1 /* SPDX-License-Identifier: LGPL-2.1+ */
6 #include <sys/eventfd.h>
9 #include <sys/personality.h>
10 #include <sys/prctl.h>
12 #include <sys/types.h>
17 #include <dlog/dlog-redirect-stdout.h>
20 #include <security/pam_appl.h>
24 #include <selinux/selinux.h>
32 #include <sys/apparmor.h>
35 #include "sd-messages.h"
38 #include "alloc-util.h"
40 #include "apparmor-util.h"
45 #include "capability-util.h"
46 #include "chown-recursive.h"
47 #include "cgroup-setup.h"
48 #include "cpu-set-util.h"
52 #include "errno-list.h"
54 #include "exit-status.h"
56 #include "format-util.h"
58 #include "glob-util.h"
65 #include "memory-util.h"
66 #include "missing_fs.h"
68 #include "namespace.h"
69 #include "parse-util.h"
70 #include "path-util.h"
71 #include "process-util.h"
72 #include "rlimit-util.h"
75 #include "seccomp-util.h"
77 #include "securebits-util.h"
78 #include "selinux-util.h"
79 #include "signal-util.h"
80 #include "smack-util.h"
81 #include "socket-util.h"
83 #include "stat-util.h"
84 #include "string-table.h"
85 #include "string-util.h"
87 #include "syslog-util.h"
88 #include "terminal-util.h"
89 #include "umask-util.h"
91 #include "user-util.h"
92 #include "utmp-wtmp.h"
94 #define IDLE_TIMEOUT_USEC (5*USEC_PER_SEC)
95 #define IDLE_TIMEOUT2_USEC (1*USEC_PER_SEC)
97 #define SNDBUF_SIZE (8*1024*1024)
99 static int shift_fds(int fds[], size_t n_fds) {
100 int start, restart_from;
105 /* Modifies the fds array! (sorts it) */
115 for (i = start; i < (int) n_fds; i++) {
118 /* Already at right index? */
122 nfd = fcntl(fds[i], F_DUPFD, i + 3);
129 /* Hmm, the fd we wanted isn't free? Then
130 * let's remember that and try again from here */
131 if (nfd != i+3 && restart_from < 0)
135 if (restart_from < 0)
138 start = restart_from;
144 static int flags_fds(const int fds[], size_t n_socket_fds, size_t n_storage_fds, bool nonblock) {
148 n_fds = n_socket_fds + n_storage_fds;
154 /* Drops/Sets O_NONBLOCK and FD_CLOEXEC from the file flags.
155 * O_NONBLOCK only applies to socket activation though. */
157 for (i = 0; i < n_fds; i++) {
159 if (i < n_socket_fds) {
160 r = fd_nonblock(fds[i], nonblock);
165 /* We unconditionally drop FD_CLOEXEC from the fds,
166 * since after all we want to pass these fds to our
169 r = fd_cloexec(fds[i], false);
177 static const char *exec_context_tty_path(const ExecContext *context) {
180 if (context->stdio_as_fds)
183 if (context->tty_path)
184 return context->tty_path;
186 return "/dev/console";
189 static void exec_context_tty_reset(const ExecContext *context, const ExecParameters *p) {
194 path = exec_context_tty_path(context);
196 if (context->tty_vhangup) {
197 if (p && p->stdin_fd >= 0)
198 (void) terminal_vhangup_fd(p->stdin_fd);
200 (void) terminal_vhangup(path);
203 if (context->tty_reset) {
204 if (p && p->stdin_fd >= 0)
205 (void) reset_terminal_fd(p->stdin_fd, true);
207 (void) reset_terminal(path);
210 if (context->tty_vt_disallocate && path)
211 (void) vt_disallocate(path);
214 static bool is_terminal_input(ExecInput i) {
217 EXEC_INPUT_TTY_FORCE,
218 EXEC_INPUT_TTY_FAIL);
221 static bool is_terminal_output(ExecOutput o) {
224 EXEC_OUTPUT_SYSLOG_AND_CONSOLE,
225 EXEC_OUTPUT_KMSG_AND_CONSOLE,
226 EXEC_OUTPUT_JOURNAL_AND_CONSOLE);
229 static bool is_syslog_output(ExecOutput o) {
232 EXEC_OUTPUT_SYSLOG_AND_CONSOLE);
235 static bool is_kmsg_output(ExecOutput o) {
238 EXEC_OUTPUT_KMSG_AND_CONSOLE);
241 static bool exec_context_needs_term(const ExecContext *c) {
244 /* Return true if the execution context suggests we should set $TERM to something useful. */
246 if (is_terminal_input(c->std_input))
249 if (is_terminal_output(c->std_output))
252 if (is_terminal_output(c->std_error))
255 return !!c->tty_path;
258 static int open_device_as(const char *devname, int flags, int nfd) {
263 fd = open(devname, flags|O_NOCTTY);
267 return move_fd(fd, nfd, false);
270 static int open_null_as(int flags, int nfd) {
271 return open_device_as("/dev/null", flags, nfd);
274 static int open_kmsg_as(int flags, int nfd) {
275 return open_device_as("/dev/kmsg", flags, nfd);
278 static int connect_journal_socket(int fd, uid_t uid, gid_t gid) {
279 static const union sockaddr_union sa = {
280 .un.sun_family = AF_UNIX,
281 .un.sun_path = "/run/systemd/journal/stdout",
283 uid_t olduid = UID_INVALID;
284 gid_t oldgid = GID_INVALID;
287 if (gid_is_valid(gid)) {
290 if (setegid(gid) < 0)
294 if (uid_is_valid(uid)) {
297 if (seteuid(uid) < 0) {
303 r = connect(fd, &sa.sa, SOCKADDR_UN_LEN(sa.un)) < 0 ? -errno : 0;
305 /* If we fail to restore the uid or gid, things will likely
306 fail later on. This should only happen if an LSM interferes. */
308 if (uid_is_valid(uid))
309 (void) seteuid(olduid);
312 if (gid_is_valid(gid))
313 (void) setegid(oldgid);
318 static int connect_logger_as(
320 const ExecContext *context,
321 const ExecParameters *params,
328 _cleanup_close_ int fd = -1;
333 assert(output < _EXEC_OUTPUT_MAX);
337 fd = socket(AF_UNIX, SOCK_STREAM, 0);
341 r = connect_journal_socket(fd, uid, gid);
345 if (shutdown(fd, SHUT_RD) < 0)
348 (void) fd_inc_sndbuf(fd, SNDBUF_SIZE);
358 context->syslog_identifier ?: ident,
359 params->flags & EXEC_PASS_LOG_UNIT ? unit->id : "",
360 context->syslog_priority,
361 !!context->syslog_level_prefix,
362 is_syslog_output(output),
363 is_kmsg_output(output),
364 is_terminal_output(output)) < 0)
367 return move_fd(TAKE_FD(fd), nfd, false);
370 static int open_terminal_as(const char *path, int flags, int nfd) {
376 fd = open_terminal(path, flags | O_NOCTTY);
380 return move_fd(fd, nfd, false);
383 static int acquire_path(const char *path, int flags, mode_t mode) {
384 union sockaddr_union sa = {};
385 _cleanup_close_ int fd = -1;
390 if (IN_SET(flags & O_ACCMODE, O_WRONLY, O_RDWR))
393 fd = open(path, flags|O_NOCTTY, mode);
397 if (errno != ENXIO) /* ENXIO is returned when we try to open() an AF_UNIX file system socket on Linux */
399 if (strlen(path) >= sizeof(sa.un.sun_path)) /* Too long, can't be a UNIX socket */
402 /* So, it appears the specified path could be an AF_UNIX socket. Let's see if we can connect to it. */
404 fd = socket(AF_UNIX, SOCK_STREAM, 0);
408 salen = sockaddr_un_set_path(&sa.un, path);
412 if (connect(fd, &sa.sa, salen) < 0)
413 return errno == EINVAL ? -ENXIO : -errno; /* Propagate initial error if we get EINVAL, i.e. we have
414 * indication that his wasn't an AF_UNIX socket after all */
416 if ((flags & O_ACCMODE) == O_RDONLY)
417 r = shutdown(fd, SHUT_WR);
418 else if ((flags & O_ACCMODE) == O_WRONLY)
419 r = shutdown(fd, SHUT_RD);
428 static int fixup_input(
429 const ExecContext *context,
431 bool apply_tty_stdin) {
437 std_input = context->std_input;
439 if (is_terminal_input(std_input) && !apply_tty_stdin)
440 return EXEC_INPUT_NULL;
442 if (std_input == EXEC_INPUT_SOCKET && socket_fd < 0)
443 return EXEC_INPUT_NULL;
445 if (std_input == EXEC_INPUT_DATA && context->stdin_data_size == 0)
446 return EXEC_INPUT_NULL;
451 static int fixup_output(ExecOutput std_output, int socket_fd) {
453 if (std_output == EXEC_OUTPUT_SOCKET && socket_fd < 0)
454 return EXEC_OUTPUT_INHERIT;
459 static int setup_input(
460 const ExecContext *context,
461 const ExecParameters *params,
463 const int named_iofds[static 3]) {
471 if (params->stdin_fd >= 0) {
472 if (dup2(params->stdin_fd, STDIN_FILENO) < 0)
475 /* Try to make this the controlling tty, if it is a tty, and reset it */
476 if (isatty(STDIN_FILENO)) {
477 (void) ioctl(STDIN_FILENO, TIOCSCTTY, context->std_input == EXEC_INPUT_TTY_FORCE);
478 (void) reset_terminal_fd(STDIN_FILENO, true);
484 i = fixup_input(context, socket_fd, params->flags & EXEC_APPLY_TTY_STDIN);
488 case EXEC_INPUT_NULL:
489 return open_null_as(O_RDONLY, STDIN_FILENO);
492 case EXEC_INPUT_TTY_FORCE:
493 case EXEC_INPUT_TTY_FAIL: {
496 fd = acquire_terminal(exec_context_tty_path(context),
497 i == EXEC_INPUT_TTY_FAIL ? ACQUIRE_TERMINAL_TRY :
498 i == EXEC_INPUT_TTY_FORCE ? ACQUIRE_TERMINAL_FORCE :
499 ACQUIRE_TERMINAL_WAIT,
504 return move_fd(fd, STDIN_FILENO, false);
507 case EXEC_INPUT_SOCKET:
508 assert(socket_fd >= 0);
510 return dup2(socket_fd, STDIN_FILENO) < 0 ? -errno : STDIN_FILENO;
512 case EXEC_INPUT_NAMED_FD:
513 assert(named_iofds[STDIN_FILENO] >= 0);
515 (void) fd_nonblock(named_iofds[STDIN_FILENO], false);
516 return dup2(named_iofds[STDIN_FILENO], STDIN_FILENO) < 0 ? -errno : STDIN_FILENO;
518 case EXEC_INPUT_DATA: {
521 fd = acquire_data_fd(context->stdin_data, context->stdin_data_size, 0);
525 return move_fd(fd, STDIN_FILENO, false);
528 case EXEC_INPUT_FILE: {
532 assert(context->stdio_file[STDIN_FILENO]);
534 rw = (context->std_output == EXEC_OUTPUT_FILE && streq_ptr(context->stdio_file[STDIN_FILENO], context->stdio_file[STDOUT_FILENO])) ||
535 (context->std_error == EXEC_OUTPUT_FILE && streq_ptr(context->stdio_file[STDIN_FILENO], context->stdio_file[STDERR_FILENO]));
537 fd = acquire_path(context->stdio_file[STDIN_FILENO], rw ? O_RDWR : O_RDONLY, 0666 & ~context->umask);
541 return move_fd(fd, STDIN_FILENO, false);
545 assert_not_reached("Unknown input type");
549 static bool can_inherit_stderr_from_stdout(
550 const ExecContext *context,
556 /* Returns true, if given the specified STDERR and STDOUT output we can directly dup() the stdout fd to the
559 if (e == EXEC_OUTPUT_INHERIT)
564 if (e == EXEC_OUTPUT_NAMED_FD)
565 return streq_ptr(context->stdio_fdname[STDOUT_FILENO], context->stdio_fdname[STDERR_FILENO]);
567 if (IN_SET(e, EXEC_OUTPUT_FILE, EXEC_OUTPUT_FILE_APPEND))
568 return streq_ptr(context->stdio_file[STDOUT_FILENO], context->stdio_file[STDERR_FILENO]);
570 /* Returns false if STDERR output is redirected to dlog */
571 if (IN_SET(e, EXEC_OUTPUT_DLOG, EXEC_OUTPUT_DLOG_OR_NULL, EXEC_OUTPUT_DLOG_OR_KMSG, EXEC_OUTPUT_DLOG_OR_JOURNAL))
577 static int wire_up_dlog(int fileno, const char *ident)
579 /* NB: dlogutil has FOO* tag filtering wildcards but not *FOO
580 * therefore the STDFOO part of the final tag goes in front.
582 * The tag can nominally be longer than 128 but dlogutil will
583 * not print a longer one anyway and it's good to limit overhead */
587 if (fileno == STDERR_FILENO) {
588 priority = 6; // ERROR
589 snprintf(tag, sizeof tag, "STDERR_%s", ident);
591 priority = 4; // INFO
592 snprintf(tag, sizeof tag, "STDOUT_%s", ident);
595 return dlog_connect_fd(2, // SYSTEM dlog buffer (consider also 0 for generic MAIN buffer)
596 fileno, tag, priority);
599 static int setup_output(
601 const ExecContext *context,
602 const ExecParameters *params,
605 const int named_iofds[static 3],
609 dev_t *journal_stream_dev,
610 ino_t *journal_stream_ino) {
620 assert(journal_stream_dev);
621 assert(journal_stream_ino);
623 if (fileno == STDOUT_FILENO && params->stdout_fd >= 0) {
625 if (dup2(params->stdout_fd, STDOUT_FILENO) < 0)
628 return STDOUT_FILENO;
631 if (fileno == STDERR_FILENO && params->stderr_fd >= 0) {
632 if (dup2(params->stderr_fd, STDERR_FILENO) < 0)
635 return STDERR_FILENO;
638 i = fixup_input(context, socket_fd, params->flags & EXEC_APPLY_TTY_STDIN);
639 o = fixup_output(context->std_output, socket_fd);
641 if (fileno == STDERR_FILENO) {
643 e = fixup_output(context->std_error, socket_fd);
645 /* This expects the input and output are already set up */
647 /* Don't change the stderr file descriptor if we inherit all
648 * the way and are not on a tty */
649 if (e == EXEC_OUTPUT_INHERIT &&
650 o == EXEC_OUTPUT_INHERIT &&
651 i == EXEC_INPUT_NULL &&
652 !is_terminal_input(context->std_input) &&
656 /* Duplicate from stdout if possible */
657 if (can_inherit_stderr_from_stdout(context, o, e))
658 return dup2(STDOUT_FILENO, fileno) < 0 ? -errno : fileno;
662 } else if (o == EXEC_OUTPUT_INHERIT) {
663 /* If input got downgraded, inherit the original value */
664 if (i == EXEC_INPUT_NULL && is_terminal_input(context->std_input))
665 return open_terminal_as(exec_context_tty_path(context), O_WRONLY, fileno);
667 /* If the input is connected to anything that's not a /dev/null or a data fd, inherit that... */
668 if (!IN_SET(i, EXEC_INPUT_NULL, EXEC_INPUT_DATA))
669 return dup2(STDIN_FILENO, fileno) < 0 ? -errno : fileno;
671 /* If we are not started from PID 1 we just inherit STDOUT from our parent process. */
675 /* We need to open /dev/null here anew, to get the right access mode. */
676 return open_null_as(O_WRONLY, fileno);
681 case EXEC_OUTPUT_NULL:
682 return open_null_as(O_WRONLY, fileno);
684 case EXEC_OUTPUT_TTY:
685 if (is_terminal_input(i))
686 return dup2(STDIN_FILENO, fileno) < 0 ? -errno : fileno;
688 /* We don't reset the terminal if this is just about output */
689 return open_terminal_as(exec_context_tty_path(context), O_WRONLY, fileno);
691 case EXEC_OUTPUT_DLOG_OR_JOURNAL:
692 r = wire_up_dlog(fileno, ident);
696 o = EXEC_OUTPUT_JOURNAL; // used inside `connect_logger_as`, but let's not involve dlog there
699 case EXEC_OUTPUT_SYSLOG:
700 case EXEC_OUTPUT_SYSLOG_AND_CONSOLE:
701 case EXEC_OUTPUT_KMSG:
702 case EXEC_OUTPUT_KMSG_AND_CONSOLE:
703 case EXEC_OUTPUT_JOURNAL:
704 case EXEC_OUTPUT_JOURNAL_AND_CONSOLE:
705 r = connect_logger_as(unit, context, params, o, ident, fileno, uid, gid);
707 log_unit_warning_errno(unit, r, "Failed to connect %s to the journal socket, ignoring: %m", fileno == STDOUT_FILENO ? "stdout" : "stderr");
708 r = open_null_as(O_WRONLY, fileno);
712 /* If we connected this fd to the journal via a stream, patch the device/inode into the passed
713 * parameters, but only then. This is useful so that we can set $JOURNAL_STREAM that permits
714 * services to detect whether they are connected to the journal or not.
