core: add new PrivateUsers= option to service execution

This setting adds minimal user namespacing support to a service. When set the invoked
processes will run in their own user namespace. Only a trivial mapping will be
set up: the root user/group is mapped to root, and the user/group of the
service will be mapped to itself, everything else is mapped to nobody.

If this setting is used the service runs with no capabilities on the host, but
configurable capabilities within the service.

This setting is particularly useful in conjunction with RootDirectory= as the
need to synchronize /etc/passwd and /etc/group between the host and the service
OS tree is reduced, as only three UID/GIDs need to match: root, nobody and the
user of the service itself. But even outside the RootDirectory= case this
setting is useful to substantially reduce the attack surface of a service.

Example command to test this:

        systemd-run -p PrivateUsers=1 -p User=foobar -t /bin/sh

This runs a shell as user "foobar". When typing "ps" only processes owned by
"root", by "foobar", and by "nobody" should be visible.

4 files changed, 174 insertions, 5 deletions

diff --git a/src/core/dbus-execute.c b/src/core/dbus-execute.c
index 9c50cd93e5..4b3bbfbc7d 100644
--- a/src/core/dbus-execute.c
+++ b/src/core/dbus-execute.c
@@ -705,8 +705,9 @@ const sd_bus_vtable bus_exec_vtable[] = {
         SD_BUS_PROPERTY("InaccessiblePaths", "as", NULL, offsetof(ExecContext, inaccessible_paths), SD_BUS_VTABLE_PROPERTY_CONST),
         SD_BUS_PROPERTY("MountFlags", "t", bus_property_get_ulong, offsetof(ExecContext, mount_flags), SD_BUS_VTABLE_PROPERTY_CONST),
         SD_BUS_PROPERTY("PrivateTmp", "b", bus_property_get_bool, offsetof(ExecContext, private_tmp), SD_BUS_VTABLE_PROPERTY_CONST),
-        SD_BUS_PROPERTY("PrivateNetwork", "b", bus_property_get_bool, offsetof(ExecContext, private_network), SD_BUS_VTABLE_PROPERTY_CONST),
         SD_BUS_PROPERTY("PrivateDevices", "b", bus_property_get_bool, offsetof(ExecContext, private_devices), SD_BUS_VTABLE_PROPERTY_CONST),
+        SD_BUS_PROPERTY("PrivateNetwork", "b", bus_property_get_bool, offsetof(ExecContext, private_network), SD_BUS_VTABLE_PROPERTY_CONST),
+        SD_BUS_PROPERTY("PrivateUsers", "b", bus_property_get_bool, offsetof(ExecContext, private_users), SD_BUS_VTABLE_PROPERTY_CONST),
         SD_BUS_PROPERTY("ProtectHome", "s", bus_property_get_protect_home, offsetof(ExecContext, protect_home), SD_BUS_VTABLE_PROPERTY_CONST),
         SD_BUS_PROPERTY("ProtectSystem", "s", bus_property_get_protect_system, offsetof(ExecContext, protect_system), SD_BUS_VTABLE_PROPERTY_CONST),
         SD_BUS_PROPERTY("SameProcessGroup", "b", bus_property_get_bool, offsetof(ExecContext, same_pgrp), SD_BUS_VTABLE_PROPERTY_CONST),
@@ -1068,7 +1069,7 @@ int bus_exec_context_set_transient_property(
 
         } else if (STR_IN_SET(name,
                               "IgnoreSIGPIPE", "TTYVHangup", "TTYReset",
-                              "PrivateTmp", "PrivateDevices", "PrivateNetwork",
+                              "PrivateTmp", "PrivateDevices", "PrivateNetwork", "PrivateUsers",
                               "NoNewPrivileges", "SyslogLevelPrefix", "MemoryDenyWriteExecute",
                               "RestrictRealtime", "DynamicUser")) {
                 int b;
@@ -1090,6 +1091,8 @@ int bus_exec_context_set_transient_property(
                                 c->private_devices = b;
                         else if (streq(name, "PrivateNetwork"))
                                 c->private_network = b;
+                        else if (streq(name, "PrivateUsers"))
+                                c->private_users = b;
                         else if (streq(name, "NoNewPrivileges"))
                                 c->no_new_privileges = b;
                         else if (streq(name, "SyslogLevelPrefix"))
diff --git a/src/core/execute.c b/src/core/execute.c
index 26e9cd5339..cec3b3cf40 100644
--- a/src/core/execute.c
+++ b/src/core/execute.c
@@ -25,6 +25,7 @@
 #include <signal.h>
 #include <string.h>
 #include <sys/capability.h>
+#include <sys/eventfd.h>
 #include <sys/mman.h>
 #include <sys/personality.h>
 #include <sys/prctl.h>
@@ -1526,6 +1527,159 @@ static bool exec_needs_mount_namespace(
         return false;
 }
 
