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"
25 #include "bpf-restrict-fs.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"
64 #define IDLE_TIMEOUT_USEC (5*USEC_PER_SEC)
65 #define IDLE_TIMEOUT2_USEC (1*USEC_PER_SEC)
67 #define SNDBUF_SIZE (8*1024*1024)
77 assert(fds
|| n_fds
== 0);
79 /* Drops/Sets O_NONBLOCK and FD_CLOEXEC from the file flags.
80 * O_NONBLOCK only applies to socket activation though. */
82 for (size_t i
= 0; i
< n_fds
; i
++) {
84 if (i
< n_socket_fds
) {
85 r
= fd_nonblock(fds
[i
], nonblock
);
90 /* We unconditionally drop FD_CLOEXEC from the fds,
91 * since after all we want to pass these fds to our
94 r
= fd_cloexec(fds
[i
], false);
102 static bool is_terminal_input(ExecInput i
) {
105 EXEC_INPUT_TTY_FORCE
,
106 EXEC_INPUT_TTY_FAIL
);
109 static bool is_terminal_output(ExecOutput o
) {
112 EXEC_OUTPUT_KMSG_AND_CONSOLE
,
113 EXEC_OUTPUT_JOURNAL_AND_CONSOLE
);
116 static bool is_kmsg_output(ExecOutput o
) {
119 EXEC_OUTPUT_KMSG_AND_CONSOLE
);
122 static bool exec_context_needs_term(const ExecContext
*c
) {
125 /* Return true if the execution context suggests we should set $TERM to something useful. */
127 if (is_terminal_input(c
->std_input
))
130 if (is_terminal_output(c
->std_output
))
133 if (is_terminal_output(c
->std_error
))
136 return !!c
->tty_path
;
139 static int open_null_as(int flags
, int nfd
) {
144 fd
= open("/dev/null", flags
|O_NOCTTY
);
148 return move_fd(fd
, nfd
, false);
151 static int connect_journal_socket(
153 const char *log_namespace
,
157 uid_t olduid
= UID_INVALID
;
158 gid_t oldgid
= GID_INVALID
;
163 strjoina("/run/systemd/journal.", log_namespace
, "/stdout") :
164 "/run/systemd/journal/stdout";
166 if (gid_is_valid(gid
)) {
169 if (setegid(gid
) < 0)
173 if (uid_is_valid(uid
)) {
176 if (seteuid(uid
) < 0) {
182 r
= connect_unix_path(fd
, AT_FDCWD
, j
);
184 /* If we fail to restore the uid or gid, things will likely fail later on. This should only happen if
185 an LSM interferes. */
187 if (uid_is_valid(uid
))
188 (void) seteuid(olduid
);
191 if (gid_is_valid(gid
))
192 (void) setegid(oldgid
);
197 static int connect_logger_as(
198 const ExecContext
*context
,
199 const ExecParameters
*params
,
206 _cleanup_close_
int fd
= -EBADF
;
211 assert(output
< _EXEC_OUTPUT_MAX
);
215 fd
= socket(AF_UNIX
, SOCK_STREAM
, 0);
219 r
= connect_journal_socket(fd
, context
->log_namespace
, uid
, gid
);
223 if (shutdown(fd
, SHUT_RD
) < 0)
226 (void) fd_inc_sndbuf(fd
, SNDBUF_SIZE
);
236 context
->syslog_identifier
?: ident
,
237 params
->flags
& EXEC_PASS_LOG_UNIT
? params
->unit_id
: "",
238 context
->syslog_priority
,
239 !!context
->syslog_level_prefix
,
241 is_kmsg_output(output
),
242 is_terminal_output(output
)) < 0)
245 return move_fd(TAKE_FD(fd
), nfd
, false);
248 static int open_terminal_as(const char *path
, int flags
, int nfd
) {
254 fd
= open_terminal(path
, flags
| O_NOCTTY
);
258 return move_fd(fd
, nfd
, false);
261 static int acquire_path(const char *path
, int flags
, mode_t mode
) {
262 _cleanup_close_
int fd
= -EBADF
;
267 if (IN_SET(flags
& O_ACCMODE
, O_WRONLY
, O_RDWR
))
270 fd
= open(path
, flags
|O_NOCTTY
, mode
);
274 if (errno
!= ENXIO
) /* ENXIO is returned when we try to open() an AF_UNIX file system socket on Linux */
277 /* So, it appears the specified path could be an AF_UNIX socket. Let's see if we can connect to it. */
279 fd
= socket(AF_UNIX
