1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
3 #include <sys/eventfd.h>
9 #include <security/pam_appl.h>
10 #include <security/pam_misc.h>
14 #include <sys/apparmor.h>
17 #include "sd-messages.h"
20 #include "apparmor-util.h"
22 #include "argv-util.h"
24 #include "bpf-dlopen.h"
26 #include "btrfs-util.h"
27 #include "capability-util.h"
28 #include "cgroup-setup.h"
30 #include "chattr-util.h"
31 #include "chown-recursive.h"
33 #include "data-fd-util.h"
36 #include "exec-credential.h"
37 #include "exec-invoke.h"
39 #include "exit-status.h"
41 #include "hexdecoct.h"
43 #include "iovec-util.h"
44 #include "missing_ioprio.h"
45 #include "missing_prctl.h"
46 #include "missing_securebits.h"
47 #include "missing_syscall.h"
48 #include "mkdir-label.h"
49 #include "proc-cmdline.h"
50 #include "process-util.h"
52 #include "rlimit-util.h"
53 #include "seccomp-util.h"
54 #include "selinux-util.h"
55 #include "signal-util.h"
56 #include "smack-util.h"
57 #include "socket-util.h"
58 #include "string-table.h"
60 #include "terminal-util.h"
61 #include "utmp-wtmp.h"
63 #define IDLE_TIMEOUT_USEC (5*USEC_PER_SEC)
64 #define IDLE_TIMEOUT2_USEC (1*USEC_PER_SEC)
66 #define SNDBUF_SIZE (8*1024*1024)
68 static int shift_fds(int fds
[], size_t n_fds
) {
72 /* Modifies the fds array! (sorts it) */
76 for (int start
= 0;;) {
77 int restart_from
= -1;
79 for (int i
= start
; i
< (int) n_fds
; i
++) {
82 /* Already at right index? */
86 nfd
= fcntl(fds
[i
], F_DUPFD
, i
+ 3);
93 /* Hmm, the fd we wanted isn't free? Then
94 * let's remember that and try again from here */
95 if (nfd
!= i
+3 && restart_from
< 0)
102 start
= restart_from
;
116 assert(fds
|| n_fds
== 0);
118 /* Drops/Sets O_NONBLOCK and FD_CLOEXEC from the file flags.
119 * O_NONBLOCK only applies to socket activation though. */
121 for (size_t i
= 0; i
< n_fds
; i
++) {
123 if (i
< n_socket_fds
) {
124 r
= fd_nonblock(fds
[i
], nonblock
);
129 /* We unconditionally drop FD_CLOEXEC from the fds,
130 * since after all we want to pass these fds to our
133 r
= fd_cloexec(fds
[i
], false);
141 static bool is_terminal_input(ExecInput i
) {
144 EXEC_INPUT_TTY_FORCE
,
145 EXEC_INPUT_TTY_FAIL
);
148 static bool is_terminal_output(ExecOutput o
) {
151 EXEC_OUTPUT_KMSG_AND_CONSOLE
,
152 EXEC_OUTPUT_JOURNAL_AND_CONSOLE
);
155 static bool is_kmsg_output(ExecOutput o
) {
158 EXEC_OUTPUT_KMSG_AND_CONSOLE
);
161 static bool exec_context_needs_term(const ExecContext
*c
) {
164 /* Return true if the execution context suggests we should set $TERM to something useful. */
166 if (is_terminal_input(c
->std_input
))
169 if (is_terminal_output(c
->std_output
))
172 if (is_terminal_output(c
->std_error
))
175 return !!c
->tty_path
;
178 static int open_null_as(int flags
, int nfd
) {
183 fd
= open("/dev/null", flags
|O_NOCTTY
);
187 return move_fd(fd
, nfd
, false);
190 static int connect_journal_socket(
192 const char *log_namespace
,
196 uid_t olduid
= UID_INVALID
;
197 gid_t oldgid
= GID_INVALID
;
202 strjoina("/run/systemd/journal.", log_namespace
, "/stdout") :
203 "/run/systemd/journal/stdout";
205 if (gid_is_valid(gid
)) {
208 if (setegid(gid
) < 0)
212 if (uid_is_valid(uid
)) {
215 if (seteuid(uid
) < 0) {
221 r
= connect_unix_path(fd
, AT_FDCWD
, j
);
223 /* If we fail to restore the uid or gid, things will likely fail later on. This should only happen if
224 an LSM interferes. */
226 if (uid_is_valid(uid
))
227 (void) seteuid(olduid
);
230 if (gid_is_valid(gid
))
231 (void) setegid(oldgid
);
236 static int connect_logger_as(
237 const ExecContext
*context
,
238 const ExecParameters
*params
,
245 _cleanup_close_
int fd
= -EBADF
;
250 assert(output
< _EXEC_OUTPUT_MAX
);
254 fd
= socket(AF_UNIX
, SOCK_STREAM
, 0);
258 r
= connect_journal_socket(fd
, context
->log_namespace
, uid
, gid
);
262 if (shutdown(fd
, SHUT_RD
) < 0)
265 (void) fd_inc_sndbuf(fd
, SNDBUF_SIZE
);
275 context
->syslog_identifier
?: ident
,
276 params
->flags
& EXEC_PASS_LOG_UNIT
? params
->unit_id
: "",
277 context
->syslog_priority
,
278 !!context
->syslog_level_prefix
,
280 is_kmsg_output(output
),
281 is_terminal_output(output
)) < 0)
284 return move_fd(TAKE_FD(fd
), nfd
, false);
287 static int open_terminal_as(const char *path
, int flags
, int nfd
) {
293 fd
= open_terminal(path
, flags
| O_NOCTTY
);
297 return move_fd(fd
, nfd
, false);
300 static int acquire_path(const char *path
, int flags
, mode_t mode
) {
301 _cleanup_close_
int fd
= -EBADF
;
306 if (IN_SET(flags
& O_ACCMODE
, O_WRONLY
, O_RDWR
))
309 fd
= open(path
, flags
|O_NOCTTY
, mode
);
313 if (errno
!= ENXIO
) /* ENXIO is returned when we try to open() an AF_UNIX file system socket on Linux */
316 /* So, it appears the specified path could be an AF_UNIX socket. Let's see if we can connect to it. */
318 fd
= socket(AF_UNIX
, SOCK_STREAM
, 0);
322 r
= connect_unix_path(fd
, AT_FDCWD
, path
);
323 if (IN_SET(r
, -ENOTSOCK
, -EINVAL
))
324 /* Propagate initial error if we get ENOTSOCK or EINVAL, i.e. we have indication that this
325 * wasn't an AF_UNIX socket after all */
330 if ((flags
& O_ACCMODE
) == O_RDONLY
)
331 r
= shutdown(fd
, SHUT_WR
);
332 else if ((flags
& O_ACCMODE
) == O_WRONLY
)
333 r
= shutdown(fd
, SHUT_RD
);
342 static int fixup_input(
343 const ExecContext
*context
,
345 bool apply_tty_stdin
) {
351 std_input
= context
->std_input
;
353 if (is_terminal_input(std_input
) && !apply_tty_stdin
)
354 return EXEC_INPUT_NULL
;
356 if (std_input
== EXEC_INPUT_SOCKET
&& socket_fd
< 0)
357 return EXEC_INPUT_NULL
;
359 if (std_input
== EXEC_INPUT_DATA
&& context
->stdin_data_size
== 0)
360 return EXEC_INPUT_NULL
;
365 static int fixup_output(ExecOutput output
, int socket_fd
) {
367 if (output
== EXEC_OUTPUT_SOCKET
&& socket_fd
< 0)
368 return EXEC_OUTPUT_INHERIT
;
373 static int setup_input(
374 const ExecContext
*context
,
375 const ExecParameters
*params
,
377 const int named_iofds
[static 3]) {
386 if (params
->stdin_fd
>= 0) {
387 if (dup2(params
->stdin_fd
, STDIN_FILENO
) < 0)
390 /* Try to make this the controlling tty, if it is a tty, and reset it */
391 if (isatty(STDIN_FILENO
)) {
392 (void) ioctl(STDIN_FILENO
, TIOCSCTTY
, context
->std_input
== EXEC_INPUT_TTY_FORCE
);
394 if (context
->tty_reset
)
395 (void) reset_terminal_fd(STDIN_FILENO
, /* switch_to_text= */ true);
397 (void) exec_context_apply_tty_size(context
, STDIN_FILENO
, /* tty_path= */ NULL
);
403 i
= fixup_input(context
, socket_fd
, params
->flags
& EXEC_APPLY_TTY_STDIN
);
407 case EXEC_INPUT_NULL
:
408 return open_null_as(O_RDONLY
, STDIN_FILENO
);
411 case EXEC_INPUT_TTY_FORCE
:
412 case EXEC_INPUT_TTY_FAIL
: {
413 _cleanup_close_
int tty_fd
= -EBADF
;
414 const char *tty_path
;
416 tty_path
= ASSERT_PTR(exec_context_tty_path(context
));
418 tty_fd
= acquire_terminal(tty_path
,
419 i
== EXEC_INPUT_TTY_FAIL
? ACQUIRE_TERMINAL_TRY
:
420 i
== EXEC_INPUT_TTY_FORCE
? ACQUIRE_TERMINAL_FORCE
:
421 ACQUIRE_TERMINAL_WAIT
,
426 r
= exec_context_apply_tty_size(context
, tty_fd
, tty_path
);
430 r
= move_fd(tty_fd
, STDIN_FILENO
, /* cloexec= */ false);
438 case EXEC_INPUT_SOCKET
:
439 assert(socket_fd
>= 0);
441 return RET_NERRNO(dup2(socket_fd
, STDIN_FILENO
));
443 case EXEC_INPUT_NAMED_FD
:
444 assert(named_iofds
[STDIN_FILENO
] >= 0);
446 (void) fd_nonblock(named_iofds
[STDIN_FILENO
], false);
447 return RET_NERRNO(dup2(named_iofds
[STDIN_FILENO
], STDIN_FILENO
));
449 case EXEC_INPUT_DATA
: {
452 fd
= acquire_data_fd(context
->stdin_data
, context
->stdin_data_size
, 0);
456 return move_fd(fd
, STDIN_FILENO
, false);
459 case EXEC_INPUT_FILE
: {
463 assert(context
->stdio_file
[STDIN_FILENO
]);
465 rw
= (context
->std_output
== EXEC_OUTPUT_FILE
&& streq_ptr(context
->stdio_file
[STDIN_FILENO
], context
->stdio_file
[STDOUT_FILENO
])) ||
466 (context
->std_error
== EXEC_OUTPUT_FILE
&& streq_ptr(context
->stdio_file
[STDIN_FILENO
], context
->stdio_file
[STDERR_FILENO
]));
468 fd
= acquire_path(context
->stdio_file
[STDIN_FILENO
], rw
? O_RDWR
: O_RDONLY
, 0666 & ~context
->umask
);
472 return move_fd(fd
, STDIN_FILENO
, false);
476 assert_not_reached();
480 static bool can_inherit_stderr_from_stdout(
481 const ExecContext
*context
,
487 /* Returns true, if given the specified STDERR and STDOUT output we can directly dup() the stdout fd to the
490 if (e
== EXEC_OUTPUT_INHERIT
)
495 if (e
== EXEC_OUTPUT_NAMED_FD
)
496 return streq_ptr(context
->stdio_fdname
[STDOUT_FILENO
], context
->stdio_fdname
[STDERR_FILENO
]);
498 if (IN_SET(e
, EXEC_OUTPUT_FILE
, EXEC_OUTPUT_FILE_APPEND
, EXEC_OUTPUT_FILE_TRUNCATE
))
499 return streq_ptr(context
->stdio_file
[STDOUT_FILENO
], context
->stdio_file
[STDERR_FILENO
]);
504 static int setup_output(
505 const ExecContext
*context
,
506 const ExecParameters
*params
,
509 const int named_iofds
[static 3],
513 dev_t
*journal_stream_dev
,
514 ino_t
*journal_stream_ino
) {
523 assert(journal_stream_dev
);
524 assert(journal_stream_ino
);
526 if (fileno
== STDOUT_FILENO
&& params
->stdout_fd
>= 0) {
528 if (dup2(params
->stdout_fd
, STDOUT_FILENO
) < 0)
531 return STDOUT_FILENO
;
534 if (fileno
== STDERR_FILENO
&& params
->stderr_fd
>= 0) {
535 if (dup2(params
->stderr_fd
, STDERR_FILENO
) < 0)
538 return STDERR_FILENO
;
541 i
= fixup_input(context
, socket_fd
, params
->flags
& EXEC_APPLY_TTY_STDIN
);
542 o
= fixup_output(context
->std_output
, socket_fd
);
544 if (fileno
== STDERR_FILENO
) {
546 e
= fixup_output(context
->std_error
, socket_fd
);
548 /* This expects the input and output are already set up */
550 /* Don't change the stderr file descriptor if we inherit all
551 * the way and are not on a tty */
552 if (e
== EXEC_OUTPUT_INHERIT
&&
553 o
== EXEC_OUTPUT_INHERIT
&&
554 i
== EXEC_INPUT_NULL
&&
555 !is_terminal_input(context
->std_input
) &&
559 /* Duplicate from stdout if possible */
560 if (can_inherit_stderr_from_stdout(context
, o
, e
))
561 return RET_NERRNO(dup2(STDOUT_FILENO
, fileno
));
565 } else if (o
== EXEC_OUTPUT_INHERIT
) {
566 /* If input got downgraded, inherit the original value */
567 if (i
== EXEC_INPUT_NULL
&& is_terminal_input(context
->std_input
))
568 return open_terminal_as(exec_context_tty_path(context
), O_WRONLY
, fileno
);
570 /* If the input is connected to anything that's not a /dev/null or a data fd, inherit that... */
571 if (!IN_SET(i
, EXEC_INPUT_NULL
, EXEC_INPUT_DATA
))
572 return RET_NERRNO(dup2(STDIN_FILENO
, fileno
));
574 /* If we are not started from PID 1 we just inherit STDOUT from our parent process. */
578 /* We need to open /dev/null here anew, to get the right access mode. */
579 return open_null_as(O_WRONLY
, fileno
);
584 case EXEC_OUTPUT_NULL
:
585 return open_null_as(O_WRONLY
, fileno
);
587 case EXEC_OUTPUT_TTY
:
588 if (is_terminal_input(i
))
589 return RET_NERRNO(dup2(STDIN_FILENO
, fileno
));
591 /* We don't reset the terminal if this is just about output */
592 return open_terminal_as(exec_context_tty_path(context
), O_WRONLY
, fileno
);
594 case EXEC_OUTPUT_KMSG
:
595 case EXEC_OUTPUT_KMSG_AND_CONSOLE
:
596 case EXEC_OUTPUT_JOURNAL
:
597 case EXEC_OUTPUT_JOURNAL_AND_CONSOLE
:
598 r
= connect_logger_as(context
, params
, o
, ident
, fileno
, uid
, gid
);
600 log_exec_warning_errno(context
,
603 "Failed to connect %s to the journal socket, ignoring: %m",
604 fileno
== STDOUT_FILENO
? "stdout" : "stderr");
605 r
= open_null_as(O_WRONLY
, fileno
);
609 /* If we connected this fd to the journal via a stream, patch the device/inode into the passed
610 * parameters, but only then. This is useful so that we can set $JOURNAL_STREAM that permits
611 * services to detect whether they are connected to the journal or not.
613 * If both stdout and stderr are connected to a stream then let's make sure to store the data
614 * about STDERR as that's usually the best way to do logging. */
616 if (fstat(fileno
, &st
) >= 0 &&
617 (*journal_stream_ino
== 0 || fileno
== STDERR_FILENO
)) {
618 *journal_stream_dev
= st
.st_dev
;
619 *journal_stream_ino
= st
.st_ino
;
624 case EXEC_OUTPUT_SOCKET
:
625 assert(socket_fd
>= 0);
627 return RET_NERRNO(dup2(socket_fd
, fileno
));
629 case EXEC_OUTPUT_NAMED_FD
:
630 assert(named_iofds
[fileno
] >= 0);
632 (void) fd_nonblock(named_iofds
[fileno
], false);
633 return RET_NERRNO(dup2(named_iofds
[fileno
], fileno
));
635 case EXEC_OUTPUT_FILE
:
636 case EXEC_OUTPUT_FILE_APPEND
:
637 case EXEC_OUTPUT_FILE_TRUNCATE
: {
641 assert(context
->stdio_file
[fileno
]);
643 rw
= context
->std_input
== EXEC_INPUT_FILE
&&
644 streq_ptr(context
->stdio_file
[fileno
], context
->stdio_file
[STDIN_FILENO
]);
647 return RET_NERRNO(dup2(STDIN_FILENO
, fileno
));
650 if (o
== EXEC_OUTPUT_FILE_APPEND
)
652 else if (o
== EXEC_OUTPUT_FILE_TRUNCATE
)
655 fd
= acquire_path(context
->stdio_file
[fileno
], flags
, 0666 & ~context
->umask
);
659 return move_fd(fd
, fileno
, 0);
663 assert_not_reached();
667 static int chown_terminal(int fd
, uid_t uid
) {
672 /* Before we chown/chmod the TTY, let's ensure this is actually a tty */
673 if (!isatty_safe(fd
))
676 /* This might fail. What matters are the results. */
677 r
= fchmod_and_chown(fd
, TTY_MODE
, uid
, GID_INVALID
);
684 static int setup_confirm_stdio(
685 const ExecContext
*context
,
687 int *ret_saved_stdin
,
688 int *ret_saved_stdout
) {
690 _cleanup_close_
int fd
= -EBADF
, saved_stdin
= -EBADF
, saved_stdout
= -EBADF
;
693 assert(ret_saved_stdin
);
694 assert(ret_saved_stdout
);
696 saved_stdin
= fcntl(STDIN_FILENO
, F_DUPFD
, 3);
700 saved_stdout
= fcntl(STDOUT_FILENO
, F_DUPFD
, 3);
701 if (saved_stdout
< 0)
704 fd
= acquire_terminal(vc
, ACQUIRE_TERMINAL_WAIT
, DEFAULT_CONFIRM_USEC
);
708 r
= chown_terminal(fd
, getuid());
712 r
= reset_terminal_fd(fd
, /* switch_to_text= */ true);
716 r
= exec_context_apply_tty_size(context
, fd
, vc
);
720 r
= rearrange_stdio(fd
, fd
, STDERR_FILENO
); /* Invalidates 'fd' also on failure */
725 *ret_saved_stdin
= TAKE_FD(saved_stdin
);
726 *ret_saved_stdout
= TAKE_FD(saved_stdout
);
730 static void write_confirm_error_fd(int err
, int fd
, const char *unit_id
) {
734 if (err
== -ETIMEDOUT
)
735 dprintf(fd
, "Confirmation question timed out for %s, assuming positive response.\n", unit_id
);
738 dprintf(fd
, "Couldn't ask confirmation for %s: %m, assuming positive response.\n", unit_id
);
742 static void write_confirm_error(int err
, const char *vc
, const char *unit_id
) {
743 _cleanup_close_
int fd
= -EBADF
;
747 fd
= open_terminal(vc
, O_WRONLY
|O_NOCTTY
|O_CLOEXEC
);
751 write_confirm_error_fd(err
, fd
, unit_id
);
754 static int restore_confirm_stdio(int *saved_stdin
, int *saved_stdout
) {
758 assert(saved_stdout
);
762 if (*saved_stdin
>= 0)
763 if (dup2(*saved_stdin
, STDIN_FILENO
) < 0)
766 if (*saved_stdout
>= 0)
767 if (dup2(*saved_stdout
, STDOUT_FILENO
) < 0)
770 *saved_stdin
= safe_close(*saved_stdin
);
771 *saved_stdout
= safe_close(*saved_stdout
);
777 CONFIRM_PRETEND_FAILURE
= -1,
778 CONFIRM_PRETEND_SUCCESS
= 0,
782 static bool confirm_spawn_disabled(void) {
783 return access("/run/systemd/confirm_spawn_disabled", F_OK
) >= 0;
786 static int ask_for_confirmation(const ExecContext
*context
, const ExecParameters
*params
, const char *cmdline
) {
787 int saved_stdout
= -1, saved_stdin
= -1, r
;
788 _cleanup_free_
char *e
= NULL
;
794 /* For any internal errors, assume a positive response. */
795 r
= setup_confirm_stdio(context
, params
->confirm_spawn
, &saved_stdin
, &saved_stdout
);
797 write_confirm_error(r
, params
->confirm_spawn
, params
->unit_id
);
798 return CONFIRM_EXECUTE
;
801 /* confirm_spawn might have been disabled while we were sleeping. */
802 if (!params
->confirm_spawn
|| confirm_spawn_disabled()) {
807 e
= ellipsize(cmdline
, 60, 100);
815 r
= ask_char(&c
, "yfshiDjcn", "Execute %s? [y, f, s – h for help] ", e
);
817 write_confirm_error_fd(r
, STDOUT_FILENO
, params
->unit_id
);
824 printf("Resuming normal execution.\n");
825 manager_disable_confirm_spawn();
829 printf(" Unit: %s\n",
831 exec_context_dump(context
, stdout
, " ");
832 exec_params_dump(params
, stdout
, " ");
833 continue; /* ask again */
835 printf("Failing execution.\n");
836 r
= CONFIRM_PRETEND_FAILURE
;
839 printf(" c - continue, proceed without asking anymore\n"
840 " D - dump, show the state of the unit\n"
841 " f - fail, don't execute the command and pretend it failed\n"
843 " i - info, show a short summary of the unit\n"
844 " j - jobs, show jobs that are in progress\n"
845 " s - skip, don't execute the command and pretend it succeeded\n"
846 " y - yes, execute the command\n");
847 continue; /* ask again */
851 params
->unit_id
, cmdline
);
852 continue; /* ask again */
854 if (sigqueue(getppid(),
856 (const union sigval
) { .sival_int
= MANAGER_SIGNAL_COMMAND_DUMP_JOBS
}) < 0)
859 continue; /* ask again */
861 /* 'n' was removed in favor of 'f'. */
862 printf("Didn't understand 'n', did you mean 'f'?\n");
863 continue; /* ask again */
865 printf("Skipping execution.\n");
866 r
= CONFIRM_PRETEND_SUCCESS
;
872 assert_not_reached();
878 restore_confirm_stdio(&saved_stdin
, &saved_stdout
);
882 static int get_fixed_user(
883 const char *user_or_uid
,
884 const char **ret_username
,
887 const char **ret_home
,
888 const char **ret_shell
) {
893 assert(ret_username
);
895 /* Note that we don't set $HOME or $SHELL if they are not particularly enlightening anyway
896 * (i.e. are "/" or "/bin/nologin"). */
898 r
= get_user_creds(&user_or_uid
, ret_uid
, ret_gid
, ret_home
, ret_shell
, USER_CREDS_CLEAN
);
902 /* user_or_uid is normalized by get_user_creds to username */
903 *ret_username
= user_or_uid
;
908 static int get_fixed_group(
909 const char *group_or_gid
,
910 const char **ret_groupname
,
915 assert(group_or_gid
);
916 assert(ret_groupname
);
918 r
= get_group_creds(&group_or_gid
, ret_gid
, /* flags = */ 0);
922 /* group_or_gid is normalized by get_group_creds to groupname */
923 *ret_groupname
= group_or_gid
;
928 static int get_supplementary_groups(const ExecContext
*c
, const char *user
,
929 const char *group
, gid_t gid
,
930 gid_t
**supplementary_gids
, int *ngids
) {
933 bool keep_groups
= false;
934 gid_t
*groups
= NULL
;
935 _cleanup_free_ gid_t
*l_gids
= NULL
;
940 * If user is given, then lookup GID and supplementary groups list.