716 * If both stdout and stderr are connected to a stream then let's make sure to store the data
717 * about STDERR as that's usually the best way to do logging. */
719 if (fstat(fileno, &st) >= 0 &&
720 (*journal_stream_ino == 0 || fileno == STDERR_FILENO)) {
721 *journal_stream_dev = st.st_dev;
722 *journal_stream_ino = st.st_ino;
727 case EXEC_OUTPUT_DLOG:
728 r = wire_up_dlog(fileno, ident);
729 return r < 0 ? r : fileno;
731 case EXEC_OUTPUT_DLOG_OR_NULL:
732 r = wire_up_dlog(fileno, ident);
736 return open_null_as(O_WRONLY, fileno);
738 case EXEC_OUTPUT_DLOG_OR_KMSG:
739 r = wire_up_dlog(fileno, ident);
743 return open_kmsg_as(O_WRONLY, fileno);
745 case EXEC_OUTPUT_SOCKET:
746 assert(socket_fd >= 0);
748 return dup2(socket_fd, fileno) < 0 ? -errno : fileno;
750 case EXEC_OUTPUT_NAMED_FD:
751 assert(named_iofds[fileno] >= 0);
753 (void) fd_nonblock(named_iofds[fileno], false);
754 return dup2(named_iofds[fileno], fileno) < 0 ? -errno : fileno;
756 case EXEC_OUTPUT_FILE:
757 case EXEC_OUTPUT_FILE_APPEND: {
761 assert(context->stdio_file[fileno]);
763 rw = context->std_input == EXEC_INPUT_FILE &&
764 streq_ptr(context->stdio_file[fileno], context->stdio_file[STDIN_FILENO]);
767 return dup2(STDIN_FILENO, fileno) < 0 ? -errno : fileno;
770 if (o == EXEC_OUTPUT_FILE_APPEND)
773 fd = acquire_path(context->stdio_file[fileno], flags, 0666 & ~context->umask);
777 return move_fd(fd, fileno, 0);
781 assert_not_reached("Unknown error type");
785 static int chown_terminal(int fd, uid_t uid) {
790 /* Before we chown/chmod the TTY, let's ensure this is actually a tty */
791 if (isatty(fd) < 1) {
792 if (IN_SET(errno, EINVAL, ENOTTY))
793 return 0; /* not a tty */
798 /* This might fail. What matters are the results. */
799 r = fchmod_and_chown(fd, TTY_MODE, uid, -1);
806 static int setup_confirm_stdio(const char *vc, int *_saved_stdin, int *_saved_stdout) {
807 _cleanup_close_ int fd = -1, saved_stdin = -1, saved_stdout = -1;
810 assert(_saved_stdin);
811 assert(_saved_stdout);
813 saved_stdin = fcntl(STDIN_FILENO, F_DUPFD, 3);
817 saved_stdout = fcntl(STDOUT_FILENO, F_DUPFD, 3);
818 if (saved_stdout < 0)
821 fd = acquire_terminal(vc, ACQUIRE_TERMINAL_WAIT, DEFAULT_CONFIRM_USEC);
825 r = chown_terminal(fd, getuid());
829 r = reset_terminal_fd(fd, true);
833 r = rearrange_stdio(fd, fd, STDERR_FILENO);
838 *_saved_stdin = saved_stdin;
839 *_saved_stdout = saved_stdout;
841 saved_stdin = saved_stdout = -1;
846 static void write_confirm_error_fd(int err, int fd, const Unit *u) {
849 if (err == -ETIMEDOUT)
850 dprintf(fd, "Confirmation question timed out for %s, assuming positive response.\n", u->id);
853 dprintf(fd, "Couldn't ask confirmation for %s: %m, assuming positive response.\n", u->id);
857 static void write_confirm_error(int err, const char *vc, const Unit *u) {
858 _cleanup_close_ int fd = -1;
862 fd = open_terminal(vc, O_WRONLY|O_NOCTTY|O_CLOEXEC);
866 write_confirm_error_fd(err, fd, u);
869 static int restore_confirm_stdio(int *saved_stdin, int *saved_stdout) {
873 assert(saved_stdout);
877 if (*saved_stdin >= 0)
878 if (dup2(*saved_stdin, STDIN_FILENO) < 0)
881 if (*saved_stdout >= 0)
882 if (dup2(*saved_stdout, STDOUT_FILENO) < 0)
885 *saved_stdin = safe_close(*saved_stdin);
886 *saved_stdout = safe_close(*saved_stdout);
892 CONFIRM_PRETEND_FAILURE = -1,
893 CONFIRM_PRETEND_SUCCESS = 0,
897 static int ask_for_confirmation(const char *vc, Unit *u, const char *cmdline) {
898 int saved_stdout = -1, saved_stdin = -1, r;
899 _cleanup_free_ char *e = NULL;
902 /* For any internal errors, assume a positive response. */
903 r = setup_confirm_stdio(vc, &saved_stdin, &saved_stdout);
905 write_confirm_error(r, vc, u);
906 return CONFIRM_EXECUTE;
909 /* confirm_spawn might have been disabled while we were sleeping. */
910 if (manager_is_confirm_spawn_disabled(u->manager)) {
915 e = ellipsize(cmdline, 60, 100);
923 r = ask_char(&c, "yfshiDjcn", "Execute %s? [y, f, s – h for help] ", e);
925 write_confirm_error_fd(r, STDOUT_FILENO, u);
932 printf("Resuming normal execution.\n");
933 manager_disable_confirm_spawn();
937 unit_dump(u, stdout, " ");
938 continue; /* ask again */
940 printf("Failing execution.\n");
941 r = CONFIRM_PRETEND_FAILURE;
944 printf(" c - continue, proceed without asking anymore\n"
945 " D - dump, show the state of the unit\n"
946 " f - fail, don't execute the command and pretend it failed\n"
948 " i - info, show a short summary of the unit\n"
949 " j - jobs, show jobs that are in progress\n"
950 " s - skip, don't execute the command and pretend it succeeded\n"
951 " y - yes, execute the command\n");
952 continue; /* ask again */
954 printf(" Description: %s\n"
957 u->id, u->description, cmdline);
958 continue; /* ask again */
960 manager_dump_jobs(u->manager, stdout, " ");
961 continue; /* ask again */
963 /* 'n' was removed in favor of 'f'. */
964 printf("Didn't understand 'n', did you mean 'f'?\n");
965 continue; /* ask again */
967 printf("Skipping execution.\n");
968 r = CONFIRM_PRETEND_SUCCESS;
974 assert_not_reached("Unhandled choice");
980 restore_confirm_stdio(&saved_stdin, &saved_stdout);
984 static int get_fixed_user(const ExecContext *c, const char **user,
985 uid_t *uid, gid_t *gid,
986 const char **home, const char **shell) {
995 /* Note that we don't set $HOME or $SHELL if they are not particularly enlightening anyway
996 * (i.e. are "/" or "/bin/nologin"). */
999 r = get_user_creds(&name, uid, gid, home, shell, USER_CREDS_CLEAN);
1007 static int get_fixed_group(const ExecContext *c, const char **group, gid_t *gid) {
1017 r = get_group_creds(&name, gid, 0);
1025 static int get_supplementary_groups(const ExecContext *c, const char *user,
1026 const char *group, gid_t gid,
1027 gid_t **supplementary_gids, int *ngids) {
1031 bool keep_groups = false;
1032 gid_t *groups = NULL;
1033 _cleanup_free_ gid_t *l_gids = NULL;
1038 * If user is given, then lookup GID and supplementary groups list.
1039 * We avoid NSS lookups for gid=0. Also we have to initialize groups
1040 * here and as early as possible so we keep the list of supplementary
1041 * groups of the caller.
1043 if (user && gid_is_valid(gid) && gid != 0) {
1044 /* First step, initialize groups from /etc/groups */
1045 if (initgroups(user, gid) < 0)
1051 if (strv_isempty(c->supplementary_groups))
1055 * If SupplementaryGroups= was passed then NGROUPS_MAX has to
1056 * be positive, otherwise fail.
1059 ngroups_max = (int) sysconf(_SC_NGROUPS_MAX);
1060 if (ngroups_max <= 0)
1061 return errno_or_else(EOPNOTSUPP);
1063 l_gids = new(gid_t, ngroups_max);
1069 * Lookup the list of groups that the user belongs to, we
1070 * avoid NSS lookups here too for gid=0.
1073 if (getgrouplist(user, gid, l_gids, &k) < 0)
1078 STRV_FOREACH(i, c->supplementary_groups) {
1081 if (k >= ngroups_max)
1085 r = get_group_creds(&g, l_gids+k, 0);
1093 * Sets ngids to zero to drop all supplementary groups, happens
1094 * when we are under root and SupplementaryGroups= is empty.
1101 /* Otherwise get the final list of supplementary groups */
1102 groups = memdup(l_gids, sizeof(gid_t) * k);
1106 *supplementary_gids = groups;
1114 static int enforce_groups(gid_t gid, const gid_t *supplementary_gids, int ngids) {
1117 /* Handle SupplementaryGroups= if it is not empty */
1119 r = maybe_setgroups(ngids, supplementary_gids);
1124 if (gid_is_valid(gid)) {
1125 /* Then set our gids */
1126 if (setresgid(gid, gid, gid) < 0)
1133 static int enforce_user(const ExecContext *context, uid_t uid) {
1136 if (!uid_is_valid(uid))
1139 /* Sets (but doesn't look up) the uid and make sure we keep the
1140 * capabilities while doing so. */
1142 if (context->capabilities || context->capability_ambient_set != 0) {
1144 /* First step: If we need to keep capabilities but
1145 * drop privileges we need to make sure we keep our
1146 * caps, while we drop privileges. */
1148 int sb = context->secure_bits | 1<<SECURE_KEEP_CAPS;
1150 if (prctl(PR_GET_SECUREBITS) != sb)
1151 if (prctl(PR_SET_SECUREBITS, sb) < 0)
1154 /* Second step: set the capabilities. This will reduce
1155 * the capabilities to the minimum we need. */
1157 if (context->capabilities) {
1158 _cleanup_cap_free_ cap_t d = NULL;
1159 static const cap_value_t bits[] = {
1160 CAP_SETUID, /* Necessary so that we can run setresuid() below */
1161 CAP_SETPCAP /* Necessary so that we can set PR_SET_SECUREBITS later on */
1164 d = cap_dup(context->capabilities);
1168 if (cap_set_flag(d, CAP_EFFECTIVE, ELEMENTSOF(bits), bits, CAP_SET) < 0 ||
1169 cap_set_flag(d, CAP_PERMITTED, ELEMENTSOF(bits), bits, CAP_SET) < 0)
1172 if (cap_set_proc(d) < 0)
1177 /* Third step: actually set the uids */
1178 if (setresuid(uid, uid, uid) < 0)
1181 /* At this point we should have all necessary capabilities but
1182 are otherwise a normal user. However, the caps might got
1183 corrupted due to the setresuid() so we need clean them up
1184 later. This is done outside of this call. */
1191 static int null_conv(
1193 const struct pam_message **msg,
1194 struct pam_response **resp,
1195 void *appdata_ptr) {
1197 /* We don't support conversations */
1199 return PAM_CONV_ERR;
1204 static int setup_pam(
1211 int fds[], size_t n_fds) {
1215 static const struct pam_conv conv = {
1220 _cleanup_(barrier_destroy) Barrier barrier = BARRIER_NULL;
1221 pam_handle_t *handle = NULL;
1223 int pam_code = PAM_SUCCESS, r;
1224 char **nv, **e = NULL;
1225 bool close_session = false;
1226 pid_t pam_pid = 0, parent_pid;
1233 /* We set up PAM in the parent process, then fork. The child
1234 * will then stay around until killed via PR_GET_PDEATHSIG or
1235 * systemd via the cgroup logic. It will then remove the PAM
1236 * session again. The parent process will exec() the actual
1237 * daemon. We do things this way to ensure that the main PID
1238 * of the daemon is the one we initially fork()ed. */
1240 r = barrier_create(&barrier);
1244 if (log_get_max_level() < LOG_DEBUG)
1245 flags |= PAM_SILENT;
1247 pam_code = pam_start(name, user, &conv, &handle);
1248 if (pam_code != PAM_SUCCESS) {
1254 _cleanup_free_ char *q = NULL;
1256 /* Hmm, so no TTY was explicitly passed, but an fd passed to us directly might be a TTY. Let's figure
1257 * out if that's the case, and read the TTY off it. */
1259 if (getttyname_malloc(STDIN_FILENO, &q) >= 0)
1260 tty = strjoina("/dev/", q);
1264 pam_code = pam_set_item(handle, PAM_TTY, tty);
1265 if (pam_code != PAM_SUCCESS)
1269 STRV_FOREACH(nv, *env) {
1270 pam_code = pam_putenv(handle, *nv);
1271 if (pam_code != PAM_SUCCESS)
1275 pam_code = pam_acct_mgmt(handle, flags);
1276 if (pam_code != PAM_SUCCESS)
1279 pam_code = pam_open_session(handle, flags);
1280 if (pam_code != PAM_SUCCESS)
1283 close_session = true;
1285 e = pam_getenvlist(handle);
1287 pam_code = PAM_BUF_ERR;
1291 /* Block SIGTERM, so that we know that it won't get lost in
1294 assert_se(sigprocmask_many(SIG_BLOCK, &old_ss, SIGTERM, -1) >= 0);
1296 parent_pid = getpid_cached();
1298 r = safe_fork("(sd-pam)", 0, &pam_pid);
1302 int sig, ret = EXIT_PAM;
1304 /* The child's job is to reset the PAM session on
1306 barrier_set_role(&barrier, BARRIER_CHILD);
1308 /* Make sure we don't keep open the passed fds in this child. We assume that otherwise only those fds
1309 * are open here that have been opened by PAM. */
1310 (void) close_many(fds, n_fds);
1312 /* Drop privileges - we don't need any to pam_close_session
1313 * and this will make PR_SET_PDEATHSIG work in most cases.
1314 * If this fails, ignore the error - but expect sd-pam threads
1315 * to fail to exit normally */
1317 r = maybe_setgroups(0, NULL);
1319 log_warning_errno(r, "Failed to setgroups() in sd-pam: %m");
1320 if (setresgid(gid, gid, gid) < 0)
1321 log_warning_errno(errno, "Failed to setresgid() in sd-pam: %m");
1322 if (setresuid(uid, uid, uid) < 0)
1323 log_warning_errno(errno, "Failed to setresuid() in sd-pam: %m");
1325 (void) ignore_signals(SIGPIPE, -1);
1327 /* Wait until our parent died. This will only work if
1328 * the above setresuid() succeeds, otherwise the kernel
1329 * will not allow unprivileged parents kill their privileged
1330 * children this way. We rely on the control groups kill logic
1331 * to do the rest for us. */
1332 if (prctl(PR_SET_PDEATHSIG, SIGTERM) < 0)
1335 /* Tell the parent that our setup is done. This is especially
1336 * important regarding dropping privileges. Otherwise, unit
1337 * setup might race against our setresuid(2) call.
1339 * If the parent aborted, we'll detect this below, hence ignore
1340 * return failure here. */
1341 (void) barrier_place(&barrier);
1343 /* Check if our parent process might already have died? */
1344 if (getppid() == parent_pid) {
1347 assert_se(sigemptyset(&ss) >= 0);
1348 assert_se(sigaddset(&ss, SIGTERM) >= 0);
1351 if (sigwait(&ss, &sig) < 0) {
1358 assert(sig == SIGTERM);
1363 /* If our parent died we'll end the session */
1364 if (getppid() != parent_pid) {
1365 pam_code = pam_close_session(handle, flags);
1366 if (pam_code != PAM_SUCCESS)
1373 pam_end(handle, pam_code | flags);
1377 barrier_set_role(&barrier, BARRIER_PARENT);
1379 /* If the child was forked off successfully it will do all the
1380 * cleanups, so forget about the handle here. */
1383 /* Unblock SIGTERM again in the parent */
1384 assert_se(sigprocmask(SIG_SETMASK, &old_ss, NULL) >= 0);
1386 /* We close the log explicitly here, since the PAM modules
1387 * might have opened it, but we don't want this fd around. */
1390 /* Synchronously wait for the child to initialize. We don't care for
1391 * errors as we cannot recover. However, warn loudly if it happens. */
1392 if (!barrier_place_and_sync(&barrier))
1393 log_error("PAM initialization failed");
1395 return strv_free_and_replace(*env, e);
1398 if (pam_code != PAM_SUCCESS) {
1399 log_error("PAM failed: %s", pam_strerror(handle, pam_code));
1400 r = -EPERM; /* PAM errors do not map to errno */
1402 log_error_errno(r, "PAM failed: %m");
1406 pam_code = pam_close_session(handle, flags);
1408 pam_end(handle, pam_code | flags);
1420 static void rename_process_from_path(const char *path) {
1421 char process_name[11];
1425 /* This resulting string must fit in 10 chars (i.e. the length
1426 * of "/sbin/init") to look pretty in /bin/ps */
1430 rename_process("(...)");
1436 /* The end of the process name is usually more
1437 * interesting, since the first bit might just be
1443 process_name[0] = '(';
1444 memcpy(process_name+1, p, l);
1445 process_name[1+l] = ')';
1446 process_name[1+l+1] = 0;
1448 rename_process(process_name);
1451 static bool context_has_address_families(const ExecContext *c) {
1454 return c->address_families_whitelist ||
1455 !set_isempty(c->address_families);
1458 static bool context_has_syscall_filters(const ExecContext *c) {
1461 return c->syscall_whitelist ||
1462 !hashmap_isempty(c->syscall_filter);
1465 static bool context_has_no_new_privileges(const ExecContext *c) {
1468 if (c->no_new_privileges)
1471 if (have_effective_cap(CAP_SYS_ADMIN)) /* if we are privileged, we don't need NNP */
1474 /* We need NNP if we have any form of seccomp and are unprivileged */
1475 return context_has_address_families(c) ||
1476 c->memory_deny_write_execute ||
1477 c->restrict_realtime ||
1478 c->restrict_suid_sgid ||
1479 exec_context_restrict_namespaces_set(c) ||
1480 c->protect_kernel_tunables ||
1481 c->protect_kernel_modules ||
1482 c->protect_kernel_logs ||
1483 c->private_devices ||
1484 context_has_syscall_filters(c) ||
1485 !set_isempty(c->syscall_archs) ||
1486 c->lock_personality ||
1487 c->protect_hostname;
1492 static bool skip_seccomp_unavailable(const Unit* u, const char* msg) {
1494 if (is_seccomp_available())
1497 log_unit_debug(u, "SECCOMP features not detected in the kernel, skipping %s", msg);
1501 static int apply_syscall_filter(const Unit* u, const ExecContext *c, bool needs_ambient_hack) {
1502 uint32_t negative_action, default_action, action;
1508 if (!context_has_syscall_filters(c))
1511 if (skip_seccomp_unavailable(u, "SystemCallFilter="))
1514 negative_action = c->syscall_errno == 0 ? scmp_act_kill_process() : SCMP_ACT_ERRNO(c->syscall_errno);
1516 if (c->syscall_whitelist) {
1517 default_action = negative_action;
1518 action = SCMP_ACT_ALLOW;
1520 default_action = SCMP_ACT_ALLOW;
1521 action = negative_action;
1524 if (needs_ambient_hack) {
1525 r = seccomp_filter_set_add(c->syscall_filter, c->syscall_whitelist, syscall_filter_sets + SYSCALL_FILTER_SET_SETUID);
1530 return seccomp_load_syscall_filter_set_raw(default_action, c->syscall_filter, action, false);
1533 static int apply_syscall_archs(const Unit *u, const ExecContext *c) {
1537 if (set_isempty(c->syscall_archs))
1540 if (skip_seccomp_unavailable(u, "SystemCallArchitectures="))
1543 return seccomp_restrict_archs(c->syscall_archs);