, SOCK_STREAM
, 0);
283 r
= connect_unix_path(fd
, AT_FDCWD
, path
);
284 if (IN_SET(r
, -ENOTSOCK
, -EINVAL
))
285 /* Propagate initial error if we get ENOTSOCK or EINVAL, i.e. we have indication that this
286 * wasn't an AF_UNIX socket after all */
291 if ((flags
& O_ACCMODE
) == O_RDONLY
)
292 r
= shutdown(fd
, SHUT_WR
);
293 else if ((flags
& O_ACCMODE
) == O_WRONLY
)
294 r
= shutdown(fd
, SHUT_RD
);
303 static int fixup_input(
304 const ExecContext
*context
,
306 bool apply_tty_stdin
) {
312 std_input
= context
->std_input
;
314 if (is_terminal_input(std_input
) && !apply_tty_stdin
)
315 return EXEC_INPUT_NULL
;
317 if (std_input
== EXEC_INPUT_SOCKET
&& socket_fd
< 0)
318 return EXEC_INPUT_NULL
;
320 if (std_input
== EXEC_INPUT_DATA
&& context
->stdin_data_size
== 0)
321 return EXEC_INPUT_NULL
;
326 static int fixup_output(ExecOutput output
, int socket_fd
) {
328 if (output
== EXEC_OUTPUT_SOCKET
&& socket_fd
< 0)
329 return EXEC_OUTPUT_INHERIT
;
334 static int setup_input(
335 const ExecContext
*context
,
336 const ExecParameters
*params
,
338 const int named_iofds
[static 3]) {
347 if (params
->stdin_fd
>= 0) {
348 if (dup2(params
->stdin_fd
, STDIN_FILENO
) < 0)
351 /* Try to make this the controlling tty, if it is a tty, and reset it */
352 if (isatty(STDIN_FILENO
)) {
353 (void) ioctl(STDIN_FILENO
, TIOCSCTTY
, context
->std_input
== EXEC_INPUT_TTY_FORCE
);
355 if (context
->tty_reset
)
356 (void) reset_terminal_fd(STDIN_FILENO
, /* switch_to_text= */ true);
358 (void) exec_context_apply_tty_size(context
, STDIN_FILENO
, /* tty_path= */ NULL
);
364 i
= fixup_input(context
, socket_fd
, params
->flags
& EXEC_APPLY_TTY_STDIN
);
368 case EXEC_INPUT_NULL
:
369 return open_null_as(O_RDONLY
, STDIN_FILENO
);
372 case EXEC_INPUT_TTY_FORCE
:
373 case EXEC_INPUT_TTY_FAIL
: {
374 _cleanup_close_
int tty_fd
= -EBADF
;
375 const char *tty_path
;
377 tty_path
= ASSERT_PTR(exec_context_tty_path(context
));
379 tty_fd
= acquire_terminal(tty_path
,
380 i
== EXEC_INPUT_TTY_FAIL
? ACQUIRE_TERMINAL_TRY
:
381 i
== EXEC_INPUT_TTY_FORCE
? ACQUIRE_TERMINAL_FORCE
:
382 ACQUIRE_TERMINAL_WAIT
,
387 r
= exec_context_apply_tty_size(context
, tty_fd
, tty_path
);
391 r
= move_fd(tty_fd
, STDIN_FILENO
, /* cloexec= */ false);
399 case EXEC_INPUT_SOCKET
:
400 assert(socket_fd
>= 0);
402 return RET_NERRNO(dup2(socket_fd
, STDIN_FILENO
));
404 case EXEC_INPUT_NAMED_FD
:
405 assert(named_iofds
[STDIN_FILENO
] >= 0);
407 (void) fd_nonblock(named_iofds
[STDIN_FILENO
], false);
408 return RET_NERRNO(dup2(named_iofds
[STDIN_FILENO
], STDIN_FILENO
));
410 case EXEC_INPUT_DATA
: {
413 fd
= acquire_data_fd(context
->stdin_data
, context
->stdin_data_size
, 0);
417 return move_fd(fd
, STDIN_FILENO
, false);
420 case EXEC_INPUT_FILE
: {
424 assert(context
->stdio_file
[STDIN_FILENO
]);
426 rw
= (context
->std_output
== EXEC_OUTPUT_FILE
&& streq_ptr(context
->stdio_file
[STDIN_FILENO
], context
->stdio_file
[STDOUT_FILENO
])) ||
427 (context
->std_error
== EXEC_OUTPUT_FILE
&& streq_ptr(context
->stdio_file
[STDIN_FILENO
], context
->stdio_file
[STDERR_FILENO
]));
429 fd
= acquire_path(context
->stdio_file
[STDIN_FILENO
], rw
? O_RDWR
: O_RDONLY
, 0666 & ~context
->umask
);
433 return move_fd(fd
, STDIN_FILENO
, false);
437 assert_not_reached();
441 static bool can_inherit_stderr_from_stdout(
442 const ExecContext
*context
,