941 * We avoid NSS lookups for gid=0. Also we have to initialize groups
942 * here and as early as possible so we keep the list of supplementary
943 * groups of the caller.
945 if (user
&& gid_is_valid(gid
) && gid
!= 0) {
946 /* First step, initialize groups from /etc/groups */
947 if (initgroups(user
, gid
) < 0)
953 if (strv_isempty(c
->supplementary_groups
))
957 * If SupplementaryGroups= was passed then NGROUPS_MAX has to
958 * be positive, otherwise fail.
961 ngroups_max
= (int) sysconf(_SC_NGROUPS_MAX
);
962 if (ngroups_max
<= 0)
963 return errno_or_else(EOPNOTSUPP
);
965 l_gids
= new(gid_t
, ngroups_max
);
971 * Lookup the list of groups that the user belongs to, we
972 * avoid NSS lookups here too for gid=0.
975 if (getgrouplist(user
, gid
, l_gids
, &k
) < 0)
980 STRV_FOREACH(i
, c
->supplementary_groups
) {
983 if (k
>= ngroups_max
)
987 r
= get_group_creds(&g
, l_gids
+k
, 0);
995 * Sets ngids to zero to drop all supplementary groups, happens
996 * when we are under root and SupplementaryGroups= is empty.
1003 /* Otherwise get the final list of supplementary groups */
1004 groups
= memdup(l_gids
, sizeof(gid_t
) * k
);
1008 *supplementary_gids
= groups
;
1016 static int enforce_groups(gid_t gid
, const gid_t
*supplementary_gids
, int ngids
) {
1019 /* Handle SupplementaryGroups= if it is not empty */
1021 r
= maybe_setgroups(ngids
, supplementary_gids
);
1026 if (gid_is_valid(gid
)) {
1027 /* Then set our gids */
1028 if (setresgid(gid
, gid
, gid
) < 0)
1035 static int set_securebits(unsigned bits
, unsigned mask
) {
1039 current
= prctl(PR_GET_SECUREBITS
);
1043 /* Clear all securebits defined in mask and set bits */
1044 applied
= ((unsigned) current
& ~mask
) | bits
;
1045 if ((unsigned) current
== applied
)
1048 if (prctl(PR_SET_SECUREBITS
, applied
) < 0)
1054 static int enforce_user(
1055 const ExecContext
*context
,
1057 uint64_t capability_ambient_set
) {
1061 if (!uid_is_valid(uid
))
1064 /* Sets (but doesn't look up) the UIS and makes sure we keep the capabilities while doing so. For
1065 * setting secure bits the capability CAP_SETPCAP is required, so we also need keep-caps in this
1068 if ((capability_ambient_set
!= 0 || context
->secure_bits
!= 0) && uid
!= 0) {
1070 /* First step: If we need to keep capabilities but drop privileges we need to make sure we
1071 * keep our caps, while we drop privileges. Add KEEP_CAPS to the securebits */
1072 r
= set_securebits(1U << SECURE_KEEP_CAPS
, 0);
1077 /* Second step: actually set the uids */
1078 if (setresuid(uid
, uid
, uid
) < 0)
1081 /* At this point we should have all necessary capabilities but are otherwise a normal user. However,
1082 * the caps might got corrupted due to the setresuid() so we need clean them up later. This is done
1083 * outside of this call. */
1089 static int null_conv(
1091 const struct pam_message
**msg
,
1092 struct pam_response
**resp
,
1093 void *appdata_ptr
) {
1095 /* We don't support conversations */
1097 return PAM_CONV_ERR
;
1102 static int setup_pam(
1108 char ***env
, /* updated on success */
1109 const int fds
[], size_t n_fds
) {
1113 static const struct pam_conv conv
= {
1118 _cleanup_(barrier_destroy
) Barrier barrier
= BARRIER_NULL
;
1119 _cleanup_strv_free_
char **e
= NULL
;
1120 pam_handle_t
*handle
= NULL
;
1122 int pam_code
= PAM_SUCCESS
, r
;
1123 bool close_session
= false;
1124 pid_t pam_pid
= 0, parent_pid
;
1131 /* We set up PAM in the parent process, then fork. The child
1132 * will then stay around until killed via PR_GET_PDEATHSIG or
1133 * systemd via the cgroup logic. It will then remove the PAM
1134 * session again. The parent process will exec() the actual
1135 * daemon. We do things this way to ensure that the main PID
1136 * of the daemon is the one we initially fork()ed. */
1138 r
= barrier_create(&barrier
);
1142 if (log_get_max_level() < LOG_DEBUG
)
1143 flags
|= PAM_SILENT
;
1145 pam_code
= pam_start(name
, user
, &conv
, &handle
);
1146 if (pam_code
!= PAM_SUCCESS
) {
1152 _cleanup_free_
char *q
= NULL
;
1154 /* Hmm, so no TTY was explicitly passed, but an fd passed to us directly might be a TTY. Let's figure
1155 * out if that's the case, and read the TTY off it. */
1157 if (getttyname_malloc(STDIN_FILENO
, &q
) >= 0)
1158 tty
= strjoina("/dev/", q
);
1162 pam_code
= pam_set_item(handle
, PAM_TTY
, tty
);
1163 if (pam_code
!= PAM_SUCCESS
)
1167 STRV_FOREACH(nv
, *env
) {
1168 pam_code
= pam_putenv(handle
, *nv
);
1169 if (pam_code
!= PAM_SUCCESS
)
1173 pam_code
= pam_acct_mgmt(handle
, flags
);
1174 if (pam_code
!= PAM_SUCCESS
)
1177 pam_code
= pam_setcred(handle
, PAM_ESTABLISH_CRED
| flags
);
1178 if (pam_code
!= PAM_SUCCESS
)
1179 log_debug("pam_setcred() failed, ignoring: %s", pam_strerror(handle
, pam_code
));
1181 pam_code
= pam_open_session(handle
, flags
);
1182 if (pam_code
!= PAM_SUCCESS
)
1185 close_session
= true;
1187 e
= pam_getenvlist(handle
);
1189 pam_code
= PAM_BUF_ERR
;
1193 /* Block SIGTERM, so that we know that it won't get lost in the child */
1195 assert_se(sigprocmask_many(SIG_BLOCK
, &old_ss
, SIGTERM
, -1) >= 0);
1197 parent_pid
= getpid_cached();
1199 r
= safe_fork("(sd-pam)", 0, &pam_pid
);
1205 /* The child's job is to reset the PAM session on termination */
1206 barrier_set_role(&barrier
, BARRIER_CHILD
);
1208 /* Make sure we don't keep open the passed fds in this child. We assume that otherwise only
1209 * those fds are open here that have been opened by PAM. */
1210 (void) close_many(fds
, n_fds
);
1212 /* Drop privileges - we don't need any to pam_close_session and this will make
1213 * PR_SET_PDEATHSIG work in most cases. If this fails, ignore the error - but expect sd-pam
1214 * threads to fail to exit normally */
1216 r
= fully_set_uid_gid(uid
, gid
, /* supplementary_gids= */ NULL
, /* n_supplementary_gids= */ 0);
1218 log_warning_errno(r
, "Failed to drop privileges in sd-pam: %m");
1220 (void) ignore_signals(SIGPIPE
);
1222 /* Wait until our parent died. This will only work if the above setresuid() succeeds,
1223 * otherwise the kernel will not allow unprivileged parents kill their privileged children
1224 * this way. We rely on the control groups kill logic to do the rest for us. */
1225 if (prctl(PR_SET_PDEATHSIG
, SIGTERM
) < 0)
1228 /* Tell the parent that our setup is done. This is especially important regarding dropping
1229 * privileges. Otherwise, unit setup might race against our setresuid(2) call.
1231 * If the parent aborted, we'll detect this below, hence ignore return failure here. */
1232 (void) barrier_place(&barrier
);
1234 /* Check if our parent process might already have died? */
1235 if (getppid() == parent_pid
) {
1239 assert_se(sigemptyset(&ss
) >= 0);
1240 assert_se(sigaddset(&ss
, SIGTERM
) >= 0);
1242 assert_se(sigwait(&ss
, &sig
) == 0);
1243 assert(sig
== SIGTERM
);
1246 pam_code
= pam_setcred(handle
, PAM_DELETE_CRED
| flags
);
1247 if (pam_code
!= PAM_SUCCESS
)
1250 /* If our parent died we'll end the session */
1251 if (getppid() != parent_pid
) {
1252 pam_code
= pam_close_session(handle
, flags
);
1253 if (pam_code
!= PAM_SUCCESS
)
1260 /* NB: pam_end() when called in child processes should set PAM_DATA_SILENT to let the module
1261 * know about this. See pam_end(3) */
1262 (void) pam_end(handle
, pam_code
| flags
| PAM_DATA_SILENT
);
1266 barrier_set_role(&barrier
, BARRIER_PARENT
);
1268 /* If the child was forked off successfully it will do all the cleanups, so forget about the handle
1272 /* Unblock SIGTERM again in the parent */
1273 assert_se(sigprocmask(SIG_SETMASK
, &old_ss
, NULL
) >= 0);
1275 /* We close the log explicitly here, since the PAM modules might have opened it, but we don't want
1276 * this fd around. */
1279 /* Synchronously wait for the child to initialize. We don't care for errors as we cannot
1280 * recover. However, warn loudly if it happens. */
1281 if (!barrier_place_and_sync(&barrier
))
1282 log_error("PAM initialization failed");
1284 return strv_free_and_replace(*env
, e
);
1287 if (pam_code
!= PAM_SUCCESS
) {
1288 log_error("PAM failed: %s", pam_strerror(handle
, pam_code
));
1289 r
= -EPERM
; /* PAM errors do not map to errno */
1291 log_error_errno(r
, "PAM failed: %m");
1295 pam_code
= pam_close_session(handle
, flags
);
1297 (void) pam_end(handle
, pam_code
| flags
);
1307 static void rename_process_from_path(const char *path
) {
1308 _cleanup_free_
char *buf
= NULL
;
1313 /* This resulting string must fit in 10 chars (i.e. the length of "/sbin/init") to look pretty in
1316 if (path_extract_filename(path
, &buf
) < 0) {
1317 rename_process("(...)");
1321 size_t l
= strlen(buf
);
1323 /* The end of the process name is usually more interesting, since the first bit might just be
1330 char process_name
[11];
1331 process_name
[0] = '(';
1332 memcpy(process_name
+1, p
, l
);
1333 process_name
[1+l
] = ')';
1334 process_name
[1+l
+1] = 0;
1336 (void) rename_process(process_name
);
1339 static bool context_has_address_families(const ExecContext
*c
) {
1342 return c
->address_families_allow_list
||
1343 !set_isempty(c
->address_families
);
1346 static bool context_has_syscall_filters(const ExecContext
*c
) {
1349 return c
->syscall_allow_list
||
1350 !hashmap_isempty(c
->syscall_filter
);
1353 static bool context_has_syscall_logs(const ExecContext
*c
) {
1356 return c
->syscall_log_allow_list
||
1357 !hashmap_isempty(c
->syscall_log
);
1360 static bool context_has_seccomp(const ExecContext
*c
) {
1361 /* We need NNP if we have any form of seccomp and are unprivileged */
1362 return c
->lock_personality
||
1363 c
->memory_deny_write_execute
||
1364 c
->private_devices
||
1366 c
->protect_hostname
||
1367 c
->protect_kernel_tunables
||
1368 c
->protect_kernel_modules
||
1369 c
->protect_kernel_logs
||
1370 context_has_address_families(c
) ||
1371 exec_context_restrict_namespaces_set(c
) ||
1372 c
->restrict_realtime
||
1373 c
->restrict_suid_sgid
||
1374 !set_isempty(c
->syscall_archs
) ||
1375 context_has_syscall_filters(c
) ||
1376 context_has_syscall_logs(c
);
1379 static bool context_has_no_new_privileges(const ExecContext
*c
) {
1382 if (c
->no_new_privileges
)
1385 if (have_effective_cap(CAP_SYS_ADMIN
) > 0) /* if we are privileged, we don't need NNP */
1388 return context_has_seccomp(c
);
1393 static bool seccomp_allows_drop_privileges(const ExecContext
*c
) {
1395 bool has_capget
= false, has_capset
= false, has_prctl
= false;
1399 /* No syscall filter, we are allowed to drop privileges */
1400 if (hashmap_isempty(c
->syscall_filter
))
1403 HASHMAP_FOREACH_KEY(val
, id
, c
->syscall_filter
) {
1404 _cleanup_free_
char *name
= NULL
;
1406 name
= seccomp_syscall_resolve_num_arch(SCMP_ARCH_NATIVE
, PTR_TO_INT(id
) - 1);
1408 if (streq(name
, "capget"))
1410 else if (streq(name
, "capset"))
1412 else if (streq(name
, "prctl"))
1416 if (c
->syscall_allow_list
)
1417 return has_capget
&& has_capset
&& has_prctl
;
1419 return !(has_capget
|| has_capset
|| has_prctl
);
1422 static bool skip_seccomp_unavailable(const ExecContext
*c
, const ExecParameters
*p
, const char* msg
) {
1424 if (is_seccomp_available())
1427 log_exec_debug(c
, p
, "SECCOMP features not detected in the kernel, skipping %s", msg
);
1431 static int apply_syscall_filter(const ExecContext
*c
, const ExecParameters
*p
, bool needs_ambient_hack
) {
1432 uint32_t negative_action
, default_action
, action
;
1438 if (!context_has_syscall_filters(c
))
1441 if (skip_seccomp_unavailable(c
, p
, "SystemCallFilter="))
1444 negative_action
= c
->syscall_errno
== SECCOMP_ERROR_NUMBER_KILL
? scmp_act_kill_process() : SCMP_ACT_ERRNO(c
->syscall_errno
);
1446 if (c
->syscall_allow_list
) {
1447 default_action
= negative_action
;
1448 action
= SCMP_ACT_ALLOW
;
1450 default_action
= SCMP_ACT_ALLOW
;
1451 action
= negative_action
;
1454 if (needs_ambient_hack
) {
1455 r
= seccomp_filter_set_add(c
->syscall_filter
, c
->syscall_allow_list
, syscall_filter_sets
+ SYSCALL_FILTER_SET_SETUID
);
1460 return seccomp_load_syscall_filter_set_raw(default_action
, c
->syscall_filter
, action
, false);
1463 static int apply_syscall_log(const ExecContext
*c
, const ExecParameters
*p
) {
1465 uint32_t default_action
, action
;
1471 if (!context_has_syscall_logs(c
))
1475 if (skip_seccomp_unavailable(c
, p
, "SystemCallLog="))
1478 if (c
->syscall_log_allow_list
) {
1479 /* Log nothing but the ones listed */
1480 default_action
= SCMP_ACT_ALLOW
;
1481 action
= SCMP_ACT_LOG
;
1483 /* Log everything but the ones listed */
1484 default_action
= SCMP_ACT_LOG
;
1485 action
= SCMP_ACT_ALLOW
;
1488 return seccomp_load_syscall_filter_set_raw(default_action
, c
->syscall_log
, action
, false);
1490 /* old libseccomp */
1491 log_exec_debug(c
, p
, "SECCOMP feature SCMP_ACT_LOG not available, skipping SystemCallLog=");
1496 static int apply_syscall_archs(const ExecContext
*c
, const ExecParameters
*p
) {
1500 if (set_isempty(c
->syscall_archs
))
1503 if (skip_seccomp_unavailable(c
, p
, "SystemCallArchitectures="))
1506 return seccomp_restrict_archs(c
->syscall_archs
);
1509 static int apply_address_families(const ExecContext
*c
, const ExecParameters
*p
) {
1513 if (!context_has_address_families(c
))
1516 if (skip_seccomp_unavailable(c
, p
, "RestrictAddressFamilies="))
1519 return seccomp_restrict_address_families(c
->address_families
, c
->address_families_allow_list
);
1522 static int apply_memory_deny_write_execute(const ExecContext
*c
, const ExecParameters
*p
) {
1528 if (!c
->memory_deny_write_execute
)
1531 /* use prctl() if kernel supports it (6.3) */
1532 r
= prctl(PR_SET_MDWE
, PR_MDWE_REFUSE_EXEC_GAIN
, 0, 0, 0);
1534 log_exec_debug(c
, p
, "Enabled MemoryDenyWriteExecute= with PR_SET_MDWE");
1537 if (r
< 0 && errno
!= EINVAL
)
1538 return log_exec_debug_errno(c
,
1541 "Failed to enable MemoryDenyWriteExecute= with PR_SET_MDWE: %m");
1542 /* else use seccomp */
1543 log_exec_debug(c
, p
, "Kernel doesn't support PR_SET_MDWE: falling back to seccomp");
1545 if (skip_seccomp_unavailable(c
, p
, "MemoryDenyWriteExecute="))
1548 return seccomp_memory_deny_write_execute();
1551 static int apply_restrict_realtime(const ExecContext
*c
, const ExecParameters
*p
) {
1555 if (!c
->restrict_realtime
)
1558 if (skip_seccomp_unavailable(c
, p
, "RestrictRealtime="))
1561 return seccomp_restrict_realtime();
1564 static int apply_restrict_suid_sgid(const ExecContext
*c
, const ExecParameters
*p
) {
1568 if (!c
->restrict_suid_sgid
)
1571 if (skip_seccomp_unavailable(c
, p
, "RestrictSUIDSGID="))
1574 return seccomp_restrict_suid_sgid();
1577 static int apply_protect_sysctl(const ExecContext
*c
, const ExecParameters
*p
) {
1581 /* Turn off the legacy sysctl() system call. Many distributions turn this off while building the kernel, but
1582 * let's protect even those systems where this is left on in the kernel. */
1584 if (!c
->protect_kernel_tunables
)
1587 if (skip_seccomp_unavailable(c
, p
, "ProtectKernelTunables="))
1590 return seccomp_protect_sysctl();
1593 static int apply_protect_kernel_modules(const ExecContext
*c
, const ExecParameters
*p
) {
1597 /* Turn off module syscalls on ProtectKernelModules=yes */
1599 if (!c
->protect_kernel_modules
)
1602 if (skip_seccomp_unavailable(c
, p
, "ProtectKernelModules="))
1605 return seccomp_load_syscall_filter_set(SCMP_ACT_ALLOW
, syscall_filter_sets
+ SYSCALL_FILTER_SET_MODULE
, SCMP_ACT_ERRNO(EPERM
), false);
1608 static int apply_protect_kernel_logs(const ExecContext
*c
, const ExecParameters
*p
) {
1612 if (!c
->protect_kernel_logs
)
1615 if (skip_seccomp_unavailable(c
, p
, "ProtectKernelLogs="))
1618 return seccomp_protect_syslog();
1621 static int apply_protect_clock(const ExecContext
*c
, const ExecParameters
*p
) {
1625 if (!c
->protect_clock
)
1628 if (skip_seccomp_unavailable(c
, p
, "ProtectClock="))
1631 return seccomp_load_syscall_filter_set(SCMP_ACT_ALLOW
, syscall_filter_sets
+ SYSCALL_FILTER_SET_CLOCK
, SCMP_ACT_ERRNO(EPERM
), false);
1634 static int apply_private_devices(const ExecContext
*c
, const ExecParameters
*p
) {
1638 /* If PrivateDevices= is set, also turn off iopl and all @raw-io syscalls. */
1640 if (!c
->private_devices
)
1643 if (skip_seccomp_unavailable(c
, p
, "PrivateDevices="))
1646 return seccomp_load_syscall_filter_set(SCMP_ACT_ALLOW
, syscall_filter_sets
+ SYSCALL_FILTER_SET_RAW_IO
, SCMP_ACT_ERRNO(EPERM
), false);
1649 static int apply_restrict_namespaces(const ExecContext
*c
, const ExecParameters
*p
) {
1653 if (!exec_context_restrict_namespaces_set(c
))
1656 if (skip_seccomp_unavailable(c
, p
, "RestrictNamespaces="))
1659 return seccomp_restrict_namespaces(c
->restrict_namespaces
);
1662 static int apply_lock_personality(const ExecContext
*c
, const ExecParameters
*p
) {
1663 unsigned long personality
;
1669 if (!c
->lock_personality
)
1672 if (skip_seccomp_unavailable(c
, p
, "LockPersonality="))
1675 personality
= c
->personality
;
1677 /* If personality is not specified, use either PER_LINUX or PER_LINUX32 depending on what is currently set. */
1678 if (personality
== PERSONALITY_INVALID
) {
1680 r
= opinionated_personality(&personality
);
1685 return seccomp_lock_personality(personality
);
1691 static int apply_restrict_filesystems(const ExecContext
*c
, const ExecParameters
*p
) {
1697 if (!exec_context_restrict_filesystems_set(c
))
1700 if (p
->bpf_outer_map_fd
< 0) {
1701 /* LSM BPF is unsupported or lsm_bpf_setup failed */
1702 log_exec_debug(c
, p
, "LSM BPF not supported, skipping RestrictFileSystems=");
1706 /* We are in a new binary, so dl-open again */