1546 static int apply_address_families(const Unit* u, const ExecContext *c) {
1550 if (!context_has_address_families(c))
1553 if (skip_seccomp_unavailable(u, "RestrictAddressFamilies="))
1556 return seccomp_restrict_address_families(c->address_families, c->address_families_whitelist);
1559 static int apply_memory_deny_write_execute(const Unit* u, const ExecContext *c) {
1563 if (!c->memory_deny_write_execute)
1566 if (skip_seccomp_unavailable(u, "MemoryDenyWriteExecute="))
1569 return seccomp_memory_deny_write_execute();
1572 static int apply_restrict_realtime(const Unit* u, const ExecContext *c) {
1576 if (!c->restrict_realtime)
1579 if (skip_seccomp_unavailable(u, "RestrictRealtime="))
1582 return seccomp_restrict_realtime();
1585 static int apply_restrict_suid_sgid(const Unit* u, const ExecContext *c) {
1589 if (!c->restrict_suid_sgid)
1592 if (skip_seccomp_unavailable(u, "RestrictSUIDSGID="))
1595 return seccomp_restrict_suid_sgid();
1598 static int apply_protect_sysctl(const Unit *u, const ExecContext *c) {
1602 /* Turn off the legacy sysctl() system call. Many distributions turn this off while building the kernel, but
1603 * let's protect even those systems where this is left on in the kernel. */
1605 if (!c->protect_kernel_tunables)
1608 if (skip_seccomp_unavailable(u, "ProtectKernelTunables="))
1611 return seccomp_protect_sysctl();
1614 static int apply_protect_kernel_modules(const Unit *u, const ExecContext *c) {
1618 /* Turn off module syscalls on ProtectKernelModules=yes */
1620 if (!c->protect_kernel_modules)
1623 if (skip_seccomp_unavailable(u, "ProtectKernelModules="))
1626 return seccomp_load_syscall_filter_set(SCMP_ACT_ALLOW, syscall_filter_sets + SYSCALL_FILTER_SET_MODULE, SCMP_ACT_ERRNO(EPERM), false);
1629 static int apply_protect_kernel_logs(const Unit *u, const ExecContext *c) {
1633 if (!c->protect_kernel_logs)
1636 if (skip_seccomp_unavailable(u, "ProtectKernelLogs="))
1639 return seccomp_protect_syslog();
1642 static int apply_private_devices(const Unit *u, const ExecContext *c) {
1646 /* If PrivateDevices= is set, also turn off iopl and all @raw-io syscalls. */
1648 if (!c->private_devices)
1651 if (skip_seccomp_unavailable(u, "PrivateDevices="))
1654 return seccomp_load_syscall_filter_set(SCMP_ACT_ALLOW, syscall_filter_sets + SYSCALL_FILTER_SET_RAW_IO, SCMP_ACT_ERRNO(EPERM), false);
1657 static int apply_restrict_namespaces(const Unit *u, const ExecContext *c) {
1661 if (!exec_context_restrict_namespaces_set(c))
1664 if (skip_seccomp_unavailable(u, "RestrictNamespaces="))
1667 return seccomp_restrict_namespaces(c->restrict_namespaces);
1670 static int apply_lock_personality(const Unit* u, const ExecContext *c) {
1671 unsigned long personality;
1677 if (!c->lock_personality)
1680 if (skip_seccomp_unavailable(u, "LockPersonality="))
1683 personality = c->personality;
1685 /* If personality is not specified, use either PER_LINUX or PER_LINUX32 depending on what is currently set. */
1686 if (personality == PERSONALITY_INVALID) {
1688 r = opinionated_personality(&personality);
1693 return seccomp_lock_personality(personality);
1698 static void do_idle_pipe_dance(int idle_pipe[static 4]) {
1701 idle_pipe[1] = safe_close(idle_pipe[1]);
1702 idle_pipe[2] = safe_close(idle_pipe[2]);
1704 if (idle_pipe[0] >= 0) {
1707 r = fd_wait_for_event(idle_pipe[0], POLLHUP, IDLE_TIMEOUT_USEC);
1709 if (idle_pipe[3] >= 0 && r == 0 /* timeout */) {
1712 /* Signal systemd that we are bored and want to continue. */
1713 n = write(idle_pipe[3], "x", 1);
1715 /* Wait for systemd to react to the signal above. */
1716 (void) fd_wait_for_event(idle_pipe[0], POLLHUP, IDLE_TIMEOUT2_USEC);
1719 idle_pipe[0] = safe_close(idle_pipe[0]);
1723 idle_pipe[3] = safe_close(idle_pipe[3]);
1726 static const char *exec_directory_env_name_to_string(ExecDirectoryType t);
1728 static int build_environment(
1730 const ExecContext *c,
1731 const ExecParameters *p,
1734 const char *username,
1736 dev_t journal_stream_dev,
1737 ino_t journal_stream_ino,
1740 _cleanup_strv_free_ char **our_env = NULL;
1741 ExecDirectoryType t;
1750 our_env = new0(char*, 14 + _EXEC_DIRECTORY_TYPE_MAX);
1755 _cleanup_free_ char *joined = NULL;
1757 if (asprintf(&x, "LISTEN_PID="PID_FMT, getpid_cached()) < 0)
1759 our_env[n_env++] = x;
1761 if (asprintf(&x, "LISTEN_FDS=%zu", n_fds) < 0)
1763 our_env[n_env++] = x;
1765 joined = strv_join(p->fd_names, ":");
1769 x = strjoin("LISTEN_FDNAMES=", joined);
1772 our_env[n_env++] = x;
1775 if ((p->flags & EXEC_SET_WATCHDOG) && p->watchdog_usec > 0) {
1776 if (asprintf(&x, "WATCHDOG_PID="PID_FMT, getpid_cached()) < 0)
1778 our_env[n_env++] = x;
1780 if (asprintf(&x, "WATCHDOG_USEC="USEC_FMT, p->watchdog_usec) < 0)
1782 our_env[n_env++] = x;
1785 /* If this is D-Bus, tell the nss-systemd module, since it relies on being able to use D-Bus look up dynamic
1786 * users via PID 1, possibly dead-locking the dbus daemon. This way it will not use D-Bus to resolve names, but
1787 * check the database directly. */
1788 if (p->flags & EXEC_NSS_BYPASS_BUS) {
1789 x = strdup("SYSTEMD_NSS_BYPASS_BUS=1");
1792 our_env[n_env++] = x;
1796 x = strjoin("HOME=", home);
1800 path_simplify(x + 5, true);
1801 our_env[n_env++] = x;
1805 x = strjoin("LOGNAME=", username);
1808 our_env[n_env++] = x;
1810 x = strjoin("USER=", username);
1813 our_env[n_env++] = x;
1817 x = strjoin("SHELL=", shell);
1821 path_simplify(x + 6, true);
1822 our_env[n_env++] = x;
1825 if (!sd_id128_is_null(u->invocation_id)) {
1826 if (asprintf(&x, "INVOCATION_ID=" SD_ID128_FORMAT_STR, SD_ID128_FORMAT_VAL(u->invocation_id)) < 0)
1829 our_env[n_env++] = x;
1832 if (exec_context_needs_term(c)) {
1833 const char *tty_path, *term = NULL;
1835 tty_path = exec_context_tty_path(c);
1837 /* If we are forked off PID 1 and we are supposed to operate on /dev/console, then let's try to inherit
1838 * the $TERM set for PID 1. This is useful for containers so that the $TERM the container manager
1839 * passes to PID 1 ends up all the way in the console login shown. */
1841 if (path_equal(tty_path, "/dev/console") && getppid() == 1)
1842 term = getenv("TERM");
1844 term = default_term_for_tty(tty_path);
1846 x = strjoin("TERM=", term);
1849 our_env[n_env++] = x;
1852 if (journal_stream_dev != 0 && journal_stream_ino != 0) {
1853 if (asprintf(&x, "JOURNAL_STREAM=" DEV_FMT ":" INO_FMT, journal_stream_dev, journal_stream_ino) < 0)
1856 our_env[n_env++] = x;
1859 for (t = 0; t < _EXEC_DIRECTORY_TYPE_MAX; t++) {
1860 _cleanup_free_ char *pre = NULL, *joined = NULL;
1866 if (strv_isempty(c->directories[t].paths))
1869 n = exec_directory_env_name_to_string(t);
1873 pre = strjoin(p->prefix[t], "/");
1877 joined = strv_join_prefix(c->directories[t].paths, ":", pre);
1881 x = strjoin(n, "=", joined);
1885 our_env[n_env++] = x;
1888 our_env[n_env++] = NULL;
1889 assert(n_env <= 14 + _EXEC_DIRECTORY_TYPE_MAX);
1891 *ret = TAKE_PTR(our_env);
1896 static int build_pass_environment(const ExecContext *c, char ***ret) {
1897 _cleanup_strv_free_ char **pass_env = NULL;
1898 size_t n_env = 0, n_bufsize = 0;
1901 STRV_FOREACH(i, c->pass_environment) {
1902 _cleanup_free_ char *x = NULL;
1908 x = strjoin(*i, "=", v);
1912 if (!GREEDY_REALLOC(pass_env, n_bufsize, n_env + 2))
1915 pass_env[n_env++] = TAKE_PTR(x);
1916 pass_env[n_env] = NULL;
1919 *ret = TAKE_PTR(pass_env);
1924 static bool exec_needs_mount_namespace(
1925 const ExecContext *context,
1926 const ExecParameters *params,
1927 const ExecRuntime *runtime) {
1932 if (context->root_image)
1935 if (!strv_isempty(context->read_write_paths) ||
1936 !strv_isempty(context->read_only_paths) ||
1937 !strv_isempty(context->inaccessible_paths))
1940 if (context->n_bind_mounts > 0)
1943 if (context->n_temporary_filesystems > 0)
1946 if (!IN_SET(context->mount_flags, 0, MS_SHARED))
1949 if (context->private_tmp && runtime && (runtime->tmp_dir || runtime->var_tmp_dir))
1952 if (context->private_devices ||
1953 context->private_mounts ||
1954 context->protect_system != PROTECT_SYSTEM_NO ||
1955 context->protect_home != PROTECT_HOME_NO ||
1956 context->protect_kernel_tunables ||
1957 context->protect_kernel_modules ||
1958 context->protect_kernel_logs ||
1959 context->protect_control_groups)
1962 if (context->root_directory) {
1963 ExecDirectoryType t;
1965 if (context->mount_apivfs)
1968 for (t = 0; t < _EXEC_DIRECTORY_TYPE_MAX; t++) {
1969 if (!params->prefix[t])
1972 if (!strv_isempty(context->directories[t].paths))
1977 if (context->dynamic_user &&
1978 (!strv_isempty(context->directories[EXEC_DIRECTORY_STATE].paths) ||
1979 !strv_isempty(context->directories[EXEC_DIRECTORY_CACHE].paths) ||
1980 !strv_isempty(context->directories[EXEC_DIRECTORY_LOGS].paths)))
1986 static int setup_private_users(uid_t uid, gid_t gid) {
1987 _cleanup_free_ char *uid_map = NULL, *gid_map = NULL;
1988 _cleanup_close_pair_ int errno_pipe[2] = { -1, -1 };
1989 _cleanup_close_ int unshare_ready_fd = -1;
1990 _cleanup_(sigkill_waitp) pid_t pid = 0;
1995 /* Set up a user namespace and map root to root, the selected UID/GID to itself, and everything else to
1996 * nobody. In order to be able to write this mapping we need CAP_SETUID in the original user namespace, which
1997 * we however lack after opening the user namespace. To work around this we fork() a temporary child process,
1998 * which waits for the parent to create the new user namespace while staying in the original namespace. The
1999 * child then writes the UID mapping, under full privileges. The parent waits for the child to finish and
2000 * continues execution normally. */
2002 if (uid != 0 && uid_is_valid(uid)) {
2003 r = asprintf(&uid_map,
2004 "0 0 1\n" /* Map root → root */
2005 UID_FMT " " UID_FMT " 1\n", /* Map $UID → $UID */
2010 uid_map = strdup("0 0 1\n"); /* The case where the above is the same */
2015 if (gid != 0 && gid_is_valid(gid)) {
2016 r = asprintf(&gid_map,
2017 "0 0 1\n" /* Map root → root */
2018 GID_FMT " " GID_FMT " 1\n", /* Map $GID → $GID */
2023 gid_map = strdup("0 0 1\n"); /* The case where the above is the same */
2028 /* Create a communication channel so that the parent can tell the child when it finished creating the user
2030 unshare_ready_fd = eventfd(0, EFD_CLOEXEC);
2031 if (unshare_ready_fd < 0)
2034 /* Create a communication channel so that the child can tell the parent a proper error code in case it
2036 if (pipe2(errno_pipe, O_CLOEXEC) < 0)
2039 r = safe_fork("(sd-userns)", FORK_RESET_SIGNALS|FORK_DEATHSIG, &pid);
2043 _cleanup_close_ int fd = -1;
2047 /* Child process, running in the original user namespace. Let's update the parent's UID/GID map from
2048 * here, after the parent opened its own user namespace. */
2051 errno_pipe[0] = safe_close(errno_pipe[0]);
2053 /* Wait until the parent unshared the user namespace */
2054 if (read(unshare_ready_fd, &c, sizeof(c)) < 0) {
2059 /* Disable the setgroups() system call in the child user namespace, for good. */
2060 a = procfs_file_alloca(ppid, "setgroups");
2061 fd = open(a, O_WRONLY|O_CLOEXEC);
2063 if (errno != ENOENT) {
2068 /* If the file is missing the kernel is too old, let's continue anyway. */
2070 if (write(fd, "deny\n", 5) < 0) {
2075 fd = safe_close(fd);
2078 /* First write the GID map */
2079 a = procfs_file_alloca(ppid, "gid_map");
2080 fd = open(a, O_WRONLY|O_CLOEXEC);
2085 if (write(fd, gid_map, strlen(gid_map)) < 0) {
2089 fd = safe_close(fd);
2091 /* The write the UID map */
2092 a = procfs_file_alloca(ppid, "uid_map");
2093 fd = open(a, O_WRONLY|O_CLOEXEC);
2098 if (write(fd, uid_map, strlen(uid_map)) < 0) {
2103 _exit(EXIT_SUCCESS);
2106 (void) write(errno_pipe[1], &r, sizeof(r));
2107 _exit(EXIT_FAILURE);
2110 errno_pipe[1] = safe_close(errno_pipe[1]);
2112 if (unshare(CLONE_NEWUSER) < 0)
2115 /* Let the child know that the namespace is ready now */
2116 if (write(unshare_ready_fd, &c, sizeof(c)) < 0)
2119 /* Try to read an error code from the child */
2120 n = read(errno_pipe[0], &r, sizeof(r));
2123 if (n == sizeof(r)) { /* an error code was sent to us */
2128 if (n != 0) /* on success we should have read 0 bytes */
2131 r = wait_for_terminate_and_check("(sd-userns)", pid, 0);
2135 if (r != EXIT_SUCCESS) /* If something strange happened with the child, let's consider this fatal, too */
2141 static bool exec_directory_is_private(const ExecContext *context, ExecDirectoryType type) {
2142 if (!context->dynamic_user)
2145 if (type == EXEC_DIRECTORY_CONFIGURATION)
2148 if (type == EXEC_DIRECTORY_RUNTIME && context->runtime_directory_preserve_mode == EXEC_PRESERVE_NO)
2154 static int setup_exec_directory(
2155 const ExecContext *context,
2156 const ExecParameters *params,
2159 ExecDirectoryType type,
2162 static const int exit_status_table[_EXEC_DIRECTORY_TYPE_MAX] = {
2163 [EXEC_DIRECTORY_RUNTIME] = EXIT_RUNTIME_DIRECTORY,
2164 [EXEC_DIRECTORY_STATE] = EXIT_STATE_DIRECTORY,
2165 [EXEC_DIRECTORY_CACHE] = EXIT_CACHE_DIRECTORY,
2166 [EXEC_DIRECTORY_LOGS] = EXIT_LOGS_DIRECTORY,
2167 [EXEC_DIRECTORY_CONFIGURATION] = EXIT_CONFIGURATION_DIRECTORY,
2174 assert(type >= 0 && type < _EXEC_DIRECTORY_TYPE_MAX);
2175 assert(exit_status);
2177 if (!params->prefix[type])
2180 if (params->flags & EXEC_CHOWN_DIRECTORIES) {
2181 if (!uid_is_valid(uid))
2183 if (!gid_is_valid(gid))
2187 STRV_FOREACH(rt, context->directories[type].paths) {
2188 _cleanup_free_ char *p = NULL, *pp = NULL;
2190 p = path_join(params->prefix[type], *rt);
2196 r = mkdir_parents_label(p, 0755);
2200 if (exec_directory_is_private(context, type)) {
2201 _cleanup_free_ char *private_root = NULL;
2203 /* So, here's one extra complication when dealing with DynamicUser=1 units. In that
2204 * case we want to avoid leaving a directory around fully accessible that is owned by
2205 * a dynamic user whose UID is later on reused. To lock this down we use the same
2206 * trick used by container managers to prohibit host users to get access to files of
2207 * the same UID in containers: we place everything inside a directory that has an
2208 * access mode of 0700 and is owned root:root, so that it acts as security boundary
2209 * for unprivileged host code. We then use fs namespacing to make this directory
2210 * permeable for the service itself.
2212 * Specifically: for a service which wants a special directory "foo/" we first create
2213 * a directory "private/" with access mode 0700 owned by root:root. Then we place
2214 * "foo" inside of that directory (i.e. "private/foo/"), and make "foo" a symlink to
2215 * "private/foo". This way, privileged host users can access "foo/" as usual, but
2216 * unprivileged host users can't look into it. Inside of the namespace of the unit
2217 * "private/" is replaced by a more liberally accessible tmpfs, into which the host's
2218 * "private/foo/" is mounted under the same name, thus disabling the access boundary
2219 * for the service and making sure it only gets access to the dirs it needs but no
2220 * others. Tricky? Yes, absolutely, but it works!
2222 * Note that we don't do this for EXEC_DIRECTORY_CONFIGURATION as that's assumed not
2223 * to be owned by the service itself.
2225 * Also, note that we don't do this for EXEC_DIRECTORY_RUNTIME as that's often used
2226 * for sharing files or sockets with other services. */
2228 private_root = path_join(params->prefix[type], "private");
2229 if (!private_root) {
2234 /* First set up private root if it doesn't exist yet, with access mode 0700 and owned by root:root */
2235 r = mkdir_safe_label(private_root, 0700, 0, 0, MKDIR_WARN_MODE);
2239 pp = path_join(private_root, *rt);
2245 /* Create all directories between the configured directory and this private root, and mark them 0755 */
2246 r = mkdir_parents_label(pp, 0755);
2250 if (is_dir(p, false) > 0 &&
2251 (laccess(pp, F_OK) < 0 && errno == ENOENT)) {
2253 /* Hmm, the private directory doesn't exist yet, but the normal one exists? If so, move
2254 * it over. Most likely the service has been upgraded from one that didn't use
2255 * DynamicUser=1, to one that does. */
2257 log_info("Found pre-existing public %s= directory %s, migrating to %s.\n"
2258 "Apparently, service previously had DynamicUser= turned off, and has now turned it on.",
2259 exec_directory_type_to_string(type), p, pp);
2261 if (rename(p, pp) < 0) {
2266 /* Otherwise, create the actual directory for the service */
2268 r = mkdir_label(pp, context->directories[type].mode);
2269 if (r < 0 && r != -EEXIST)
2273 /* And link it up from the original place */
2274 r = symlink_idempotent(pp, p, true);
2279 _cleanup_free_ char *target = NULL;
2281 if (type != EXEC_DIRECTORY_CONFIGURATION &&
2282 readlink_and_make_absolute(p, &target) >= 0) {
2283 _cleanup_free_ char *q = NULL;
2285 /* This already exists and is a symlink? Interesting. Maybe it's one created
2286 * by DynamicUser=1 (see above)?