448 /* Returns true, if given the specified STDERR and STDOUT output we can directly dup() the stdout fd to the
451 if (e
== EXEC_OUTPUT_INHERIT
)
456 if (e
== EXEC_OUTPUT_NAMED_FD
)
457 return streq_ptr(context
->stdio_fdname
[STDOUT_FILENO
], context
->stdio_fdname
[STDERR_FILENO
]);
459 if (IN_SET(e
, EXEC_OUTPUT_FILE
, EXEC_OUTPUT_FILE_APPEND
, EXEC_OUTPUT_FILE_TRUNCATE
))
460 return streq_ptr(context
->stdio_file
[STDOUT_FILENO
], context
->stdio_file
[STDERR_FILENO
]);
465 static int setup_output(
466 const ExecContext
*context
,
467 const ExecParameters
*params
,
470 const int named_iofds
[static 3],
474 dev_t
*journal_stream_dev
,
475 ino_t
*journal_stream_ino
) {
484 assert(journal_stream_dev
);
485 assert(journal_stream_ino
);
487 if (fileno
== STDOUT_FILENO
&& params
->stdout_fd
>= 0) {
489 if (dup2(params
->stdout_fd
, STDOUT_FILENO
) < 0)
492 return STDOUT_FILENO
;
495 if (fileno
== STDERR_FILENO
&& params
->stderr_fd
>= 0) {
496 if (dup2(params
->stderr_fd
, STDERR_FILENO
) < 0)
499 return STDERR_FILENO
;
502 i
= fixup_input(context
, socket_fd
, params
->flags
& EXEC_APPLY_TTY_STDIN
);
503 o
= fixup_output(context
->std_output
, socket_fd
);
505 if (fileno
== STDERR_FILENO
) {
507 e
= fixup_output(context
->std_error
, socket_fd
);
509 /* This expects the input and output are already set up */
511 /* Don't change the stderr file descriptor if we inherit all
512 * the way and are not on a tty */
513 if (e
== EXEC_OUTPUT_INHERIT
&&
514 o
== EXEC_OUTPUT_INHERIT
&&
515 i
== EXEC_INPUT_NULL
&&
516 !is_terminal_input(context
->std_input
) &&
520 /* Duplicate from stdout if possible */
521 if (can_inherit_stderr_from_stdout(context
, o
, e
))
522 return RET_NERRNO(dup2(STDOUT_FILENO
, fileno
));
526 } else if (o
== EXEC_OUTPUT_INHERIT
) {
527 /* If input got downgraded, inherit the original value */
528 if (i
== EXEC_INPUT_NULL
&& is_terminal_input(context
->std_input
))
529 return open_terminal_as(exec_context_tty_path(context
), O_WRONLY
, fileno
);
531 /* If the input is connected to anything that's not a /dev/null or a data fd, inherit that... */
532 if (!IN_SET(i
, EXEC_INPUT_NULL
, EXEC_INPUT_DATA
))
533 return RET_NERRNO(dup2(STDIN_FILENO
, fileno
));
535 /* If we are not started from PID 1 we just inherit STDOUT from our parent process. */
539 /* We need to open /dev/null here anew, to get the right access mode. */
540 return open_null_as(O_WRONLY
, fileno
);
545 case EXEC_OUTPUT_NULL
:
546 return open_null_as(O_WRONLY
, fileno
);
548 case EXEC_OUTPUT_TTY
:
549 if (is_terminal_input(i
))
550 return RET_NERRNO(dup2(STDIN_FILENO
, fileno
));
552 /* We don't reset the terminal if this is just about output */
553 return open_terminal_as(exec_context_tty_path(context
), O_WRONLY
, fileno
);
555 case EXEC_OUTPUT_KMSG
:
556 case EXEC_OUTPUT_KMSG_AND_CONSOLE
:
557 case EXEC_OUTPUT_JOURNAL
:
558 case EXEC_OUTPUT_JOURNAL_AND_CONSOLE
:
559 r
= connect_logger_as(context
, params
, o
, ident
, fileno
, uid
, gid
);
561 log_exec_warning_errno(context
,
564 "Failed to connect %s to the journal socket, ignoring: %m",
565 fileno
== STDOUT_FILENO
? "stdout" : "stderr");
566 r
= open_null_as(O_WRONLY
, fileno
);
570 /* If we connected this fd to the journal via a stream, patch the device/inode into the passed
571 * parameters, but only then. This is useful so that we can set $JOURNAL_STREAM that permits
572 * services to detect whether they are connected to the journal or not.
