libct/nsenter: become root after joining userns

libcontainer/nsenter/nsexec.c

@@ -505,6 +505,17 @@ void join_namespaces(char *nslist)
 		if (setns(ns->fd, flag) < 0)
 			bail("failed to setns into %s namespace", ns->type);
 
+		/*
+		 * If we change user namespaces, make sure we switch to root in the
+		 * namespace (this matches the logic for unshare(CLONE_NEWUSER)), lots
+		 * of things can break if we aren't the right user. See
+		 * <https://github.com/opencontainers/runc/issues/4466> for one example.
+		 */
+		if (flag == CLONE_NEWUSER) {
+			if (setresuid(0, 0, 0) < 0)
+				bail("failed to become root in user namespace");
+		}
+
 		close(ns->fd);
 	}
 

tests/integration/userns.bats

@@ -125,6 +125,26 @@ function teardown() {
 	fi
 }
 
+# issue: https://github.com/opencontainers/runc/issues/4466
+@test "userns join other container userns[selinux enabled]" {
+	if ! selinuxenabled; then
+		skip "requires SELinux enabled and in enforcing mode"
+	fi
+	# Create a detached container with the id-mapping we want.
+	update_config '.process.args = ["sleep", "infinity"]'
+	runc run -d --console-socket "$CONSOLE_SOCKET" target_userns
+	[ "$status" -eq 0 ]
+
+	# Configure our container to attach to the first container's userns.
+	target_pid="$(__runc state target_userns | jq .pid)"
+	update_config '.linux.namespaces |= map(if .type == "user" then (.path = "/proc/'"$target_pid"'/ns/" + .type) else . end)
+		| del(.linux.uidMappings)
+		| del(.linux.gidMappings)
+		| .linux.mountLabel="system_u:object_r:container_file_t:s0:c344,c805"'
+	runc run -d --console-socket "$CONSOLE_SOCKET" in_userns
+	[ "$status" -eq 0 ]
+}
+
 @test "userns join other container userns [bind-mounted nsfd]" {
 	requires root