+static int setup_private_users(uid_t uid, gid_t gid) {
+        _cleanup_free_ char *uid_map = NULL, *gid_map = NULL;
+        _cleanup_close_pair_ int errno_pipe[2] = { -1, -1 };
+        _cleanup_close_ int unshare_ready_fd = -1;
+        _cleanup_(sigkill_waitp) pid_t pid = 0;
+        uint64_t c = 1;
+        siginfo_t si;
+        ssize_t n;
+        int r;
+
+        /* Set up a user namespace and map root to root, the selected UID/GID to itself, and everything else to
+         * nobody. In order to be able to write this mapping we need CAP_SETUID in the original user namespace, which
+         * we however lack after opening the user namespace. To work around this we fork() a temporary child process,
+         * which waits for the parent to create the new user namespace while staying in the original namespace. The
+         * child then writes the UID mapping, under full privileges. The parent waits for the child to finish and
+         * continues execution normally. */
+
+        if (uid != 0 && uid_is_valid(uid))
+                asprintf(&uid_map,
+                         "0 0 1\n"                      /* Map root → root */
+                         UID_FMT " " UID_FMT " 1\n",    /* Map $UID → $UID */
+                         uid, uid);                     /* The case where the above is the same */
+        else
+                uid_map = strdup("0 0 1\n");
+        if (!uid_map)
+                return -ENOMEM;
+
+        if (gid != 0 && gid_is_valid(gid))
+                asprintf(&gid_map,
+                         "0 0 1\n"                      /* Map root → root */
+                         GID_FMT " " GID_FMT " 1\n",    /* Map $GID → $GID */
+                         gid, gid);
+        else
+                gid_map = strdup("0 0 1\n");            /* The case where the above is the same */
+        if (!gid_map)
+                return -ENOMEM;
+
+        /* Create a communication channel so that the parent can tell the child when it finished creating the user
+         * namespace. */
+        unshare_ready_fd = eventfd(0, EFD_CLOEXEC);
+        if (unshare_ready_fd < 0)
+                return -errno;
+
+        /* Create a communication channel so that the child can tell the parent a proper error code in case it
+         * failed. */
+        if (pipe2(errno_pipe, O_CLOEXEC) < 0)
+                return -errno;
+
+        pid = fork();
+        if (pid < 0)
+                return -errno;
+
+        if (pid == 0) {
+                _cleanup_close_ int fd = -1;
+                const char *a;
+                pid_t ppid;
+
+                /* Child process, running in the original user namespace. Let's update the parent's UID/GID map from
+                 * here, after the parent opened its own user namespace. */
+
+                ppid = getppid();
+                errno_pipe[0] = safe_close(errno_pipe[0]);
+
+                /* Wait until the parent unshared the user namespace */
+                if (read(unshare_ready_fd, &c, sizeof(c)) < 0) {
+                        r = -errno;
+                        goto child_fail;
+                }
+
+                /* Disable the setgroups() system call in the child user namespace, for good. */
+                a = procfs_file_alloca(ppid, "setgroups");
+                fd = open(a, O_WRONLY|O_CLOEXEC);
+                if (fd < 0) {
+                        if (errno != ENOENT) {
+                                r = -errno;
+                                goto child_fail;
+                        }
+
+                        /* If the file is missing the kernel is too old, let's continue anyway. */
+                } else {
+                        if (write(fd, "deny\n", 5) < 0) {
+                                r = -errno;
+                                goto child_fail;
+                        }
+
+                        fd = safe_close(fd);
+                }
+
+                /* First write the GID map */
+                a = procfs_file_alloca(ppid, "gid_map");
+                fd = open(a, O_WRONLY|O_CLOEXEC);
+                if (fd < 0) {
+                        r = -errno;
+                        goto child_fail;
+                }
+                if (write(fd, gid_map, strlen(gid_map)) < 0) {
+                        r = -errno;
+                        goto child_fail;
+                }
+                fd = safe_close(fd);
+
+                /* The write the UID map */
+                a = procfs_file_alloca(ppid, "uid_map");
+                fd = open(a, O_WRONLY|O_CLOEXEC);
+                if (fd < 0) {
+                        r = -errno;
+                        goto child_fail;
+                }
+                if (write(fd, uid_map, strlen(uid_map)) < 0) {
+                        r = -errno;
+                        goto child_fail;
+                }
+
+                _exit(EXIT_SUCCESS);
+
+        child_fail:
+                (void) write(errno_pipe[1], &r, sizeof(r));
+                _exit(EXIT_FAILURE);
+        }
+
+        errno_pipe[1] = safe_close(errno_pipe[1]);
+
+        if (unshare(CLONE_NEWUSER) < 0)
+                return -errno;
+
+        /* Let the child know that the namespace is ready now */
+        if (write(unshare_ready_fd, &c, sizeof(c)) < 0)
+                return -errno;
+
+        /* Try to read an error code from the child */
+        n = read(errno_pipe[0], &r, sizeof(r));
+        if (n < 0)
+                return -errno;
+        if (n == sizeof(r)) { /* an error code was sent to us */
+                if (r < 0)
+                        return r;
+                return -EIO;
+        }
+        if (n != 0) /* on success we should have read 0 bytes */
+                return -EIO;
+
+        r = wait_for_terminate(pid, &si);
+        if (r < 0)
+                return r;
+        pid = 0;
+
+        /* If something strange happened with the child, let's consider this fatal, too */
+        if (si.si_code != CLD_EXITED || si.si_status != 0)
+                return -EIO;
+
+        return 0;
+}
+
 static void append_socket_pair(int *array, unsigned *n, int pair[2]) {
         assert(array);
         assert(n);
@@ -2037,6 +2191,14 @@ static int exec_child(
         }
 #endif
 
