Merge pull request #13184 from poettering/nspawn-usr-root

Allow "systemd-nspawn -D / --volatile=yes" to work

2 files changed, 65 insertions, 18 deletions

diff --git a/src/nspawn/nspawn.c b/src/nspawn/nspawn.c
index fed5c651ce..2aec8041f0 100644
--- a/src/nspawn/nspawn.c
+++ b/src/nspawn/nspawn.c
@@ -3264,8 +3264,24 @@ static int outer_child(
                          "Selected user namespace base " UID_FMT " and range " UID_FMT ".", arg_uid_shift, arg_uid_range);
         }
 
-        if (!dissected_image) {
-                /* Turn directory into bind mount */
+        if (path_equal(directory, "/")) {
+                /* If the directory we shall boot is the host, let's operate on a bind mount at a different
+                 * place, so that we can make changes to its mount structure (for example, to implement
+                 * --volatile=) without this interfering with our ability to access files such as
+                 * /etc/localtime to copy into the container. Note that we use a fixed place for this
+                 * (instead of a temporary directory, since we are living in our own mount namspace here
+                 * already, and thus don't need to be afraid of colliding with anyone else's mounts).*/
+                (void) mkdir_p("/run/systemd/nspawn-root", 0755);
+
+                r = mount_verbose(LOG_ERR, "/", "/run/systemd/nspawn-root", NULL, MS_BIND|MS_REC, NULL);
+                if (r < 0)
+                        return r;
+
+                directory = "/run/systemd/nspawn-root";
+
+        } else if (!dissected_image) {
+                /* Turn directory into bind mount (we need that so that we can move the bind mount to root
+                 * later on). */
                 r = mount_verbose(LOG_ERR, directory, directory, NULL, MS_BIND|MS_REC, NULL);
                 if (r < 0)
                         return r;
@@ -4018,7 +4034,7 @@ static int load_settings(void) {
                         p = file_in_same_dir(arg_image, fn);
                         if (!p)
                                 return log_oom();
-                } else if (arg_directory) {
+                } else if (arg_directory && !path_equal(arg_directory, "/")) {
                         p = file_in_same_dir(arg_directory, fn);
                         if (!p)
                                 return log_oom();
@@ -4740,8 +4756,12 @@ static int run(int argc, char *argv[]) {
         if (arg_directory) {
                 assert(!arg_image);
 
-                if (path_equal(arg_directory, "/") && !arg_ephemeral) {
-                        log_error("Spawning container on root directory is not supported. Consider using --ephemeral.");
+                /* Safety precaution: let's not allow running images from the live host OS image, as long as
+                 * /var from the host will propagate into container dynamically (because bad things happen if
+                 * two systems write to the same /var). Let's allow it for the special cases where /var is
+                 * either copied (i.e. --ephemeral) or replaced (i.e. --volatile=yes|state). */
+                if (path_equal(arg_directory, "/") && !(arg_ephemeral || IN_SET(arg_volatile_mode, VOLATILE_YES, VOLATILE_STATE))) {
+                        log_error("Spawning container on root directory is not supported. Consider using --ephemeral, --volatile=yes or --volatile=state.");
                         r = -EINVAL;
                         goto finish;
                 }
@@ -4770,7 +4790,9 @@ static int run(int argc, char *argv[]) {
                                 goto finish;
                         }
 
-                        r = image_path_lock(np, (arg_read_only ? LOCK_SH : LOCK_EX) | LOCK_NB, &tree_global_lock, &tree_local_lock);
+                        /* We take an exclusive lock on this image, since it's our private, ephemeral copy
+                         * only owned by us and noone else. */
+                        r = image_path_lock(np, LOCK_EX|LOCK_NB, &tree_global_lock, &tree_local_lock);
                         if (r < 0) {
                                 log_error_errno(r, "Failed to lock %s: %m", np);
                                 goto finish;
@@ -4890,7 +4912,8 @@ static int run(int argc, char *argv[]) {
                                 goto finish;
                         }
 