2288 * We do this for all directory types except for ConfigurationDirectory=,
2289 * since they all support the private/ symlink logic at least in some
2290 * configurations, see above. */
2292 q = path_join(params->prefix[type], "private", *rt);
2298 if (path_equal(q, target)) {
2300 /* Hmm, apparently DynamicUser= was once turned on for this service,
2301 * but is no longer. Let's move the directory back up. */
2303 log_info("Found pre-existing private %s= directory %s, migrating to %s.\n"
2304 "Apparently, service previously had DynamicUser= turned on, and has now turned it off.",
2305 exec_directory_type_to_string(type), q, p);
2307 if (unlink(p) < 0) {
2312 if (rename(q, p) < 0) {
2319 r = mkdir_label(p, context->directories[type].mode);
2324 if (type == EXEC_DIRECTORY_CONFIGURATION) {
2327 /* Don't change the owner/access mode of the configuration directory,
2328 * as in the common case it is not written to by a service, and shall
2329 * not be writable. */
2331 if (stat(p, &st) < 0) {
2336 /* Still complain if the access mode doesn't match */
2337 if (((st.st_mode ^ context->directories[type].mode) & 07777) != 0)
2338 log_warning("%s \'%s\' already exists but the mode is different. "
2339 "(File system: %o %sMode: %o)",
2340 exec_directory_type_to_string(type), *rt,
2341 st.st_mode & 07777, exec_directory_type_to_string(type), context->directories[type].mode & 07777);
2348 /* Lock down the access mode (we use chmod_and_chown() to make this idempotent. We don't
2349 * specify UID/GID here, so that path_chown_recursive() can optimize things depending on the
2350 * current UID/GID ownership.) */
2351 r = chmod_and_chown(pp ?: p, context->directories[type].mode, UID_INVALID, GID_INVALID);
2355 /* Then, change the ownership of the whole tree, if necessary. When dynamic users are used we
2356 * drop the suid/sgid bits, since we really don't want SUID/SGID files for dynamic UID/GID
2357 * assignments to exist.*/
2358 r = path_chown_recursive(pp ?: p, uid, gid, context->dynamic_user ? 01777 : 07777);
2366 *exit_status = exit_status_table[type];
2371 static int setup_smack(
2372 const Manager *manager,
2373 const ExecContext *context,
2374 const ExecCommand *command) {
2381 if (context->smack_process_label) {
2382 r = mac_smack_apply_pid(0, context->smack_process_label);
2385 } else if (manager->default_smack_process_label) {
2386 _cleanup_free_ char *exec_label = NULL;
2388 r = mac_smack_read(command->path, SMACK_ATTR_EXEC, &exec_label);
2389 if (r < 0 && !IN_SET(r, -ENODATA, -EOPNOTSUPP))
2392 r = mac_smack_apply_pid(0, exec_label ? : manager->default_smack_process_label);
2401 static int compile_bind_mounts(
2402 const ExecContext *context,
2403 const ExecParameters *params,
2404 BindMount **ret_bind_mounts,
2405 size_t *ret_n_bind_mounts,
2406 char ***ret_empty_directories) {
2408 _cleanup_strv_free_ char **empty_directories = NULL;
2409 BindMount *bind_mounts;
2411 ExecDirectoryType t;
2416 assert(ret_bind_mounts);
2417 assert(ret_n_bind_mounts);
2418 assert(ret_empty_directories);
2420 n = context->n_bind_mounts;
2421 for (t = 0; t < _EXEC_DIRECTORY_TYPE_MAX; t++) {
2422 if (!params->prefix[t])
2425 n += strv_length(context->directories[t].paths);
2429 *ret_bind_mounts = NULL;
2430 *ret_n_bind_mounts = 0;
2431 *ret_empty_directories = NULL;
2435 bind_mounts = new(BindMount, n);
2439 for (i = 0; i < context->n_bind_mounts; i++) {
2440 BindMount *item = context->bind_mounts + i;
2443 s = strdup(item->source);
2449 d = strdup(item->destination);
2456 bind_mounts[h++] = (BindMount) {
2459 .read_only = item->read_only,
2460 .recursive = item->recursive,
2461 .ignore_enoent = item->ignore_enoent,
2465 for (t = 0; t < _EXEC_DIRECTORY_TYPE_MAX; t++) {
2468 if (!params->prefix[t])
2471 if (strv_isempty(context->directories[t].paths))
2474 if (exec_directory_is_private(context, t) &&
2475 !(context->root_directory || context->root_image)) {
2478 /* So this is for a dynamic user, and we need to make sure the process can access its own
2479 * directory. For that we overmount the usually inaccessible "private" subdirectory with a
2480 * tmpfs that makes it accessible and is empty except for the submounts we do this for. */
2482 private_root = path_join(params->prefix[t], "private");
2483 if (!private_root) {
2488 r = strv_consume(&empty_directories, private_root);
2493 STRV_FOREACH(suffix, context->directories[t].paths) {
2496 if (exec_directory_is_private(context, t))
2497 s = path_join(params->prefix[t], "private", *suffix);
2499 s = path_join(params->prefix[t], *suffix);
2505 if (exec_directory_is_private(context, t) &&
2506 (context->root_directory || context->root_image))
2507 /* When RootDirectory= or RootImage= are set, then the symbolic link to the private
2508 * directory is not created on the root directory. So, let's bind-mount the directory
2509 * on the 'non-private' place. */
2510 d = path_join(params->prefix[t], *suffix);
2519 bind_mounts[h++] = (BindMount) {
2523 .nosuid = context->dynamic_user, /* don't allow suid/sgid when DynamicUser= is on */
2525 .ignore_enoent = false,
2532 *ret_bind_mounts = bind_mounts;
2533 *ret_n_bind_mounts = n;
2534 *ret_empty_directories = TAKE_PTR(empty_directories);
2539 bind_mount_free_many(bind_mounts, h);
2543 static bool insist_on_sandboxing(
2544 const ExecContext *context,
2545 const char *root_dir,
2546 const char *root_image,
2547 const BindMount *bind_mounts,
2548 size_t n_bind_mounts) {
2553 assert(n_bind_mounts == 0 || bind_mounts);
2555 /* Checks whether we need to insist on fs namespacing. i.e. whether we have settings configured that
2556 * would alter the view on the file system beyond making things read-only or invisble, i.e. would
2557 * rearrange stuff in a way we cannot ignore gracefully. */
2559 if (context->n_temporary_filesystems > 0)
2562 if (root_dir || root_image)
2565 if (context->dynamic_user)
2568 /* If there are any bind mounts set that don't map back onto themselves, fs namespacing becomes
2570 for (i = 0; i < n_bind_mounts; i++)
2571 if (!path_equal(bind_mounts[i].source, bind_mounts[i].destination))
2577 static int apply_mount_namespace(
2579 const ExecCommand *command,
2580 const ExecContext *context,
2581 const ExecParameters *params,
2582 const ExecRuntime *runtime,
2583 char **error_path) {
2585 _cleanup_strv_free_ char **empty_directories = NULL;
2586 char *tmp = NULL, *var = NULL;
2587 const char *root_dir = NULL, *root_image = NULL;
2588 NamespaceInfo ns_info;
2589 bool needs_sandboxing;
2590 BindMount *bind_mounts = NULL;
2591 size_t n_bind_mounts = 0;
2596 /* The runtime struct only contains the parent of the private /tmp,
2597 * which is non-accessible to world users. Inside of it there's a /tmp
2598 * that is sticky, and that's the one we want to use here. */
2600 if (context->private_tmp && runtime) {
2601 if (runtime->tmp_dir)
2602 tmp = strjoina(runtime->tmp_dir, "/tmp");
2603 if (runtime->var_tmp_dir)
2604 var = strjoina(runtime->var_tmp_dir, "/tmp");
2607 if (params->flags & EXEC_APPLY_CHROOT) {
2608 root_image = context->root_image;
2611 root_dir = context->root_directory;
2614 r = compile_bind_mounts(context, params, &bind_mounts, &n_bind_mounts, &empty_directories);
2618 needs_sandboxing = (params->flags & EXEC_APPLY_SANDBOXING) && !(command->flags & EXEC_COMMAND_FULLY_PRIVILEGED);
2619 if (needs_sandboxing)
2620 ns_info = (NamespaceInfo) {
2621 .ignore_protect_paths = false,
2622 .private_dev = context->private_devices,
2623 .protect_control_groups = context->protect_control_groups,
2624 .protect_kernel_tunables = context->protect_kernel_tunables,
2625 .protect_kernel_modules = context->protect_kernel_modules,
2626 .protect_kernel_logs = context->protect_kernel_logs,
2627 .protect_hostname = context->protect_hostname,
2628 .mount_apivfs = context->mount_apivfs,
2629 .private_mounts = context->private_mounts,
2631 else if (!context->dynamic_user && root_dir)
2633 * If DynamicUser=no and RootDirectory= is set then lets pass a relaxed
2634 * sandbox info, otherwise enforce it, don't ignore protected paths and
2635 * fail if we are enable to apply the sandbox inside the mount namespace.
2637 ns_info = (NamespaceInfo) {
2638 .ignore_protect_paths = true,
2641 ns_info = (NamespaceInfo) {};
2643 if (context->mount_flags == MS_SHARED)
2644 log_unit_debug(u, "shared mount propagation hidden by other fs namespacing unit settings: ignoring");
2646 r = setup_namespace(root_dir, root_image,
2647 &ns_info, context->read_write_paths,
2648 needs_sandboxing ? context->read_only_paths : NULL,
2649 needs_sandboxing ? context->inaccessible_paths : NULL,
2653 context->temporary_filesystems,
2654 context->n_temporary_filesystems,
2657 needs_sandboxing ? context->protect_home : PROTECT_HOME_NO,
2658 needs_sandboxing ? context->protect_system : PROTECT_SYSTEM_NO,
2659 context->mount_flags,
2660 DISSECT_IMAGE_DISCARD_ON_LOOP,
2663 /* If we couldn't set up the namespace this is probably due to a missing capability. setup_namespace() reports
2664 * that with a special, recognizable error ENOANO. In this case, silently proceed, but only if exclusively
2665 * sandboxing options were used, i.e. nothing such as RootDirectory= or BindMount= that would result in a
2666 * completely different execution environment. */
2668 if (insist_on_sandboxing(
2670 root_dir, root_image,
2673 log_unit_debug(u, "Failed to set up namespace, and refusing to continue since the selected namespacing options alter mount environment non-trivially.\n"
2674 "Bind mounts: %zu, temporary filesystems: %zu, root directory: %s, root image: %s, dynamic user: %s",
2675 n_bind_mounts, context->n_temporary_filesystems, yes_no(root_dir), yes_no(root_image), yes_no(context->dynamic_user));
2679 log_unit_debug(u, "Failed to set up namespace, assuming containerized execution and ignoring.");
2684 bind_mount_free_many(bind_mounts, n_bind_mounts);
2688 static int apply_working_directory(
2689 const ExecContext *context,
2690 const ExecParameters *params,
2697 assert(exit_status);
2699 if (context->working_directory_home) {
2702 *exit_status = EXIT_CHDIR;
2708 } else if (context->working_directory)
2709 wd = context->working_directory;
2713 if (params->flags & EXEC_APPLY_CHROOT)
2716 d = prefix_roota(context->root_directory, wd);
2718 if (chdir(d) < 0 && !context->working_directory_missing_ok) {
2719 *exit_status = EXIT_CHDIR;
2726 static int apply_root_directory(
2727 const ExecContext *context,
2728 const ExecParameters *params,
2729 const bool needs_mount_ns,
2733 assert(exit_status);
2735 if (params->flags & EXEC_APPLY_CHROOT) {
2736 if (!needs_mount_ns && context->root_directory)
2737 if (chroot(context->root_directory) < 0) {
2738 *exit_status = EXIT_CHROOT;
2746 static int setup_keyring(
2748 const ExecContext *context,
2749 const ExecParameters *p,
2750 uid_t uid, gid_t gid) {
2752 key_serial_t keyring;
2761 /* Let's set up a new per-service "session" kernel keyring for each system service. This has the benefit that
2762 * each service runs with its own keyring shared among all processes of the service, but with no hook-up beyond
2763 * that scope, and in particular no link to the per-UID keyring. If we don't do this the keyring will be
2764 * automatically created on-demand and then linked to the per-UID keyring, by the kernel. The kernel's built-in
2765 * on-demand behaviour is very appropriate for login users, but probably not so much for system services, where
2766 * UIDs are not necessarily specific to a service but reused (at least in the case of UID 0). */
2768 if (context->keyring_mode == EXEC_KEYRING_INHERIT)
2771 /* Acquiring a reference to the user keyring is nasty. We briefly change identity in order to get things set up
2772 * properly by the kernel. If we don't do that then we can't create it atomically, and that sucks for parallel
2773 * execution. This mimics what pam_keyinit does, too. Setting up session keyring, to be owned by the right user
2774 * & group is just as nasty as acquiring a reference to the user keyring. */
2776 saved_uid = getuid();
2777 saved_gid = getgid();
2779 if (gid_is_valid(gid) && gid != saved_gid) {
2780 if (setregid(gid, -1) < 0)
2781 return log_unit_error_errno(u, errno, "Failed to change GID for user keyring: %m");
2784 if (uid_is_valid(uid) && uid != saved_uid) {
2785 if (setreuid(uid, -1) < 0) {
2786 r = log_unit_error_errno(u, errno, "Failed to change UID for user keyring: %m");
2791 keyring = keyctl(KEYCTL_JOIN_SESSION_KEYRING, 0, 0, 0, 0);
2792 if (keyring == -1) {
2793 if (errno == ENOSYS)
2794 log_unit_debug_errno(u, errno, "Kernel keyring not supported, ignoring.");
2795 else if (IN_SET(errno, EACCES, EPERM))
2796 log_unit_debug_errno(u, errno, "Kernel keyring access prohibited, ignoring.");
2797 else if (errno == EDQUOT)
2798 log_unit_debug_errno(u, errno, "Out of kernel keyrings to allocate, ignoring.");
2800 r = log_unit_error_errno(u, errno, "Setting up kernel keyring failed: %m");
2805 /* When requested link the user keyring into the session keyring. */
2806 if (context->keyring_mode == EXEC_KEYRING_SHARED) {
2808 if (keyctl(KEYCTL_LINK,
2809 KEY_SPEC_USER_KEYRING,
2810 KEY_SPEC_SESSION_KEYRING, 0, 0) < 0) {
2811 r = log_unit_error_errno(u, errno, "Failed to link user keyring into session keyring: %m");
2816 /* Restore uid/gid back */
2817 if (uid_is_valid(uid) && uid != saved_uid) {
2818 if (setreuid(saved_uid, -1) < 0) {
2819 r = log_unit_error_errno(u, errno, "Failed to change UID back for user keyring: %m");
2824 if (gid_is_valid(gid) && gid != saved_gid) {
2825 if (setregid(saved_gid, -1) < 0)
2826 return log_unit_error_errno(u, errno, "Failed to change GID back for user keyring: %m");
2829 /* Populate they keyring with the invocation ID by default, as original saved_uid. */
2830 if (!sd_id128_is_null(u->invocation_id)) {
2833 key = add_key("user", "invocation_id", &u->invocation_id, sizeof(u->invocation_id), KEY_SPEC_SESSION_KEYRING);
2835 log_unit_debug_errno(u, errno, "Failed to add invocation ID to keyring, ignoring: %m");
2837 if (keyctl(KEYCTL_SETPERM, key,
2838 KEY_POS_VIEW|KEY_POS_READ|KEY_POS_SEARCH|
2839 KEY_USR_VIEW|KEY_USR_READ|KEY_USR_SEARCH, 0, 0) < 0)
2840 r = log_unit_error_errno(u, errno, "Failed to restrict invocation ID permission: %m");
2845 /* Revert back uid & gid for the the last time, and exit */
2846 /* no extra logging, as only the first already reported error matters */
2847 if (getuid() != saved_uid)
2848 (void) setreuid(saved_uid, -1);
2850 if (getgid() != saved_gid)
2851 (void) setregid(saved_gid, -1);
2856 static void append_socket_pair(int *array, size_t *n, const int pair[static 2]) {
2862 array[(*n)++] = pair[0];
2864 array[(*n)++] = pair[1];
2867 static int close_remaining_fds(
2868 const ExecParameters *params,
2869 const ExecRuntime *runtime,
2870 const DynamicCreds *dcreds,
2874 int *fds, size_t n_fds) {
2876 size_t n_dont_close = 0;
2877 int dont_close[n_fds + 12];
2881 if (params->stdin_fd >= 0)
2882 dont_close[n_dont_close++] = params->stdin_fd;
2883 if (params->stdout_fd >= 0)
2884 dont_close[n_dont_close++] = params->stdout_fd;
2885 if (params->stderr_fd >= 0)
2886 dont_close[n_dont_close++] = params->stderr_fd;
2889 dont_close[n_dont_close++] = socket_fd;
2891 dont_close[n_dont_close++] = exec_fd;
2893 memcpy(dont_close + n_dont_close, fds, sizeof(int) * n_fds);
2894 n_dont_close += n_fds;
2898 append_socket_pair(dont_close, &n_dont_close, runtime->netns_storage_socket);
2902 append_socket_pair(dont_close, &n_dont_close, dcreds->user->storage_socket);
2904 append_socket_pair(dont_close, &n_dont_close, dcreds->group->storage_socket);
2907 if (user_lookup_fd >= 0)
2908 dont_close[n_dont_close++] = user_lookup_fd;
2910 return close_all_fds(dont_close, n_dont_close);
2913 static int send_user_lookup(
2921 /* Send the resolved UID/GID to PID 1 after we learnt it. We send a single datagram, containing the UID/GID
2922 * data as well as the unit name. Note that we suppress sending this if no user/group to resolve was
2925 if (user_lookup_fd < 0)
2928 if (!uid_is_valid(uid) && !gid_is_valid(gid))
2931 if (writev(user_lookup_fd,
2933 IOVEC_INIT(&uid, sizeof(uid)),
2934 IOVEC_INIT(&gid, sizeof(gid)),
2935 IOVEC_INIT_STRING(unit->id) }, 3) < 0)
2941 static int acquire_home(const ExecContext *c, uid_t uid, const char** home, char **buf) {
2948 /* If WorkingDirectory=~ is set, try to acquire a usable home directory. */
2953 if (!c->working_directory_home)
2956 r = get_home_dir(buf);
2964 static int compile_suggested_paths(const ExecContext *c, const ExecParameters *p, char ***ret) {
2965 _cleanup_strv_free_ char ** list = NULL;
2966 ExecDirectoryType t;
2973 assert(c->dynamic_user);
2975 /* Compile a list of paths that it might make sense to read the owning UID from to use as initial candidate for
2976 * dynamic UID allocation, in order to save us from doing costly recursive chown()s of the special
2979 for (t = 0; t < _EXEC_DIRECTORY_TYPE_MAX; t++) {
2982 if (t == EXEC_DIRECTORY_CONFIGURATION)
2988 STRV_FOREACH(i, c->directories[t].paths) {
2991 if (exec_directory_is_private(c, t))
2992 e = path_join(p->prefix[t], "private", *i);
2994 e = path_join(p->prefix[t], *i);
2998 r = strv_consume(&list, e);
3004 *ret = TAKE_PTR(list);
3009 static char *exec_command_line(char **argv);
3011 static int exec_parameters_get_cgroup_path(const ExecParameters *params, char **ret) {
3012 bool using_subcgroup;
3018 if (!params->cgroup_path)
3021 /* If we are called for a unit where cgroup delegation is on, and the payload created its own populated
3022 * subcgroup (which we expect it to do, after all it asked for delegation), then we cannot place the control
3023 * processes started after the main unit's process in the unit's main cgroup because it is now an inner one,
3024 * and inner cgroups may not contain processes. Hence, if delegation is on, and this is a control process,
3025 * let's use ".control" as subcgroup instead. Note that we do so only for ExecStartPost=, ExecReload=,
3026 * ExecStop=, ExecStopPost=, i.e. for the commands where the main process is already forked. For ExecStartPre=
3027 * this is not necessary, the cgroup is still empty. We distinguish these cases with the EXEC_CONTROL_CGROUP
3028 * flag, which is only passed for the former statements, not for the latter. */
3030 using_subcgroup = FLAGS_SET(params->flags, EXEC_CONTROL_CGROUP|EXEC_CGROUP_DELEGATE|EXEC_IS_CONTROL);
3031 if (using_subcgroup)
3032 p = path_join(params->cgroup_path, ".control");
3034 p = strdup(params->cgroup_path);
3039 return using_subcgroup;
3042 static int exec_child(
3044 const ExecCommand *command,
3045 const ExecContext *context,
3046 const ExecParameters *params,
3047 ExecRuntime *runtime,
3048 DynamicCreds *dcreds,
3050 const int named_iofds[static 3],
3052 size_t n_socket_fds,
3053 size_t n_storage_fds,
3058 _cleanup_strv_free_ char **our_env = NULL, **pass_env = NULL, **accum_env = NULL, **replaced_argv = NULL;
3059 int *fds_with_exec_fd, n_fds_with_exec_fd, r, ngids = 0, exec_fd = -1;
3060 _cleanup_free_ gid_t *supplementary_gids = NULL;
3061 const char *username = NULL, *groupname = NULL;
3062 _cleanup_free_ char *home_buffer = NULL;
3063 const char *home = NULL, *shell = NULL;
3064 char **final_argv = NULL;
3065 dev_t journal_stream_dev = 0;
3066 ino_t journal_stream_ino = 0;
3067 bool needs_sandboxing, /* Do we need to set up full sandboxing? (i.e. all namespacing, all MAC stuff, caps, yadda yadda */
3068 needs_setuid, /* Do we need to do the actual setresuid()/setresgid() calls? */
3069 needs_mount_namespace, /* Do we need to set up a mount namespace for this kernel? */
3070 needs_ambient_hack; /* Do we need to apply the ambient capabilities hack? */
3072 _cleanup_free_ char *mac_selinux_context_net = NULL;
3073 bool use_selinux = false;
3076 bool use_smack = false;
3079 bool use_apparmor = false;
3081 uid_t uid = UID_INVALID;
3082 gid_t gid = GID_INVALID;
3084 ExecDirectoryType dt;
3091 assert(exit_status);
3093 rename_process_from_path(command->path);
3095 /* We reset exactly these signals, since they are the
3096 * only ones we set to SIG_IGN in the main daemon. All
3097 * others we leave untouched because we set them to
3098 * SIG_DFL or a valid handler initially, both of which
3099 * will be demoted to SIG_DFL. */
3100 (void) default_signals(SIGNALS_CRASH_HANDLER,
3101 SIGNALS_IGNORE, -1);
3103 if (context->ignore_sigpipe)
3104 (void) ignore_signals(SIGPIPE, -1);
3106 r = reset_signal_mask();
3108 *exit_status = EXIT_SIGNAL_MASK;
3109 return log_unit_error_errno(unit, r, "Failed to set process signal mask: %m");
3112 if (params->idle_pipe)
3113 do_idle_pipe_dance(params->idle_pipe);
3115 /* Close fds we don't need very early to make sure we don't block init reexecution because it cannot bind its
3116 * sockets. Among the fds we close are the logging fds, and we want to keep them closed, so that we don't have
3117 * any fds open we don't really want open during the transition. In order to make logging work, we switch the
3118 * log subsystem into open_when_needed mode, so that it reopens the logs on every single log call. */
3121 log_set_open_when_needed(true);
3123 /* In case anything used libc syslog(), close this here, too */
3126 n_fds = n_socket_fds + n_storage_fds;
3127 r = close_remaining_fds(params, runtime, dcreds, user_lookup_fd, socket_fd, params->exec_fd, fds, n_fds);
3129 *exit_status = EXIT_FDS;
3130 return log_unit_error_errno(unit, r, "Failed to close unwanted file descriptors: %m");
3133 if (!context->same_pgrp)
3135 *exit_status = EXIT_SETSID;
3136 return log_unit_error_errno(unit, errno, "Failed to create new process session: %m");
3139 exec_context_tty_reset(context, params);
3141 if (unit_shall_confirm_spawn(unit)) {
3142 const char *vc = params->confirm_spawn;
3143 _cleanup_free_ char *cmdline = NULL;
3145 cmdline = exec_command_line(command->argv);
3147 *exit_status = EXIT_MEMORY;
3151 r = ask_for_confirmation(vc, unit, cmdline);
3152 if (r != CONFIRM_EXECUTE) {
3153 if (r == CONFIRM_PRETEND_SUCCESS) {
3154 *exit_status = EXIT_SUCCESS;
3157 *exit_status = EXIT_CONFIRM;
3158 log_unit_error(unit, "Execution cancelled by the user");
3163 /* We are about to invoke NSS and PAM modules. Let's tell them what we are doing here, maybe they care. This is
3164 * used by nss-resolve to disable itself when we are about to start systemd-resolved, to avoid deadlocks. Note
3165 * that these env vars do not survive the execve(), which means they really only apply to the PAM and NSS
3166 * invocations themselves. Also note that while we'll only invoke NSS modules involved in user management they
3167 * might internally call into other NSS modules that are involved in hostname resolution, we never know. */
3168 if (setenv("SYSTEMD_ACTIVATION_UNIT", unit->id, true) != 0 ||
3169 setenv("SYSTEMD_ACTIVATION_SCOPE", MANAGER_IS_SYSTEM(unit->manager) ? "system" : "user", true) != 0) {
3170 *exit_status = EXIT_MEMORY;
3171 return log_unit_error_errno(unit, errno, "Failed to update environment: %m");
3174 if (context->dynamic_user && dcreds) {
3175 _cleanup_strv_free_ char **suggested_paths = NULL;
3177 /* On top of that, make sure we bypass our own NSS module nss-systemd comprehensively for any NSS
3178 * checks, if DynamicUser=1 is used, as we shouldn't create a feedback loop with ourselves here.*/
3179 if (putenv((char*) "SYSTEMD_NSS_DYNAMIC_BYPASS=1") != 0) {
3180 *exit_status = EXIT_USER;
3181 return log_unit_error_errno(unit, errno, "Failed to update environment: %m");
3184 r = compile_suggested_paths(context, params, &suggested_paths);
3186 *exit_status = EXIT_MEMORY;
3190 r = dynamic_creds_realize(dcreds, suggested_paths, &uid, &gid);
3192 *exit_status = EXIT_USER;
3194 log_unit_error(unit, "Failed to update dynamic user credentials: User or group with specified name already exists.");
3197 return log_unit_error_errno(unit, r, "Failed to update dynamic user credentials: %m");
3200 if (!uid_is_valid(uid)) {
3201 *exit_status = EXIT_USER;
3202 log_unit_error(unit, "UID validation failed for \""UID_FMT"\"", uid);
3206 if (!gid_is_valid(gid)) {
3207 *exit_status = EXIT_USER;
3208 log_unit_error(unit, "GID validation failed for \""GID_FMT"\"", gid);
3213 username = dcreds->user->name;
3216 r = get_fixed_user(context, &username, &uid, &gid, &home, &shell);
3218 *exit_status = EXIT_USER;
3219 return log_unit_error_errno(unit, r, "Failed to determine user credentials: %m");
3222 r = get_fixed_group(context, &groupname, &gid);
3224 *exit_status = EXIT_GROUP;
3225 return log_unit_error_errno(unit, r, "Failed to determine group credentials: %m");
3229 /* Initialize user supplementary groups and get SupplementaryGroups= ones */
3230 r = get_supplementary_groups(context, username, groupname, gid,
3231 &supplementary_gids, &ngids);
3233 *exit_status = EXIT_GROUP;
3234 return log_unit_error_errno(unit, r, "Failed to determine supplementary groups: %m");
3237 r = send_user_lookup(unit, user_lookup_fd, uid, gid);
3239 *exit_status = EXIT_USER;
3240 return log_unit_error_errno(unit, r, "Failed to send user credentials to PID1: %m");
3243 user_lookup_fd = safe_close(user_lookup_fd);
3245 r = acquire_home(context, uid, &home, &home_buffer);
3247 *exit_status = EXIT_CHDIR;
3248 return log_unit_error_errno(unit, r, "Failed to determine $HOME for user: %m");
3251 /* If a socket is connected to STDIN/STDOUT/STDERR, we
3252 * must sure to drop O_NONBLOCK */
3254 (void) fd_nonblock(socket_fd, false);
3256 /* Journald will try to look-up our cgroup in order to populate _SYSTEMD_CGROUP and _SYSTEMD_UNIT fields.