1711 return lsm_bpf_restrict_filesystems(c
->restrict_filesystems
, p
->cgroup_id
, p
->bpf_outer_map_fd
, c
->restrict_filesystems_allow_list
);
1715 static int apply_protect_hostname(const ExecContext
*c
, const ExecParameters
*p
, int *ret_exit_status
) {
1719 if (!c
->protect_hostname
)
1722 if (ns_type_supported(NAMESPACE_UTS
)) {
1723 if (unshare(CLONE_NEWUTS
) < 0) {
1724 if (!ERRNO_IS_NOT_SUPPORTED(errno
) && !ERRNO_IS_PRIVILEGE(errno
)) {
1725 *ret_exit_status
= EXIT_NAMESPACE
;
1726 return log_exec_error_errno(c
,
1729 "Failed to set up UTS namespacing: %m");
1734 "ProtectHostname=yes is configured, but UTS namespace setup is "
1735 "prohibited (container manager?), ignoring namespace setup.");
1740 "ProtectHostname=yes is configured, but the kernel does not "
1741 "support UTS namespaces, ignoring namespace setup.");
1746 if (skip_seccomp_unavailable(c
, p
, "ProtectHostname="))
1749 r
= seccomp_protect_hostname();
1751 *ret_exit_status
= EXIT_SECCOMP
;
1752 return log_exec_error_errno(c
, p
, r
, "Failed to apply hostname restrictions: %m");
1759 static void do_idle_pipe_dance(int idle_pipe
[static 4]) {
1762 idle_pipe
[1] = safe_close(idle_pipe
[1]);
1763 idle_pipe
[2] = safe_close(idle_pipe
[2]);
1765 if (idle_pipe
[0] >= 0) {
1768 r
= fd_wait_for_event(idle_pipe
[0], POLLHUP
, IDLE_TIMEOUT_USEC
);
1770 if (idle_pipe
[3] >= 0 && r
== 0 /* timeout */) {
1773 /* Signal systemd that we are bored and want to continue. */
1774 n
= write(idle_pipe
[3], "x", 1);
1776 /* Wait for systemd to react to the signal above. */
1777 (void) fd_wait_for_event(idle_pipe
[0], POLLHUP
, IDLE_TIMEOUT2_USEC
);
1780 idle_pipe
[0] = safe_close(idle_pipe
[0]);
1784 idle_pipe
[3] = safe_close(idle_pipe
[3]);
1787 static const char *exec_directory_env_name_to_string(ExecDirectoryType t
);
1789 /* And this table also maps ExecDirectoryType, to the environment variable we pass the selected directory to
1790 * the service payload in. */
1791 static const char* const exec_directory_env_name_table
[_EXEC_DIRECTORY_TYPE_MAX
] = {
1792 [EXEC_DIRECTORY_RUNTIME
] = "RUNTIME_DIRECTORY",
1793 [EXEC_DIRECTORY_STATE
] = "STATE_DIRECTORY",
1794 [EXEC_DIRECTORY_CACHE
] = "CACHE_DIRECTORY",
1795 [EXEC_DIRECTORY_LOGS
] = "LOGS_DIRECTORY",
1796 [EXEC_DIRECTORY_CONFIGURATION
] = "CONFIGURATION_DIRECTORY",
1799 DEFINE_PRIVATE_STRING_TABLE_LOOKUP_TO_STRING(exec_directory_env_name
, ExecDirectoryType
);
1801 static int build_environment(
1802 const ExecContext
*c
,
1803 const ExecParameters
*p
,
1804 const CGroupContext
*cgroup_context
,
1807 const char *username
,
1809 dev_t journal_stream_dev
,
1810 ino_t journal_stream_ino
,
1811 const char *memory_pressure_path
,
1814 _cleanup_strv_free_
char **our_env
= NULL
;
1823 #define N_ENV_VARS 19
1824 our_env
= new0(char*, N_ENV_VARS
+ _EXEC_DIRECTORY_TYPE_MAX
);
1829 _cleanup_free_
char *joined
= NULL
;
1831 if (asprintf(&x
, "LISTEN_PID="PID_FMT
, getpid_cached()) < 0)
1833 our_env
[n_env
++] = x
;
1835 if (asprintf(&x
, "LISTEN_FDS=%zu", n_fds
) < 0)
1837 our_env
[n_env
++] = x
;
1839 joined
= strv_join(p
->fd_names
, ":");
1843 x
= strjoin("LISTEN_FDNAMES=", joined
);
1846 our_env
[n_env
++] = x
;
1849 if ((p
->flags
& EXEC_SET_WATCHDOG
) && p
->watchdog_usec
> 0) {
1850 if (asprintf(&x
, "WATCHDOG_PID="PID_FMT
, getpid_cached()) < 0)
1852 our_env
[n_env
++] = x
;
1854 if (asprintf(&x
, "WATCHDOG_USEC="USEC_FMT
, p
->watchdog_usec
) < 0)
1856 our_env
[n_env
++] = x
;
1859 /* If this is D-Bus, tell the nss-systemd module, since it relies on being able to use blocking
1860 * Varlink calls back to us for look up dynamic users in PID 1. Break the deadlock between D-Bus and
1861 * PID 1 by disabling use of PID1' NSS interface for looking up dynamic users. */
1862 if (p
->flags
& EXEC_NSS_DYNAMIC_BYPASS
) {
1863 x
= strdup("SYSTEMD_NSS_DYNAMIC_BYPASS=1");
1866 our_env
[n_env
++] = x
;
1869 /* We query "root" if this is a system unit and User= is not specified. $USER is always set. $HOME
1870 * could cause problem for e.g. getty, since login doesn't override $HOME, and $LOGNAME and $SHELL don't
1871 * really make much sense since we're not logged in. Hence we conditionalize the three based on
1872 * SetLoginEnvironment= switch. */
1873 if (!c
->user
&& !c
->dynamic_user
&& p
->runtime_scope
== RUNTIME_SCOPE_SYSTEM
) {
1874 r
= get_fixed_user("root", &username
, NULL
, NULL
, &home
, &shell
);
1876 return log_exec_debug_errno(c
,
1879 "Failed to determine user credentials for root: %m");
1882 bool set_user_login_env
= exec_context_get_set_login_environment(c
);
1885 x
= strjoin("USER=", username
);
1888 our_env
[n_env
++] = x
;
1890 if (set_user_login_env
) {
1891 x
= strjoin("LOGNAME=", username
);
1894 our_env
[n_env
++] = x
;
1898 if (home
&& set_user_login_env
) {
1899 x
= strjoin("HOME=", home
);
1903 path_simplify(x
+ 5);
1904 our_env
[n_env
++] = x
;
1907 if (shell
&& set_user_login_env
) {
1908 x
= strjoin("SHELL=", shell
);
1912 path_simplify(x
+ 6);
1913 our_env
[n_env
++] = x
;
1916 if (!sd_id128_is_null(p
->invocation_id
)) {
1917 assert(p
->invocation_id_string
);
1919 x
= strjoin("INVOCATION_ID=", p
->invocation_id_string
);
1923 our_env
[n_env
++] = x
;
1926 if (exec_context_needs_term(c
)) {
1927 _cleanup_free_
char *cmdline
= NULL
;
1928 const char *tty_path
, *term
= NULL
;
1930 tty_path
= exec_context_tty_path(c
);
1932 /* If we are forked off PID 1 and we are supposed to operate on /dev/console, then let's try
1933 * to inherit the $TERM set for PID 1. This is useful for containers so that the $TERM the
1934 * container manager passes to PID 1 ends up all the way in the console login shown. */
1936 if (path_equal_ptr(tty_path
, "/dev/console") && getppid() == 1)
1937 term
= getenv("TERM");
1938 else if (tty_path
&& in_charset(skip_dev_prefix(tty_path
), ALPHANUMERICAL
)) {
1939 _cleanup_free_
char *key
= NULL
;
1941 key
= strjoin("systemd.tty.term.", skip_dev_prefix(tty_path
));
1945 r
= proc_cmdline_get_key(key
, 0, &cmdline
);
1947 log_exec_debug_errno(c
,
1950 "Failed to read %s from kernel cmdline, ignoring: %m",
1957 term
= default_term_for_tty(tty_path
);
1959 x
= strjoin("TERM=", term
);
1962 our_env
[n_env
++] = x
;
1965 if (journal_stream_dev
!= 0 && journal_stream_ino
!= 0) {
1966 if (asprintf(&x
, "JOURNAL_STREAM=" DEV_FMT
":" INO_FMT
, journal_stream_dev
, journal_stream_ino
) < 0)
1969 our_env
[n_env
++] = x
;
1972 if (c
->log_namespace
) {
1973 x
= strjoin("LOG_NAMESPACE=", c
->log_namespace
);
1977 our_env
[n_env
++] = x
;
1980 for (ExecDirectoryType t
= 0; t
< _EXEC_DIRECTORY_TYPE_MAX
; t
++) {
1981 _cleanup_free_
char *joined
= NULL
;
1987 if (c
->directories
[t
].n_items
== 0)
1990 n
= exec_directory_env_name_to_string(t
);
1994 for (size_t i
= 0; i
< c
->directories
[t
].n_items
; i
++) {
1995 _cleanup_free_
char *prefixed
= NULL
;
1997 prefixed
= path_join(p
->prefix
[t
], c
->directories
[t
].items
[i
].path
);
2001 if (!strextend_with_separator(&joined
, ":", prefixed
))
2005 x
= strjoin(n
, "=", joined
);
2009 our_env
[n_env
++] = x
;
2012 _cleanup_free_
char *creds_dir
= NULL
;
2013 r
= exec_context_get_credential_directory(c
, p
, p
->unit_id
, &creds_dir
);
2017 x
= strjoin("CREDENTIALS_DIRECTORY=", creds_dir
);
2021 our_env
[n_env
++] = x
;
2024 if (asprintf(&x
, "SYSTEMD_EXEC_PID=" PID_FMT
, getpid_cached()) < 0)
2027 our_env
[n_env
++] = x
;
2029 if (memory_pressure_path
) {
2030 x
= strjoin("MEMORY_PRESSURE_WATCH=", memory_pressure_path
);
2034 our_env
[n_env
++] = x
;
2036 if (cgroup_context
&& !path_equal(memory_pressure_path
, "/dev/null")) {
2037 _cleanup_free_
char *b
= NULL
, *e
= NULL
;
2039 if (asprintf(&b
, "%s " USEC_FMT
" " USEC_FMT
,
2040 MEMORY_PRESSURE_DEFAULT_TYPE
,
2041 cgroup_context
->memory_pressure_threshold_usec
== USEC_INFINITY
? MEMORY_PRESSURE_DEFAULT_THRESHOLD_USEC
:
2042 CLAMP(cgroup_context
->memory_pressure_threshold_usec
, 1U, MEMORY_PRESSURE_DEFAULT_WINDOW_USEC
),
2043 MEMORY_PRESSURE_DEFAULT_WINDOW_USEC
) < 0)
2046 if (base64mem(b
, strlen(b
) + 1, &e
) < 0)
2049 x
= strjoin("MEMORY_PRESSURE_WRITE=", e
);
2053 our_env
[n_env
++] = x
;
2057 assert(n_env
< N_ENV_VARS
+ _EXEC_DIRECTORY_TYPE_MAX
);
2060 *ret
= TAKE_PTR(our_env
);
2065 static int build_pass_environment(const ExecContext
*c
, char ***ret
) {
2066 _cleanup_strv_free_
char **pass_env
= NULL
;
2069 STRV_FOREACH(i
, c
->pass_environment
) {
2070 _cleanup_free_
char *x
= NULL
;
2076 x
= strjoin(*i
, "=", v
);
2080 if (!GREEDY_REALLOC(pass_env
, n_env
+ 2))
2083 pass_env
[n_env
++] = TAKE_PTR(x
);
2084 pass_env
[n_env
] = NULL
;
2087 *ret
= TAKE_PTR(pass_env
);
2092 static int setup_private_users(uid_t ouid
, gid_t ogid
, uid_t uid
, gid_t gid
) {
2093 _cleanup_free_
char *uid_map
= NULL
, *gid_map
= NULL
;
2094 _cleanup_close_pair_
int errno_pipe
[2] = EBADF_PAIR
;
2095 _cleanup_close_
int unshare_ready_fd
= -EBADF
;
2096 _cleanup_(sigkill_waitp
) pid_t pid
= 0;
2101 /* Set up a user namespace and map the original UID/GID (IDs from before any user or group changes, i.e.
2102 * the IDs from the user or system manager(s)) to itself, the selected UID/GID to itself, and everything else to
2103 * nobody. In order to be able to write this mapping we need CAP_SETUID in the original user namespace, which
2104 * we however lack after opening the user namespace. To work around this we fork() a temporary child process,
2105 * which waits for the parent to create the new user namespace while staying in the original namespace. The
2106 * child then writes the UID mapping, under full privileges. The parent waits for the child to finish and
2107 * continues execution normally.
2108 * For unprivileged users (i.e. without capabilities), the root to root mapping is excluded. As such, it
2109 * does not need CAP_SETUID to write the single line mapping to itself. */
2111 /* Can only set up multiple mappings with CAP_SETUID. */
2112 if (have_effective_cap(CAP_SETUID
) > 0 && uid
!= ouid
&& uid_is_valid(uid
))
2113 r
= asprintf(&uid_map
,
2114 UID_FMT
" " UID_FMT
" 1\n" /* Map $OUID → $OUID */
2115 UID_FMT
" " UID_FMT
" 1\n", /* Map $UID → $UID */
2116 ouid
, ouid
, uid
, uid
);
2118 r
= asprintf(&uid_map
,
2119 UID_FMT
" " UID_FMT
" 1\n", /* Map $OUID → $OUID */
2125 /* Can only set up multiple mappings with CAP_SETGID. */
2126 if (have_effective_cap(CAP_SETGID
) > 0 && gid
!= ogid
&& gid_is_valid(gid
))
2127 r
= asprintf(&gid_map
,
2128 GID_FMT
" " GID_FMT
" 1\n" /* Map $OGID → $OGID */
2129 GID_FMT
" " GID_FMT
" 1\n", /* Map $GID → $GID */
2130 ogid
, ogid
, gid
, gid
);
2132 r
= asprintf(&gid_map
,
2133 GID_FMT
" " GID_FMT
" 1\n", /* Map $OGID -> $OGID */
2139 /* Create a communication channel so that the parent can tell the child when it finished creating the user
2141 unshare_ready_fd
= eventfd(0, EFD_CLOEXEC
);
2142 if (unshare_ready_fd
< 0)
2145 /* Create a communication channel so that the child can tell the parent a proper error code in case it
2147 if (pipe2(errno_pipe
, O_CLOEXEC
) < 0)
2150 r
= safe_fork("(sd-userns)", FORK_RESET_SIGNALS
|FORK_DEATHSIG_SIGKILL
, &pid
);
2154 _cleanup_close_
int fd
= -EBADF
;
2158 /* Child process, running in the original user namespace. Let's update the parent's UID/GID map from
2159 * here, after the parent opened its own user namespace. */
2162 errno_pipe
[0] = safe_close(errno_pipe
[0]);
2164 /* Wait until the parent unshared the user namespace */
2165 if (read(unshare_ready_fd
, &c
, sizeof(c
)) < 0) {
2170 /* Disable the setgroups() system call in the child user namespace, for good. */
2171 a
= procfs_file_alloca(ppid
, "setgroups");
2172 fd
= open(a
, O_WRONLY
|O_CLOEXEC
);
2174 if (errno
!= ENOENT
) {
2179 /* If the file is missing the kernel is too old, let's continue anyway. */
2181 if (write(fd
, "deny\n", 5) < 0) {
2186 fd
= safe_close(fd
);
2189 /* First write the GID map */
2190 a
= procfs_file_alloca(ppid
, "gid_map");
2191 fd
= open(a
, O_WRONLY
|O_CLOEXEC
);
2196 if (write(fd
, gid_map
, strlen(gid_map
)) < 0) {
2200 fd
= safe_close(fd
);
2202 /* The write the UID map */
2203 a
= procfs_file_alloca(ppid
, "uid_map");
2204 fd
= open(a
, O_WRONLY
|O_CLOEXEC
);
2209 if (write(fd
, uid_map
, strlen(uid_map
)) < 0) {
2214 _exit(EXIT_SUCCESS
);
2217 (void) write(errno_pipe
[1], &r
, sizeof(r
));
2218 _exit(EXIT_FAILURE
);
2221 errno_pipe
[1] = safe_close(errno_pipe
[1]);
2223 if (unshare(CLONE_NEWUSER
) < 0)
2226 /* Let the child know that the namespace is ready now */
2227 if (write(unshare_ready_fd
, &c
, sizeof(c
)) < 0)
2230 /* Try to read an error code from the child */
2231 n
= read(errno_pipe
[0], &r
, sizeof(r
));
2234 if (n
== sizeof(r
)) { /* an error code was sent to us */
2239 if (n
!= 0) /* on success we should have read 0 bytes */
2242 r
= wait_for_terminate_and_check("(sd-userns)", TAKE_PID(pid
), 0);
2245 if (r
!= EXIT_SUCCESS
) /* If something strange happened with the child, let's consider this fatal, too */
2251 static int create_many_symlinks(const char *root
, const char *source
, char **symlinks
) {
2252 _cleanup_free_
char *src_abs
= NULL
;
2257 src_abs
= path_join(root
, source
);
2261 STRV_FOREACH(dst
, symlinks
) {
2262 _cleanup_free_
char *dst_abs
= NULL
;
2264 dst_abs
= path_join(root
, *dst
);
2268 r
= mkdir_parents_label(dst_abs
, 0755);
2272 r
= symlink_idempotent(src_abs
, dst_abs
, true);
2280 static int setup_exec_directory(
2281 const ExecContext
*context
,
2282 const ExecParameters
*params
,
2285 ExecDirectoryType type
,
2286 bool needs_mount_namespace
,
2289 static const int exit_status_table
[_EXEC_DIRECTORY_TYPE_MAX
] = {
2290 [EXEC_DIRECTORY_RUNTIME
] = EXIT_RUNTIME_DIRECTORY
,
2291 [EXEC_DIRECTORY_STATE
] = EXIT_STATE_DIRECTORY
,
2292 [EXEC_DIRECTORY_CACHE
] = EXIT_CACHE_DIRECTORY
,
2293 [EXEC_DIRECTORY_LOGS
] = EXIT_LOGS_DIRECTORY
,
2294 [EXEC_DIRECTORY_CONFIGURATION
] = EXIT_CONFIGURATION_DIRECTORY
,
2300 assert(type
>= 0 && type
< _EXEC_DIRECTORY_TYPE_MAX
);
2301 assert(exit_status
);
2303 if (!params
->prefix
[type
])
2306 if (params
->flags
& EXEC_CHOWN_DIRECTORIES
) {
2307 if (!uid_is_valid(uid
))
2309 if (!gid_is_valid(gid
))
2313 for (size_t i
= 0; i
< context
->directories
[type
].n_items
; i
++) {
2314 _cleanup_free_
char *p
= NULL
, *pp
= NULL
;
2316 p
= path_join(params
->prefix
[type
], context
->directories
[type
].items
[i
].path
);
2322 r
= mkdir_parents_label(p
, 0755);
2326 if (IN_SET(type
, EXEC_DIRECTORY_STATE
, EXEC_DIRECTORY_LOGS
) && params
->runtime_scope
== RUNTIME_SCOPE_USER
) {
2328 /* If we are in user mode, and a configuration directory exists but a state directory
2329 * doesn't exist, then we likely are upgrading from an older systemd version that
2330 * didn't know the more recent addition to the xdg-basedir spec: the $XDG_STATE_HOME
2331 * directory. In older systemd versions EXEC_DIRECTORY_STATE was aliased to
2332 * EXEC_DIRECTORY_CONFIGURATION, with the advent of $XDG_STATE_HOME is is now
2333 * separated. If a service has both dirs configured but only the configuration dir
2334 * exists and the state dir does not, we assume we are looking at an update
2335 * situation. Hence, create a compatibility symlink, so that all expectations are
2338 * (We also do something similar with the log directory, which still doesn't exist in
2339 * the xdg basedir spec. We'll make it a subdir of the state dir.) */
2341 /* this assumes the state dir is always created before the configuration dir */
2342 assert_cc(EXEC_DIRECTORY_STATE
< EXEC_DIRECTORY_LOGS
);
2343 assert_cc(EXEC_DIRECTORY_LOGS
< EXEC_DIRECTORY_CONFIGURATION
);
2345 r
= laccess(p
, F_OK
);
2347 _cleanup_free_
char *q
= NULL
;
2349 /* OK, we know that the state dir does not exist. Let's see if the dir exists
2350 * under the configuration hierarchy. */
2352 if (type
== EXEC_DIRECTORY_STATE
)
2353 q
= path_join(params
->prefix
[EXEC_DIRECTORY_CONFIGURATION
], context
->directories
[type
].items
[i
].path
);
2354 else if (type
== EXEC_DIRECTORY_LOGS
)
2355 q
= path_join(params
->prefix
[EXEC_DIRECTORY_CONFIGURATION
], "log", context
->directories
[type
].items
[i
].path
);
2357 assert_not_reached();
2363 r
= laccess(q
, F_OK
);
2365 /* It does exist! This hence looks like an update. Symlink the
2366 * configuration directory into the state directory. */
2368 r
= symlink_idempotent(q
, p
, /* make_relative= */ true);
2372 log_exec_notice(context
, params
, "Unit state directory %s missing but matching configuration directory %s exists, assuming update from systemd 253 or older, creating compatibility symlink.", p
, q
);
2374 } else if (r
!= -ENOENT
)
2375 log_exec_warning_errno(context
, params
, r
, "Unable to detect whether unit configuration directory '%s' exists, assuming not: %m", q
);
2378 log_exec_warning_errno(context
, params
, r
, "Unable to detect whether unit state directory '%s' is missing, assuming it is: %m", p
);
2381 if (exec_directory_is_private(context
, type
)) {
2382 /* So, here's one extra complication when dealing with DynamicUser=1 units. In that
2383 * case we want to avoid leaving a directory around fully accessible that is owned by
2384 * a dynamic user whose UID is later on reused. To lock this down we use the same
2385 * trick used by container managers to prohibit host users to get access to files of
2386 * the same UID in containers: we place everything inside a directory that has an
2387 * access mode of 0700 and is owned root:root, so that it acts as security boundary
2388 * for unprivileged host code. We then use fs namespacing to make this directory
2389 * permeable for the service itself.
