nspawn: Move cgroup mount stuff from nspawn-mount.c to nspawn-cgroup.c

1 files changed, 417 insertions, 0 deletions

diff --git a/src/nspawn/nspawn-cgroup.c b/src/nspawn/nspawn-cgroup.c
index d8a39a6959..0be911999a 100644
--- a/src/nspawn/nspawn-cgroup.c
+++ b/src/nspawn/nspawn-cgroup.c
@@ -5,12 +5,16 @@
 #include "alloc-util.h"
 #include "fd-util.h"
 #include "fileio.h"
+#include "fs-util.h"
 #include "mkdir.h"
 #include "mount-util.h"
 #include "nspawn-cgroup.h"
+#include "nspawn-mount.h"
+#include "path-util.h"
 #include "rm-rf.h"
 #include "string-util.h"
 #include "strv.h"
+#include "user-util.h"
 #include "util.h"
 
 static int chown_cgroup_path(const char *path, uid_t uid_shift) {
@@ -188,3 +192,416 @@ int create_subcgroup(pid_t pid, bool keep_unit, CGroupUnified unified_requested)
         (void) cg_enable_everywhere(supported, supported, cgroup);
         return 0;
 }
+
+/* Retrieve existing subsystems. This function is called in a new cgroup
+ * namespace.
+ */
+static int get_process_controllers(Set **ret) {
+        _cleanup_set_free_free_ Set *controllers = NULL;
+        _cleanup_fclose_ FILE *f = NULL;
+        int r;
+
+        assert(ret);
+
+        controllers = set_new(&string_hash_ops);
+        if (!controllers)
+                return -ENOMEM;
+
+        f = fopen("/proc/self/cgroup", "re");
+        if (!f)
+                return errno == ENOENT ? -ESRCH : -errno;
+
+        for (;;) {
+                _cleanup_free_ char *line = NULL;
+                char *e, *l;
+
+                r = read_line(f, LONG_LINE_MAX, &line);
+                if (r < 0)
+                        return r;
+                if (r == 0)
+                        break;
+
+                l = strchr(line, ':');
+                if (!l)
+                        continue;
+
+                l++;
+                e = strchr(l, ':');
+                if (!e)
+                        continue;
+
+                *e = 0;
+
+                if (STR_IN_SET(l, "", "name=systemd", "name=unified"))
+                        continue;
+
+                r = set_put_strdup(controllers, l);
+                if (r < 0)
+                        return r;
+        }
+
+        *ret = TAKE_PTR(controllers);
+
+        return 0;
+}
+
+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;
+}
+
+/* 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);
+                        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;
+}
+
+/* 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);
+                        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");
+}
+
+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);
+}
+
+int mount_cgroups(
+                const char *dest,
+                CGroupUnified unified_requested,
+                bool userns,
+                uid_t uid_shift,
+                uid_t uid_range,
+                const char *selinux_apifs_context,
+                bool use_cgns) {
+
+        if (unified_requested >= CGROUP_UNIFIED_ALL)
+                return mount_unified_cgroups(dest);
+        if (use_cgns)
+                return mount_legacy_cgns_supported(dest, unified_requested, userns, uid_shift, uid_range, selinux_apifs_context);
+
+        return mount_legacy_cgns_unsupported(dest, unified_requested, userns, uid_shift, uid_range, selinux_apifs_context);
+}
+
+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);
+}
+
+int mount_systemd_cgroup_writable(
+                const char *dest,
+                CGroupUnified unified_requested) {
+
+        _cleanup_free_ char *own_cgroup_path = NULL;
+        const char *root, *own;
+        int r;
+
+        assert(dest);
+
+        r = cg_pid_get_path(NULL, 0, &own_cgroup_path);
+        if (r < 0)
+                return log_error_errno(r, "Failed to determine our own cgroup path: %m");
+
+        /* If we are living in the top-level, then there's nothing to do... */
+        if (path_equal(own_cgroup_path, "/"))
+                return 0;
+
+        if (unified_requested >= CGROUP_UNIFIED_ALL) {
+
+                root = prefix_roota(dest, "/sys/fs/cgroup");
+                own = strjoina(root, own_cgroup_path);
+
+        } else {
+
+                if (unified_requested >= CGROUP_UNIFIED_SYSTEMD) {
+                        root = prefix_roota(dest, "/sys/fs/cgroup/unified");
+                        own = strjoina(root, own_cgroup_path);
+
+                        r = mount_systemd_cgroup_writable_one(root, own);
+                        if (r < 0)
+                                return r;
+                }
+
+                root = prefix_roota(dest, "/sys/fs/cgroup/systemd");
+                own = strjoina(root, own_cgroup_path);
+        }
+
+        return mount_systemd_cgroup_writable_one(root, own);
+}