3257 * Hence we need to migrate to the target cgroup from init.scope before connecting to journald */
3258 if (params->cgroup_path) {
3259 _cleanup_free_ char *p = NULL;
3261 r = exec_parameters_get_cgroup_path(params, &p);
3263 *exit_status = EXIT_CGROUP;
3264 return log_unit_error_errno(unit, r, "Failed to acquire cgroup path: %m");
3267 r = cg_attach_everywhere(params->cgroup_supported, p, 0, NULL, NULL);
3269 *exit_status = EXIT_CGROUP;
3270 return log_unit_error_errno(unit, r, "Failed to attach to cgroup %s: %m", p);
3274 if (context->network_namespace_path && runtime && runtime->netns_storage_socket[0] >= 0) {
3275 r = open_netns_path(runtime->netns_storage_socket, context->network_namespace_path);
3277 *exit_status = EXIT_NETWORK;
3278 return log_unit_error_errno(unit, r, "Failed to open network namespace path %s: %m", context->network_namespace_path);
3282 r = setup_input(context, params, socket_fd, named_iofds);
3284 *exit_status = EXIT_STDIN;
3285 return log_unit_error_errno(unit, r, "Failed to set up standard input: %m");
3288 r = setup_output(unit, context, params, STDOUT_FILENO, socket_fd, named_iofds, basename(command->path), uid, gid, &journal_stream_dev, &journal_stream_ino);
3290 *exit_status = EXIT_STDOUT;
3291 return log_unit_error_errno(unit, r, "Failed to set up standard output: %m");
3294 r = setup_output(unit, context, params, STDERR_FILENO, socket_fd, named_iofds, basename(command->path), uid, gid, &journal_stream_dev, &journal_stream_ino);
3296 *exit_status = EXIT_STDERR;
3297 return log_unit_error_errno(unit, r, "Failed to set up standard error output: %m");
3300 if (context->oom_score_adjust_set) {
3301 /* When we can't make this change due to EPERM, then let's silently skip over it. User namespaces
3302 * prohibit write access to this file, and we shouldn't trip up over that. */
3303 r = set_oom_score_adjust(context->oom_score_adjust);
3304 if (IN_SET(r, -EPERM, -EACCES))
3305 log_unit_debug_errno(unit, r, "Failed to adjust OOM setting, assuming containerized execution, ignoring: %m");
3307 *exit_status = EXIT_OOM_ADJUST;
3308 return log_unit_error_errno(unit, r, "Failed to adjust OOM setting: %m");
3312 if (context->nice_set)
3313 if (setpriority(PRIO_PROCESS, 0, context->nice) < 0) {
3314 *exit_status = EXIT_NICE;
3315 return log_unit_error_errno(unit, errno, "Failed to set up process scheduling priority (nice level): %m");
3318 if (context->cpu_sched_set) {
3319 struct sched_param param = {
3320 .sched_priority = context->cpu_sched_priority,
3323 r = sched_setscheduler(0,
3324 context->cpu_sched_policy |
3325 (context->cpu_sched_reset_on_fork ?
3326 SCHED_RESET_ON_FORK : 0),
3329 *exit_status = EXIT_SETSCHEDULER;
3330 return log_unit_error_errno(unit, errno, "Failed to set up CPU scheduling: %m");
3334 if (context->cpu_set.set)
3335 if (sched_setaffinity(0, context->cpu_set.allocated, context->cpu_set.set) < 0) {
3336 *exit_status = EXIT_CPUAFFINITY;
3337 return log_unit_error_errno(unit, errno, "Failed to set up CPU affinity: %m");
3340 if (mpol_is_valid(numa_policy_get_type(&context->numa_policy))) {
3341 r = apply_numa_policy(&context->numa_policy);
3342 if (r == -EOPNOTSUPP)
3343 log_unit_debug_errno(unit, r, "NUMA support not available, ignoring.");
3345 *exit_status = EXIT_NUMA_POLICY;
3346 return log_unit_error_errno(unit, r, "Failed to set NUMA memory policy: %m");
3350 if (context->ioprio_set)
3351 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, context->ioprio) < 0) {
3352 *exit_status = EXIT_IOPRIO;
3353 return log_unit_error_errno(unit, errno, "Failed to set up IO scheduling priority: %m");
3356 if (context->timer_slack_nsec != NSEC_INFINITY)
3357 if (prctl(PR_SET_TIMERSLACK, context->timer_slack_nsec) < 0) {
3358 *exit_status = EXIT_TIMERSLACK;
3359 return log_unit_error_errno(unit, errno, "Failed to set up timer slack: %m");
3362 if (context->personality != PERSONALITY_INVALID) {
3363 r = safe_personality(context->personality);
3365 *exit_status = EXIT_PERSONALITY;
3366 return log_unit_error_errno(unit, r, "Failed to set up execution domain (personality): %m");
3370 if (context->utmp_id)
3371 utmp_put_init_process(context->utmp_id, getpid_cached(), getsid(0),
3373 context->utmp_mode == EXEC_UTMP_INIT ? INIT_PROCESS :
3374 context->utmp_mode == EXEC_UTMP_LOGIN ? LOGIN_PROCESS :
3378 if (uid_is_valid(uid)) {
3379 r = chown_terminal(STDIN_FILENO, uid);
3381 *exit_status = EXIT_STDIN;
3382 return log_unit_error_errno(unit, r, "Failed to change ownership of terminal: %m");
3386 /* If delegation is enabled we'll pass ownership of the cgroup to the user of the new process. On cgroup v1
3387 * this is only about systemd's own hierarchy, i.e. not the controller hierarchies, simply because that's not
3388 * safe. On cgroup v2 there's only one hierarchy anyway, and delegation is safe there, hence in that case only
3389 * touch a single hierarchy too. */
3390 if (params->cgroup_path && context->user && (params->flags & EXEC_CGROUP_DELEGATE)) {
3391 r = cg_set_access(SYSTEMD_CGROUP_CONTROLLER, params->cgroup_path, uid, gid);
3393 *exit_status = EXIT_CGROUP;
3394 return log_unit_error_errno(unit, r, "Failed to adjust control group access: %m");
3398 for (dt = 0; dt < _EXEC_DIRECTORY_TYPE_MAX; dt++) {
3399 r = setup_exec_directory(context, params, uid, gid, dt, exit_status);
3401 return log_unit_error_errno(unit, r, "Failed to set up special execution directory in %s: %m", params->prefix[dt]);
3404 r = build_environment(
3416 *exit_status = EXIT_MEMORY;
3420 r = build_pass_environment(context, &pass_env);
3422 *exit_status = EXIT_MEMORY;
3426 accum_env = strv_env_merge(5,
3427 params->environment,
3430 context->environment,
3434 *exit_status = EXIT_MEMORY;
3437 accum_env = strv_env_clean(accum_env);
3439 (void) umask(context->umask);
3441 r = setup_keyring(unit, context, params, uid, gid);
3443 *exit_status = EXIT_KEYRING;
3444 return log_unit_error_errno(unit, r, "Failed to set up kernel keyring: %m");
3447 /* We need sandboxing if the caller asked us to apply it and the command isn't explicitly excepted from it */
3448 needs_sandboxing = (params->flags & EXEC_APPLY_SANDBOXING) && !(command->flags & EXEC_COMMAND_FULLY_PRIVILEGED);
3450 /* We need the ambient capability hack, if the caller asked us to apply it and the command is marked for it, and the kernel doesn't actually support ambient caps */
3451 needs_ambient_hack = (params->flags & EXEC_APPLY_SANDBOXING) && (command->flags & EXEC_COMMAND_AMBIENT_MAGIC) && !ambient_capabilities_supported();
3453 /* We need setresuid() if the caller asked us to apply sandboxing and the command isn't explicitly excepted from either whole sandboxing or just setresuid() itself, and the ambient hack is not desired */
3454 if (needs_ambient_hack)
3455 needs_setuid = false;
3457 needs_setuid = (params->flags & EXEC_APPLY_SANDBOXING) && !(command->flags & (EXEC_COMMAND_FULLY_PRIVILEGED|EXEC_COMMAND_NO_SETUID));
3459 if (needs_sandboxing) {
3460 /* MAC enablement checks need to be done before a new mount ns is created, as they rely on /sys being
3461 * present. The actual MAC context application will happen later, as late as possible, to avoid
3462 * impacting our own code paths. */
3465 use_selinux = mac_selinux_use();
3468 use_smack = mac_smack_use();
3471 use_apparmor = mac_apparmor_use();
3475 if (needs_sandboxing) {
3478 /* Let's set the resource limits before we call into PAM, so that pam_limits wins over what
3479 * is set here. (See below.) */
3481 r = setrlimit_closest_all((const struct rlimit* const *) context->rlimit, &which_failed);
3483 *exit_status = EXIT_LIMITS;
3484 return log_unit_error_errno(unit, r, "Failed to adjust resource limit RLIMIT_%s: %m", rlimit_to_string(which_failed));
3490 /* Let's call into PAM after we set up our own idea of resource limits to that pam_limits
3491 * wins here. (See above.) */
3493 if (context->pam_name && username) {
3494 r = setup_pam(context->pam_name, username, uid, gid, context->tty_path, &accum_env, fds, n_fds);
3496 *exit_status = EXIT_PAM;
3497 return log_unit_error_errno(unit, r, "Failed to set up PAM session: %m");
3502 if ((context->private_network || context->network_namespace_path) && runtime && runtime->netns_storage_socket[0] >= 0) {
3504 if (ns_type_supported(NAMESPACE_NET)) {
3505 r = setup_netns(runtime->netns_storage_socket);
3507 *exit_status = EXIT_NETWORK;
3508 return log_unit_error_errno(unit, r, "Failed to set up network namespacing: %m");
3510 } else if (context->network_namespace_path) {
3511 *exit_status = EXIT_NETWORK;
3512 return log_unit_error_errno(unit, SYNTHETIC_ERRNO(EOPNOTSUPP), "NetworkNamespacePath= is not supported, refusing.");
3514 log_unit_warning(unit, "PrivateNetwork=yes is configured, but the kernel does not support network namespaces, ignoring.");
3517 needs_mount_namespace = exec_needs_mount_namespace(context, params, runtime);
3518 if (needs_mount_namespace) {
3519 _cleanup_free_ char *error_path = NULL;
3521 r = apply_mount_namespace(unit, command, context, params, runtime, &error_path);
3523 *exit_status = EXIT_NAMESPACE;
3524 return log_unit_error_errno(unit, r, "Failed to set up mount namespacing%s%s: %m",
3525 error_path ? ": " : "", strempty(error_path));
3529 if (context->protect_hostname) {
3530 if (ns_type_supported(NAMESPACE_UTS)) {
3531 if (unshare(CLONE_NEWUTS) < 0) {
3532 if (!ERRNO_IS_NOT_SUPPORTED(errno) && !ERRNO_IS_PRIVILEGE(errno)) {
3533 *exit_status = EXIT_NAMESPACE;
3534 return log_unit_error_errno(unit, errno, "Failed to set up UTS namespacing: %m");
3537 log_unit_warning(unit, "ProtectHostname=yes is configured, but UTS namespace setup is prohibited (container manager?), ignoring namespace setup.");
3540 log_unit_warning(unit, "ProtectHostname=yes is configured, but the kernel does not support UTS namespaces, ignoring namespace setup.");
3542 r = seccomp_protect_hostname();
3544 *exit_status = EXIT_SECCOMP;
3545 return log_unit_error_errno(unit, r, "Failed to apply hostname restrictions: %m");
3550 /* Drop groups as early as possbile */
3552 r = enforce_groups(gid, supplementary_gids, ngids);
3554 *exit_status = EXIT_GROUP;
3555 return log_unit_error_errno(unit, r, "Changing group credentials failed: %m");
3559 if (needs_sandboxing) {
3561 if (use_selinux && params->selinux_context_net && socket_fd >= 0) {
3562 r = mac_selinux_get_child_mls_label(socket_fd, command->path, context->selinux_context, &mac_selinux_context_net);
3564 *exit_status = EXIT_SELINUX_CONTEXT;
3565 return log_unit_error_errno(unit, r, "Failed to determine SELinux context: %m");
3570 if (context->private_users) {
3571 r = setup_private_users(uid, gid);
3573 *exit_status = EXIT_USER;
3574 return log_unit_error_errno(unit, r, "Failed to set up user namespacing: %m");
3579 /* We repeat the fd closing here, to make sure that nothing is leaked from the PAM modules. Note that we are
3580 * more aggressive this time since socket_fd and the netns fds we don't need anymore. We do keep the exec_fd
3581 * however if we have it as we want to keep it open until the final execve(). */
3583 if (params->exec_fd >= 0) {
3584 exec_fd = params->exec_fd;
3586 if (exec_fd < 3 + (int) n_fds) {
3589 /* Let's move the exec fd far up, so that it's outside of the fd range we want to pass to the
3590 * process we are about to execute. */
3592 moved_fd = fcntl(exec_fd, F_DUPFD_CLOEXEC, 3 + (int) n_fds);
3594 *exit_status = EXIT_FDS;
3595 return log_unit_error_errno(unit, errno, "Couldn't move exec fd up: %m");
3598 safe_close(exec_fd);
3601 /* This fd should be FD_CLOEXEC already, but let's make sure. */
3602 r = fd_cloexec(exec_fd, true);
3604 *exit_status = EXIT_FDS;
3605 return log_unit_error_errno(unit, r, "Failed to make exec fd FD_CLOEXEC: %m");
3609 fds_with_exec_fd = newa(int, n_fds + 1);
3610 memcpy_safe(fds_with_exec_fd, fds, n_fds * sizeof(int));
3611 fds_with_exec_fd[n_fds] = exec_fd;
3612 n_fds_with_exec_fd = n_fds + 1;
3614 fds_with_exec_fd = fds;
3615 n_fds_with_exec_fd = n_fds;
3618 r = close_all_fds(fds_with_exec_fd, n_fds_with_exec_fd);
3620 r = shift_fds(fds, n_fds);
3622 r = flags_fds(fds, n_socket_fds, n_storage_fds, context->non_blocking);
3624 *exit_status = EXIT_FDS;
3625 return log_unit_error_errno(unit, r, "Failed to adjust passed file descriptors: %m");
3628 /* At this point, the fds we want to pass to the program are all ready and set up, with O_CLOEXEC turned off
3629 * and at the right fd numbers. The are no other fds open, with one exception: the exec_fd if it is defined,
3630 * and it has O_CLOEXEC set, after all we want it to be closed by the execve(), so that our parent knows we
3633 secure_bits = context->secure_bits;
3635 if (needs_sandboxing) {
3638 /* Set the RTPRIO resource limit to 0, but only if nothing else was explicitly
3639 * requested. (Note this is placed after the general resource limit initialization, see
3640 * above, in order to take precedence.) */
3641 if (context->restrict_realtime && !context->rlimit[RLIMIT_RTPRIO]) {
3642 if (setrlimit(RLIMIT_RTPRIO, &RLIMIT_MAKE_CONST(0)) < 0) {
3643 *exit_status = EXIT_LIMITS;
3644 return log_unit_error_errno(unit, errno, "Failed to adjust RLIMIT_RTPRIO resource limit: %m");
3649 /* LSM Smack needs the capability CAP_MAC_ADMIN to change the current execution security context of the
3650 * process. This is the latest place before dropping capabilities. Other MAC context are set later. */
3652 r = setup_smack(unit->manager, context, command);
3654 *exit_status = EXIT_SMACK_PROCESS_LABEL;
3655 return log_unit_error_errno(unit, r, "Failed to set SMACK process label: %m");
3660 bset = context->capability_bounding_set;
3661 /* If the ambient caps hack is enabled (which means the kernel can't do them, and the user asked for
3662 * our magic fallback), then let's add some extra caps, so that the service can drop privs of its own,
3663 * instead of us doing that */
3664 if (needs_ambient_hack)
3665 bset |= (UINT64_C(1) << CAP_SETPCAP) |
3666 (UINT64_C(1) << CAP_SETUID) |
3667 (UINT64_C(1) << CAP_SETGID);
3669 if (!cap_test_all(bset)) {
3670 r = capability_bounding_set_drop(bset, false);
3672 *exit_status = EXIT_CAPABILITIES;
3673 return log_unit_error_errno(unit, r, "Failed to drop capabilities: %m");
3677 /* This is done before enforce_user, but ambient set
3678 * does not survive over setresuid() if keep_caps is not set. */
3679 if (!needs_ambient_hack &&
3680 context->capability_ambient_set != 0) {
3681 r = capability_ambient_set_apply(context->capability_ambient_set, true);
3683 *exit_status = EXIT_CAPABILITIES;
3684 return log_unit_error_errno(unit, r, "Failed to apply ambient capabilities (before UID change): %m");
3687 if (context->capabilities) {
3689 /* The capabilities in ambient set need to be also in the inherited
3690 * set. If they aren't, trying to get them will fail. Add the ambient
3691 * set inherited capabilities to the capability set in the context.