2391 * Specifically: for a service which wants a special directory "foo/" we first create
2392 * a directory "private/" with access mode 0700 owned by root:root. Then we place
2393 * "foo" inside of that directory (i.e. "private/foo/"), and make "foo" a symlink to
2394 * "private/foo". This way, privileged host users can access "foo/" as usual, but
2395 * unprivileged host users can't look into it. Inside of the namespace of the unit
2396 * "private/" is replaced by a more liberally accessible tmpfs, into which the host's
2397 * "private/foo/" is mounted under the same name, thus disabling the access boundary
2398 * for the service and making sure it only gets access to the dirs it needs but no
2399 * others. Tricky? Yes, absolutely, but it works!
2401 * Note that we don't do this for EXEC_DIRECTORY_CONFIGURATION as that's assumed not
2402 * to be owned by the service itself.
2404 * Also, note that we don't do this for EXEC_DIRECTORY_RUNTIME as that's often used
2405 * for sharing files or sockets with other services. */
2407 pp
= path_join(params
->prefix
[type
], "private");
2413 /* First set up private root if it doesn't exist yet, with access mode 0700 and owned by root:root */
2414 r
= mkdir_safe_label(pp
, 0700, 0, 0, MKDIR_WARN_MODE
);
2418 if (!path_extend(&pp
, context
->directories
[type
].items
[i
].path
)) {
2423 /* Create all directories between the configured directory and this private root, and mark them 0755 */
2424 r
= mkdir_parents_label(pp
, 0755);
2428 if (is_dir(p
, false) > 0 &&
2429 (laccess(pp
, F_OK
) == -ENOENT
)) {
2431 /* Hmm, the private directory doesn't exist yet, but the normal one exists? If so, move
2432 * it over. Most likely the service has been upgraded from one that didn't use
2433 * DynamicUser=1, to one that does. */
2435 log_exec_info(context
,
2437 "Found pre-existing public %s= directory %s, migrating to %s.\n"
2438 "Apparently, service previously had DynamicUser= turned off, and has now turned it on.",
2439 exec_directory_type_to_string(type
), p
, pp
);
2441 r
= RET_NERRNO(rename(p
, pp
));
2445 /* Otherwise, create the actual directory for the service */
2447 r
= mkdir_label(pp
, context
->directories
[type
].mode
);
2448 if (r
< 0 && r
!= -EEXIST
)
2452 if (!context
->directories
[type
].items
[i
].only_create
) {
2453 /* And link it up from the original place.
2455 * 1) If a mount namespace is going to be used, then this symlink remains on
2456 * the host, and a new one for the child namespace will be created later.
2457 * 2) It is not necessary to create this symlink when one of its parent
2458 * directories is specified and already created. E.g.
2459 * StateDirectory=foo foo/bar
2460 * In that case, the inode points to pp and p for "foo/bar" are the same:
2461 * pp = "/var/lib/private/foo/bar"
2462 * p = "/var/lib/foo/bar"
2463 * and, /var/lib/foo is a symlink to /var/lib/private/foo. So, not only
2464 * we do not need to create the symlink, but we cannot create the symlink.
2465 * See issue #24783. */
2466 r
= symlink_idempotent(pp
, p
, true);
2472 _cleanup_free_
char *target
= NULL
;
2474 if (type
!= EXEC_DIRECTORY_CONFIGURATION
&&
2475 readlink_and_make_absolute(p
, &target
) >= 0) {
2476 _cleanup_free_
char *q
= NULL
, *q_resolved
= NULL
, *target_resolved
= NULL
;
2478 /* This already exists and is a symlink? Interesting. Maybe it's one created
2479 * by DynamicUser=1 (see above)?
2481 * We do this for all directory types except for ConfigurationDirectory=,
2482 * since they all support the private/ symlink logic at least in some
2483 * configurations, see above. */
2485 r
= chase(target
, NULL
, 0, &target_resolved
, NULL
);
2489 q
= path_join(params
->prefix
[type
], "private", context
->directories
[type
].items
[i
].path
);
2495 /* /var/lib or friends may be symlinks. So, let's chase them also. */
2496 r
= chase(q
, NULL
, CHASE_NONEXISTENT
, &q_resolved
, NULL
);
2500 if (path_equal(q_resolved
, target_resolved
)) {
2502 /* Hmm, apparently DynamicUser= was once turned on for this service,
2503 * but is no longer. Let's move the directory back up. */
2505 log_exec_info(context
,
2507 "Found pre-existing private %s= directory %s, migrating to %s.\n"
2508 "Apparently, service previously had DynamicUser= turned on, and has now turned it off.",
2509 exec_directory_type_to_string(type
), q
, p
);
2511 r
= RET_NERRNO(unlink(p
));
2515 r
= RET_NERRNO(rename(q
, p
));
2521 r
= mkdir_label(p
, context
->directories
[type
].mode
);
2526 if (type
== EXEC_DIRECTORY_CONFIGURATION
) {
2529 /* Don't change the owner/access mode of the configuration directory,
2530 * as in the common case it is not written to by a service, and shall
2531 * not be writable. */
2533 r
= RET_NERRNO(stat(p
, &st
));
2537 /* Still complain if the access mode doesn't match */
2538 if (((st
.st_mode
^ context
->directories
[type
].mode
) & 07777) != 0)
2539 log_exec_warning(context
,
2541 "%s \'%s\' already exists but the mode is different. "
2542 "(File system: %o %sMode: %o)",
2543 exec_directory_type_to_string(type
), context
->directories
[type
].items
[i
].path
,
2544 st
.st_mode
& 07777, exec_directory_type_to_string(type
), context
->directories
[type
].mode
& 07777);
2551 /* Lock down the access mode (we use chmod_and_chown() to make this idempotent. We don't
2552 * specify UID/GID here, so that path_chown_recursive() can optimize things depending on the
2553 * current UID/GID ownership.) */
2554 r
= chmod_and_chown(pp
?: p
, context
->directories
[type
].mode
, UID_INVALID
, GID_INVALID
);
2558 /* Skip the rest (which deals with ownership) in user mode, since ownership changes are not
2559 * available to user code anyway */
2560 if (params
->runtime_scope
!= RUNTIME_SCOPE_SYSTEM
)
2563 /* Then, change the ownership of the whole tree, if necessary. When dynamic users are used we
2564 * drop the suid/sgid bits, since we really don't want SUID/SGID files for dynamic UID/GID
2565 * assignments to exist. */
2566 r
= path_chown_recursive(pp
?: p
, uid
, gid
, context
->dynamic_user
? 01777 : 07777, AT_SYMLINK_FOLLOW
);
2571 /* If we are not going to run in a namespace, set up the symlinks - otherwise
2572 * they are set up later, to allow configuring empty var/run/etc. */
2573 if (!needs_mount_namespace
)
2574 for (size_t i
= 0; i
< context
->directories
[type
].n_items
; i
++) {
2575 r
= create_many_symlinks(params
->prefix
[type
],
2576 context
->directories
[type
].items
[i
].path
,
2577 context
->directories
[type
].items
[i
].symlinks
);
2585 *exit_status
= exit_status_table
[type
];
2590 static int setup_smack(
2591 const ExecParameters
*params
,
2592 const ExecContext
*context
,
2593 int executable_fd
) {
2597 assert(executable_fd
>= 0);
2599 if (context
->smack_process_label
) {
2600 r
= mac_smack_apply_pid(0, context
->smack_process_label
);
2603 } else if (params
->fallback_smack_process_label
) {
2604 _cleanup_free_
char *exec_label
= NULL
;
2606 r
= mac_smack_read_fd(executable_fd
, SMACK_ATTR_EXEC
, &exec_label
);
2607 if (r
< 0 && !ERRNO_IS_XATTR_ABSENT(r
))
2610 r
= mac_smack_apply_pid(0, exec_label
?: params
->fallback_smack_process_label
);
2619 static int compile_bind_mounts(
2620 const ExecContext
*context
,
2621 const ExecParameters
*params
,
2622 BindMount
**ret_bind_mounts
,
2623 size_t *ret_n_bind_mounts
,
2624 char ***ret_empty_directories
) {
2626 _cleanup_strv_free_
char **empty_directories
= NULL
;
2627 BindMount
*bind_mounts
= NULL
;
2633 assert(ret_bind_mounts
);
2634 assert(ret_n_bind_mounts
);
2635 assert(ret_empty_directories
);
2637 CLEANUP_ARRAY(bind_mounts
, h
, bind_mount_free_many
);
2639 n
= context
->n_bind_mounts
;
2640 for (ExecDirectoryType t
= 0; t
< _EXEC_DIRECTORY_TYPE_MAX
; t
++) {
2641 if (!params
->prefix
[t
])
2644 for (size_t i
= 0; i
< context
->directories
[t
].n_items
; i
++)
2645 n
+= !context
->directories
[t
].items
[i
].only_create
;
2649 *ret_bind_mounts
= NULL
;
2650 *ret_n_bind_mounts
= 0;
2651 *ret_empty_directories
= NULL
;
2655 bind_mounts
= new(BindMount
, n
);
2659 for (size_t i
= 0; i
< context
->n_bind_mounts
; i
++) {
2660 BindMount
*item
= context
->bind_mounts
+ i
;
2661 _cleanup_free_
char *s
= NULL
, *d
= NULL
;
2663 s
= strdup(item
->source
);
2667 d
= strdup(item
->destination
);
2671 bind_mounts
[h
++] = (BindMount
) {
2672 .source
= TAKE_PTR(s
),
2673 .destination
= TAKE_PTR(d
),
2674 .read_only
= item
->read_only
,
2675 .recursive
= item
->recursive
,
2676 .ignore_enoent
= item
->ignore_enoent
,
2680 for (ExecDirectoryType t
= 0; t
< _EXEC_DIRECTORY_TYPE_MAX
; t
++) {
2681 if (!params
->prefix
[t
])
2684 if (context
->directories
[t
].n_items
== 0)
2687 if (exec_directory_is_private(context
, t
) &&
2688 !exec_context_with_rootfs(context
)) {
2691 /* So this is for a dynamic user, and we need to make sure the process can access its own
2692 * directory. For that we overmount the usually inaccessible "private" subdirectory with a
2693 * tmpfs that makes it accessible and is empty except for the submounts we do this for. */
2695 private_root
= path_join(params
->prefix
[t
], "private");
2699 r
= strv_consume(&empty_directories
, private_root
);
2704 for (size_t i
= 0; i
< context
->directories
[t
].n_items
; i
++) {
2705 _cleanup_free_
char *s
= NULL
, *d
= NULL
;
2707 /* When one of the parent directories is in the list, we cannot create the symlink
2708 * for the child directory. See also the comments in setup_exec_directory(). */
2709 if (context
->directories
[t
].items
[i
].only_create
)
2712 if (exec_directory_is_private(context
, t
))
2713 s
= path_join(params
->prefix
[t
], "private", context
->directories
[t
].items
[i
].path
);
2715 s
= path_join(params
->prefix
[t
], context
->directories
[t
].items
[i
].path
);
2719 if (exec_directory_is_private(context
, t
) &&
2720 exec_context_with_rootfs(context
))
2721 /* When RootDirectory= or RootImage= are set, then the symbolic link to the private
2722 * directory is not created on the root directory. So, let's bind-mount the directory
2723 * on the 'non-private' place. */
2724 d
= path_join(params
->prefix
[t
], context
->directories
[t
].items
[i
].path
);
2730 bind_mounts
[h
++] = (BindMount
) {
2731 .source
= TAKE_PTR(s
),
2732 .destination
= TAKE_PTR(d
),
2734 .nosuid
= context
->dynamic_user
, /* don't allow suid/sgid when DynamicUser= is on */
2736 .ignore_enoent
= false,
2743 *ret_bind_mounts
= TAKE_PTR(bind_mounts
);
2744 *ret_n_bind_mounts
= n
;
2745 *ret_empty_directories
= TAKE_PTR(empty_directories
);
2750 /* ret_symlinks will contain a list of pairs src:dest that describes
2751 * the symlinks to create later on. For example, the symlinks needed
2752 * to safely give private directories to DynamicUser=1 users. */
2753 static int compile_symlinks(
2754 const ExecContext
*context
,
2755 const ExecParameters
*params
,
2756 bool setup_os_release_symlink
,
2757 char ***ret_symlinks
) {
2759 _cleanup_strv_free_
char **symlinks
= NULL
;
2764 assert(ret_symlinks
);
2766 for (ExecDirectoryType dt
= 0; dt
< _EXEC_DIRECTORY_TYPE_MAX
; dt
++) {
2767 for (size_t i
= 0; i
< context
->directories
[dt
].n_items
; i
++) {
2768 _cleanup_free_
char *private_path
= NULL
, *path
= NULL
;
2770 STRV_FOREACH(symlink
, context
->directories
[dt
].items
[i
].symlinks
) {
2771 _cleanup_free_
char *src_abs
= NULL
, *dst_abs
= NULL
;
2773 src_abs
= path_join(params
->prefix
[dt
], context
->directories
[dt
].items
[i
].path
);
2774 dst_abs
= path_join(params
->prefix
[dt
], *symlink
);
2775 if (!src_abs
|| !dst_abs
)
2778 r
= strv_consume_pair(&symlinks
, TAKE_PTR(src_abs
), TAKE_PTR(dst_abs
));
2783 if (!exec_directory_is_private(context
, dt
) ||
2784 exec_context_with_rootfs(context
) ||
2785 context
->directories
[dt
].items
[i
].only_create
)
2788 private_path
= path_join(params
->prefix
[dt
], "private", context
->directories
[dt
].items
[i
].path
);
2792 path
= path_join(params
->prefix
[dt
], context
->directories
[dt
].items
[i
].path
);
2796 r
= strv_consume_pair(&symlinks
, TAKE_PTR(private_path
), TAKE_PTR(path
));
2802 /* We make the host's os-release available via a symlink, so that we can copy it atomically
2803 * and readers will never get a half-written version. Note that, while the paths specified here are
2804 * absolute, when they are processed in namespace.c they will be made relative automatically, i.e.:
2805 * 'os-release -> .os-release-stage/os-release' is what will be created. */
2806 if (setup_os_release_symlink
) {
2807 r
= strv_extend(&symlinks
, "/run/host/.os-release-stage/os-release");
2811 r
= strv_extend(&symlinks
, "/run/host/os-release");
2816 *ret_symlinks
= TAKE_PTR(symlinks
);
2821 static bool insist_on_sandboxing(
2822 const ExecContext
*context
,
2823 const char *root_dir
,
2824 const char *root_image
,
2825 const BindMount
*bind_mounts
,
2826 size_t n_bind_mounts
) {
2829 assert(n_bind_mounts
== 0 || bind_mounts
);
2831 /* Checks whether we need to insist on fs namespacing. i.e. whether we have settings configured that
2832 * would alter the view on the file system beyond making things read-only or invisible, i.e. would
2833 * rearrange stuff in a way we cannot ignore gracefully. */
2835 if (context
->n_temporary_filesystems
> 0)
2838 if (root_dir
|| root_image
)
2841 if (context
->n_mount_images
> 0)
2844 if (context
->dynamic_user
)
2847 if (context
->n_extension_images
> 0 || !strv_isempty(context
->extension_directories
))
2850 /* If there are any bind mounts set that don't map back onto themselves, fs namespacing becomes
2852 for (size_t i
= 0; i
< n_bind_mounts
; i
++)
2853 if (!path_equal(bind_mounts
[i
].source
, bind_mounts
[i
].destination
))
2856 if (context
->log_namespace
)
2862 static int setup_ephemeral(const ExecContext
*context
, ExecRuntime
*runtime
) {
2863 _cleanup_close_
int fd
= -EBADF
;
2866 if (!runtime
|| !runtime
->ephemeral_copy
)
2869 r
= posix_lock(runtime
->ephemeral_storage_socket
[0], LOCK_EX
);
2871 return log_debug_errno(r
, "Failed to lock ephemeral storage socket: %m");
2873 CLEANUP_POSIX_UNLOCK(runtime
->ephemeral_storage_socket
[0]);
2875 fd
= receive_one_fd(runtime
->ephemeral_storage_socket
[0], MSG_PEEK
|MSG_DONTWAIT
);
2877 /* We got an fd! That means ephemeral has already been set up, so nothing to do here. */
2881 return log_debug_errno(fd
, "Failed to receive file descriptor queued on ephemeral storage socket: %m");
2883 log_debug("Making ephemeral snapshot of %s to %s",
2884 context
->root_image
?: context
->root_directory
, runtime
->ephemeral_copy
);
2886 if (context
->root_image
)
2887 fd
= copy_file(context
->root_image
, runtime
->ephemeral_copy
, O_EXCL
, 0600,
2888 COPY_LOCK_BSD
|COPY_REFLINK
|COPY_CRTIME
);
2890 fd
= btrfs_subvol_snapshot_at(AT_FDCWD
, context
->root_directory
,
2891 AT_FDCWD
, runtime
->ephemeral_copy
,
2892 BTRFS_SNAPSHOT_FALLBACK_COPY
|
2893 BTRFS_SNAPSHOT_FALLBACK_DIRECTORY
|
2894 BTRFS_SNAPSHOT_RECURSIVE
|
2895 BTRFS_SNAPSHOT_LOCK_BSD
);
2897 return log_debug_errno(fd
, "Failed to snapshot %s to %s: %m",
2898 context
->root_image
?: context
->root_directory
, runtime
->ephemeral_copy
);
2900 if (context
->root_image
) {
2901 /* A root image might be subject to lots of random writes so let's try to disable COW on it
2902 * which tends to not perform well in combination with lots of random writes.
2904 * Note: btrfs actually isn't impressed by us setting the flag after making the reflink'ed
2905 * copy, but we at least want to make the intention clear.
