but directly a single cgroup hierarchy */
r = mount_legacy_cgroup_hierarchy(dest, controller, controller, true);
if (r < 0)
return r;
} else if (r < 0)
return log_error_errno(r, "Failed to read link %s: %m", origin);
else {
_cleanup_free_ char *target = NULL;
target = prefix_root(dest, origin);
if (!target)
return log_oom();
/* A symbolic link, a combination of controllers in one hierarchy */
if (!filename_is_valid(combined)) {
log_warning("Ignoring invalid combined hierarchy %s.", combined);
continue;
}
r = mount_legacy_cgroup_hierarchy(dest, combined, combined, true);
if (r < 0)
return r;
r = symlink_idempotent(combined, target, false);
if (r == -EINVAL)
return log_error_errno(r, "Invalid existing symlink for combined hierarchy: %m");
if (r < 0)
return log_error_errno(r, "Failed to create symlink for combined hierarchy: %m");
}
}
skip_controllers:
if (unified_requested >= CGROUP_UNIFIED_SYSTEMD) {
r = mount_legacy_cgroup_hierarchy(dest, SYSTEMD_CGROUP_CONTROLLER_HYBRID, "unified", false);
if (r < 0)
return r;
}
r = mount_legacy_cgroup_hierarchy(dest, SYSTEMD_CGROUP_CONTROLLER_LEGACY, "systemd", false);
if (r < 0)
return r;
return mount_verbose(LOG_ERR, NULL, cgroup_root, NULL,
MS_REMOUNT|MS_NOSUID|MS_NOEXEC|MS_NODEV|MS_STRICTATIME|MS_RDONLY, "mode=755");
}
/* Mount a legacy cgroup hierarchy when cgroup namespaces are supported. */
static int mount_legacy_cgns_supported(
const char *dest,
CGroupUnified unified_requested,
bool userns,
uid_t uid_shift,
uid_t uid_range,
const char *selinux_apifs_context) {
_cleanup_set_free_free_ Set *controllers = NULL;
const char *cgroup_root = "/sys/fs/cgroup", *c;
int r;
(void) mkdir_p(cgroup_root, 0755);
/* Mount a tmpfs to /sys/fs/cgroup if it's not mounted there yet. */
r = path_is_mount_point(cgroup_root, dest, AT_SYMLINK_FOLLOW);
if (r < 0)
return log_error_errno(r, "Failed to determine if /sys/fs/cgroup is already mounted: %m");
if (r == 0) {
_cleanup_free_ char *options = NULL;
/* When cgroup namespaces are enabled and user namespaces are
* used then the mount of the cgroupfs is done *inside* the new
* user namespace. We're root in the new user namespace and the
* kernel will happily translate our uid/gid to the correct
* uid/gid as seen from e.g. /proc/1/mountinfo. So we simply
* pass uid 0 and not uid_shift to tmpfs_patch_options().
*/
r = tmpfs_patch_options("mode=755", 0, selinux_apifs_context, &options);
if (r < 0)
return log_oom();
r = mount_verbose(LOG_ERR, "tmpfs", cgroup_root, "tmpfs",
MS_NOSUID|MS_NOEXEC|MS_NODEV|MS_STRICTATIME, options);
if (r < 0)
return r;
}
r = cg_all_unified();
if (r < 0)
return r;
if (r > 0)
goto skip_controllers;
r = get_process_controllers(&controllers);
if (r < 0)
return log_error_errno(r, "Failed to determine cgroup controllers: %m");
for (;;) {
_cleanup_free_ const char *controller = NULL;
controller = set_steal_first(controllers);
if (!controller)
break;
r = mount_legacy_cgroup_hierarchy("", controller, controller, !userns);
if (r < 0)
return r;
/* When multiple hierarchies are co-mounted, make their
* constituting individual hierarchies a symlink to the
* co-mount.
*/
c = controller;
for (;;) {
_cleanup_free_ char *target = NULL, *tok = NULL;
r = extract_first_word(&c, &tok, ",", 0);
if (r < 0)
return log_error_errno(r, "Failed to extract co-mounted cgroup controller: %m");
if (r == 0)
break;
if (streq(controller, tok))
break;
target = prefix_root("/sys/fs/cgroup/", tok);
if (!target)
return log_oom();
r = symlink_idempotent(controller, target, false);
if (r == -EINVAL)
return log_error_errno(r, "Invalid existing symlink for combined hierarchy: %m");
if (r < 0)
return log_error_errno(r, "Failed to create symlink for combined hierarchy: %m");
}
}
skip_controllers:
if (unified_requested >= CGROUP_UNIFIED_SYSTEMD) {
r = mount_legacy_cgroup_hierarchy("", SYSTEMD_CGROUP_CONTROLLER_HYBRID, "unified", false);
if (r < 0)
return r;
}
r = mount_legacy_cgroup_hierarchy("", SYSTEMD_CGROUP_CONTROLLER_LEGACY, "systemd", false);
if (r < 0)
return r;
if (!userns)
return mount_verbose(LOG_ERR, NULL, cgroup_root, NULL,
MS_REMOUNT|MS_NOSUID|MS_NOEXEC|MS_NODEV|MS_STRICTATIME|MS_RDONLY, "mode=755");
return 0;
}