3692 * This is needed because if capabilities are set (using "Capabilities="
3693 * keyword), they will override whatever we set now. */
3695 r = capability_update_inherited_set(context->capabilities, context->capability_ambient_set);
3697 *exit_status = EXIT_CAPABILITIES;
3705 /* chroot to root directory first, before we lose the ability to chroot */
3706 r = apply_root_directory(context, params, needs_mount_namespace, exit_status);
3708 return log_unit_error_errno(unit, r, "Chrooting to the requested root directory failed: %m");
3711 if (uid_is_valid(uid)) {
3712 r = enforce_user(context, uid);
3714 *exit_status = EXIT_USER;
3715 return log_unit_error_errno(unit, r, "Failed to change UID to " UID_FMT ": %m", uid);
3718 if (!needs_ambient_hack &&
3719 context->capability_ambient_set != 0) {
3721 /* Fix the ambient capabilities after user change. */
3722 r = capability_ambient_set_apply(context->capability_ambient_set, false);
3724 *exit_status = EXIT_CAPABILITIES;
3725 return log_unit_error_errno(unit, r, "Failed to apply ambient capabilities (after UID change): %m");
3728 /* If we were asked to change user and ambient capabilities
3729 * were requested, we had to add keep-caps to the securebits
3730 * so that we would maintain the inherited capability set
3731 * through the setresuid(). Make sure that the bit is added
3732 * also to the context secure_bits so that we don't try to
3733 * drop the bit away next. */
3735 secure_bits |= 1<<SECURE_KEEP_CAPS;
3740 /* Apply working directory here, because the working directory might be on NFS and only the user running
3741 * this service might have the correct privilege to change to the working directory */
3742 r = apply_working_directory(context, params, home, exit_status);
3744 return log_unit_error_errno(unit, r, "Changing to the requested working directory failed: %m");
3746 if (needs_sandboxing) {
3747 /* Apply other MAC contexts late, but before seccomp syscall filtering, as those should really be last to
3748 * influence our own codepaths as little as possible. Moreover, applying MAC contexts usually requires
3749 * syscalls that are subject to seccomp filtering, hence should probably be applied before the syscalls
3750 * are restricted. */
3754 char *exec_context = mac_selinux_context_net ?: context->selinux_context;
3757 r = setexeccon(exec_context);
3759 *exit_status = EXIT_SELINUX_CONTEXT;
3760 return log_unit_error_errno(unit, r, "Failed to change SELinux context to %s: %m", exec_context);
3767 if (use_apparmor && context->apparmor_profile) {
3768 r = aa_change_onexec(context->apparmor_profile);
3769 if (r < 0 && !context->apparmor_profile_ignore) {
3770 *exit_status = EXIT_APPARMOR_PROFILE;
3771 return log_unit_error_errno(unit, errno, "Failed to prepare AppArmor profile change to %s: %m", context->apparmor_profile);
3776 /* PR_GET_SECUREBITS is not privileged, while PR_SET_SECUREBITS is. So to suppress potential EPERMs
3777 * we'll try not to call PR_SET_SECUREBITS unless necessary. */
3778 if (prctl(PR_GET_SECUREBITS) != secure_bits)
3779 if (prctl(PR_SET_SECUREBITS, secure_bits) < 0) {
3780 *exit_status = EXIT_SECUREBITS;
3781 return log_unit_error_errno(unit, errno, "Failed to set process secure bits: %m");
3784 if (context->capabilities)
3785 if (cap_set_proc(context->capabilities) < 0) {
3786 *exit_status = EXIT_CAPABILITIES;
3790 if (context_has_no_new_privileges(context))
3791 if (prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0) < 0) {
3792 *exit_status = EXIT_NO_NEW_PRIVILEGES;
3793 return log_unit_error_errno(unit, errno, "Failed to disable new privileges: %m");
3797 r = apply_address_families(unit, context);
3799 *exit_status = EXIT_ADDRESS_FAMILIES;
3800 return log_unit_error_errno(unit, r, "Failed to restrict address families: %m");
3803 r = apply_memory_deny_write_execute(unit, context);
3805 *exit_status = EXIT_SECCOMP;
3806 return log_unit_error_errno(unit, r, "Failed to disable writing to executable memory: %m");
3809 r = apply_restrict_realtime(unit, context);
3811 *exit_status = EXIT_SECCOMP;
3812 return log_unit_error_errno(unit, r, "Failed to apply realtime restrictions: %m");
3815 r = apply_restrict_suid_sgid(unit, context);
3817 *exit_status = EXIT_SECCOMP;
3818 return log_unit_error_errno(unit, r, "Failed to apply SUID/SGID restrictions: %m");
3821 r = apply_restrict_namespaces(unit, context);
3823 *exit_status = EXIT_SECCOMP;
3824 return log_unit_error_errno(unit, r, "Failed to apply namespace restrictions: %m");
3827 r = apply_protect_sysctl(unit, context);
3829 *exit_status = EXIT_SECCOMP;
3830 return log_unit_error_errno(unit, r, "Failed to apply sysctl restrictions: %m");
3833 r = apply_protect_kernel_modules(unit, context);
3835 *exit_status = EXIT_SECCOMP;
3836 return log_unit_error_errno(unit, r, "Failed to apply module loading restrictions: %m");
3839 r = apply_protect_kernel_logs(unit, context);
3841 *exit_status = EXIT_SECCOMP;
3842 return log_unit_error_errno(unit, r, "Failed to apply kernel log restrictions: %m");
3845 r = apply_private_devices(unit, context);
3847 *exit_status = EXIT_SECCOMP;
3848 return log_unit_error_errno(unit, r, "Failed to set up private devices: %m");
3851 r = apply_syscall_archs(unit, context);
3853 *exit_status = EXIT_SECCOMP;
3854 return log_unit_error_errno(unit, r, "Failed to apply syscall architecture restrictions: %m");
3857 r = apply_lock_personality(unit, context);
3859 *exit_status = EXIT_SECCOMP;
3860 return log_unit_error_errno(unit, r, "Failed to lock personalities: %m");
3863 /* This really should remain the last step before the execve(), to make sure our own code is unaffected
3864 * by the filter as little as possible. */
3865 r = apply_syscall_filter(unit, context, needs_ambient_hack);
3867 *exit_status = EXIT_SECCOMP;
3868 return log_unit_error_errno(unit, r, "Failed to apply system call filters: %m");
3873 if (!strv_isempty(context->unset_environment)) {
3876 ee = strv_env_delete(accum_env, 1, context->unset_environment);
3878 *exit_status = EXIT_MEMORY;
3882 strv_free_and_replace(accum_env, ee);
3885 if (!FLAGS_SET(command->flags, EXEC_COMMAND_NO_ENV_EXPAND)) {
3886 replaced_argv = replace_env_argv(command->argv, accum_env);
3887 if (!replaced_argv) {
3888 *exit_status = EXIT_MEMORY;
3891 final_argv = replaced_argv;
3893 final_argv = command->argv;
3895 if (DEBUG_LOGGING) {
3896 _cleanup_free_ char *line;
3898 line = exec_command_line(final_argv);
3900 log_struct(LOG_DEBUG,
3901 "EXECUTABLE=%s", command->path,
3902 LOG_UNIT_MESSAGE(unit, "Executing: %s", line),
3904 LOG_UNIT_INVOCATION_ID(unit));
3910 /* We have finished with all our initializations. Let's now let the manager know that. From this point
3911 * on, if the manager sees POLLHUP on the exec_fd, then execve() was successful. */
3913 if (write(exec_fd, &hot, sizeof(hot)) < 0) {
3914 *exit_status = EXIT_EXEC;
3915 return log_unit_error_errno(unit, errno, "Failed to enable exec_fd: %m");
3919 execve(command->path, final_argv, accum_env);
3925 /* The execve() failed. This means the exec_fd is still open. Which means we need to tell the manager
3926 * that POLLHUP on it no longer means execve() succeeded. */
3928 if (write(exec_fd, &hot, sizeof(hot)) < 0) {
3929 *exit_status = EXIT_EXEC;
3930 return log_unit_error_errno(unit, errno, "Failed to disable exec_fd: %m");
3934 if (r == -ENOENT && (command->flags & EXEC_COMMAND_IGNORE_FAILURE)) {
3935 log_struct_errno(LOG_INFO, r,
3936 "MESSAGE_ID=" SD_MESSAGE_SPAWN_FAILED_STR,
3938 LOG_UNIT_INVOCATION_ID(unit),
3939 LOG_UNIT_MESSAGE(unit, "Executable %s missing, skipping: %m",
3941 "EXECUTABLE=%s", command->path);
3945 *exit_status = EXIT_EXEC;
3946 return log_unit_error_errno(unit, r, "Failed to execute command: %m");
3949 static int exec_context_load_environment(const Unit *unit, const ExecContext *c, char ***l);
3950 static int exec_context_named_iofds(const ExecContext *c, const ExecParameters *p, int named_iofds[static 3]);
3952 int exec_spawn(Unit *unit,
3953 ExecCommand *command,
3954 const ExecContext *context,
3955 const ExecParameters *params,
3956 ExecRuntime *runtime,
3957 DynamicCreds *dcreds,
3960 int socket_fd, r, named_iofds[3] = { -1, -1, -1 }, *fds = NULL;
3961 _cleanup_free_ char *subcgroup_path = NULL;
3962 _cleanup_strv_free_ char **files_env = NULL;
3963 size_t n_storage_fds = 0, n_socket_fds = 0;
3964 _cleanup_free_ char *line = NULL;
3972 assert(params->fds || (params->n_socket_fds + params->n_storage_fds <= 0));
3974 if (context->std_input == EXEC_INPUT_SOCKET ||
3975 context->std_output == EXEC_OUTPUT_SOCKET ||
3976 context->std_error == EXEC_OUTPUT_SOCKET) {
3978 if (params->n_socket_fds > 1) {
3979 log_unit_error(unit, "Got more than one socket.");
3983 if (params->n_socket_fds == 0) {
3984 log_unit_error(unit, "Got no socket.");
3988 socket_fd = params->fds[0];
3992 n_socket_fds = params->n_socket_fds;
3993 n_storage_fds = params->n_storage_fds;
3996 r = exec_context_named_iofds(context, params, named_iofds);
3998 return log_unit_error_errno(unit, r, "Failed to load a named file descriptor: %m");
4000 r = exec_context_load_environment(unit, context, &files_env);
4002 return log_unit_error_errno(unit, r, "Failed to load environment files: %m");
4004 line = exec_command_line(command->argv);
4008 log_struct(LOG_DEBUG,
4009 LOG_UNIT_MESSAGE(unit, "About to execute: %s", line),
4010 "EXECUTABLE=%s", command->path,
4012 LOG_UNIT_INVOCATION_ID(unit));
4014 if (params->cgroup_path) {
4015 r = exec_parameters_get_cgroup_path(params, &subcgroup_path);
4017 return log_unit_error_errno(unit, r, "Failed to acquire subcgroup path: %m");
4018 if (r > 0) { /* We are using a child cgroup */
4019 r = cg_create(SYSTEMD_CGROUP_CONTROLLER, subcgroup_path);
4021 return log_unit_error_errno(unit, r, "Failed to create control group '%s': %m", subcgroup_path);
4027 return log_unit_error_errno(unit, errno, "Failed to fork: %m");
4030 int exit_status = EXIT_SUCCESS;
4032 r = exec_child(unit,
4044 unit->manager->user_lookup_fds[1],
4048 const char *status =
4049 exit_status_to_string(exit_status,
4050 EXIT_STATUS_LIBC | EXIT_STATUS_SYSTEMD);
4052 log_struct_errno(LOG_ERR, r,
4053 "MESSAGE_ID=" SD_MESSAGE_SPAWN_FAILED_STR,
4055 LOG_UNIT_INVOCATION_ID(unit),
4056 LOG_UNIT_MESSAGE(unit, "Failed at step %s spawning %s: %m",
4057 status, command->path),
4058 "EXECUTABLE=%s", command->path);
4064 log_unit_debug(unit, "Forked %s as "PID_FMT, command->path, pid);
4066 /* We add the new process to the cgroup both in the child (so that we can be sure that no user code is ever
4067 * executed outside of the cgroup) and in the parent (so that we can be sure that when we kill the cgroup the
4068 * process will be killed too). */
4070 (void) cg_attach(SYSTEMD_CGROUP_CONTROLLER, subcgroup_path, pid);
4072 exec_status_start(&command->exec_status, pid);
4078 void exec_context_init(ExecContext *c) {
4079 ExecDirectoryType i;
4084 c->ioprio = IOPRIO_PRIO_VALUE(IOPRIO_CLASS_BE, 0);
4085 c->cpu_sched_policy = SCHED_OTHER;
4086 c->syslog_priority = LOG_DAEMON|LOG_INFO;
4087 c->syslog_level_prefix = true;
4088 c->ignore_sigpipe = true;
4089 c->timer_slack_nsec = NSEC_INFINITY;
4090 c->personality = PERSONALITY_INVALID;
4091 for (i = 0; i < _EXEC_DIRECTORY_TYPE_MAX; i++)
4092 c->directories[i].mode = 0755;
4093 c->timeout_clean_usec = USEC_INFINITY;
4094 c->capability_bounding_set = CAP_ALL;
4095 assert_cc(NAMESPACE_FLAGS_INITIAL != NAMESPACE_FLAGS_ALL);
4096 c->restrict_namespaces = NAMESPACE_FLAGS_INITIAL;
4097 c->log_level_max = -1;
4098 numa_policy_reset(&c->numa_policy);
4101 void exec_context_done(ExecContext *c) {
4102 ExecDirectoryType i;
4107 c->environment = strv_free(c->environment);
4108 c->environment_files = strv_free(c->environment_files);
4109 c->pass_environment = strv_free(c->pass_environment);
4110 c->unset_environment = strv_free(c->unset_environment);
4112 rlimit_free_all(c->rlimit);
4114 for (l = 0; l < 3; l++) {
4115 c->stdio_fdname[l] = mfree(c->stdio_fdname[l]);
4116 c->stdio_file[l] = mfree(c->stdio_file[l]);
4119 c->working_directory = mfree(c->working_directory);
4120 c->root_directory = mfree(c->root_directory);
4121 c->root_image = mfree(c->root_image);
4122 c->tty_path = mfree(c->tty_path);
4123 c->syslog_identifier = mfree(c->syslog_identifier);
4124 c->user = mfree(c->user);
4125 c->group = mfree(c->group);
4127 c->supplementary_groups = strv_free(c->supplementary_groups);
4129 c->pam_name = mfree(c->pam_name);
4131 if (c->capabilities) {
4132 cap_free(c->capabilities);
4133 c->capabilities = NULL;
4136 c->read_only_paths = strv_free(c->read_only_paths);
4137 c->read_write_paths = strv_free(c->read_write_paths);
4138 c->inaccessible_paths = strv_free(c->inaccessible_paths);
4140 bind_mount_free_many(c->bind_mounts, c->n_bind_mounts);
4141 c->bind_mounts = NULL;
4142 c->n_bind_mounts = 0;
4143 temporary_filesystem_free_many(c->temporary_filesystems, c->n_temporary_filesystems);
4144 c->temporary_filesystems = NULL;
4145 c->n_temporary_filesystems = 0;
4147 cpu_set_reset(&c->cpu_set);
4148 numa_policy_reset(&c->numa_policy);
4150 c->utmp_id = mfree(c->utmp_id);
4151 c->selinux_context = mfree(c->selinux_context);
4152 c->apparmor_profile = mfree(c->apparmor_profile);
4153 c->smack_process_label = mfree(c->smack_process_label);
4155 c->syscall_filter = hashmap_free(c->syscall_filter);
4156 c->syscall_archs = set_free(c->syscall_archs);
4157 c->address_families = set_free(c->address_families);
4159 for (i = 0; i < _EXEC_DIRECTORY_TYPE_MAX; i++)
4160 c->directories[i].paths = strv_free(c->directories[i].paths);
4162 c->log_level_max = -1;
4164 exec_context_free_log_extra_fields(c);
4166 c->log_ratelimit_interval_usec = 0;
4167 c->log_ratelimit_burst = 0;
4169 c->stdin_data = mfree(c->stdin_data);
4170 c->stdin_data_size = 0;
4172 c->network_namespace_path = mfree(c->network_namespace_path);
4175 int exec_context_destroy_runtime_directory(const ExecContext *c, const char *runtime_prefix) {
4180 if (!runtime_prefix)
4183 STRV_FOREACH(i, c->directories[EXEC_DIRECTORY_RUNTIME].paths) {
4184 _cleanup_free_ char *p;
4186 if (exec_directory_is_private(c, EXEC_DIRECTORY_RUNTIME))
4187 p = path_join(runtime_prefix, "private", *i);
4189 p = path_join(runtime_prefix, *i);
4193 /* We execute this synchronously, since we need to be sure this is gone when we start the
4195 (void) rm_rf(p, REMOVE_ROOT);
4201 static void exec_command_done(ExecCommand *c) {
4204 c->path = mfree(c->path);
4205 c->argv = strv_free(c->argv);
4208 void exec_command_done_array(ExecCommand *c, size_t n) {
4211 for (i = 0; i < n; i++)
4212 exec_command_done(c+i);
4215 ExecCommand* exec_command_free_list(ExecCommand *c) {
4219 LIST_REMOVE(command, c, i);
4220 exec_command_done(i);
4227 void exec_command_free_array(ExecCommand **c, size_t n) {
4230 for (i = 0; i < n; i++)
4231 c[i] = exec_command_free_list(c[i]);
4234 void exec_command_reset_status_array(ExecCommand *c, size_t n) {
4237 for (i = 0; i < n; i++)
4238 exec_status_reset(&c[i].exec_status);
4241 void exec_command_reset_status_list_array(ExecCommand **c, size_t n) {
4244 for (i = 0; i < n; i++) {
4247 LIST_FOREACH(command, z, c[i])
4248 exec_status_reset(&z->exec_status);
4252 typedef struct InvalidEnvInfo {
4257 static void invalid_env(const char *p, void *userdata) {
4258 InvalidEnvInfo *info = userdata;
4260 log_unit_error(info->unit, "Ignoring invalid environment assignment '%s': %s", p, info->path);
4263 const char* exec_context_fdname(const ExecContext *c, int fd_index) {