2907 r
= chattr_fd(fd
, FS_NOCOW_FL
, FS_NOCOW_FL
, NULL
);
2909 log_debug_errno(fd
, "Failed to disable copy-on-write for %s, ignoring: %m", runtime
->ephemeral_copy
);
2912 r
= send_one_fd(runtime
->ephemeral_storage_socket
[1], fd
, MSG_DONTWAIT
);
2914 return log_debug_errno(r
, "Failed to queue file descriptor on ephemeral storage socket: %m");
2919 static int verity_settings_prepare(
2920 VeritySettings
*verity
,
2921 const char *root_image
,
2922 const void *root_hash
,
2923 size_t root_hash_size
,
2924 const char *root_hash_path
,
2925 const void *root_hash_sig
,
2926 size_t root_hash_sig_size
,
2927 const char *root_hash_sig_path
,
2928 const char *verity_data_path
) {
2937 d
= memdup(root_hash
, root_hash_size
);
2941 free_and_replace(verity
->root_hash
, d
);
2942 verity
->root_hash_size
= root_hash_size
;
2943 verity
->designator
= PARTITION_ROOT
;
2946 if (root_hash_sig
) {
2949 d
= memdup(root_hash_sig
, root_hash_sig_size
);
2953 free_and_replace(verity
->root_hash_sig
, d
);
2954 verity
->root_hash_sig_size
= root_hash_sig_size
;
2955 verity
->designator
= PARTITION_ROOT
;
2958 if (verity_data_path
) {
2959 r
= free_and_strdup(&verity
->data_path
, verity_data_path
);
2964 r
= verity_settings_load(
2968 root_hash_sig_path
);
2970 return log_debug_errno(r
, "Failed to load root hash: %m");
2975 static int apply_mount_namespace(
2976 ExecCommandFlags command_flags
,
2977 const ExecContext
*context
,
2978 const ExecParameters
*params
,
2979 ExecRuntime
*runtime
,
2980 const char *memory_pressure_path
,
2981 char **error_path
) {
2983 _cleanup_(verity_settings_done
) VeritySettings verity
= VERITY_SETTINGS_DEFAULT
;
2984 _cleanup_strv_free_
char **empty_directories
= NULL
, **symlinks
= NULL
,
2985 **read_write_paths_cleanup
= NULL
;
2986 _cleanup_free_
char *creds_path
= NULL
, *incoming_dir
= NULL
, *propagate_dir
= NULL
,
2987 *extension_dir
= NULL
, *host_os_release_stage
= NULL
;
2988 const char *root_dir
= NULL
, *root_image
= NULL
, *tmp_dir
= NULL
, *var_tmp_dir
= NULL
;
2989 char **read_write_paths
;
2990 bool needs_sandboxing
, setup_os_release_symlink
;
2991 BindMount
*bind_mounts
= NULL
;
2992 size_t n_bind_mounts
= 0;
2997 CLEANUP_ARRAY(bind_mounts
, n_bind_mounts
, bind_mount_free_many
);
2999 if (params
->flags
& EXEC_APPLY_CHROOT
) {
3000 r
= setup_ephemeral(context
, runtime
);
3004 if (context
->root_image
)
3005 root_image
= (runtime
? runtime
->ephemeral_copy
: NULL
) ?: context
->root_image
;
3007 root_dir
= (runtime
? runtime
->ephemeral_copy
: NULL
) ?: context
->root_directory
;
3010 r
= compile_bind_mounts(context
, params
, &bind_mounts
, &n_bind_mounts
, &empty_directories
);
3014 /* We need to make the pressure path writable even if /sys/fs/cgroups is made read-only, as the
3015 * service will need to write to it in order to start the notifications. */
3016 if (context
->protect_control_groups
&& memory_pressure_path
&& !streq(memory_pressure_path
, "/dev/null")) {
3017 read_write_paths_cleanup
= strv_copy(context
->read_write_paths
);
3018 if (!read_write_paths_cleanup
)
3021 r
= strv_extend(&read_write_paths_cleanup
, memory_pressure_path
);
3025 read_write_paths
= read_write_paths_cleanup
;
3027 read_write_paths
= context
->read_write_paths
;
3029 needs_sandboxing
= (params
->flags
& EXEC_APPLY_SANDBOXING
) && !(command_flags
& EXEC_COMMAND_FULLY_PRIVILEGED
);
3030 if (needs_sandboxing
) {
3031 /* The runtime struct only contains the parent of the private /tmp, which is non-accessible
3032 * to world users. Inside of it there's a /tmp that is sticky, and that's the one we want to
3033 * use here. This does not apply when we are using /run/systemd/empty as fallback. */
3035 if (context
->private_tmp
&& runtime
&& runtime
->shared
) {
3036 if (streq_ptr(runtime
->shared
->tmp_dir
, RUN_SYSTEMD_EMPTY
))
3037 tmp_dir
= runtime
->shared
->tmp_dir
;
3038 else if (runtime
->shared
->tmp_dir
)
3039 tmp_dir
= strjoina(runtime
->shared
->tmp_dir
, "/tmp");
3041 if (streq_ptr(runtime
->shared
->var_tmp_dir
, RUN_SYSTEMD_EMPTY
))
3042 var_tmp_dir
= runtime
->shared
->var_tmp_dir
;
3043 else if (runtime
->shared
->var_tmp_dir
)
3044 var_tmp_dir
= strjoina(runtime
->shared
->var_tmp_dir
, "/tmp");
3048 /* Symlinks (exec dirs, os-release) are set up after other mounts, before they are made read-only. */
3049 setup_os_release_symlink
= needs_sandboxing
&& exec_context_get_effective_mount_apivfs(context
) && (root_dir
|| root_image
);
3050 r
= compile_symlinks(context
, params
, setup_os_release_symlink
, &symlinks
);
3054 if (context
->mount_propagation_flag
== MS_SHARED
)
3055 log_exec_debug(context
,
3057 "shared mount propagation hidden by other fs namespacing unit settings: ignoring");
3059 if (FLAGS_SET(params
->flags
, EXEC_WRITE_CREDENTIALS
)) {
3060 r
= exec_context_get_credential_directory(context
, params
, params
->unit_id
, &creds_path
);
3065 if (params
->runtime_scope
== RUNTIME_SCOPE_SYSTEM
) {
3066 propagate_dir
= path_join("/run/systemd/propagate/", params
->unit_id
);
3070 incoming_dir
= strdup("/run/systemd/incoming");
3074 extension_dir
= strdup("/run/systemd/unit-extensions");
3078 /* If running under a different root filesystem, propagate the host's os-release. We make a
3079 * copy rather than just bind mounting it, so that it can be updated on soft-reboot. */
3080 if (setup_os_release_symlink
) {
3081 host_os_release_stage
= strdup("/run/systemd/propagate/.os-release-stage");
3082 if (!host_os_release_stage
)
3086 assert(params
->runtime_scope
== RUNTIME_SCOPE_USER
);
3088 if (asprintf(&extension_dir
, "/run/user/" UID_FMT
"/systemd/unit-extensions", geteuid()) < 0)
3091 if (setup_os_release_symlink
) {
3092 if (asprintf(&host_os_release_stage
,
3093 "/run/user/" UID_FMT
"/systemd/propagate/.os-release-stage",
3100 r
= verity_settings_prepare(
3103 context
->root_hash
, context
->root_hash_size
, context
->root_hash_path
,
3104 context
->root_hash_sig
, context
->root_hash_sig_size
, context
->root_hash_sig_path
,
3105 context
->root_verity
);
3110 NamespaceParameters parameters
= {
3111 .runtime_scope
= params
->runtime_scope
,
3113 .root_directory
= root_dir
,
3114 .root_image
= root_image
,
3115 .root_image_options
= context
->root_image_options
,
3116 .root_image_policy
= context
->root_image_policy
?: &image_policy_service
,
3118 .read_write_paths
= read_write_paths
,
3119 .read_only_paths
= needs_sandboxing
? context
->read_only_paths
: NULL
,
3120 .inaccessible_paths
= needs_sandboxing
? context
->inaccessible_paths
: NULL
,
3122 .exec_paths
= needs_sandboxing
? context
->exec_paths
: NULL
,
3123 .no_exec_paths
= needs_sandboxing
? context
->no_exec_paths
: NULL
,
3125 .empty_directories
= empty_directories
,
3126 .symlinks
= symlinks
,
3128 .bind_mounts
= bind_mounts
,
3129 .n_bind_mounts
= n_bind_mounts
,
3131 .temporary_filesystems
= context
->temporary_filesystems
,
3132 .n_temporary_filesystems
= context
->n_temporary_filesystems
,
3134 .mount_images
= context
->mount_images
,
3135 .n_mount_images
= context
->n_mount_images
,
3136 .mount_image_policy
= context
->mount_image_policy
?: &image_policy_service
,
3139 .var_tmp_dir
= var_tmp_dir
,
3141 .creds_path
= creds_path
,
3142 .log_namespace
= context
->log_namespace
,
3143 .mount_propagation_flag
= context
->mount_propagation_flag
,
3147 .extension_images
= context
->extension_images
,
3148 .n_extension_images
= context
->n_extension_images
,
3149 .extension_image_policy
= context
->extension_image_policy
?: &image_policy_sysext
,
3150 .extension_directories
= context
->extension_directories
,
3152 .propagate_dir
= propagate_dir
,
3153 .incoming_dir
= incoming_dir
,
3154 .extension_dir
= extension_dir
,
3155 .notify_socket
= root_dir
|| root_image
? params
->notify_socket
: NULL
,
3156 .host_os_release_stage
= host_os_release_stage
,
3158 /* If DynamicUser=no and RootDirectory= is set then lets pass a relaxed sandbox info,
3159 * otherwise enforce it, don't ignore protected paths and fail if we are enable to apply the
3160 * sandbox inside the mount namespace. */
3161 .ignore_protect_paths
= !needs_sandboxing
&& !context
->dynamic_user
&& root_dir
,
3163 .protect_control_groups
= needs_sandboxing
&& context
->protect_control_groups
,
3164 .protect_kernel_tunables
= needs_sandboxing
&& context
->protect_kernel_tunables
,
3165 .protect_kernel_modules
= needs_sandboxing
&& context
->protect_kernel_modules
,
3166 .protect_kernel_logs
= needs_sandboxing
&& context
->protect_kernel_logs
,
3167 .protect_hostname
= needs_sandboxing
&& context
->protect_hostname
,
3169 .private_dev
= needs_sandboxing
&& context
->private_devices
,
3170 .private_network
= needs_sandboxing
&& exec_needs_network_namespace(context
),
3171 .private_ipc
= needs_sandboxing
&& exec_needs_ipc_namespace(context
),
3173 .mount_apivfs
= needs_sandboxing
&& exec_context_get_effective_mount_apivfs(context
),
3175 /* If NNP is on, we can turn on MS_NOSUID, since it won't have any effect anymore. */
3176 .mount_nosuid
= needs_sandboxing
&& context
->no_new_privileges
&& !mac_selinux_use(),
3178 .protect_home
= needs_sandboxing
? context
->protect_home
: false,
3179 .protect_system
= needs_sandboxing
? context
->protect_system
: false,
3180 .protect_proc
= needs_sandboxing
? context
->protect_proc
: false,
3181 .proc_subset
= needs_sandboxing
? context
->proc_subset
: false,
3184 r
= setup_namespace(¶meters
, error_path
);
3185 /* If we couldn't set up the namespace this is probably due to a missing capability. setup_namespace() reports
3186 * that with a special, recognizable error ENOANO. In this case, silently proceed, but only if exclusively
3187 * sandboxing options were used, i.e. nothing such as RootDirectory= or BindMount= that would result in a
3188 * completely different execution environment. */
3190 if (insist_on_sandboxing(
3192 root_dir
, root_image
,
3195 return log_exec_debug_errno(context
,
3197 SYNTHETIC_ERRNO(EOPNOTSUPP
),
3198 "Failed to set up namespace, and refusing to continue since "
3199 "the selected namespacing options alter mount environment non-trivially.\n"
3200 "Bind mounts: %zu, temporary filesystems: %zu, root directory: %s, root image: %s, dynamic user: %s",
3202 context
->n_temporary_filesystems
,
3205 yes_no(context
->dynamic_user
));
3207 log_exec_debug(context
, params
, "Failed to set up namespace, assuming containerized execution and ignoring.");
3214 static int apply_working_directory(
3215 const ExecContext
*context
,
3216 const ExecParameters
*params
,
3217 ExecRuntime
*runtime
,
3224 assert(exit_status
);
3226 if (context
->working_directory_home
) {
3229 *exit_status
= EXIT_CHDIR
;
3236 wd
= empty_to_root(context
->working_directory
);
3238 if (params
->flags
& EXEC_APPLY_CHROOT
)
3241 d
= prefix_roota((runtime
? runtime
->ephemeral_copy
: NULL
) ?: context
->root_directory
, wd
);
3243 if (chdir(d
) < 0 && !context
->working_directory_missing_ok
) {
3244 *exit_status
= EXIT_CHDIR
;
3251 static int apply_root_directory(
3252 const ExecContext
*context
,
3253 const ExecParameters
*params
,
3254 ExecRuntime
*runtime
,
3255 const bool needs_mount_ns
,
3259 assert(exit_status
);
3261 if (params
->flags
& EXEC_APPLY_CHROOT
)
3262 if (!needs_mount_ns
&& context
->root_directory
)
3263 if (chroot((runtime
? runtime
->ephemeral_copy
: NULL
) ?: context
->root_directory
) < 0) {
3264 *exit_status
= EXIT_CHROOT
;
3271 static int setup_keyring(
3272 const ExecContext
*context
,
3273 const ExecParameters
*p
,
3274 uid_t uid
, gid_t gid
) {
3276 key_serial_t keyring
;
3284 /* Let's set up a new per-service "session" kernel keyring for each system service. This has the benefit that
3285 * each service runs with its own keyring shared among all processes of the service, but with no hook-up beyond
3286 * that scope, and in particular no link to the per-UID keyring. If we don't do this the keyring will be
3287 * automatically created on-demand and then linked to the per-UID keyring, by the kernel. The kernel's built-in
3288 * on-demand behaviour is very appropriate for login users, but probably not so much for system services, where
3289 * UIDs are not necessarily specific to a service but reused (at least in the case of UID 0). */
3291 if (context
->keyring_mode
== EXEC_KEYRING_INHERIT
)
3294 /* Acquiring a reference to the user keyring is nasty. We briefly change identity in order to get things set up
3295 * properly by the kernel. If we don't do that then we can't create it atomically, and that sucks for parallel
3296 * execution. This mimics what pam_keyinit does, too. Setting up session keyring, to be owned by the right user
3297 * & group is just as nasty as acquiring a reference to the user keyring. */
3299 saved_uid
= getuid();
3300 saved_gid
= getgid();
3302 if (gid_is_valid(gid
) && gid
!= saved_gid
) {
3303 if (setregid(gid
, -1) < 0)
3304 return log_exec_error_errno(context
,
3307 "Failed to change GID for user keyring: %m");
3310 if (uid_is_valid(uid
) && uid
!= saved_uid
) {
3311 if (setreuid(uid
, -1) < 0) {
3312 r
= log_exec_error_errno(context
,
3315 "Failed to change UID for user keyring: %m");
3320 keyring
= keyctl(KEYCTL_JOIN_SESSION_KEYRING
, 0, 0, 0, 0);
3321 if (keyring
== -1) {
3322 if (errno
== ENOSYS
)
3323 log_exec_debug_errno(context
,
3326 "Kernel keyring not supported, ignoring.");
3327 else if (ERRNO_IS_PRIVILEGE(errno
))
3328 log_exec_debug_errno(context
,
3331 "Kernel keyring access prohibited, ignoring.");
3332 else if (errno
== EDQUOT
)
3333 log_exec_debug_errno(context
,
3336 "Out of kernel keyrings to allocate, ignoring.");
3338 r
= log_exec_error_errno(context
,
3341 "Setting up kernel keyring failed: %m");
3346 /* When requested link the user keyring into the session keyring. */
3347 if (context
->keyring_mode
== EXEC_KEYRING_SHARED
) {
3349 if (keyctl(KEYCTL_LINK
,
3350 KEY_SPEC_USER_KEYRING
,
3351 KEY_SPEC_SESSION_KEYRING
, 0, 0) < 0) {
3352 r
= log_exec_error_errno(context
,
3355 "Failed to link user keyring into session keyring: %m");
3360 /* Restore uid/gid back */
3361 if (uid_is_valid(uid
) && uid
!= saved_uid
) {
3362 if (setreuid(saved_uid
, -1) < 0) {
3363 r
= log_exec_error_errno(context
,
3366 "Failed to change UID back for user keyring: %m");
3371 if (gid_is_valid(gid
) && gid
!= saved_gid
) {
3372 if (setregid(saved_gid
, -1) < 0)
3373 return log_exec_error_errno(context
,
3376 "Failed to change GID back for user keyring: %m");
3379 /* Populate they keyring with the invocation ID by default, as original saved_uid. */
3380 if (!sd_id128_is_null(p
->invocation_id
)) {
3383 key
= add_key("user",
3386 sizeof(p
->invocation_id
),
3387 KEY_SPEC_SESSION_KEYRING
);
3389 log_exec_debug_errno(context
,
3392 "Failed to add invocation ID to keyring, ignoring: %m");
3394 if (keyctl(KEYCTL_SETPERM
, key
,
3395 KEY_POS_VIEW
|KEY_POS_READ
|KEY_POS_SEARCH
|
3396 KEY_USR_VIEW
|KEY_USR_READ
|KEY_USR_SEARCH
, 0, 0) < 0)
3397 r
= log_exec_error_errno(context
,
3400 "Failed to restrict invocation ID permission: %m");
3405 /* Revert back uid & gid for the last time, and exit */
3406 /* no extra logging, as only the first already reported error matters */
3407 if (getuid() != saved_uid
)
3408 (void) setreuid(saved_uid
, -1);
3410 if (getgid() != saved_gid
)
3411 (void) setregid(saved_gid
, -1);
3416 static void append_socket_pair(int *array
, size_t *n
, const int pair
[static 2]) {
3422 array
[(*n
)++] = pair
[0];
3424 array
[(*n
)++] = pair
[1];
3427 static int close_remaining_fds(
3428 const ExecParameters
*params
,
3429 const ExecRuntime
*runtime
,
3431 const int *fds
, size_t n_fds
) {
3433 size_t n_dont_close
= 0;
3434 int dont_close
[n_fds
+ 14];
3438 if (params
->stdin_fd
>= 0)
3439 dont_close
[n_dont_close
++] = params
->stdin_fd
;
3440 if (params
->stdout_fd
>= 0)
3441 dont_close
[n_dont_close
++] = params
->stdout_fd
;
3442 if (params
->stderr_fd
>= 0)
3443 dont_close
[n_dont_close
++] = params
->stderr_fd
;
3446 dont_close
[n_dont_close
++] = socket_fd
;
3448 memcpy(dont_close
+ n_dont_close
, fds
, sizeof(int) * n_fds
);
3449 n_dont_close
+= n_fds
;
3453 append_socket_pair(dont_close
, &n_dont_close
, runtime
->ephemeral_storage_socket
);
3455 if (runtime
&& runtime
->shared
) {
3456 append_socket_pair(dont_close
, &n_dont_close
, runtime
->shared
->netns_storage_socket
);
3457 append_socket_pair(dont_close
, &n_dont_close
, runtime
->shared
->ipcns_storage_socket
);
3460 if (runtime
&& runtime
->dynamic_creds
) {
3461 if (runtime
->dynamic_creds
->user
)
3462 append_socket_pair(dont_close
, &n_dont_close
, runtime
->dynamic_creds
->user
->storage_socket
);
3463 if (runtime
->dynamic_creds
->group
)
3464 append_socket_pair(dont_close
, &n_dont_close
, runtime
->dynamic_creds
->group
->storage_socket
);
3467 if (params
->user_lookup_fd
>= 0)
3468 dont_close
[n_dont_close
++] = params
->user_lookup_fd
;
3470 return close_all_fds(dont_close
, n_dont_close
);
3473 static int send_user_lookup(
3474 const char *unit_id
,
3481 /* Send the resolved UID/GID to PID 1 after we learnt it. We send a single datagram, containing the UID/GID
3482 * data as well as the unit name. Note that we suppress sending this if no user/group to resolve was
3485 if (user_lookup_fd
< 0)
3488 if (!uid_is_valid(uid
) && !gid_is_valid(gid
))
3491 if (writev(user_lookup_fd
,
3493 IOVEC_MAKE(&uid
, sizeof(uid
)),
3494 IOVEC_MAKE(&gid
, sizeof(gid
)),
3495 IOVEC_MAKE_STRING(unit_id
) }, 3) < 0)
3501 static int acquire_home(const ExecContext
*c
, uid_t uid
, const char** home
, char **buf
) {
3508 /* If WorkingDirectory=~ is set, try to acquire a usable home directory. */
3513 if (!c
->working_directory_home
)
3516 r
= get_home_dir(buf
);
3524 static int compile_suggested_paths(const ExecContext
*c
, const ExecParameters
*p
, char ***ret
) {
3525 _cleanup_strv_free_
char ** list
= NULL
;
3532 assert(c
->dynamic_user
);
3534 /* Compile a list of paths that it might make sense to read the owning UID from to use as initial candidate for
3535 * dynamic UID allocation, in order to save us from doing costly recursive chown()s of the special
3538 for (ExecDirectoryType t
= 0; t
< _EXEC_DIRECTORY_TYPE_MAX
; t
++) {
3539 if (t
== EXEC_DIRECTORY_CONFIGURATION
