fs: don't copy to userspace under namespace semaphore

Don't copy mount ids to userspace while holding the namespace semaphore.
We really shouldn't do that and I've gone through lenghts avoiding that
in statmount() already.

Limit the number of mounts that can be retrieved in one go to 1 million
mount ids. That's effectively 10 times the default limt of 100000 mounts
that we put on each mount namespace by default. Since listmount() is an
iterator limiting the number of mounts retrievable in one go isn't a
problem as userspace can just pick up where they left off.

Karel menti_ned that libmount will probably be reading the mount table
in "in small steps, 512 nodes per request. Nobody likes a tool that
takes too long in the kernel, and huge servers are unusual use cases.
Libmount will very probably provide API to define size of the step (IDs
per request)."

Reported-by: Mateusz Guzik <mjguzik@gmail.com>
Link: https://lore.kernel.org/r/20240610-frettchen-liberal-a9a5c53865f8@brauner
Signed-off-by: Christian Brauner <brauner@kernel.org>

fs/namespace.c

@@ -5047,89 +5047,103 @@ static struct mount *listmnt_next(struct mount *curr)
 	return node_to_mount(rb_next(&curr->mnt_node));
 }
 
-static ssize_t do_listmount(struct mount *first, struct path *orig,
-			    u64 mnt_parent_id, u64 __user *mnt_ids,
-			    size_t nr_mnt_ids, const struct path *root)
+static ssize_t do_listmount(u64 mnt_parent_id, u64 last_mnt_id, u64 *mnt_ids,
+			    size_t nr_mnt_ids)
 {
-	struct mount *r;
+	struct path root;
+	struct mnt_namespace *ns = current->nsproxy->mnt_ns;
+	struct path orig;
+	struct mount *r, *first;
 	ssize_t ret;
 
+	rwsem_assert_held(&namespace_sem);
+
+	get_fs_root(current->fs, &root);
+	if (mnt_parent_id == LSMT_ROOT) {
+		orig = root;
+	} else {
+		orig.mnt = lookup_mnt_in_ns(mnt_parent_id, ns);
+		if (!orig.mnt) {
+			ret = -ENOENT;
+			goto err;
+		}
+		orig.dentry = orig.mnt->mnt_root;
+	}
+
 	/*
 	 * Don't trigger audit denials. We just want to determine what
 	 * mounts to show users.
 	 */
-	if (!is_path_reachable(real_mount(orig->mnt), orig->dentry, root) &&
-	    !ns_capable_noaudit(&init_user_ns, CAP_SYS_ADMIN))
-		return -EPERM;
+	if (!is_path_reachable(real_mount(orig.mnt), orig.dentry, &root) &&
+	    !ns_capable_noaudit(&init_user_ns, CAP_SYS_ADMIN)) {
+		ret = -EPERM;
+		goto err;
+	}
 
-	ret = security_sb_statfs(orig->dentry);
+	ret = security_sb_statfs(orig.dentry);
 	if (ret)
-		return ret;
+		goto err;
+
+	if (!last_mnt_id)
+		first = node_to_mount(rb_first(&ns->mounts));
+	else
+		first = mnt_find_id_at(ns, last_mnt_id + 1);
 
 	for (ret = 0, r = first; r && nr_mnt_ids; r = listmnt_next(r)) {
 		if (r->mnt_id_unique == mnt_parent_id)
 			continue;
-		if (!is_path_reachable(r, r->mnt.mnt_root, orig))
+		if (!is_path_reachable(r, r->mnt.mnt_root, &orig))
 			continue;
-		if (put_user(r->mnt_id_unique, mnt_ids))
-			return -EFAULT;
+		*mnt_ids = r->mnt_id_unique;
 		mnt_ids++;
 		nr_mnt_ids--;
 		ret++;
 	}
+err:
+	path_put(&root);
 	return ret;
 }
 
-SYSCALL_DEFINE4(listmount, const struct mnt_id_req __user *, req, u64 __user *,
-		mnt_ids, size_t, nr_mnt_ids, unsigned int, flags)
+SYSCALL_DEFINE4(listmount, const struct mnt_id_req __user *, req,
+		u64 __user *, mnt_ids, size_t, nr_mnt_ids, unsigned int, flags)
 {
-	struct mnt_namespace *ns = current->nsproxy->mnt_ns;
+	u64 *kmnt_ids __free(kvfree) = NULL;
+	const size_t maxcount = 1000000;
 	struct mnt_id_req kreq;
-	struct mount *first;
-	struct path root, orig;
-	u64 mnt_parent_id, last_mnt_id;
-	const size_t maxcount = (size_t)-1 >> 3;
 	ssize_t ret;
 
 	if (flags)
 		return -EINVAL;
 
+	/*
+	 * If the mount namespace really has more than 1 million mounts the
+	 * caller must iterate over the mount namespace (and reconsider their
+	 * system design...).
+	 */
 	if (unlikely(nr_mnt_ids > maxcount))
-		return -EFAULT;
+		return -EOVERFLOW;
 
 	if (!access_ok(mnt_ids, nr_mnt_ids * sizeof(*mnt_ids)))
 		return -EFAULT;
 
 	ret = copy_mnt_id_req(req, &kreq);
 	if (ret)
 		return ret;
-	mnt_parent_id = kreq.mnt_id;
-	last_mnt_id = kreq.param;
 
-	down_read(&namespace_sem);
-	get_fs_root(current->fs, &root);
-	if (mnt_parent_id == LSMT_ROOT) {
-		orig = root;
-	} else {
-		ret = -ENOENT;
-		orig.mnt = lookup_mnt_in_ns(mnt_parent_id, ns);
-		if (!orig.mnt)
-			goto err;
-		orig.dentry = orig.mnt->mnt_root;
-	}
-	if (!last_mnt_id)
-		first = node_to_mount(rb_first(&ns->mounts));
-	else
-		first = mnt_find_id_at(ns, last_mnt_id + 1);
+	kmnt_ids = kvmalloc_array(nr_mnt_ids, sizeof(*kmnt_ids),
+				  GFP_KERNEL_ACCOUNT);
+	if (!kmnt_ids)
+		return -ENOMEM;
+
+	scoped_guard(rwsem_read, &namespace_sem)
+		ret = do_listmount(kreq.mnt_id, kreq.param, kmnt_ids, nr_mnt_ids);
+
+	if (copy_to_user(mnt_ids, kmnt_ids, ret * sizeof(*mnt_ids)))
+		return -EFAULT;
 
-	ret = do_listmount(first, &orig, mnt_parent_id, mnt_ids, nr_mnt_ids, &root);
-err:
-	path_put(&root);
-	up_read(&namespace_sem);
 	return ret;
 }
 
-
 static void __init init_mount_tree(void)
 {
 	struct vfsmount *mnt;