4269 if (c->std_input != EXEC_INPUT_NAMED_FD)
4272 return c->stdio_fdname[STDIN_FILENO] ?: "stdin";
4275 if (c->std_output != EXEC_OUTPUT_NAMED_FD)
4278 return c->stdio_fdname[STDOUT_FILENO] ?: "stdout";
4281 if (c->std_error != EXEC_OUTPUT_NAMED_FD)
4284 return c->stdio_fdname[STDERR_FILENO] ?: "stderr";
4291 static int exec_context_named_iofds(
4292 const ExecContext *c,
4293 const ExecParameters *p,
4294 int named_iofds[static 3]) {
4297 const char* stdio_fdname[3];
4302 assert(named_iofds);
4304 targets = (c->std_input == EXEC_INPUT_NAMED_FD) +
4305 (c->std_output == EXEC_OUTPUT_NAMED_FD) +
4306 (c->std_error == EXEC_OUTPUT_NAMED_FD);
4308 for (i = 0; i < 3; i++)
4309 stdio_fdname[i] = exec_context_fdname(c, i);
4311 n_fds = p->n_storage_fds + p->n_socket_fds;
4313 for (i = 0; i < n_fds && targets > 0; i++)
4314 if (named_iofds[STDIN_FILENO] < 0 &&
4315 c->std_input == EXEC_INPUT_NAMED_FD &&
4316 stdio_fdname[STDIN_FILENO] &&
4317 streq(p->fd_names[i], stdio_fdname[STDIN_FILENO])) {
4319 named_iofds[STDIN_FILENO] = p->fds[i];
4322 } else if (named_iofds[STDOUT_FILENO] < 0 &&
4323 c->std_output == EXEC_OUTPUT_NAMED_FD &&
4324 stdio_fdname[STDOUT_FILENO] &&
4325 streq(p->fd_names[i], stdio_fdname[STDOUT_FILENO])) {
4327 named_iofds[STDOUT_FILENO] = p->fds[i];
4330 } else if (named_iofds[STDERR_FILENO] < 0 &&
4331 c->std_error == EXEC_OUTPUT_NAMED_FD &&
4332 stdio_fdname[STDERR_FILENO] &&
4333 streq(p->fd_names[i], stdio_fdname[STDERR_FILENO])) {
4335 named_iofds[STDERR_FILENO] = p->fds[i];
4339 return targets == 0 ? 0 : -ENOENT;
4342 static int exec_context_load_environment(const Unit *unit, const ExecContext *c, char ***l) {
4343 char **i, **r = NULL;
4348 STRV_FOREACH(i, c->environment_files) {
4352 bool ignore = false;
4354 _cleanup_globfree_ glob_t pglob = {};
4363 if (!path_is_absolute(fn)) {
4371 /* Filename supports globbing, take all matching files */
4372 k = safe_glob(fn, 0, &pglob);
4381 /* When we don't match anything, -ENOENT should be returned */
4382 assert(pglob.gl_pathc > 0);
4384 for (n = 0; n < pglob.gl_pathc; n++) {
4385 k = load_env_file(NULL, pglob.gl_pathv[n], &p);
4393 /* Log invalid environment variables with filename */
4395 InvalidEnvInfo info = {
4397 .path = pglob.gl_pathv[n]
4400 p = strv_env_clean_with_callback(p, invalid_env, &info);
4408 m = strv_env_merge(2, r, p);
4424 static bool tty_may_match_dev_console(const char *tty) {
4425 _cleanup_free_ char *resolved = NULL;
4430 tty = skip_dev_prefix(tty);
4432 /* trivial identity? */
4433 if (streq(tty, "console"))
4436 if (resolve_dev_console(&resolved) < 0)
4437 return true; /* if we could not resolve, assume it may */
4439 /* "tty0" means the active VC, so it may be the same sometimes */
4440 return path_equal(resolved, tty) || (streq(resolved, "tty0") && tty_is_vc(tty));
4443 static bool exec_context_may_touch_tty(const ExecContext *ec) {
4446 return ec->tty_reset ||
4448 ec->tty_vt_disallocate ||
4449 is_terminal_input(ec->std_input) ||
4450 is_terminal_output(ec->std_output) ||
4451 is_terminal_output(ec->std_error);
4454 bool exec_context_may_touch_console(const ExecContext *ec) {
4456 return exec_context_may_touch_tty(ec) &&
4457 tty_may_match_dev_console(exec_context_tty_path(ec));
4460 static void strv_fprintf(FILE *f, char **l) {
4466 fprintf(f, " %s", *g);
4469 void exec_context_dump(const ExecContext *c, FILE* f, const char *prefix) {
4470 char **e, **d, buf_clean[FORMAT_TIMESPAN_MAX];
4471 ExecDirectoryType dt;
4478 prefix = strempty(prefix);
4482 "%sWorkingDirectory: %s\n"
4483 "%sRootDirectory: %s\n"
4484 "%sNonBlocking: %s\n"
4485 "%sPrivateTmp: %s\n"
4486 "%sPrivateDevices: %s\n"
4487 "%sProtectKernelTunables: %s\n"
4488 "%sProtectKernelModules: %s\n"
4489 "%sProtectKernelLogs: %s\n"
4490 "%sProtectControlGroups: %s\n"
4491 "%sPrivateNetwork: %s\n"
4492 "%sPrivateUsers: %s\n"
4493 "%sProtectHome: %s\n"
4494 "%sProtectSystem: %s\n"
4495 "%sMountAPIVFS: %s\n"
4496 "%sIgnoreSIGPIPE: %s\n"
4497 "%sMemoryDenyWriteExecute: %s\n"
4498 "%sRestrictRealtime: %s\n"
4499 "%sRestrictSUIDSGID: %s\n"
4500 "%sKeyringMode: %s\n"
4501 "%sProtectHostname: %s\n",
4503 prefix, c->working_directory ? c->working_directory : "/",
4504 prefix, c->root_directory ? c->root_directory : "/",
4505 prefix, yes_no(c->non_blocking),
4506 prefix, yes_no(c->private_tmp),
4507 prefix, yes_no(c->private_devices),
4508 prefix, yes_no(c->protect_kernel_tunables),
4509 prefix, yes_no(c->protect_kernel_modules),
4510 prefix, yes_no(c->protect_kernel_logs),
4511 prefix, yes_no(c->protect_control_groups),
4512 prefix, yes_no(c->private_network),
4513 prefix, yes_no(c->private_users),
4514 prefix, protect_home_to_string(c->protect_home),
4515 prefix, protect_system_to_string(c->protect_system),
4516 prefix, yes_no(c->mount_apivfs),
4517 prefix, yes_no(c->ignore_sigpipe),
4518 prefix, yes_no(c->memory_deny_write_execute),
4519 prefix, yes_no(c->restrict_realtime),
4520 prefix, yes_no(c->restrict_suid_sgid),
4521 prefix, exec_keyring_mode_to_string(c->keyring_mode),
4522 prefix, yes_no(c->protect_hostname));
4525 fprintf(f, "%sRootImage: %s\n", prefix, c->root_image);
4527 STRV_FOREACH(e, c->environment)
4528 fprintf(f, "%sEnvironment: %s\n", prefix, *e);
4530 STRV_FOREACH(e, c->environment_files)
4531 fprintf(f, "%sEnvironmentFile: %s\n", prefix, *e);
4533 STRV_FOREACH(e, c->pass_environment)
4534 fprintf(f, "%sPassEnvironment: %s\n", prefix, *e);
4536 STRV_FOREACH(e, c->unset_environment)
4537 fprintf(f, "%sUnsetEnvironment: %s\n", prefix, *e);
4539 fprintf(f, "%sRuntimeDirectoryPreserve: %s\n", prefix, exec_preserve_mode_to_string(c->runtime_directory_preserve_mode));
4541 for (dt = 0; dt < _EXEC_DIRECTORY_TYPE_MAX; dt++) {
4542 fprintf(f, "%s%sMode: %04o\n", prefix, exec_directory_type_to_string(dt), c->directories[dt].mode);
4544 STRV_FOREACH(d, c->directories[dt].paths)
4545 fprintf(f, "%s%s: %s\n", prefix, exec_directory_type_to_string(dt), *d);
4549 "%sTimeoutCleanSec: %s\n",
4550 prefix, format_timespan(buf_clean, sizeof(buf_clean), c->timeout_clean_usec, USEC_PER_SEC));
4557 if (c->oom_score_adjust_set)
4559 "%sOOMScoreAdjust: %i\n",
4560 prefix, c->oom_score_adjust);
4562 for (i = 0; i < RLIM_NLIMITS; i++)
4564 fprintf(f, "%sLimit%s: " RLIM_FMT "\n",
4565 prefix, rlimit_to_string(i), c->rlimit[i]->rlim_max);
4566 fprintf(f, "%sLimit%sSoft: " RLIM_FMT "\n",
4567 prefix, rlimit_to_string(i), c->rlimit[i]->rlim_cur);
4570 if (c->ioprio_set) {
4571 _cleanup_free_ char *class_str = NULL;
4573 r = ioprio_class_to_string_alloc(IOPRIO_PRIO_CLASS(c->ioprio), &class_str);
4575 fprintf(f, "%sIOSchedulingClass: %s\n", prefix, class_str);
4577 fprintf(f, "%sIOPriority: %lu\n", prefix, IOPRIO_PRIO_DATA(c->ioprio));
4580 if (c->cpu_sched_set) {
4581 _cleanup_free_ char *policy_str = NULL;
4583 r = sched_policy_to_string_alloc(c->cpu_sched_policy, &policy_str);
4585 fprintf(f, "%sCPUSchedulingPolicy: %s\n", prefix, policy_str);
4588 "%sCPUSchedulingPriority: %i\n"
4589 "%sCPUSchedulingResetOnFork: %s\n",
4590 prefix, c->cpu_sched_priority,
4591 prefix, yes_no(c->cpu_sched_reset_on_fork));
4594 if (c->cpu_set.set) {
4595 _cleanup_free_ char *affinity = NULL;
4597 affinity = cpu_set_to_range_string(&c->cpu_set);
4598 fprintf(f, "%sCPUAffinity: %s\n", prefix, affinity);
4601 if (mpol_is_valid(numa_policy_get_type(&c->numa_policy))) {
4602 _cleanup_free_ char *nodes = NULL;
4604 nodes = cpu_set_to_range_string(&c->numa_policy.nodes);
4605 fprintf(f, "%sNUMAPolicy: %s\n", prefix, mpol_to_string(numa_policy_get_type(&c->numa_policy)));
4606 fprintf(f, "%sNUMAMask: %s\n", prefix, strnull(nodes));
4609 if (c->timer_slack_nsec != NSEC_INFINITY)
4610 fprintf(f, "%sTimerSlackNSec: "NSEC_FMT "\n", prefix, c->timer_slack_nsec);
4613 "%sStandardInput: %s\n"
4614 "%sStandardOutput: %s\n"
4615 "%sStandardError: %s\n",
4616 prefix, exec_input_to_string(c->std_input),
4617 prefix, exec_output_to_string(c->std_output),
4618 prefix, exec_output_to_string(c->std_error));
4620 if (c->std_input == EXEC_INPUT_NAMED_FD)
4621 fprintf(f, "%sStandardInputFileDescriptorName: %s\n", prefix, c->stdio_fdname[STDIN_FILENO]);
4622 if (c->std_output == EXEC_OUTPUT_NAMED_FD)
4623 fprintf(f, "%sStandardOutputFileDescriptorName: %s\n", prefix, c->stdio_fdname[STDOUT_FILENO]);
4624 if (c->std_error == EXEC_OUTPUT_NAMED_FD)
4625 fprintf(f, "%sStandardErrorFileDescriptorName: %s\n", prefix, c->stdio_fdname[STDERR_FILENO]);
4627 if (c->std_input == EXEC_INPUT_FILE)
4628 fprintf(f, "%sStandardInputFile: %s\n", prefix, c->stdio_file[STDIN_FILENO]);
4629 if (c->std_output == EXEC_OUTPUT_FILE)
4630 fprintf(f, "%sStandardOutputFile: %s\n", prefix, c->stdio_file[STDOUT_FILENO]);
4631 if (c->std_output == EXEC_OUTPUT_FILE_APPEND)
4632 fprintf(f, "%sStandardOutputFileToAppend: %s\n", prefix, c->stdio_file[STDOUT_FILENO]);
4633 if (c->std_error == EXEC_OUTPUT_FILE)
4634 fprintf(f, "%sStandardErrorFile: %s\n", prefix, c->stdio_file[STDERR_FILENO]);
4635 if (c->std_error == EXEC_OUTPUT_FILE_APPEND)
4636 fprintf(f, "%sStandardErrorFileToAppend: %s\n", prefix, c->stdio_file[STDERR_FILENO]);
4642 "%sTTYVHangup: %s\n"
4643 "%sTTYVTDisallocate: %s\n",
4644 prefix, c->tty_path,
4645 prefix, yes_no(c->tty_reset),
4646 prefix, yes_no(c->tty_vhangup),
4647 prefix, yes_no(c->tty_vt_disallocate));
4649 if (IN_SET(c->std_output,
4652 EXEC_OUTPUT_JOURNAL,
4653 EXEC_OUTPUT_SYSLOG_AND_CONSOLE,
4654 EXEC_OUTPUT_KMSG_AND_CONSOLE,
4655 EXEC_OUTPUT_JOURNAL_AND_CONSOLE) ||
4656 IN_SET(c->std_error,
4659 EXEC_OUTPUT_JOURNAL,
4660 EXEC_OUTPUT_SYSLOG_AND_CONSOLE,
4661 EXEC_OUTPUT_KMSG_AND_CONSOLE,
4662 EXEC_OUTPUT_JOURNAL_AND_CONSOLE)) {
4664 _cleanup_free_ char *fac_str = NULL, *lvl_str = NULL;
4666 r = log_facility_unshifted_to_string_alloc(c->syslog_priority >> 3, &fac_str);
4668 fprintf(f, "%sSyslogFacility: %s\n", prefix, fac_str);
4670 r = log_level_to_string_alloc(LOG_PRI(c->syslog_priority), &lvl_str);
4672 fprintf(f, "%sSyslogLevel: %s\n", prefix, lvl_str);
4675 if (c->capabilities) {
4676 _cleanup_cap_free_charp_ char *t;
4678 t = cap_to_text(c->capabilities, NULL);
4680 fprintf(f, "%sCapabilities: %s\n", prefix, t);
4683 if (c->log_level_max >= 0) {
4684 _cleanup_free_ char *t = NULL;
4686 (void) log_level_to_string_alloc(c->log_level_max, &t);
4688 fprintf(f, "%sLogLevelMax: %s\n", prefix, strna(t));
4691 if (c->log_ratelimit_interval_usec > 0) {
4692 char buf_timespan[FORMAT_TIMESPAN_MAX];
4695 "%sLogRateLimitIntervalSec: %s\n",
4696 prefix, format_timespan(buf_timespan, sizeof(buf_timespan), c->log_ratelimit_interval_usec, USEC_PER_SEC));
4699 if (c->log_ratelimit_burst > 0)
4700 fprintf(f, "%sLogRateLimitBurst: %u\n", prefix, c->log_ratelimit_burst);
4702 if (c->n_log_extra_fields > 0) {
4705 for (j = 0; j < c->n_log_extra_fields; j++) {
4706 fprintf(f, "%sLogExtraFields: ", prefix);
4707 fwrite(c->log_extra_fields[j].iov_base,
4708 1, c->log_extra_fields[j].iov_len,
4714 if (c->secure_bits) {
4715 _cleanup_free_ char *str = NULL;
4717 r = secure_bits_to_string_alloc(c->secure_bits, &str);
4719 fprintf(f, "%sSecure Bits: %s\n", prefix, str);
4722 if (c->capability_bounding_set != CAP_ALL) {
4723 _cleanup_free_ char *str = NULL;
4725 r = capability_set_to_string_alloc(c->capability_bounding_set, &str);
4727 fprintf(f, "%sCapabilityBoundingSet: %s\n", prefix, str);
4730 if (c->capability_ambient_set != 0) {
4731 _cleanup_free_ char *str = NULL;
4733 r = capability_set_to_string_alloc(c->capability_ambient_set, &str);
4735 fprintf(f, "%sAmbientCapabilities: %s\n", prefix, str);
4739 fprintf(f, "%sUser: %s\n", prefix, c->user);
4741 fprintf(f, "%sGroup: %s\n", prefix, c->group);
4743 fprintf(f, "%sDynamicUser: %s\n", prefix, yes_no(c->dynamic_user));
4745 if (!strv_isempty(c->supplementary_groups)) {
4746 fprintf(f, "%sSupplementaryGroups:", prefix);
4747 strv_fprintf(f, c->supplementary_groups);
4752 fprintf(f, "%sPAMName: %s\n", prefix, c->pam_name);
4754 if (!strv_isempty(c->read_write_paths)) {
4755 fprintf(f, "%sReadWritePaths:", prefix);
4756 strv_fprintf(f, c->read_write_paths);
4760 if (!strv_isempty(c->read_only_paths)) {
4761 fprintf(f, "%sReadOnlyPaths:", prefix);
4762 strv_fprintf(f, c->read_only_paths);
4766 if (!strv_isempty(c->inaccessible_paths)) {
4767 fprintf(f, "%sInaccessiblePaths:", prefix);
4768 strv_fprintf(f, c->inaccessible_paths);
4772 if (c->n_bind_mounts > 0)
4773 for (i = 0; i < c->n_bind_mounts; i++)
4774 fprintf(f, "%s%s: %s%s:%s:%s\n", prefix,
4775 c->bind_mounts[i].read_only ? "BindReadOnlyPaths" : "BindPaths",
4776 c->bind_mounts[i].ignore_enoent ? "-": "",
4777 c->bind_mounts[i].source,
4778 c->bind_mounts[i].destination,
4779 c->bind_mounts[i].recursive ? "rbind" : "norbind");
4781 if (c->n_temporary_filesystems > 0)
4782 for (i = 0; i < c->n_temporary_filesystems; i++) {
4783 TemporaryFileSystem *t = c->temporary_filesystems + i;
4785 fprintf(f, "%sTemporaryFileSystem: %s%s%s\n", prefix,
4787 isempty(t->options) ? "" : ":",
4788 strempty(t->options));
4793 "%sUtmpIdentifier: %s\n",
4794 prefix, c->utmp_id);
4796 if (c->selinux_context)
4798 "%sSELinuxContext: %s%s\n",
4799 prefix, c->selinux_context_ignore ? "-" : "", c->selinux_context);
4801 if (c->apparmor_profile)
4803 "%sAppArmorProfile: %s%s\n",
4804 prefix, c->apparmor_profile_ignore ? "-" : "", c->apparmor_profile);
4806 if (c->smack_process_label)
4808 "%sSmackProcessLabel: %s%s\n",
4809 prefix, c->smack_process_label_ignore ? "-" : "", c->smack_process_label);
4811 if (c->personality != PERSONALITY_INVALID)
4813 "%sPersonality: %s\n",
4814 prefix, strna(personality_to_string(c->personality)));
4817 "%sLockPersonality: %s\n",
4818 prefix, yes_no(c->lock_personality));
4820 if (c->syscall_filter) {
4828 "%sSystemCallFilter: ",
4831 if (!c->syscall_whitelist)
4835 HASHMAP_FOREACH_KEY(val, id, c->syscall_filter, j) {
4836 _cleanup_free_ char *name = NULL;
4837 const char *errno_name = NULL;
4838 int num = PTR_TO_INT(val);
4845 name = seccomp_syscall_resolve_num_arch(SCMP_ARCH_NATIVE, PTR_TO_INT(id) - 1);
4846 fputs(strna(name), f);
4849 errno_name = errno_to_name(num);
4851 fprintf(f, ":%s", errno_name);
4853 fprintf(f, ":%d", num);
4861 if (c->syscall_archs) {
4868 "%sSystemCallArchitectures:",
4872 SET_FOREACH(id, c->syscall_archs, j)
4873 fprintf(f, " %s", strna(seccomp_arch_to_string(PTR_TO_UINT32(id) - 1)));
4878 if (exec_context_restrict_namespaces_set(c)) {
4879 _cleanup_free_ char *s = NULL;
4881 r = namespace_flags_to_string(c->restrict_namespaces, &s);
4883 fprintf(f, "%sRestrictNamespaces: %s\n",
4887 if (c->network_namespace_path)
4889 "%sNetworkNamespacePath: %s\n",
4890 prefix, c->network_namespace_path);
4892 if (c->syscall_errno > 0) {
4893 const char *errno_name;
4895 fprintf(f, "%sSystemCallErrorNumber: ", prefix);
4897 errno_name = errno_to_name(c->syscall_errno);
4899 fprintf(f, "%s\n", errno_name);
4901 fprintf(f, "%d\n", c->syscall_errno);
4905 bool exec_context_maintains_privileges(const ExecContext *c) {
4908 /* Returns true if the process forked off would run under
4909 * an unchanged UID or as root. */
4914 if (streq(c->user, "root") || streq(c->user, "0"))
4920 int exec_context_get_effective_ioprio(const ExecContext *c) {
4928 p = ioprio_get(IOPRIO_WHO_PROCESS, 0);
4930 return IOPRIO_PRIO_VALUE(IOPRIO_CLASS_BE, 4);
4935 void exec_context_free_log_extra_fields(ExecContext *c) {
4940 for (l = 0; l < c->n_log_extra_fields; l++)
4941 free(c->log_extra_fields[l].iov_base);
4942 c->log_extra_fields = mfree(c->log_extra_fields);
4943 c->n_log_extra_fields = 0;
4946 void exec_context_revert_tty(ExecContext *c) {