+        if (params->apply_permissions && context->private_users) {
+                r = setup_private_users(uid, gid);
+                if (r < 0) {
+                        *exit_status = EXIT_USER;
+                        return r;
+                }
+        }
+
         /* We repeat the fd closing here, to make sure that
          * nothing is leaked from the PAM modules. Note that
          * we are more aggressive this time since socket_fd
@@ -2598,8 +2760,9 @@ void exec_context_dump(ExecContext *c, FILE* f, const char *prefix) {
                 "%sRootDirectory: %s\n"
                 "%sNonBlocking: %s\n"
                 "%sPrivateTmp: %s\n"
-                "%sPrivateNetwork: %s\n"
                 "%sPrivateDevices: %s\n"
+                "%sPrivateNetwork: %s\n"
+                "%sPrivateUsers: %s\n"
                 "%sProtectHome: %s\n"
                 "%sProtectSystem: %s\n"
                 "%sIgnoreSIGPIPE: %s\n"
@@ -2610,8 +2773,9 @@ void exec_context_dump(ExecContext *c, FILE* f, const char *prefix) {
                 prefix, c->root_directory ? c->root_directory : "/",
                 prefix, yes_no(c->non_blocking),
                 prefix, yes_no(c->private_tmp),
-                prefix, yes_no(c->private_network),
                 prefix, yes_no(c->private_devices),
+                prefix, yes_no(c->private_network),
+                prefix, yes_no(c->private_users),
                 prefix, protect_home_to_string(c->protect_home),
                 prefix, protect_system_to_string(c->protect_system),
                 prefix, yes_no(c->ignore_sigpipe),
diff --git a/src/core/execute.h b/src/core/execute.h
index 48cc18fbb3..5fac3e85e8 100644
--- a/src/core/execute.h
+++ b/src/core/execute.h
@@ -171,6 +171,7 @@ struct ExecContext {
         bool private_tmp;
         bool private_network;
         bool private_devices;
+        bool private_users;
         ProtectSystem protect_system;
         ProtectHome protect_home;
 
diff --git a/src/core/load-fragment-gperf.gperf.m4 b/src/core/load-fragment-gperf.gperf.m4
index 396f847213..251155b428 100644
--- a/src/core/load-fragment-gperf.gperf.m4
+++ b/src/core/load-fragment-gperf.gperf.m4
@@ -88,8 +88,9 @@ $1.ReadWritePaths,               config_parse_namespace_path_strv,   0,
 $1.ReadOnlyPaths,                config_parse_namespace_path_strv,   0,                             offsetof($1, exec_context.read_only_paths)
 $1.InaccessiblePaths,            config_parse_namespace_path_strv,   0,                             offsetof($1, exec_context.inaccessible_paths)
 $1.PrivateTmp,                   config_parse_bool,                  0,                             offsetof($1, exec_context.private_tmp)
-$1.PrivateNetwork,               config_parse_bool,                  0,                             offsetof($1, exec_context.private_network)
 $1.PrivateDevices,               config_parse_bool,                  0,                             offsetof($1, exec_context.private_devices)
+$1.PrivateNetwork,               config_parse_bool,                  0,                             offsetof($1, exec_context.private_network)
+$1.PrivateUsers,                 config_parse_bool,                  0,                             offsetof($1, exec_context.private_users)
 $1.ProtectSystem,                config_parse_protect_system,        0,                             offsetof($1, exec_context)
 $1.ProtectHome,                  config_parse_protect_home,          0,                             offsetof($1, exec_context)
 $1.MountFlags,                   config_parse_exec_mount_flags,      0,                             offsetof($1, exec_context)