mount: hold namespace_sem across copy in create_new_namespace()

Fix an oversight when creating a new mount namespace. If someone had the
bright idea to make the real rootfs a shared or dependent mount and it
is later copied the copy will become a peer of the old real rootfs mount
or a dependent mount of it. The namespace semaphore is dropped and we
use mount lock exact to lock the new real root mount. If that fails or
the subsequent do_loopback() fails we rely on the copy of the real root
mount to be cleaned up by path_put(). The problem is that this doesn't
deal with mount propagation and will leave the mounts linked in the
propagation lists.

When creating a new mount namespace create_new_namespace() first
acquires namespace_sem to clone the nullfs root, drops it, then
reacquires it via LOCK_MOUNT_EXACT which takes inode_lock first to
respect the inode_lock -> namespace_sem lock ordering. This
drop-and-reacquire pattern is fragile and was the source of the
propagation cleanup bug fixed in the preceding commit.

Extend lock_mount_exact() with a copy_mount mode that clones the mount
under the locks atomically. When copy_mount is true, path_overmounted()
is skipped since we're copying the mount, not mounting on top of it -
the nullfs root always has rootfs mounted on top so the check would
always fail. If clone_mnt() fails after get_mountpoint() has pinned the
mountpoint, __unlock_mount() is used to properly unpin the mountpoint
and release both locks.

This allows create_new_namespace() to use LOCK_MOUNT_EXACT_COPY which
takes inode_lock and namespace_sem once and holds them throughout the
clone and subsequent mount operations, eliminating the
drop-and-reacquire pattern entirely.

Reported-by: [e-mail]
Fixes: 9b8a0ba68246 ("mount: add OPEN_TREE_NAMESPACE") # mainline only
Link: https://lore.kernel.org/[e-mail]
Signed-off-by: Christian Brauner <[e-mail]>

fs/namespace.c

···
 }
 
 static void lock_mount_exact(const struct path *path,
-			     struct pinned_mountpoint *mp);
+			     struct pinned_mountpoint *mp, bool copy_mount,
+			     unsigned int copy_flags);
 
 #define LOCK_MOUNT_MAYBE_BENEATH(mp, path, beneath) \
 	struct pinned_mountpoint mp __cleanup(unlock_mount) = {}; \
···
 #define LOCK_MOUNT(mp, path) LOCK_MOUNT_MAYBE_BENEATH(mp, (path), false)
 #define LOCK_MOUNT_EXACT(mp, path) \
 	struct pinned_mountpoint mp __cleanup(unlock_mount) = {}; \
-	lock_mount_exact((path), &mp)
+	lock_mount_exact((path), &mp, false, 0)
+#define LOCK_MOUNT_EXACT_COPY(mp, path, copy_flags) \
+	struct pinned_mountpoint mp __cleanup(unlock_mount) = {}; \
+	lock_mount_exact((path), &mp, true, (copy_flags))
 
 static int graft_tree(struct mount *mnt, const struct pinned_mountpoint *mp)
 {
···
 	return file;
 }
 
-DEFINE_FREE(put_empty_mnt_ns, struct mnt_namespace *,
-	    if (!IS_ERR_OR_NULL(_T)) free_mnt_ns(_T))
-
 static struct mnt_namespace *create_new_namespace(struct path *path, unsigned int flags)
 {
-	struct mnt_namespace *new_ns __free(put_empty_mnt_ns) = NULL;
-	struct path to_path __free(path_put) = {};
 	struct mnt_namespace *ns = current->nsproxy->mnt_ns;
 	struct user_namespace *user_ns = current_user_ns();
-	struct mount *new_ns_root;
+	struct mnt_namespace *new_ns;
+	struct mount *new_ns_root, *old_ns_root;
+	struct path to_path;
 	struct mount *mnt;
 	unsigned int copy_flags = 0;
 	bool locked = false;
···
 	if (IS_ERR(new_ns))
 		return ERR_CAST(new_ns);
 
-	scoped_guard(namespace_excl) {
-		new_ns_root = clone_mnt(ns->root, ns->root->mnt.mnt_root, copy_flags);
-		if (IS_ERR(new_ns_root))
-			return ERR_CAST(new_ns_root);
+	old_ns_root = ns->root;
+	to_path.mnt = &old_ns_root->mnt;
+	to_path.dentry = old_ns_root->mnt.mnt_root;
 