)
3545 for (size_t i
= 0; i
< c
->directories
[t
].n_items
; i
++) {
3548 if (exec_directory_is_private(c
, t
))
3549 e
= path_join(p
->prefix
[t
], "private", c
->directories
[t
].items
[i
].path
);
3551 e
= path_join(p
->prefix
[t
], c
->directories
[t
].items
[i
].path
);
3555 r
= strv_consume(&list
, e
);
3561 *ret
= TAKE_PTR(list
);
3566 static int exec_context_cpu_affinity_from_numa(const ExecContext
*c
, CPUSet
*ret
) {
3567 _cleanup_(cpu_set_reset
) CPUSet s
= {};
3573 if (!c
->numa_policy
.nodes
.set
) {
3574 log_debug("Can't derive CPU affinity mask from NUMA mask because NUMA mask is not set, ignoring");
3578 r
= numa_to_cpu_set(&c
->numa_policy
, &s
);
3584 return cpu_set_add_all(ret
, &s
);
3587 static int add_shifted_fd(int *fds
, size_t fds_size
, size_t *n_fds
, int *fd
) {
3592 assert(*n_fds
< fds_size
);
3598 if (*fd
< 3 + (int) *n_fds
) {
3599 /* Let's move the fd up, so that it's outside of the fd range we will use to store
3600 * the fds we pass to the process (or which are closed only during execve). */
3602 r
= fcntl(*fd
, F_DUPFD_CLOEXEC
, 3 + (int) *n_fds
);
3606 close_and_replace(*fd
, r
);
3609 fds
[(*n_fds
)++] = *fd
;
3613 static int connect_unix_harder(const ExecContext
*c
, const ExecParameters
*p
, const OpenFile
*of
, int ofd
) {
3614 union sockaddr_union addr
= {
3615 .un
.sun_family
= AF_UNIX
,
3618 static const int socket_types
[] = { SOCK_DGRAM
, SOCK_STREAM
, SOCK_SEQPACKET
};
3626 r
= sockaddr_un_set_path(&addr
.un
, FORMAT_PROC_FD_PATH(ofd
));
3628 return log_exec_error_errno(c
, p
, r
, "Failed to set sockaddr for %s: %m", of
->path
);
3632 for (size_t i
= 0; i
< ELEMENTSOF(socket_types
); i
++) {
3633 _cleanup_close_
int fd
= -EBADF
;
3635 fd
= socket(AF_UNIX
, socket_types
[i
] | SOCK_CLOEXEC
, 0);
3637 return log_exec_error_errno(c
,
3640 "Failed to create socket for %s: %m",
3643 r
= RET_NERRNO(connect(fd
, &addr
.sa
, sa_len
));
3644 if (r
== -EPROTOTYPE
)
3647 return log_exec_error_errno(c
,
3650 "Failed to connect socket for %s: %m",
3656 return log_exec_error_errno(c
,
3658 SYNTHETIC_ERRNO(EPROTOTYPE
), "Failed to connect socket for \"%s\".",
3662 static int get_open_file_fd(const ExecContext
*c
, const ExecParameters
*p
, const OpenFile
*of
) {
3664 _cleanup_close_
int fd
= -EBADF
, ofd
= -EBADF
;
3670 ofd
= open(of
->path
, O_PATH
| O_CLOEXEC
);
3672 return log_exec_error_errno(c
, p
, errno
, "Could not open \"%s\": %m", of
->path
);
3674 if (fstat(ofd
, &st
) < 0)
3675 return log_exec_error_errno(c
, p
, errno
, "Failed to stat %s: %m", of
->path
);
3677 if (S_ISSOCK(st
.st_mode
)) {
3678 fd
= connect_unix_harder(c
, p
, of
, ofd
);
3682 if (FLAGS_SET(of
->flags
, OPENFILE_READ_ONLY
) && shutdown(fd
, SHUT_WR
) < 0)
3683 return log_exec_error_errno(c
, p
, errno
, "Failed to shutdown send for socket %s: %m",
3686 log_exec_debug(c
, p
, "socket %s opened (fd=%d)", of
->path
, fd
);
3688 int flags
= FLAGS_SET(of
->flags
, OPENFILE_READ_ONLY
) ? O_RDONLY
: O_RDWR
;
3689 if (FLAGS_SET(of
->flags
, OPENFILE_APPEND
))
3691 else if (FLAGS_SET(of
->flags
, OPENFILE_TRUNCATE
))
3694 fd
= fd_reopen(ofd
, flags
| O_CLOEXEC
);
3696 return log_exec_error_errno(c
, p
, fd
, "Failed to open file %s: %m", of
->path
);
3698 log_exec_debug(c
, p
, "file %s opened (fd=%d)", of
->path
, fd
);
3704 static int collect_open_file_fds(const ExecContext
*c
, ExecParameters
*p
, size_t *n_fds
) {
3711 LIST_FOREACH(open_files
, of
, p
->open_files
) {
3712 _cleanup_close_
int fd
= -EBADF
;
3714 fd
= get_open_file_fd(c
, p
, of
);
3716 if (FLAGS_SET(of
->flags
, OPENFILE_GRACEFUL
)) {
3717 log_exec_debug_errno(c
, p
, fd
, "Failed to get OpenFile= file descriptor for %s, ignoring: %m", of
->path
);
3724 if (!GREEDY_REALLOC(p
->fds
, *n_fds
+ 1))
3727 r
= strv_extend(&p
->fd_names
, of
->fdname
);
3731 p
->fds
[*n_fds
] = TAKE_FD(fd
);
3739 static void log_command_line(
3740 const ExecContext
*context
,
3741 const ExecParameters
*params
,
3743 const char *executable
,
3754 _cleanup_free_
char *cmdline
= quote_command_line(argv
, SHELL_ESCAPE_EMPTY
);
3756 log_exec_struct(context
, params
, LOG_DEBUG
,
3757 "EXECUTABLE=%s", executable
,
3758 LOG_EXEC_MESSAGE(params
, "%s: %s", msg
, strnull(cmdline
)),
3759 LOG_EXEC_INVOCATION_ID(params
));
3762 static bool exec_context_need_unprivileged_private_users(
3763 const ExecContext
*context
,
3764 const ExecParameters
*params
) {
3769 /* These options require PrivateUsers= when used in user units, as we need to be in a user namespace
3770 * to have permission to enable them when not running as root. If we have effective CAP_SYS_ADMIN
3771 * (system manager) then we have privileges and don't need this. */
3772 if (params
->runtime_scope
!= RUNTIME_SCOPE_USER
)
3775 return context
->private_users
||
3776 context
->private_tmp
||
3777 context
->private_devices
||
3778 context
->private_network
||
3779 context
->network_namespace_path
||
3780 context
->private_ipc
||
3781 context
->ipc_namespace_path
||
3782 context
->private_mounts
> 0 ||
3783 context
->mount_apivfs
||
3784 context
->n_bind_mounts
> 0 ||
3785 context
->n_temporary_filesystems
> 0 ||
3786 context
->root_directory
||
3787 !strv_isempty(context
->extension_directories
) ||
3788 context
->protect_system
!= PROTECT_SYSTEM_NO
||
3789 context
->protect_home
!= PROTECT_HOME_NO
||
3790 context
->protect_kernel_tunables
||
3791 context
->protect_kernel_modules
||
3792 context
->protect_kernel_logs
||
3793 context
->protect_control_groups
||
3794 context
->protect_clock
||
3795 context
->protect_hostname
||
3796 !strv_isempty(context
->read_write_paths
) ||
3797 !strv_isempty(context
->read_only_paths
) ||
3798 !strv_isempty(context
->inaccessible_paths
) ||
3799 !strv_isempty(context
->exec_paths
) ||
3800 !strv_isempty(context
->no_exec_paths
);
3803 static bool exec_context_shall_confirm_spawn(const ExecContext
*context
) {
3806 if (confirm_spawn_disabled())
3809 /* For some reasons units remaining in the same process group
3810 * as PID 1 fail to acquire the console even if it's not used
3811 * by any process. So skip the confirmation question for them. */
3812 return !context
->same_pgrp
;
3815 static int exec_context_named_iofds(
3816 const ExecContext
*c
,
3817 const ExecParameters
*p
,
3818 int named_iofds
[static 3]) {
3821 const char* stdio_fdname
[3];
3826 assert(named_iofds
);
3828 targets
= (c
->std_input
== EXEC_INPUT_NAMED_FD
) +
3829 (c
->std_output
== EXEC_OUTPUT_NAMED_FD
) +
3830 (c
->std_error
== EXEC_OUTPUT_NAMED_FD
);
3832 for (size_t i
= 0; i
< 3; i
++)
3833 stdio_fdname
[i
] = exec_context_fdname(c
, i
);
3835 n_fds
= p
->n_storage_fds
+ p
->n_socket_fds
;
3837 for (size_t i
= 0; i
< n_fds
&& targets
> 0; i
++)
3838 if (named_iofds
[STDIN_FILENO
] < 0 &&
3839 c
->std_input
== EXEC_INPUT_NAMED_FD
&&
3840 stdio_fdname
[STDIN_FILENO
] &&
3841 streq(p
->fd_names
[i
], stdio_fdname
[STDIN_FILENO
])) {
3843 named_iofds
[STDIN_FILENO
] = p
->fds
[i
];
3846 } else if (named_iofds
[STDOUT_FILENO
] < 0 &&
3847 c
->std_output
== EXEC_OUTPUT_NAMED_FD
&&
3848 stdio_fdname
[STDOUT_FILENO
] &&
3849 streq(p
->fd_names
[i
], stdio_fdname
[STDOUT_FILENO
])) {
3851 named_iofds
[STDOUT_FILENO
] = p
->fds
[i
];
3854 } else if (named_iofds
[STDERR_FILENO
] < 0 &&
3855 c
->std_error
== EXEC_OUTPUT_NAMED_FD
&&
3856 stdio_fdname
[STDERR_FILENO
] &&
3857 streq(p
->fd_names
[i
], stdio_fdname
[STDERR_FILENO
])) {
3859 named_iofds
[STDERR_FILENO
] = p
->fds
[i
];
3863 return targets
== 0 ? 0 : -ENOENT
;
3866 static void exec_shared_runtime_close(ExecSharedRuntime
*shared
) {
3870 safe_close_pair(shared
->netns_storage_socket
);
3871 safe_close_pair(shared
->ipcns_storage_socket
);
3874 static void exec_runtime_close(ExecRuntime
*rt
) {
3878 safe_close_pair(rt
->ephemeral_storage_socket
);
3880 exec_shared_runtime_close(rt
->shared
);
3881 dynamic_creds_close(rt
->dynamic_creds
);
3884 static void exec_params_close(ExecParameters
*p
) {
3888 p
->stdin_fd
= safe_close(p
->stdin_fd
);
3889 p
->stdout_fd
= safe_close(p
->stdout_fd
);
3890 p
->stderr_fd
= safe_close(p
->stderr_fd
);
3894 const ExecCommand
*command
,
3895 const ExecContext
*context
,
3896 ExecParameters
*params
,
3897 ExecRuntime
*runtime
,
3898 const CGroupContext
*cgroup_context
,
3901 _cleanup_strv_free_
char **our_env
= NULL
, **pass_env
= NULL
, **joined_exec_search_path
= NULL
, **accum_env
= NULL
, **replaced_argv
= NULL
;
3903 _cleanup_free_ gid_t
*supplementary_gids
= NULL
;
3904 const char *username
= NULL
, *groupname
= NULL
;
3905 _cleanup_free_
char *home_buffer
= NULL
, *memory_pressure_path
= NULL
;
3906 const char *home
= NULL
, *shell
= NULL
;
3907 char **final_argv
= NULL
;
3908 dev_t journal_stream_dev
= 0;
3909 ino_t journal_stream_ino
= 0;
3910 bool userns_set_up
= false;
3911 bool needs_sandboxing
, /* Do we need to set up full sandboxing? (i.e. all namespacing, all MAC stuff, caps, yadda yadda */
3912 needs_setuid
, /* Do we need to do the actual setresuid()/setresgid() calls? */
3913 needs_mount_namespace
, /* Do we need to set up a mount namespace for this kernel? */
3914 needs_ambient_hack
; /* Do we need to apply the ambient capabilities hack? */
3915 bool keep_seccomp_privileges
= false;
3917 _cleanup_free_
char *mac_selinux_context_net
= NULL
;
3918 bool use_selinux
= false;
3921 bool use_smack
= false;
3924 bool use_apparmor
= false;
3927 uint64_t saved_bset
= 0;
3929 uid_t saved_uid
= getuid();
3930 gid_t saved_gid
= getgid();
3931 uid_t uid
= UID_INVALID
;
3932 gid_t gid
= GID_INVALID
;
3933 size_t n_fds
, /* fds to pass to the child */
3934 n_keep_fds
; /* total number of fds not to close */
3936 _cleanup_free_ gid_t
*gids_after_pam
= NULL
;
3937 int ngids_after_pam
= 0;
3939 int socket_fd
= -EBADF
, named_iofds
[3] = EBADF_TRIPLET
;
3940 size_t n_storage_fds
, n_socket_fds
;
3945 assert(exit_status
);
3947 /* This should be mostly redundant, as the log level is also passed as an argument of the executor,
3948 * and is already applied earlier. Just for safety. */
3949 if (context
->log_level_max
>= 0)
3950 log_set_max_level(context
->log_level_max
);
3952 /* Explicitly test for CVE-2021-4034 inspired invocations */
3953 if (!command
->path
|| strv_isempty(command
->argv
)) {
3954 *exit_status
= EXIT_EXEC
;
3955 return log_exec_error_errno(
3958 SYNTHETIC_ERRNO(EINVAL
),
3959 "Invalid command line arguments.");
3962 LOG_CONTEXT_PUSH_EXEC(context
, params
);
3964 if (context
->std_input
== EXEC_INPUT_SOCKET
||
3965 context
->std_output
== EXEC_OUTPUT_SOCKET
||
3966 context
->std_error
== EXEC_OUTPUT_SOCKET
) {
3968 if (params
->n_socket_fds
> 1)
3969 return log_exec_error_errno(context
, params
, SYNTHETIC_ERRNO(EINVAL
), "Got more than one socket.");
3971 if (params
->n_socket_fds
== 0)
3972 return log_exec_error_errno(context
, params
, SYNTHETIC_ERRNO(EINVAL
), "Got no socket.");
3974 socket_fd
= params
->fds
[0];
3975 n_storage_fds
= n_socket_fds
= 0;
3977 n_socket_fds
= params
->n_socket_fds
;
3978 n_storage_fds
= params
->n_storage_fds
;
3980 n_fds
= n_socket_fds
+ n_storage_fds
;
3982 r
= exec_context_named_iofds(context
, params
, named_iofds
);
3984 return log_exec_error_errno(context
, params
, r
, "Failed to load a named file descriptor: %m");
3986 rename_process_from_path(command
->path
);
3988 /* We reset exactly these signals, since they are the only ones we set to SIG_IGN in the main
3989 * daemon. All others we leave untouched because we set them to SIG_DFL or a valid handler initially,
3990 * both of which will be demoted to SIG_DFL. */
3991 (void) default_signals(SIGNALS_CRASH_HANDLER
,
3994 if (context
->ignore_sigpipe
)
3995 (void) ignore_signals(SIGPIPE
);
3997 r
= reset_signal_mask();
3999 *exit_status
= EXIT_SIGNAL_MASK
;
4000 return log_exec_error_errno(context
, params
, r
, "Failed to set process signal mask: %m");
4003 if (params
->idle_pipe
)
4004 do_idle_pipe_dance(params
->idle_pipe
);
4006 /* Close fds we don't need very early to make sure we don't block init reexecution because it cannot bind its
4007 * sockets. Among the fds we close are the logging fds, and we want to keep them closed, so that we don't have
4008 * any fds open we don't really want open during the transition. In order to make logging work, we switch the
4009 * log subsystem into open_when_needed mode, so that it reopens the logs on every single log call. */
4012 log_set_open_when_needed(true);
4013 log_settle_target();
4015 /* In case anything used libc syslog(), close this here, too */
4018 r
= collect_open_file_fds(context
, params
, &n_fds
);
4020 *exit_status
= EXIT_FDS
;
4021 return log_exec_error_errno(context
, params
, r
, "Failed to get OpenFile= file descriptors: %m");
4024 int keep_fds
[n_fds
+ 3];
4025 memcpy_safe(keep_fds
, params
->fds
, n_fds
* sizeof(int));
4028 r
= add_shifted_fd(keep_fds
, ELEMENTSOF(keep_fds
), &n_keep_fds
, ¶ms
->exec_fd
);
4030 *exit_status
= EXIT_FDS
;
4031 return log_exec_error_errno(context
, params
, r
, "Failed to collect shifted fd: %m");
4035 r
= add_shifted_fd(keep_fds
, ELEMENTSOF(keep_fds
), &n_keep_fds
, ¶ms
->bpf_outer_map_fd
);
4037 *exit_status
= EXIT_FDS
;
4038 return log_exec_error_errno(context
, params
, r
, "Failed to collect shifted fd: %m");
4042 r
= close_remaining_fds(params
, runtime
, socket_fd
, keep_fds
, n_keep_fds
);
4044 *exit_status
= EXIT_FDS
;
4045 return log_exec_error_errno(context
, params
, r
, "Failed to close unwanted file descriptors: %m");
4048 if (!context
->same_pgrp
&&
4050 *exit_status
= EXIT_SETSID
;
4051 return log_exec_error_errno(context
, params
, errno
, "Failed to create new process session: %m");
4054 exec_context_tty_reset(context
, params
);
4056 if (params
->shall_confirm_spawn
&& exec_context_shall_confirm_spawn(context
)) {
4057 _cleanup_free_
char *cmdline
= NULL
;
4059 cmdline
= quote_command_line(command
->argv
, SHELL_ESCAPE_EMPTY
);
4061 *exit_status
= EXIT_MEMORY
;
4065 r
= ask_for_confirmation(context
, params
, cmdline
);
4066 if (r
!= CONFIRM_EXECUTE
) {
4067 if (r
== CONFIRM_PRETEND_SUCCESS
) {
4068 *exit_status
= EXIT_SUCCESS
;
4072 *exit_status
= EXIT_CONFIRM
;
4073 return log_exec_error_errno(context
, params
, SYNTHETIC_ERRNO(ECANCELED
),
4074 "Execution cancelled by the user");
4078 /* We are about to invoke NSS and PAM modules. Let's tell them what we are doing here, maybe they care. This is
4079 * used by nss-resolve to disable itself when we are about to start systemd-resolved, to avoid deadlocks. Note
4080 * that these env vars do not survive the execve(), which means they really only apply to the PAM and NSS
4081 * invocations themselves. Also note that while we'll only invoke NSS modules involved in user management they
4082 * might internally call into other NSS modules that are involved in hostname resolution, we never know. */
4083 if (setenv("SYSTEMD_ACTIVATION_UNIT", params
->unit_id
, true) != 0 ||
4084 setenv("SYSTEMD_ACTIVATION_SCOPE", runtime_scope_to_string(params
->runtime_scope
), true) != 0) {
4085 *exit_status
= EXIT_MEMORY
;
4086 return log_exec_error_errno(context
, params
, errno
, "Failed to update environment: %m");
4089 if (context
->dynamic_user
&& runtime
&& runtime
->dynamic_creds
) {
4090 _cleanup_strv_free_
char **suggested_paths
= NULL
;
4092 /* On top of that, make sure we bypass our own NSS module nss-systemd comprehensively for any NSS
4093 * checks, if DynamicUser=1 is used, as we shouldn't create a feedback loop with ourselves here. */
4094 if (putenv((char*) "SYSTEMD_NSS_DYNAMIC_BYPASS=1") != 0) {
4095 *exit_status
= EXIT_USER
;
4096 return log_exec_error_errno(context
, params
, errno
, "Failed to update environment: %m");
4099 r
= compile_suggested_paths(context
, params
, &suggested_paths
);
4101 *exit_status
= EXIT_MEMORY
;
4105 r
= dynamic_creds_realize(runtime
->dynamic_creds
, suggested_paths
, &uid
, &gid
);
4107 *exit_status
= EXIT_USER
;
4109 return log_exec_error_errno(context
, params
, SYNTHETIC_ERRNO(EOPNOTSUPP
),
4110 "Failed to update dynamic user credentials: User or group with specified name already exists.");
4111 return log_exec_error_errno(context
, params
, r
, "Failed to update dynamic user credentials: %m");
4114 if (!uid_is_valid(uid
)) {
4115 *exit_status
= EXIT_USER
;
4116 return log_exec_error_errno(context
, params
, SYNTHETIC_ERRNO(ESRCH
), "UID validation failed for \""UID_FMT
"\"", uid
);
4119 if (!gid_is_valid(gid
)) {
4120 *exit_status
= EXIT_USER
;
4121 return log_exec_error_errno(context
, params
, SYNTHETIC_ERRNO(ESRCH
), "GID validation failed for \""GID_FMT
"\"", gid
);
4124 if (runtime
->dynamic_creds
->user
)
4125 username
= runtime
->dynamic_creds
->user
->name
;
4128 if (context
->user
) {
4129 r
= get_fixed_user(context
->user
, &username
, &uid
, &gid
, &home
, &shell
);
4131 *exit_status
= EXIT_USER
;
4132 return log_exec_error_errno(context
, params
, r
, "Failed to determine user credentials: %m");
4136 if (context
->group
) {
4137 r
= get_fixed_group(context
->group
, &groupname
, &gid
);
4139 *exit_status
= EXIT_GROUP
;
4140 return log_exec_error_errno(context
, params
, r
, "Failed to determine group credentials: %m");
4145 /* Initialize user supplementary groups and get SupplementaryGroups= ones */
4146 r
= get_supplementary_groups(context
, username
, groupname
, gid
,
4147 &supplementary_gids
, &ngids
);
4149 *exit_status
= EXIT_GROUP
;
4150 return log_exec_error_errno(context
, params
, r
, "Failed to determine supplementary groups: %m");
4153 r
= send_user_lookup(params
->unit_id
, params
->user_lookup_fd
, uid
, gid
);
4155 *exit_status
= EXIT_USER
;
4156 return log_exec_error_errno(context
, params
, r
, "Failed to send user credentials to PID1: %m");
4159 params
->user_lookup_fd
= safe_close(params
->user_lookup_fd
);
4161 r
= acquire_home(context
, uid
, &home
, &home_buffer
);
4163 *exit_status
= EXIT_CHDIR
;
4164 return log_exec_error_errno(context
, params
, r
, "Failed to determine $HOME for user: %m");
4167 /* If a socket is connected to STDIN/STDOUT/STDERR, we must drop O_NONBLOCK */
4169 (void) fd_nonblock(socket_fd
, false);
4171 /* Journald will try to look-up our cgroup in order to populate _SYSTEMD_CGROUP and _SYSTEMD_UNIT fields.