4951 /* First, reset the TTY (possibly kicking everybody else from the TTY) */
4952 exec_context_tty_reset(c, NULL);
4954 /* And then undo what chown_terminal() did earlier. Note that we only do this if we have a path
4955 * configured. If the TTY was passed to us as file descriptor we assume the TTY is opened and managed
4956 * by whoever passed it to us and thus knows better when and how to chmod()/chown() it back. */
4958 if (exec_context_may_touch_tty(c)) {
4961 path = exec_context_tty_path(c);
4963 r = chmod_and_chown(path, TTY_MODE, 0, TTY_GID);
4964 if (r < 0 && r != -ENOENT)
4965 log_warning_errno(r, "Failed to reset TTY ownership/access mode of %s, ignoring: %m", path);
4970 int exec_context_get_clean_directories(
4976 _cleanup_strv_free_ char **l = NULL;
4977 ExecDirectoryType t;
4984 for (t = 0; t < _EXEC_DIRECTORY_TYPE_MAX; t++) {
4987 if (!FLAGS_SET(mask, 1U << t))
4993 STRV_FOREACH(i, c->directories[t].paths) {
4996 j = path_join(prefix[t], *i);
5000 r = strv_consume(&l, j);
5004 /* Also remove private directories unconditionally. */
5005 if (t != EXEC_DIRECTORY_CONFIGURATION) {
5006 j = path_join(prefix[t], "private", *i);
5010 r = strv_consume(&l, j);
5021 int exec_context_get_clean_mask(ExecContext *c, ExecCleanMask *ret) {
5022 ExecCleanMask mask = 0;
5027 for (ExecDirectoryType t = 0; t < _EXEC_DIRECTORY_TYPE_MAX; t++)
5028 if (!strv_isempty(c->directories[t].paths))
5035 void exec_status_start(ExecStatus *s, pid_t pid) {
5042 dual_timestamp_get(&s->start_timestamp);
5045 void exec_status_exit(ExecStatus *s, const ExecContext *context, pid_t pid, int code, int status) {
5048 if (s->pid != pid) {
5054 dual_timestamp_get(&s->exit_timestamp);
5059 if (context && context->utmp_id)
5060 (void) utmp_put_dead_process(context->utmp_id, pid, code, status);
5063 void exec_status_reset(ExecStatus *s) {
5066 *s = (ExecStatus) {};
5069 void exec_status_dump(const ExecStatus *s, FILE *f, const char *prefix) {
5070 char buf[FORMAT_TIMESTAMP_MAX];
5078 prefix = strempty(prefix);
5081 "%sPID: "PID_FMT"\n",
5084 if (dual_timestamp_is_set(&s->start_timestamp))
5086 "%sStart Timestamp: %s\n",
5087 prefix, format_timestamp(buf, sizeof(buf), s->start_timestamp.realtime));
5089 if (dual_timestamp_is_set(&s->exit_timestamp))
5091 "%sExit Timestamp: %s\n"
5093 "%sExit Status: %i\n",
5094 prefix, format_timestamp(buf, sizeof(buf), s->exit_timestamp.realtime),
5095 prefix, sigchld_code_to_string(s->code),
5099 static char *exec_command_line(char **argv) {
5107 STRV_FOREACH(a, argv)
5115 STRV_FOREACH(a, argv) {
5122 if (strpbrk(*a, WHITESPACE)) {
5133 /* FIXME: this doesn't really handle arguments that have
5134 * spaces and ticks in them */
5139 static void exec_command_dump(ExecCommand *c, FILE *f, const char *prefix) {
5140 _cleanup_free_ char *cmd = NULL;
5141 const char *prefix2;
5146 prefix = strempty(prefix);
5147 prefix2 = strjoina(prefix, "\t");
5149 cmd = exec_command_line(c->argv);
5151 "%sCommand Line: %s\n",
5152 prefix, cmd ? cmd : strerror_safe(ENOMEM));
5154 exec_status_dump(&c->exec_status, f, prefix2);
5157 void exec_command_dump_list(ExecCommand *c, FILE *f, const char *prefix) {
5160 prefix = strempty(prefix);
5162 LIST_FOREACH(command, c, c)
5163 exec_command_dump(c, f, prefix);
5166 void exec_command_append_list(ExecCommand **l, ExecCommand *e) {
5173 /* It's kind of important, that we keep the order here */
5174 LIST_FIND_TAIL(command, *l, end);
5175 LIST_INSERT_AFTER(command, *l, end, e);
5180 int exec_command_set(ExecCommand *c, const char *path, ...) {
5188 l = strv_new_ap(path, ap);
5200 free_and_replace(c->path, p);
5202 return strv_free_and_replace(c->argv, l);
5205 int exec_command_append(ExecCommand *c, const char *path, ...) {
5206 _cleanup_strv_free_ char **l = NULL;
5214 l = strv_new_ap(path, ap);
5220 r = strv_extend_strv(&c->argv, l, false);
5227 static void *remove_tmpdir_thread(void *p) {
5228 _cleanup_free_ char *path = p;
5230 (void) rm_rf(path, REMOVE_ROOT|REMOVE_PHYSICAL);
5234 static ExecRuntime* exec_runtime_free(ExecRuntime *rt, bool destroy) {
5241 (void) hashmap_remove(rt->manager->exec_runtime_by_id, rt->id);
5243 /* When destroy is true, then rm_rf tmp_dir and var_tmp_dir. */
5244 if (destroy && rt->tmp_dir) {
5245 log_debug("Spawning thread to nuke %s", rt->tmp_dir);
5247 r = asynchronous_job(remove_tmpdir_thread, rt->tmp_dir);
5249 log_warning_errno(r, "Failed to nuke %s: %m", rt->tmp_dir);
5256 if (destroy && rt->var_tmp_dir) {
5257 log_debug("Spawning thread to nuke %s", rt->var_tmp_dir);
5259 r = asynchronous_job(remove_tmpdir_thread, rt->var_tmp_dir);
5261 log_warning_errno(r, "Failed to nuke %s: %m", rt->var_tmp_dir);
5262 free(rt->var_tmp_dir);
5265 rt->var_tmp_dir = NULL;
5268 rt->id = mfree(rt->id);
5269 rt->tmp_dir = mfree(rt->tmp_dir);
5270 rt->var_tmp_dir = mfree(rt->var_tmp_dir);
5271 safe_close_pair(rt->netns_storage_socket);
5275 static void exec_runtime_freep(ExecRuntime **rt) {
5276 (void) exec_runtime_free(*rt, false);
5279 static int exec_runtime_allocate(ExecRuntime **ret) {
5284 n = new(ExecRuntime, 1);
5288 *n = (ExecRuntime) {
5289 .netns_storage_socket = { -1, -1 },
5296 static int exec_runtime_add(
5299 const char *tmp_dir,
5300 const char *var_tmp_dir,
5301 const int netns_storage_socket[2],
5302 ExecRuntime **ret) {
5304 _cleanup_(exec_runtime_freep) ExecRuntime *rt = NULL;
5310 r = hashmap_ensure_allocated(&m->exec_runtime_by_id, &string_hash_ops);
5314 r = exec_runtime_allocate(&rt);
5318 rt->id = strdup(id);
5323 rt->tmp_dir = strdup(tmp_dir);
5327 /* When tmp_dir is set, then we require var_tmp_dir is also set. */
5328 assert(var_tmp_dir);
5329 rt->var_tmp_dir = strdup(var_tmp_dir);
5330 if (!rt->var_tmp_dir)
5334 if (netns_storage_socket) {
5335 rt->netns_storage_socket[0] = netns_storage_socket[0];
5336 rt->netns_storage_socket[1] = netns_storage_socket[1];
5339 r = hashmap_put(m->exec_runtime_by_id, rt->id, rt);
5348 /* do not remove created ExecRuntime object when the operation succeeds. */
5353 static int exec_runtime_make(Manager *m, const ExecContext *c, const char *id, ExecRuntime **ret) {
5354 _cleanup_free_ char *tmp_dir = NULL, *var_tmp_dir = NULL;
5355 _cleanup_close_pair_ int netns_storage_socket[2] = { -1, -1 };
5362 /* It is not necessary to create ExecRuntime object. */
5363 if (!c->private_network && !c->private_tmp && !c->network_namespace_path)
5366 if (c->private_tmp) {
5367 r = setup_tmp_dirs(id, &tmp_dir, &var_tmp_dir);
5372 if (c->private_network || c->network_namespace_path) {
5373 if (socketpair(AF_UNIX, SOCK_DGRAM|SOCK_CLOEXEC, 0, netns_storage_socket) < 0)
5377 r = exec_runtime_add(m, id, tmp_dir, var_tmp_dir, netns_storage_socket, ret);
5382 netns_storage_socket[0] = netns_storage_socket[1] = -1;
5386 int exec_runtime_acquire(Manager *m, const ExecContext *c, const char *id, bool create, ExecRuntime **ret) {
5394 rt = hashmap_get(m->exec_runtime_by_id, id);
5396 /* We already have a ExecRuntime object, let's increase the ref count and reuse it */
5402 /* If not found, then create a new object. */
5403 r = exec_runtime_make(m, c, id, &rt);
5405 /* When r == 0, it is not necessary to create ExecRuntime object. */
5409 /* increment reference counter. */
5415 ExecRuntime *exec_runtime_unref(ExecRuntime *rt, bool destroy) {
5419 assert(rt->n_ref > 0);
5425 return exec_runtime_free(rt, destroy);
5428 int exec_runtime_serialize(const Manager *m, FILE *f, FDSet *fds) {
5436 HASHMAP_FOREACH(rt, m->exec_runtime_by_id, i) {
5437 fprintf(f, "exec-runtime=%s", rt->id);
5440 fprintf(f, " tmp-dir=%s", rt->tmp_dir);
5442 if (rt->var_tmp_dir)
5443 fprintf(f, " var-tmp-dir=%s", rt->var_tmp_dir);
5445 if (rt->netns_storage_socket[0] >= 0) {
5448 copy = fdset_put_dup(fds, rt->netns_storage_socket[0]);
5452 fprintf(f, " netns-socket-0=%i", copy);
5455 if (rt->netns_storage_socket[1] >= 0) {
5458 copy = fdset_put_dup(fds, rt->netns_storage_socket[1]);
5462 fprintf(f, " netns-socket-1=%i", copy);
5471 int exec_runtime_deserialize_compat(Unit *u, const char *key, const char *value, FDSet *fds) {
5472 _cleanup_(exec_runtime_freep) ExecRuntime *rt_create = NULL;
5476 /* This is for the migration from old (v237 or earlier) deserialization text.
5477 * Due to the bug #7790, this may not work with the units that use JoinsNamespaceOf=.
5478 * Even if the ExecRuntime object originally created by the other unit, we cannot judge
5479 * so or not from the serialized text, then we always creates a new object owned by this. */
5485 /* Manager manages ExecRuntime objects by the unit id.
5486 * So, we omit the serialized text when the unit does not have id (yet?)... */
5487 if (isempty(u->id)) {
5488 log_unit_debug(u, "Invocation ID not found. Dropping runtime parameter.");
5492 r = hashmap_ensure_allocated(&u->manager->exec_runtime_by_id, &string_hash_ops);
5494 log_unit_debug_errno(u, r, "Failed to allocate storage for runtime parameter: %m");
5498 rt = hashmap_get(u->manager->exec_runtime_by_id, u->id);
5500 r = exec_runtime_allocate(&rt_create);
5504 rt_create->id = strdup(u->id);
5511 if (streq(key, "tmp-dir")) {
5514 copy = strdup(value);
5518 free_and_replace(rt->tmp_dir, copy);
5520 } else if (streq(key, "var-tmp-dir")) {
5523 copy = strdup(value);
5527 free_and_replace(rt->var_tmp_dir, copy);
5529 } else if (streq(key, "netns-socket-0")) {
5532 if (safe_atoi(value, &fd) < 0 || !fdset_contains(fds, fd)) {
5533 log_unit_debug(u, "Failed to parse netns socket value: %s", value);
5537 safe_close(rt->netns_storage_socket[0]);
5538 rt->netns_storage_socket[0] = fdset_remove(fds, fd);
5540 } else if (streq(key, "netns-socket-1")) {
5543 if (safe_atoi(value, &fd) < 0 || !fdset_contains(fds, fd)) {
5544 log_unit_debug(u, "Failed to parse netns socket value: %s", value);
5548 safe_close(rt->netns_storage_socket[1]);
5549 rt->netns_storage_socket[1] = fdset_remove(fds, fd);
5553 /* If the object is newly created, then put it to the hashmap which manages ExecRuntime objects. */
5555 r = hashmap_put(u->manager->exec_runtime_by_id, rt_create->id, rt_create);
5557 log_unit_debug_errno(u, r, "Failed to put runtime parameter to manager's storage: %m");
5561 rt_create->manager = u->manager;
5570 void exec_runtime_deserialize_one(Manager *m, const char *value, FDSet *fds) {
5571 char *id = NULL, *tmp_dir = NULL, *var_tmp_dir = NULL;
5572 int r, fd0 = -1, fd1 = -1;
5573 const char *p, *v = value;
5580 n = strcspn(v, " ");
5581 id = strndupa(v, n);
5586 v = startswith(p, "tmp-dir=");
5588 n = strcspn(v, " ");
5589 tmp_dir = strndupa(v, n);
5595 v = startswith(p, "var-tmp-dir=");
5597 n = strcspn(v, " ");
5598 var_tmp_dir = strndupa(v, n);
5604 v = startswith(p, "netns-socket-0=");
5608 n = strcspn(v, " ");
5609 buf = strndupa(v, n);
5610 if (safe_atoi(buf, &fd0) < 0 || !fdset_contains(fds, fd0)) {
5611 log_debug("Unable to process exec-runtime netns fd specification.");
5614 fd0 = fdset_remove(fds, fd0);
5620 v = startswith(p, "netns-socket-1=");
5624 n = strcspn(v, " ");
5625 buf = strndupa(v, n);
5626 if (safe_atoi(buf, &fd1) < 0 || !fdset_contains(fds, fd1)) {
5627 log_debug("Unable to process exec-runtime netns fd specification.");
5630 fd1 = fdset_remove(fds, fd1);
5635 r = exec_runtime_add(m, id, tmp_dir, var_tmp_dir, (int[]) { fd0, fd1 }, NULL);
5637 log_debug_errno(r, "Failed to add exec-runtime: %m");
5640 void exec_runtime_vacuum(Manager *m) {
5646 /* Free unreferenced ExecRuntime objects. This is used after manager deserialization process. */
5648 HASHMAP_FOREACH(rt, m->exec_runtime_by_id, i) {
5652 (void) exec_runtime_free(rt, false);
5656 void exec_params_clear(ExecParameters *p) {
5660 strv_free(p->environment);
5663 static const char* const exec_input_table[_EXEC_INPUT_MAX] = {
5664 [EXEC_INPUT_NULL] = "null",
5665 [EXEC_INPUT_TTY] = "tty",
5666 [EXEC_INPUT_TTY_FORCE] = "tty-force",
5667 [EXEC_INPUT_TTY_FAIL] = "tty-fail",
5668 [EXEC_INPUT_SOCKET] = "socket",
5669 [EXEC_INPUT_NAMED_FD] = "fd",
5670 [EXEC_INPUT_DATA] = "data",
5671 [EXEC_INPUT_FILE] = "file",
5674 DEFINE_STRING_TABLE_LOOKUP(exec_input, ExecInput);
5676 static const char* const exec_output_table[_EXEC_OUTPUT_MAX] = {
5677 [EXEC_OUTPUT_INHERIT] = "inherit",
5678 [EXEC_OUTPUT_NULL] = "null",
5679 [EXEC_OUTPUT_TTY] = "tty",
5680 [EXEC_OUTPUT_SYSLOG] = "syslog",
5681 [EXEC_OUTPUT_SYSLOG_AND_CONSOLE] = "syslog+console",
5682 [EXEC_OUTPUT_KMSG] = "kmsg",
5683 [EXEC_OUTPUT_KMSG_AND_CONSOLE] = "kmsg+console",
5684 [EXEC_OUTPUT_JOURNAL] = "journal",
5685 [EXEC_OUTPUT_JOURNAL_AND_CONSOLE] = "journal+console",
5686 [EXEC_OUTPUT_SOCKET] = "socket",
5687 [EXEC_OUTPUT_NAMED_FD] = "fd",
5688 [EXEC_OUTPUT_FILE] = "file",
5689 [EXEC_OUTPUT_FILE_APPEND] = "append",
5690 [EXEC_OUTPUT_DLOG] = "dlog",
5691 [EXEC_OUTPUT_DLOG_OR_NULL] = "dlog|null",
5692 [EXEC_OUTPUT_DLOG_OR_KMSG] = "dlog|kmsg",
5693 [EXEC_OUTPUT_DLOG_OR_JOURNAL] = "dlog|journal",
5696 DEFINE_STRING_TABLE_LOOKUP(exec_output, ExecOutput);
5698 static const char* const exec_utmp_mode_table[_EXEC_UTMP_MODE_MAX] = {
5699 [EXEC_UTMP_INIT] = "init",
5700 [EXEC_UTMP_LOGIN] = "login",
5701 [EXEC_UTMP_USER] = "user",
5704 DEFINE_STRING_TABLE_LOOKUP(exec_utmp_mode, ExecUtmpMode);
5706 static const char* const exec_preserve_mode_table[_EXEC_PRESERVE_MODE_MAX] = {
5707 [EXEC_PRESERVE_NO] = "no",
5708 [EXEC_PRESERVE_YES] = "yes",
5709 [EXEC_PRESERVE_RESTART] = "restart",
5712 DEFINE_STRING_TABLE_LOOKUP_WITH_BOOLEAN(exec_preserve_mode, ExecPreserveMode, EXEC_PRESERVE_YES);
5714 /* This table maps ExecDirectoryType to the setting it is configured with in the unit */
5715 static const char* const exec_directory_type_table[_EXEC_DIRECTORY_TYPE_MAX] = {
5716 [EXEC_DIRECTORY_RUNTIME] = "RuntimeDirectory",
5717 [EXEC_DIRECTORY_STATE] = "StateDirectory",
5718 [EXEC_DIRECTORY_CACHE] = "CacheDirectory",
5719 [EXEC_DIRECTORY_LOGS] = "LogsDirectory",
5720 [EXEC_DIRECTORY_CONFIGURATION] = "ConfigurationDirectory",
5723 DEFINE_STRING_TABLE_LOOKUP(exec_directory_type, ExecDirectoryType);
5725 /* And this table maps ExecDirectoryType too, but to a generic term identifying the type of resource. This
5726 * one is supposed to be generic enough to be used for unit types that don't use ExecContext and per-unit
5727 * directories, specifically .timer units with their timestamp touch file. */
5728 static const char* const exec_resource_type_table[_EXEC_DIRECTORY_TYPE_MAX] = {
5729 [EXEC_DIRECTORY_RUNTIME] = "runtime",
5730 [EXEC_DIRECTORY_STATE] = "state",
5731 [EXEC_DIRECTORY_CACHE] = "cache",
5732 [EXEC_DIRECTORY_LOGS] = "logs",
5733 [EXEC_DIRECTORY_CONFIGURATION] = "configuration",
5736 DEFINE_STRING_TABLE_LOOKUP(exec_resource_type, ExecDirectoryType);
5738 /* And this table also maps ExecDirectoryType, to the environment variable we pass the selected directory to
5739 * the service payload in. */
5740 static const char* const exec_directory_env_name_table[_EXEC_DIRECTORY_TYPE_MAX] = {
5741 [EXEC_DIRECTORY_RUNTIME] = "RUNTIME_DIRECTORY",
5742 [EXEC_DIRECTORY_STATE] = "STATE_DIRECTORY",
5743 [EXEC_DIRECTORY_CACHE] = "CACHE_DIRECTORY",
5744 [EXEC_DIRECTORY_LOGS] = "LOGS_DIRECTORY",
5745 [EXEC_DIRECTORY_CONFIGURATION] = "CONFIGURATION_DIRECTORY",
5748 DEFINE_PRIVATE_STRING_TABLE_LOOKUP_TO_STRING(exec_directory_env_name, ExecDirectoryType);
5750 static const char* const exec_keyring_mode_table[_EXEC_KEYRING_MODE_MAX] = {
5751 [EXEC_KEYRING_INHERIT] = "inherit",
5752 [EXEC_KEYRING_PRIVATE] = "private",
5753 [EXEC_KEYRING_SHARED] = "shared",
5756 DEFINE_STRING_TABLE_LOOKUP(exec_keyring_mode, ExecKeyringMode);