-		/*
-		 * If the real rootfs had a locked mount on top of it somewhere
-		 * in the stack, lock the new mount tree as well so it can't be
-		 * exposed.
-		 */
-		mnt = ns->root;
-		while (mnt->overmount) {
-			mnt = mnt->overmount;
-			if (mnt->mnt.mnt_flags & MNT_LOCKED)
-				locked = true;
-		}
+	VFS_WARN_ON_ONCE(old_ns_root->mnt.mnt_sb->s_type != &nullfs_fs_type);
+
+	LOCK_MOUNT_EXACT_COPY(mp, &to_path, copy_flags);
+	if (IS_ERR(mp.parent)) {
+		free_mnt_ns(new_ns);
+		return ERR_CAST(mp.parent);
+	}
+	new_ns_root = mp.parent;
+
+	/*
+	 * If the real rootfs had a locked mount on top of it somewhere
+	 * in the stack, lock the new mount tree as well so it can't be
+	 * exposed.
+	 */
+	mnt = old_ns_root;
+	while (mnt->overmount) {
+		mnt = mnt->overmount;
+		if (mnt->mnt.mnt_flags & MNT_LOCKED)
+			locked = true;
 	}
 
 	/*
-	 * We dropped the namespace semaphore so we can actually lock
-	 * the copy for mounting. The copied mount isn't attached to any
-	 * mount namespace and it is thus excluded from any propagation.
-	 * So realistically we're isolated and the mount can't be
-	 * overmounted.
-	 */
-
-	/* Borrow the reference from clone_mnt(). */
-	to_path.mnt = &new_ns_root->mnt;
-	to_path.dentry = dget(new_ns_root->mnt.mnt_root);
-
-	/* Now lock for actual mounting. */
-	LOCK_MOUNT_EXACT(mp, &to_path);
-	if (unlikely(IS_ERR(mp.parent)))
-		return ERR_CAST(mp.parent);
-
-	/*
-	 * We don't emulate unshare()ing a mount namespace. We stick to the
-	 * restrictions of creating detached bind-mounts. It has a lot
-	 * saner and simpler semantics.
+	 * We don't emulate unshare()ing a mount namespace. We stick
+	 * to the restrictions of creating detached bind-mounts. It
+	 * has a lot saner and simpler semantics.
 	 */
 	mnt = __do_loopback(path, flags, copy_flags);
-	if (IS_ERR(mnt))
-		return ERR_CAST(mnt);
-
 	scoped_guard(mount_writer) {
+		if (IS_ERR(mnt)) {
+			emptied_ns = new_ns;
+			umount_tree(new_ns_root, 0);
+			return ERR_CAST(mnt);
+		}
+
 		if (locked)
 			mnt->mnt.mnt_flags |= MNT_LOCKED;
 		/*
-		 * Now mount the detached tree on top of the copy of the
-		 * real rootfs we created.
+		 * now mount the detached tree on top of the copy
+		 * of the real rootfs we created.
 		 */
 		attach_mnt(mnt, new_ns_root, mp.mp);
 		if (user_ns != ns->user_ns)
 			lock_mnt_tree(new_ns_root);
 	}
 
-	/* Add all mounts to the new namespace. */
-	for (struct mount *p = new_ns_root; p; p = next_mnt(p, new_ns_root)) {
-		mnt_add_to_ns(new_ns, p);
+	for (mnt = new_ns_root; mnt; mnt = next_mnt(mnt, new_ns_root)) {
+		mnt_add_to_ns(new_ns, mnt);
 		new_ns->nr_mounts++;
 	}
 
-	new_ns->root = real_mount(no_free_ptr(to_path.mnt));
+	new_ns->root = new_ns_root;
 	ns_tree_add_raw(new_ns);
-	return no_free_ptr(new_ns);
+	return new_ns;
 }
 
 static struct file *open_new_namespace(struct path *path, unsigned int flags)
···
 }
 
 static void lock_mount_exact(const struct path *path,
-			     struct pinned_mountpoint *mp)
+			     struct pinned_mountpoint *mp, bool copy_mount,
+			     unsigned int copy_flags)
 {
 	struct dentry *dentry = path->dentry;
 	int err;
+
+	/* Assert that inode_lock() locked the correct inode. */
+	VFS_WARN_ON_ONCE(copy_mount && !path_mounted(path));
 
 	inode_lock(dentry->d_inode);
 	namespace_lock();
 	if (unlikely(cant_mount(dentry)))
 		err = -ENOENT;
-	else if (path_overmounted(path))
+	else if (!copy_mount && path_overmounted(path))
 		err = -EBUSY;
 	else
 		err = get_mountpoint(dentry, mp);
···
 		namespace_unlock();
 		inode_unlock(dentry->d_inode);
 		mp->parent = ERR_PTR(err);
-	} else {
-		mp->parent = real_mount(path->mnt);
+		return;
 	}
+
+	if (copy_mount)
+		mp->parent = clone_mnt(real_mount(path->mnt), dentry, copy_flags);
+	else
+		mp->parent = real_mount(path->mnt);
+	if (unlikely(IS_ERR(mp->parent)))
+		__unlock_mount(mp);
 }
 
 int finish_automount(struct vfsmount *__m, const struct path *path)