4172 * Hence we need to migrate to the target cgroup from init.scope before connecting to journald */
4173 if (params
->cgroup_path
) {
4174 _cleanup_free_
char *p
= NULL
;
4176 r
= exec_params_get_cgroup_path(params
, cgroup_context
, &p
);
4178 *exit_status
= EXIT_CGROUP
;
4179 return log_exec_error_errno(context
, params
, r
, "Failed to acquire cgroup path: %m");
4182 r
= cg_attach_everywhere(params
->cgroup_supported
, p
, 0, NULL
, NULL
);
4183 if (r
== -EUCLEAN
) {
4184 *exit_status
= EXIT_CGROUP
;
4185 return log_exec_error_errno(context
, params
, r
, "Failed to attach process to cgroup %s "
4186 "because the cgroup or one of its parents or "
4187 "siblings is in the threaded mode: %m", p
);
4190 *exit_status
= EXIT_CGROUP
;
4191 return log_exec_error_errno(context
, params
, r
, "Failed to attach to cgroup %s: %m", p
);
4195 if (context
->network_namespace_path
&& runtime
&& runtime
->shared
&& runtime
->shared
->netns_storage_socket
[0] >= 0) {
4196 r
= open_shareable_ns_path(runtime
->shared
->netns_storage_socket
, context
->network_namespace_path
, CLONE_NEWNET
);
4198 *exit_status
= EXIT_NETWORK
;
4199 return log_exec_error_errno(context
, params
, r
, "Failed to open network namespace path %s: %m", context
->network_namespace_path
);
4203 if (context
->ipc_namespace_path
&& runtime
&& runtime
->shared
&& runtime
->shared
->ipcns_storage_socket
[0] >= 0) {
4204 r
= open_shareable_ns_path(runtime
->shared
->ipcns_storage_socket
, context
->ipc_namespace_path
, CLONE_NEWIPC
);
4206 *exit_status
= EXIT_NAMESPACE
;
4207 return log_exec_error_errno(context
, params
, r
, "Failed to open IPC namespace path %s: %m", context
->ipc_namespace_path
);
4211 r
= setup_input(context
, params
, socket_fd
, named_iofds
);
4213 *exit_status
= EXIT_STDIN
;
4214 return log_exec_error_errno(context
, params
, r
, "Failed to set up standard input: %m");
4217 r
= setup_output(context
, params
, STDOUT_FILENO
, socket_fd
, named_iofds
, basename(command
->path
), uid
, gid
, &journal_stream_dev
, &journal_stream_ino
);
4219 *exit_status
= EXIT_STDOUT
;
4220 return log_exec_error_errno(context
, params
, r
, "Failed to set up standard output: %m");
4223 r
= setup_output(context
, params
, STDERR_FILENO
, socket_fd
, named_iofds
, basename(command
->path
), uid
, gid
, &journal_stream_dev
, &journal_stream_ino
);
4225 *exit_status
= EXIT_STDERR
;
4226 return log_exec_error_errno(context
, params
, r
, "Failed to set up standard error output: %m");
4229 if (context
->oom_score_adjust_set
) {
4230 /* When we can't make this change due to EPERM, then let's silently skip over it. User
4231 * namespaces prohibit write access to this file, and we shouldn't trip up over that. */
4232 r
= set_oom_score_adjust(context
->oom_score_adjust
);
4233 if (ERRNO_IS_NEG_PRIVILEGE(r
))
4234 log_exec_debug_errno(context
, params
, r
,
4235 "Failed to adjust OOM setting, assuming containerized execution, ignoring: %m");
4237 *exit_status
= EXIT_OOM_ADJUST
;
4238 return log_exec_error_errno(context
, params
, r
, "Failed to adjust OOM setting: %m");
4242 if (context
->coredump_filter_set
) {
4243 r
= set_coredump_filter(context
->coredump_filter
);
4244 if (ERRNO_IS_NEG_PRIVILEGE(r
))
4245 log_exec_debug_errno(context
, params
, r
, "Failed to adjust coredump_filter, ignoring: %m");
4247 *exit_status
= EXIT_LIMITS
;
4248 return log_exec_error_errno(context
, params
, r
, "Failed to adjust coredump_filter: %m");
4252 if (context
->nice_set
) {
4253 r
= setpriority_closest(context
->nice
);
4255 *exit_status
= EXIT_NICE
;
4256 return log_exec_error_errno(context
, params
, r
, "Failed to set up process scheduling priority (nice level): %m");
4260 if (context
->cpu_sched_set
) {
4261 struct sched_param param
= {
4262 .sched_priority
= context
->cpu_sched_priority
,
4265 r
= sched_setscheduler(0,
4266 context
->cpu_sched_policy
|
4267 (context
->cpu_sched_reset_on_fork
?
4268 SCHED_RESET_ON_FORK
: 0),
4271 *exit_status
= EXIT_SETSCHEDULER
;
4272 return log_exec_error_errno(context
, params
, errno
, "Failed to set up CPU scheduling: %m");
4276 if (context
->cpu_affinity_from_numa
|| context
->cpu_set
.set
) {
4277 _cleanup_(cpu_set_reset
) CPUSet converted_cpu_set
= {};
4278 const CPUSet
*cpu_set
;
4280 if (context
->cpu_affinity_from_numa
) {
4281 r
= exec_context_cpu_affinity_from_numa(context
, &converted_cpu_set
);
4283 *exit_status
= EXIT_CPUAFFINITY
;
4284 return log_exec_error_errno(context
, params
, r
, "Failed to derive CPU affinity mask from NUMA mask: %m");
4287 cpu_set
= &converted_cpu_set
;
4289 cpu_set
= &context
->cpu_set
;
4291 if (sched_setaffinity(0, cpu_set
->allocated
, cpu_set
->set
) < 0) {
4292 *exit_status
= EXIT_CPUAFFINITY
;
4293 return log_exec_error_errno(context
, params
, errno
, "Failed to set up CPU affinity: %m");
4297 if (mpol_is_valid(numa_policy_get_type(&context
->numa_policy
))) {
4298 r
= apply_numa_policy(&context
->numa_policy
);
4299 if (ERRNO_IS_NEG_NOT_SUPPORTED(r
))
4300 log_exec_debug_errno(context
, params
, r
, "NUMA support not available, ignoring.");
4302 *exit_status
= EXIT_NUMA_POLICY
;
4303 return log_exec_error_errno(context
, params
, r
, "Failed to set NUMA memory policy: %m");
4307 if (context
->ioprio_set
)
4308 if (ioprio_set(IOPRIO_WHO_PROCESS
, 0, context
->ioprio
) < 0) {
4309 *exit_status
= EXIT_IOPRIO
;
4310 return log_exec_error_errno(context
, params
, errno
, "Failed to set up IO scheduling priority: %m");
4313 if (context
->timer_slack_nsec
!= NSEC_INFINITY
)
4314 if (prctl(PR_SET_TIMERSLACK
, context
->timer_slack_nsec
) < 0) {
4315 *exit_status
= EXIT_TIMERSLACK
;
4316 return log_exec_error_errno(context
, params
, errno
, "Failed to set up timer slack: %m");
4319 if (context
->personality
!= PERSONALITY_INVALID
) {
4320 r
= safe_personality(context
->personality
);
4322 *exit_status
= EXIT_PERSONALITY
;
4323 return log_exec_error_errno(context
, params
, r
, "Failed to set up execution domain (personality): %m");
4328 if (context
->utmp_id
) {
4329 const char *line
= context
->tty_path
?
4330 (path_startswith(context
->tty_path
, "/dev/") ?: context
->tty_path
) :
4332 utmp_put_init_process(context
->utmp_id
, getpid_cached(), getsid(0),
4334 context
->utmp_mode
== EXEC_UTMP_INIT
? INIT_PROCESS
:
4335 context
->utmp_mode
== EXEC_UTMP_LOGIN
? LOGIN_PROCESS
:
4341 if (uid_is_valid(uid
)) {
4342 r
= chown_terminal(STDIN_FILENO
, uid
);
4344 *exit_status
= EXIT_STDIN
;
4345 return log_exec_error_errno(context
, params
, r
, "Failed to change ownership of terminal: %m");
4349 if (params
->cgroup_path
) {
4350 /* If delegation is enabled we'll pass ownership of the cgroup to the user of the new process. On cgroup v1
4351 * this is only about systemd's own hierarchy, i.e. not the controller hierarchies, simply because that's not
4352 * safe. On cgroup v2 there's only one hierarchy anyway, and delegation is safe there, hence in that case only
4353 * touch a single hierarchy too. */
4355 if (params
->flags
& EXEC_CGROUP_DELEGATE
) {
4356 _cleanup_free_
char *p
= NULL
;
4358 r
= cg_set_access(SYSTEMD_CGROUP_CONTROLLER
, params
->cgroup_path
, uid
, gid
);
4360 *exit_status
= EXIT_CGROUP
;
4361 return log_exec_error_errno(context
, params
, r
, "Failed to adjust control group access: %m");
4364 r
= exec_params_get_cgroup_path(params
, cgroup_context
, &p
);
4366 *exit_status
= EXIT_CGROUP
;
4367 return log_exec_error_errno(context
, params
, r
, "Failed to acquire cgroup path: %m");
4370 r
= cg_set_access_recursive(SYSTEMD_CGROUP_CONTROLLER
, p
, uid
, gid
);
4372 *exit_status
= EXIT_CGROUP
;
4373 return log_exec_error_errno(context
, params
, r
, "Failed to adjust control subgroup access: %m");
4378 if (cgroup_context
&& cg_unified() > 0 && is_pressure_supported() > 0) {
4379 if (cgroup_context_want_memory_pressure(cgroup_context
)) {
4380 r
= cg_get_path("memory", params
->cgroup_path
, "memory.pressure", &memory_pressure_path
);
4382 *exit_status
= EXIT_MEMORY
;
4386 r
= chmod_and_chown(memory_pressure_path
, 0644, uid
, gid
);
4388 log_exec_full_errno(context
, params
, r
== -ENOENT
|| ERRNO_IS_PRIVILEGE(r
) ? LOG_DEBUG
: LOG_WARNING
, r
,
4389 "Failed to adjust ownership of '%s', ignoring: %m", memory_pressure_path
);
4390 memory_pressure_path
= mfree(memory_pressure_path
);
4392 } else if (cgroup_context
->memory_pressure_watch
== CGROUP_PRESSURE_WATCH_OFF
) {
4393 memory_pressure_path
= strdup("/dev/null"); /* /dev/null is explicit indicator for turning of memory pressure watch */
4394 if (!memory_pressure_path
) {
4395 *exit_status
= EXIT_MEMORY
;
4402 needs_mount_namespace
= exec_needs_mount_namespace(context
, params
, runtime
);
4404 for (ExecDirectoryType dt
= 0; dt
< _EXEC_DIRECTORY_TYPE_MAX
; dt
++) {
4405 r
= setup_exec_directory(context
, params
, uid
, gid
, dt
, needs_mount_namespace
, exit_status
);
4407 return log_exec_error_errno(context
, params
, r
, "Failed to set up special execution directory in %s: %m", params
->prefix
[dt
]);
4410 if (FLAGS_SET(params
->flags
, EXEC_WRITE_CREDENTIALS
)) {
4411 r
= exec_setup_credentials(context
, params
, params
->unit_id
, uid
, gid
);
4413 *exit_status
= EXIT_CREDENTIALS
;
4414 return log_exec_error_errno(context
, params
, r
, "Failed to set up credentials: %m");
4418 r
= build_environment(
4428 memory_pressure_path
,
4431 *exit_status
= EXIT_MEMORY
;
4435 r
= build_pass_environment(context
, &pass_env
);
4437 *exit_status
= EXIT_MEMORY
;
4441 /* The $PATH variable is set to the default path in params->environment. However, this is overridden
4442 * if user-specified fields have $PATH set. The intention is to also override $PATH if the unit does
4443 * not specify PATH but the unit has ExecSearchPath. */
4444 if (!strv_isempty(context
->exec_search_path
)) {
4445 _cleanup_free_
char *joined
= NULL
;
4447 joined
= strv_join(context
->exec_search_path
, ":");
4449 *exit_status
= EXIT_MEMORY
;
4453 r
= strv_env_assign(&joined_exec_search_path
, "PATH", joined
);
4455 *exit_status
= EXIT_MEMORY
;
4460 accum_env
= strv_env_merge(params
->environment
,
4462 joined_exec_search_path
,
4464 context
->environment
,
4467 *exit_status
= EXIT_MEMORY
;
4470 accum_env
= strv_env_clean(accum_env
);
4472 (void) umask(context
->umask
);
4474 r
= setup_keyring(context
, params
, uid
, gid
);
4476 *exit_status
= EXIT_KEYRING
;
4477 return log_exec_error_errno(context
, params
, r
, "Failed to set up kernel keyring: %m");
4480 /* We need sandboxing if the caller asked us to apply it and the command isn't explicitly excepted
4482 needs_sandboxing
= (params
->flags
& EXEC_APPLY_SANDBOXING
) && !(command
->flags
& EXEC_COMMAND_FULLY_PRIVILEGED
);
4484 /* We need the ambient capability hack, if the caller asked us to apply it and the command is marked
4485 * for it, and the kernel doesn't actually support ambient caps. */
4486 needs_ambient_hack
= (params
->flags
& EXEC_APPLY_SANDBOXING
) && (command
->flags
& EXEC_COMMAND_AMBIENT_MAGIC
) && !ambient_capabilities_supported();
4488 /* We need setresuid() if the caller asked us to apply sandboxing and the command isn't explicitly
4489 * excepted from either whole sandboxing or just setresuid() itself, and the ambient hack is not
4491 if (needs_ambient_hack
)
4492 needs_setuid
= false;
4494 needs_setuid
= (params
->flags
& EXEC_APPLY_SANDBOXING
) && !(command
->flags
& (EXEC_COMMAND_FULLY_PRIVILEGED
|EXEC_COMMAND_NO_SETUID
));
4496 uint64_t capability_ambient_set
= context
->capability_ambient_set
;
4498 if (needs_sandboxing
) {
4499 /* MAC enablement checks need to be done before a new mount ns is created, as they rely on
4500 * /sys being present. The actual MAC context application will happen later, as late as
4501 * possible, to avoid impacting our own code paths. */
4504 use_selinux
= mac_selinux_use();
4507 use_smack
= mac_smack_use();
4510 use_apparmor
= mac_apparmor_use();
4514 if (needs_sandboxing
) {
4517 /* Let's set the resource limits before we call into PAM, so that pam_limits wins over what
4518 * is set here. (See below.) */
4520 r
= setrlimit_closest_all((const struct rlimit
* const *) context
->rlimit
, &which_failed
);
4522 *exit_status
= EXIT_LIMITS
;
4523 return log_exec_error_errno(context
, params
, r
, "Failed to adjust resource limit RLIMIT_%s: %m", rlimit_to_string(which_failed
));
4527 if (needs_setuid
&& context
->pam_name
&& username
) {
4528 /* Let's call into PAM after we set up our own idea of resource limits so that pam_limits
4529 * wins here. (See above.) */
4531 /* All fds passed in the fds array will be closed in the pam child process. */
4532 r
= setup_pam(context
->pam_name
, username
, uid
, gid
, context
->tty_path
, &accum_env
, params
->fds
, n_fds
);
4534 *exit_status
= EXIT_PAM
;
4535 return log_exec_error_errno(context
, params
, r
, "Failed to set up PAM session: %m");
4538 if (ambient_capabilities_supported()) {
4539 uint64_t ambient_after_pam
;
4541 /* PAM modules might have set some ambient caps. Query them here and merge them into
4542 * the caps we want to set in the end, so that we don't end up unsetting them. */
4543 r
= capability_get_ambient(&ambient_after_pam
);
4545 *exit_status
= EXIT_CAPABILITIES
;
4546 return log_exec_error_errno(context
, params
, r
, "Failed to query ambient caps: %m");
4549 capability_ambient_set
|= ambient_after_pam
;
4552 ngids_after_pam
= getgroups_alloc(&gids_after_pam
);
4553 if (ngids_after_pam
< 0) {
4554 *exit_status
= EXIT_GROUP
;
4555 return log_exec_error_errno(context
, params
, ngids_after_pam
, "Failed to obtain groups after setting up PAM: %m");
4559 if (needs_sandboxing
&& exec_context_need_unprivileged_private_users(context
, params
)) {
4560 /* If we're unprivileged, set up the user namespace first to enable use of the other namespaces.
4561 * Users with CAP_SYS_ADMIN can set up user namespaces last because they will be able to
4562 * set up all of the other namespaces (i.e. network, mount, UTS) without a user namespace. */
4564 r
= setup_private_users(saved_uid
, saved_gid
, uid
, gid
);
4565 /* If it was requested explicitly and we can't set it up, fail early. Otherwise, continue and let
4566 * the actual requested operations fail (or silently continue). */
4567 if (r
< 0 && context
->private_users
) {
4568 *exit_status
= EXIT_USER
;
4569 return log_exec_error_errno(context
, params
, r
, "Failed to set up user namespacing for unprivileged user: %m");
4572 log_exec_info_errno(context
, params
, r
, "Failed to set up user namespacing for unprivileged user, ignoring: %m");
4574 userns_set_up
= true;
4577 if (exec_needs_network_namespace(context
) && runtime
&& runtime
->shared
&& runtime
->shared
->netns_storage_socket
[0] >= 0) {
4579 /* Try to enable network namespacing if network namespacing is available and we have
4580 * CAP_NET_ADMIN. We need CAP_NET_ADMIN to be able to configure the loopback device in the
4581 * new network namespace. And if we don't have that, then we could only create a network
4582 * namespace without the ability to set up "lo". Hence gracefully skip things then. */
4583 if (ns_type_supported(NAMESPACE_NET
) && have_effective_cap(CAP_NET_ADMIN
) > 0) {
4584 r
= setup_shareable_ns(runtime
->shared
->netns_storage_socket
, CLONE_NEWNET
);
4585 if (ERRNO_IS_NEG_PRIVILEGE(r
))
4586 log_exec_notice_errno(context
, params
, r
,
4587 "PrivateNetwork=yes is configured, but network namespace setup not permitted, proceeding without: %m");
4589 *exit_status
= EXIT_NETWORK
;
4590 return log_exec_error_errno(context
, params
, r
, "Failed to set up network namespacing: %m");
4592 } else if (context
->network_namespace_path
) {
4593 *exit_status
= EXIT_NETWORK
;
4594 return log_exec_error_errno(context
, params
, SYNTHETIC_ERRNO(EOPNOTSUPP
),
4595 "NetworkNamespacePath= is not supported, refusing.");
4597 log_exec_notice(context
, params
, "PrivateNetwork=yes is configured, but the kernel does not support or we lack privileges for network namespace, proceeding without.");
4600 if (exec_needs_ipc_namespace(context
) && runtime
&& runtime
->shared
&& runtime
->shared
->ipcns_storage_socket
[0] >= 0) {
4602 if (ns_type_supported(NAMESPACE_IPC
)) {
4603 r
= setup_shareable_ns(runtime
->shared
->ipcns_storage_socket
, CLONE_NEWIPC
);
4605 log_exec_warning_errno(context
, params
, r
,
4606 "PrivateIPC=yes is configured, but IPC namespace setup failed, ignoring: %m");
4608 *exit_status
= EXIT_NAMESPACE
;
4609 return log_exec_error_errno(context
, params
, r
, "Failed to set up IPC namespacing: %m");
4611 } else if (context
->ipc_namespace_path
) {
4612 *exit_status
= EXIT_NAMESPACE
;
4613 return log_exec_error_errno(context
, params
, SYNTHETIC_ERRNO(EOPNOTSUPP
),
4614 "IPCNamespacePath= is not supported, refusing.");
4616 log_exec_warning(context
, params
, "PrivateIPC=yes is configured, but the kernel does not support IPC namespaces, ignoring.");
4619 if (needs_mount_namespace
) {
4620 _cleanup_free_
char *error_path
= NULL
;
4622 r
= apply_mount_namespace(command
->flags
, context
, params
, runtime
, memory_pressure_path
, &error_path
);
4624 *exit_status
= EXIT_NAMESPACE
;
4625 return log_exec_error_errno(context
, params
, r
, "Failed to set up mount namespacing%s%s: %m",
4626 error_path
? ": " : "", strempty(error_path
));
4630 if (needs_sandboxing
) {
4631 r
= apply_protect_hostname(context
, params
, exit_status
);
4636 if (context
->memory_ksm
>= 0)
4637 if (prctl(PR_SET_MEMORY_MERGE
, context
->memory_ksm
) < 0) {
4638 if (ERRNO_IS_NOT_SUPPORTED(errno
))
4639 log_exec_debug_errno(context
,
4642 "KSM support not available, ignoring.");
4644 *exit_status
= EXIT_KSM
;
4645 return log_exec_error_errno(context
, params
, errno
, "Failed to set KSM: %m");
4649 /* Drop groups as early as possible.
