From: Karel Zak <kzak@redhat.com>
Date: Mon, 30 Mar 2026 12:16:25 +0000 (+0200)
Subject: nsenter: keep PID:inode validated pidfd open
X-Git-Url: http://git.ipfire.org/gitweb.cgi?a=commitdiff_plain;h=06d05fc0bc4043c1f00a2aa4e4b7e6a40ff7f6b0;p=thirdparty%2Futil-linux.git

nsenter: keep PID:inode validated pidfd open

The original code validated the PID:inode by opening a pidfd via
ul_get_valid_pidfd_or_err(), but immediately closed it. Later,
pidfd_open() was called again for the same PID. This created a
window between close() and the second pidfd_open() where the PID
could be recycled by the kernel, defeating the purpose of the
PID:inode identity verification.

Keep the validated pidfd from PID:inode open and reuse it, so the
process identity is pinned for the entire lifetime of nsenter. As
long as the pidfd is held open, the kernel will not recycle the PID,
making subsequent operations (setns, ioctls) race-free.

Introduce ns_pid_fd to separate the pidfd used for enter_namespaces()
(setns, requires kernel >= 5.7) from the general-purpose pid_fd used
for ioctls like PIDFD_GET_USER_NAMESPACE and pidfd_getfd(). The kernel
version check gates only the setns-related usage.

Also use ul_parse_pid_str_or_err() for consistent error handling
across utilities that support the PID:inode format.

Signed-off-by: Karel Zak <kzak@redhat.com>
---

diff --git a/sys-utils/nsenter.c b/sys-utils/nsenter.c
index e10ba9cf9..c0309a905 100644
--- a/sys-utils/nsenter.c
+++ b/sys-utils/nsenter.c
@@ -576,15 +576,13 @@ static void continue_as_child(void)
 
 static int parse_pid_str(char *pidstr, pid_t *ns_target_pid)
 {
-	int rc, pfd;
+	int pfd = -1;
 	uint64_t pidfd_ino = 0;
 
-	rc = ul_parse_pid_str(pidstr, ns_target_pid, &pidfd_ino);
-	if (pidfd_ino) {
+	ul_parse_pid_str_or_err(pidstr, ns_target_pid, &pidfd_ino);
+	if (pidfd_ino)
 		pfd = ul_get_valid_pidfd_or_err(*ns_target_pid, pidfd_ino);
-		close(pfd);
-	}
-	return rc;
+	return pfd;
 }
 
 int main(int argc, char *argv[])
@@ -630,7 +628,7 @@ int main(int argc, char *argv[])
 	};
 	int excl_st[ARRAY_SIZE(excl)] = UL_EXCL_STATUS_INIT;
 
-	int c, rc, namespaces = 0, setgroups_nerrs = 0, preserve_cred = 0;
+	int c, namespaces = 0, setgroups_nerrs = 0, preserve_cred = 0;
 	bool do_rd = false, do_wd = false, do_uid = false, force_uid = false,
 	     do_gid = false, force_gid = false, do_env = false, do_all = false,
 	     do_join_cgroup = false, do_user_parent = false;
@@ -640,7 +638,7 @@ int main(int argc, char *argv[])
 	gid_t gid = 0;
 	int keepcaps = 0;
 	int sock_fd = -1;
-	int pid_fd = -1;
+	int pid_fd = -1, ns_pid_fd = -1;
 #ifdef HAVE_LIBSELINUX
 	bool selinux = 0;
 #endif
@@ -662,9 +660,9 @@ int main(int argc, char *argv[])
 			do_all = true;
 			break;
 		case 't':
-			rc = parse_pid_str(optarg, &namespace_target_pid);
-			if (rc)
-				errx(EXIT_FAILURE, _("invalid PID argument"));
+			if (namespace_target_pid)
+				errx(EXIT_FAILURE, _("option --target may be given only once"));
+			pid_fd = parse_pid_str(optarg, &namespace_target_pid);
 			break;
 		case 'm':
 			enable_namespace(CLONE_NEWNS, optarg);
@@ -795,9 +793,12 @@ int main(int argc, char *argv[])
 		 * For other cases such as sock_fd and user-parent, the global
 		 * pidfd needs to be optional.
 		 */
-		if (get_linux_version() > KERNEL_VERSION(5, 7, 0))
-			pid_fd = pidfd_open(namespace_target_pid, 0);
-		if (pid_fd < 0 && namespaces)
+		if (get_linux_version() > KERNEL_VERSION(5, 7, 0)) {
+			if (pid_fd < 0)
+				pid_fd = pidfd_open(namespace_target_pid, 0);
+			ns_pid_fd = pid_fd;
+		}
+		if (ns_pid_fd < 0 && namespaces)
 			open_namespaces(namespaces);	/* fallback */
 	}
 
@@ -850,11 +851,11 @@ int main(int argc, char *argv[])
 	 * namespace last and if we're privileging it then we enter the user
 	 * namespace first (because the initial setns will fail).
 	 */
-	enter_namespaces(pid_fd, namespaces & ~CLONE_NEWUSER, 1);	/* ignore errors */
+	enter_namespaces(ns_pid_fd, namespaces & ~CLONE_NEWUSER, 1);	/* ignore errors */
 
 	namespaces = get_namespaces();
 	if (namespaces)
-		enter_namespaces(pid_fd, namespaces, 0);		/* report errors */
+		enter_namespaces(ns_pid_fd, namespaces, 0);		/* report errors */
 
 	if (pid_fd >= 0)
 		close(pid_fd);