-                        r = image_path_lock(np, (arg_read_only ? LOCK_SH : LOCK_EX) | LOCK_NB, &tree_global_lock, &tree_local_lock);
+                        /* Always take an exclusive lock on our own ephemeral copy. */
+                        r = image_path_lock(np, LOCK_EX|LOCK_NB, &tree_global_lock, &tree_local_lock);
                         if (r < 0) {
                                 r = log_error_errno(r, "Failed to create image lock: %m");
                                 goto finish;
diff --git a/src/shared/machine-image.c b/src/shared/machine-image.c
index 07744b34b4..7007374192 100644
--- a/src/shared/machine-image.c
+++ b/src/shared/machine-image.c
@@ -989,28 +989,52 @@ int image_path_lock(const char *path, int operation, LockFile *global, LockFile
         _cleanup_free_ char *p = NULL;
         LockFile t = LOCK_FILE_INIT;
         struct stat st;
+        bool exclusive;
         int r;
 
         assert(path);
         assert(global);
         assert(local);
 
-        /* Locks an image path. This actually creates two locks: one
-         * "local" one, next to the image path itself, which might be
-         * shared via NFS. And another "global" one, in /run, that
-         * uses the device/inode number. This has the benefit that we
-         * can even lock a tree that is a mount point, correctly. */
+        /* Locks an image path. This actually creates two locks: one "local" one, next to the image path
+         * itself, which might be shared via NFS. And another "global" one, in /run, that uses the
+         * device/inode number. This has the benefit that we can even lock a tree that is a mount point,
+         * correctly. */
 
         if (!path_is_absolute(path))
                 return -EINVAL;
 
+        switch (operation & (LOCK_SH|LOCK_EX)) {
+        case LOCK_SH:
+                exclusive = false;
+                break;
+        case LOCK_EX:
+                exclusive = true;
+                break;
+        default:
+                return -EINVAL;
+        }
+
         if (getenv_bool("SYSTEMD_NSPAWN_LOCK") == 0) {
                 *local = *global = (LockFile) LOCK_FILE_INIT;
                 return 0;
         }
 
-        if (path_equal(path, "/"))
-                return -EBUSY;
+        /* Prohibit taking exclusive locks on the host image. We can't allow this, since we ourselves are
+         * running off it after all, and we don't want any images to manipulate the host image. We make an
+         * exception for shared locks however: we allow those (and make them NOPs since there's no point in
+         * taking them if there can't be exclusive locks). Strictly speaking these are questionable as well,
+         * since it means changes made to the host might propagate to the container as they happen (and a
+         * shared lock kinda suggests that no changes happen at all while it is in place), but it's too
+         * useful not to allow read-only containers off the host root, hence let's support this, and trust
+         * the user to do the right thing with this. */
+        if (path_equal(path, "/")) {
+                if (exclusive)
+                        return -EBUSY;
+
+                *local = *global = (LockFile) LOCK_FILE_INIT;
+                return 0;
+        }
 
         if (stat(path, &st) >= 0) {
                 if (S_ISBLK(st.st_mode))
@@ -1024,12 +1048,12 @@ int image_path_lock(const char *path, int operation, LockFile *global, LockFile
                         return -ENOMEM;
         }
 
-        /* For block devices we don't need the "local" lock, as the major/minor lock above should be sufficient, since
-         * block devices are device local anyway. */
-        if (!path_startswith(path, "/dev")) {
+        /* For block devices we don't need the "local" lock, as the major/minor lock above should be
+         * sufficient, since block devices are host local anyway. */
+        if (!path_startswith(path, "/dev/")) {
                 r = make_lock_file_for(path, operation, &t);
                 if (r < 0) {
-                        if ((operation & LOCK_SH) && r == -EROFS)
+                        if (!exclusive && r == -EROFS)
                                 log_debug_errno(r, "Failed to create shared lock for '%s', ignoring: %m", path);
                         else
                                 return r;