4650 * This needs to be done after PrivateDevices=yes setup as device nodes should be owned by the host's root.
4651 * For non-root in a userns, devices will be owned by the user/group before the group change, and nobody. */
4653 _cleanup_free_ gid_t
*gids_to_enforce
= NULL
;
4654 int ngids_to_enforce
= 0;
4656 ngids_to_enforce
= merge_gid_lists(supplementary_gids
,
4661 if (ngids_to_enforce
< 0) {
4662 *exit_status
= EXIT_GROUP
;
4663 return log_exec_error_errno(context
, params
,
4665 "Failed to merge group lists. Group membership might be incorrect: %m");
4668 r
= enforce_groups(gid
, gids_to_enforce
, ngids_to_enforce
);
4670 *exit_status
= EXIT_GROUP
;
4671 return log_exec_error_errno(context
, params
, r
, "Changing group credentials failed: %m");
4675 /* If the user namespace was not set up above, try to do it now.
4676 * It's preferred to set up the user namespace later (after all other namespaces) so as not to be
4677 * restricted by rules pertaining to combining user namespaces with other namespaces (e.g. in the
4678 * case of mount namespaces being less privileged when the mount point list is copied from a
4679 * different user namespace). */
4681 if (needs_sandboxing
&& context
->private_users
&& !userns_set_up
) {
4682 r
= setup_private_users(saved_uid
, saved_gid
, uid
, gid
);
4684 *exit_status
= EXIT_USER
;
4685 return log_exec_error_errno(context
, params
, r
, "Failed to set up user namespacing: %m");
4689 /* Now that the mount namespace has been set up and privileges adjusted, let's look for the thing we
4692 _cleanup_free_
char *executable
= NULL
;
4693 _cleanup_close_
int executable_fd
= -EBADF
;
4694 r
= find_executable_full(command
->path
, /* root= */ NULL
, context
->exec_search_path
, false, &executable
, &executable_fd
);
4696 if (r
!= -ENOMEM
&& (command
->flags
& EXEC_COMMAND_IGNORE_FAILURE
)) {
4697 log_exec_struct_errno(context
, params
, LOG_INFO
, r
,
4698 "MESSAGE_ID=" SD_MESSAGE_SPAWN_FAILED_STR
,
4699 LOG_EXEC_INVOCATION_ID(params
),
4700 LOG_EXEC_MESSAGE(params
,
4701 "Executable %s missing, skipping: %m",
4703 "EXECUTABLE=%s", command
->path
);
4704 *exit_status
= EXIT_SUCCESS
;
4708 *exit_status
= EXIT_EXEC
;
4709 return log_exec_struct_errno(context
, params
, LOG_INFO
, r
,
4710 "MESSAGE_ID=" SD_MESSAGE_SPAWN_FAILED_STR
,
4711 LOG_EXEC_INVOCATION_ID(params
),
4712 LOG_EXEC_MESSAGE(params
,
4713 "Failed to locate executable %s: %m",
4715 "EXECUTABLE=%s", command
->path
);
4718 r
= add_shifted_fd(keep_fds
, ELEMENTSOF(keep_fds
), &n_keep_fds
, &executable_fd
);
4720 *exit_status
= EXIT_FDS
;
4721 return log_exec_error_errno(context
, params
, r
, "Failed to collect shifted fd: %m");
4725 if (needs_sandboxing
&& use_selinux
&& params
->selinux_context_net
) {
4730 else if (params
->n_socket_fds
== 1)
4731 /* If stdin is not connected to a socket but we are triggered by exactly one socket unit then we
4732 * use context from that fd to compute the label. */
4733 fd
= params
->fds
[0];
4736 r
= mac_selinux_get_child_mls_label(fd
, executable
, context
->selinux_context
, &mac_selinux_context_net
);
4738 if (!context
->selinux_context_ignore
) {
4739 *exit_status
= EXIT_SELINUX_CONTEXT
;
4740 return log_exec_error_errno(context
,
4743 "Failed to determine SELinux context: %m");
4745 log_exec_debug_errno(context
,
4748 "Failed to determine SELinux context, ignoring: %m");
4754 /* We repeat the fd closing here, to make sure that nothing is leaked from the PAM modules. Note that
4755 * we are more aggressive this time, since we don't need socket_fd and the netns and ipcns fds any
4756 * more. We do keep exec_fd however, if we have it, since we need to keep it open until the final
4757 * execve(). But first, close the remaining sockets in the context objects. */
4759 exec_runtime_close(runtime
);
4760 exec_params_close(params
);
4762 r
= close_all_fds(keep_fds
, n_keep_fds
);
4764 r
= shift_fds(params
->fds
, n_fds
);
4766 r
= flag_fds(params
->fds
, n_socket_fds
, n_fds
, context
->non_blocking
);
4768 *exit_status
= EXIT_FDS
;
4769 return log_exec_error_errno(context
, params
, r
, "Failed to adjust passed file descriptors: %m");
4772 /* At this point, the fds we want to pass to the program are all ready and set up, with O_CLOEXEC turned off
4773 * and at the right fd numbers. The are no other fds open, with one exception: the exec_fd if it is defined,
4774 * and it has O_CLOEXEC set, after all we want it to be closed by the execve(), so that our parent knows we
4777 secure_bits
= context
->secure_bits
;
4779 if (needs_sandboxing
) {
4782 /* Set the RTPRIO resource limit to 0, but only if nothing else was explicitly requested.
4783 * (Note this is placed after the general resource limit initialization, see above, in order
4784 * to take precedence.) */
4785 if (context
->restrict_realtime
&& !context
->rlimit
[RLIMIT_RTPRIO
]) {
4786 if (setrlimit(RLIMIT_RTPRIO
, &RLIMIT_MAKE_CONST(0)) < 0) {
4787 *exit_status
= EXIT_LIMITS
;
4788 return log_exec_error_errno(context
, params
, errno
, "Failed to adjust RLIMIT_RTPRIO resource limit: %m");
4793 /* LSM Smack needs the capability CAP_MAC_ADMIN to change the current execution security context of the
4794 * process. This is the latest place before dropping capabilities. Other MAC context are set later. */
4795 if (use_smack
&& context
->smack_process_label
) {
4796 r
= setup_smack(params
, context
, executable_fd
);
4797 if (r
< 0 && !context
->smack_process_label_ignore
) {
4798 *exit_status
= EXIT_SMACK_PROCESS_LABEL
;
4799 return log_exec_error_errno(context
, params
, r
, "Failed to set SMACK process label: %m");
4804 bset
= context
->capability_bounding_set
;
4805 /* If the ambient caps hack is enabled (which means the kernel can't do them, and the user asked for
4806 * our magic fallback), then let's add some extra caps, so that the service can drop privs of its own,
4807 * instead of us doing that */
4808 if (needs_ambient_hack
)
4809 bset
|= (UINT64_C(1) << CAP_SETPCAP
) |
4810 (UINT64_C(1) << CAP_SETUID
) |
4811 (UINT64_C(1) << CAP_SETGID
);
4814 /* If the service has any form of a seccomp filter and it allows dropping privileges, we'll
4815 * keep the needed privileges to apply it even if we're not root. */
4817 uid_is_valid(uid
) &&
4818 context_has_seccomp(context
) &&
4819 seccomp_allows_drop_privileges(context
)) {
4820 keep_seccomp_privileges
= true;
4822 if (prctl(PR_SET_KEEPCAPS
, 1) < 0) {
4823 *exit_status
= EXIT_USER
;
4824 return log_exec_error_errno(context
, params
, errno
, "Failed to enable keep capabilities flag: %m");
4827 /* Save the current bounding set so we can restore it after applying the seccomp
4830 bset
|= (UINT64_C(1) << CAP_SYS_ADMIN
) |
4831 (UINT64_C(1) << CAP_SETPCAP
);
4835 if (!cap_test_all(bset
)) {
4836 r
= capability_bounding_set_drop(bset
, /* right_now= */ false);
4838 *exit_status
= EXIT_CAPABILITIES
;
4839 return log_exec_error_errno(context
, params
, r
, "Failed to drop capabilities: %m");
4843 /* Ambient capabilities are cleared during setresuid() (in enforce_user()) even with
4846 * To be able to raise the ambient capabilities after setresuid() they have to be added to
4847 * the inherited set and keep caps has to be set (done in enforce_user()). After setresuid()
4848 * the ambient capabilities can be raised as they are present in the permitted and
4849 * inhertiable set. However it is possible that someone wants to set ambient capabilities
4850 * without changing the user, so we also set the ambient capabilities here.
4852 * The requested ambient capabilities are raised in the inheritable set if the second
4853 * argument is true. */
4854 if (!needs_ambient_hack
) {
4855 r
= capability_ambient_set_apply(capability_ambient_set
, /* also_inherit= */ true);
4857 *exit_status
= EXIT_CAPABILITIES
;
4858 return log_exec_error_errno(context
, params
, r
, "Failed to apply ambient capabilities (before UID change): %m");
4863 /* chroot to root directory first, before we lose the ability to chroot */
4864 r
= apply_root_directory(context
, params
, runtime
, needs_mount_namespace
, exit_status
);
4866 return log_exec_error_errno(context
, params
, r
, "Chrooting to the requested root directory failed: %m");
4869 if (uid_is_valid(uid
)) {
4870 r
= enforce_user(context
, uid
, capability_ambient_set
);
4872 *exit_status
= EXIT_USER
;
4873 return log_exec_error_errno(context
, params
, r
, "Failed to change UID to " UID_FMT
": %m", uid
);
4876 if (keep_seccomp_privileges
) {
4877 if (!FLAGS_SET(capability_ambient_set
, (UINT64_C(1) << CAP_SETUID
))) {
4878 r
= drop_capability(CAP_SETUID
);
4880 *exit_status
= EXIT_USER
;
4881 return log_exec_error_errno(context
, params
, r
, "Failed to drop CAP_SETUID: %m");
4885 r
= keep_capability(CAP_SYS_ADMIN
);
4887 *exit_status
= EXIT_USER
;
4888 return log_exec_error_errno(context
, params
, r
, "Failed to keep CAP_SYS_ADMIN: %m");
4891 r
= keep_capability(CAP_SETPCAP
);
4893 *exit_status
= EXIT_USER
;
4894 return log_exec_error_errno(context
, params
, r
, "Failed to keep CAP_SETPCAP: %m");
4898 if (!needs_ambient_hack
&& capability_ambient_set
!= 0) {
4900 /* Raise the ambient capabilities after user change. */
4901 r
= capability_ambient_set_apply(capability_ambient_set
, /* also_inherit= */ false);
4903 *exit_status
= EXIT_CAPABILITIES
;
4904 return log_exec_error_errno(context
, params
, r
, "Failed to apply ambient capabilities (after UID change): %m");
4910 /* Apply working directory here, because the working directory might be on NFS and only the user running
4911 * this service might have the correct privilege to change to the working directory */
4912 r
= apply_working_directory(context
, params
, runtime
, home
, exit_status
);
4914 return log_exec_error_errno(context
, params
, r
, "Changing to the requested working directory failed: %m");
4916 if (needs_sandboxing
) {
4917 /* Apply other MAC contexts late, but before seccomp syscall filtering, as those should really be last to
4918 * influence our own codepaths as little as possible. Moreover, applying MAC contexts usually requires
4919 * syscalls that are subject to seccomp filtering, hence should probably be applied before the syscalls
4920 * are restricted. */
4924 char *exec_context
= mac_selinux_context_net
?: context
->selinux_context
;
4927 r
= setexeccon(exec_context
);
4929 if (!context
->selinux_context_ignore
) {
4930 *exit_status
= EXIT_SELINUX_CONTEXT
;
4931 return log_exec_error_errno(context
, params
, r
, "Failed to change SELinux context to %s: %m", exec_context
);
4933 log_exec_debug_errno(context
,
4936 "Failed to change SELinux context to %s, ignoring: %m",
4944 if (use_apparmor
&& context
->apparmor_profile
) {
4945 r
= aa_change_onexec(context
->apparmor_profile
);
4946 if (r
< 0 && !context
->apparmor_profile_ignore
) {
4947 *exit_status
= EXIT_APPARMOR_PROFILE
;
4948 return log_exec_error_errno(context
,
4951 "Failed to prepare AppArmor profile change to %s: %m",
4952 context
->apparmor_profile
);
4957 /* PR_GET_SECUREBITS is not privileged, while PR_SET_SECUREBITS is. So to suppress potential
4958 * EPERMs we'll try not to call PR_SET_SECUREBITS unless necessary. Setting securebits
4959 * requires CAP_SETPCAP. */
4960 if (prctl(PR_GET_SECUREBITS
) != secure_bits
) {
4961 /* CAP_SETPCAP is required to set securebits. This capability is raised into the
4962 * effective set here.
4964 * The effective set is overwritten during execve() with the following values:
4966 * - ambient set (for non-root processes)
4968 * - (inheritable | bounding) set for root processes)
4970 * Hence there is no security impact to raise it in the effective set before execve
4972 r
= capability_gain_cap_setpcap(/* return_caps= */ NULL
);
4974 *exit_status
= EXIT_CAPABILITIES
;
4975 return log_exec_error_errno(context
, params
, r
, "Failed to gain CAP_SETPCAP for setting secure bits");
4977 if (prctl(PR_SET_SECUREBITS
, secure_bits
) < 0) {
4978 *exit_status
= EXIT_SECUREBITS
;
4979 return log_exec_error_errno(context
, params
, errno
, "Failed to set process secure bits: %m");
4983 if (context_has_no_new_privileges(context
))
4984 if (prctl(PR_SET_NO_NEW_PRIVS
, 1, 0, 0, 0) < 0) {
4985 *exit_status
= EXIT_NO_NEW_PRIVILEGES
;
4986 return log_exec_error_errno(context
, params
, errno
, "Failed to disable new privileges: %m");
4990 r
= apply_address_families(context
, params
);
4992 *exit_status
= EXIT_ADDRESS_FAMILIES
;
4993 return log_exec_error_errno(context
, params
, r
, "Failed to restrict address families: %m");
4996 r
= apply_memory_deny_write_execute(context
, params
);
4998 *exit_status
= EXIT_SECCOMP
;
4999 return log_exec_error_errno(context
, params
, r
, "Failed to disable writing to executable memory: %m");
5002 r
= apply_restrict_realtime(context
, params
);
5004 *exit_status
= EXIT_SECCOMP
;
5005 return log_exec_error_errno(context
, params
, r
, "Failed to apply realtime restrictions: %m");
5008 r
= apply_restrict_suid_sgid(context
, params
);
5010 *exit_status
= EXIT_SECCOMP
;
5011 return log_exec_error_errno(context
, params
, r
, "Failed to apply SUID/SGID restrictions: %m");
5014 r
= apply_restrict_namespaces(context
, params
);
5016 *exit_status
= EXIT_SECCOMP
;
5017 return log_exec_error_errno(context
, params
, r
, "Failed to apply namespace restrictions: %m");
5020 r
= apply_protect_sysctl(context
, params
);
5022 *exit_status
= EXIT_SECCOMP
;
5023 return log_exec_error_errno(context
, params
, r
, "Failed to apply sysctl restrictions: %m");
5026 r
= apply_protect_kernel_modules(context
, params
);
5028 *exit_status
= EXIT_SECCOMP
;
5029 return log_exec_error_errno(context
, params
, r
, "Failed to apply module loading restrictions: %m");
5032 r
= apply_protect_kernel_logs(context
, params
);
5034 *exit_status
= EXIT_SECCOMP
;
5035 return log_exec_error_errno(context
, params
, r
, "Failed to apply kernel log restrictions: %m");
5038 r
= apply_protect_clock(context
, params
);
5040 *exit_status
= EXIT_SECCOMP
;
5041 return log_exec_error_errno(context
, params
, r
, "Failed to apply clock restrictions: %m");
5044 r
= apply_private_devices(context
, params
);
5046 *exit_status
= EXIT_SECCOMP
;
5047 return log_exec_error_errno(context
, params
, r
, "Failed to set up private devices: %m");
5050 r
= apply_syscall_archs(context
, params
);
5052 *exit_status
= EXIT_SECCOMP
;
5053 return log_exec_error_errno(context
, params
, r
, "Failed to apply syscall architecture restrictions: %m");
5056 r
= apply_lock_personality(context
, params
);
5058 *exit_status
= EXIT_SECCOMP
;
5059 return log_exec_error_errno(context
, params
, r
, "Failed to lock personalities: %m");
5062 r
= apply_syscall_log(context
, params
);
5064 *exit_status
= EXIT_SECCOMP
;
5065 return log_exec_error_errno(context
, params
, r
, "Failed to apply system call log filters: %m");
5070 r
= apply_restrict_filesystems(context
, params
);
5072 *exit_status
= EXIT_BPF
;
5073 return log_exec_error_errno(context
, params
, r
, "Failed to restrict filesystems: %m");
5078 /* This really should remain as close to the execve() as possible, to make sure our own code is affected
5079 * by the filter as little as possible. */
5080 r
= apply_syscall_filter(context
, params
, needs_ambient_hack
);
5082 *exit_status
= EXIT_SECCOMP
;
5083 return log_exec_error_errno(context
, params
, r
, "Failed to apply system call filters: %m");
5086 if (keep_seccomp_privileges
) {
5087 /* Restore the capability bounding set with what's expected from the service + the
5088 * ambient capabilities hack */
5089 if (!cap_test_all(saved_bset
)) {
5090 r
= capability_bounding_set_drop(saved_bset
, /* right_now= */ false);
5092 *exit_status
= EXIT_CAPABILITIES
;
5093 return log_exec_error_errno(context
, params
, r
, "Failed to drop bset capabilities: %m");
5097 /* Only drop CAP_SYS_ADMIN if it's not in the bounding set, otherwise we'll break
5098 * applications that use it. */
5099 if (!FLAGS_SET(saved_bset
, (UINT64_C(1) << CAP_SYS_ADMIN
))) {
5100 r
= drop_capability(CAP_SYS_ADMIN
);
5102 *exit_status
= EXIT_USER
;
5103 return log_exec_error_errno(context
, params
, r
, "Failed to drop CAP_SYS_ADMIN: %m");
5107 /* Only drop CAP_SETPCAP if it's not in the bounding set, otherwise we'll break
5108 * applications that use it. */
5109 if (!FLAGS_SET(saved_bset
, (UINT64_C(1) << CAP_SETPCAP
))) {
5110 r
= drop_capability(CAP_SETPCAP
);
5112 *exit_status
= EXIT_USER
;
5113 return log_exec_error_errno(context
, params
, r
, "Failed to drop CAP_SETPCAP: %m");
5117 if (prctl(PR_SET_KEEPCAPS
, 0) < 0) {
5118 *exit_status
= EXIT_USER
;
5119 return log_exec_error_errno(context
, params
, errno
, "Failed to drop keep capabilities flag: %m");
5126 if (!strv_isempty(context
->unset_environment
)) {
5129 ee
= strv_env_delete(accum_env
, 1, context
->unset_environment
);
5131 *exit_status
= EXIT_MEMORY
;
5135 strv_free_and_replace(accum_env
, ee
);
5138 if (!FLAGS_SET(command
->flags
, EXEC_COMMAND_NO_ENV_EXPAND
)) {
5139 _cleanup_strv_free_
char **unset_variables
= NULL
, **bad_variables
= NULL
;
5141 r
= replace_env_argv(command
->argv
, accum_env
, &replaced_argv
, &unset_variables
, &bad_variables
);
5143 *exit_status
= EXIT_MEMORY
;
5144 return log_exec_error_errno(context
,
5147 "Failed to replace environment variables: %m");
5149 final_argv
= replaced_argv
;
5151 if (!strv_isempty(unset_variables
)) {
5152 _cleanup_free_
char *ju
= strv_join(unset_variables
, ", ");
5153 log_exec_warning(context
,
5155 "Referenced but unset environment variable evaluates to an empty string: %s",
5159 if (!strv_isempty(bad_variables
)) {
5160 _cleanup_free_
char *jb
= strv_join(bad_variables
, ", ");
5161 log_exec_warning(context
,
5163 "Invalid environment variable name evaluates to an empty string: %s",
5167 final_argv
= command
->argv
;
5169 log_command_line(context
, params
, "Executing", executable
, final_argv
);
5171 if (params
->exec_fd
>= 0) {
5174 /* We have finished with all our initializations. Let's now let the manager know that. From this point
5175 * on, if the manager sees POLLHUP on the exec_fd, then execve() was successful. */
5177 if (write(params
->exec_fd
, &hot
, sizeof(hot
)) < 0) {
5178 *exit_status
= EXIT_EXEC
;
5179 return log_exec_error_errno(context
, params
, errno
, "Failed to enable exec_fd: %m");
5183 r
= fexecve_or_execve(executable_fd
, executable
, final_argv
, accum_env
);
5185 if (params
->exec_fd
>= 0) {
5188 /* The execve() failed. This means the exec_fd is still open. Which means we need to tell the manager
5189 * that POLLHUP on it no longer means execve() succeeded. */
5191 if (write(params
->exec_fd
, &hot
, sizeof(hot
)) < 0) {
5192 *exit_status
= EXIT_EXEC
;
5193 return log_exec_error_errno(context
, params
, errno
, "Failed to disable exec_fd: %m");
5197 *exit_status
= EXIT_EXEC
;
5198 return log_exec_error_errno(context
, params
, r
, "Failed to execute %s: %m", executable
);