574 * If both stdout and stderr are connected to a stream then let's make sure to store the data
575 * about STDERR as that's usually the best way to do logging. */
577 if (fstat(fileno
, &st
) >= 0 &&
578 (*journal_stream_ino
== 0 || fileno
== STDERR_FILENO
)) {
579 *journal_stream_dev
= st
.st_dev
;
580 *journal_stream_ino
= st
.st_ino
;
585 case EXEC_OUTPUT_SOCKET
:
586 assert(socket_fd
>= 0);
588 return RET_NERRNO(dup2(socket_fd
, fileno
));
590 case EXEC_OUTPUT_NAMED_FD
:
591 assert(named_iofds
[fileno
] >= 0);
593 (void) fd_nonblock(named_iofds
[fileno
], false);
594 return RET_NERRNO(dup2(named_iofds
[fileno
], fileno
));
596 case EXEC_OUTPUT_FILE
:
597 case EXEC_OUTPUT_FILE_APPEND
:
598 case EXEC_OUTPUT_FILE_TRUNCATE
: {
602 assert(context
->stdio_file
[fileno
]);
604 rw
= context
->std_input
== EXEC_INPUT_FILE
&&
605 streq_ptr(context
->stdio_file
[fileno
], context
->stdio_file
[STDIN_FILENO
]);
608 return RET_NERRNO(dup2(STDIN_FILENO
, fileno
));
611 if (o
== EXEC_OUTPUT_FILE_APPEND
)
613 else if (o
== EXEC_OUTPUT_FILE_TRUNCATE
)
616 fd
= acquire_path(context
->stdio_file
[fileno
], flags
, 0666 & ~context
->umask
);
620 return move_fd(fd
, fileno
, 0);
624 assert_not_reached();
628 static int chown_terminal(int fd
, uid_t uid
) {
633 /* Before we chown/chmod the TTY, let's ensure this is actually a tty */
634 if (!isatty_safe(fd
))
637 /* This might fail. What matters are the results. */
638 r
= fchmod_and_chown(fd
, TTY_MODE
, uid
, GID_INVALID
);
645 static int setup_confirm_stdio(
646 const ExecContext
*context
,
648 int *ret_saved_stdin
,
649 int *ret_saved_stdout
) {
651 _cleanup_close_
int fd
= -EBADF
, saved_stdin
= -EBADF
, saved_stdout
= -EBADF
;
654 assert(ret_saved_stdin
);
655 assert(ret_saved_stdout
);
657 saved_stdin
= fcntl(STDIN_FILENO
, F_DUPFD
, 3);
661 saved_stdout
= fcntl(STDOUT_FILENO
, F_DUPFD
, 3);
662 if (saved_stdout
< 0)
665 fd
= acquire_terminal(vc
, ACQUIRE_TERMINAL_WAIT
, DEFAULT_CONFIRM_USEC
);
669 r
= chown_terminal(fd
, getuid());
673 r
= reset_terminal_fd(fd
, /* switch_to_text= */ true);
677 r
= exec_context_apply_tty_size(context
, fd
, vc
);
681 r
= rearrange_stdio(fd
, fd
, STDERR_FILENO
); /* Invalidates 'fd' also on failure */
686 *ret_saved_stdin
= TAKE_FD(saved_stdin
);
687 *ret_saved_stdout
= TAKE_FD(saved_stdout
);
691 static void write_confirm_error_fd(int err
, int fd
, const char *unit_id
) {
695 if (err
== -ETIMEDOUT
)
696 dprintf(fd
, "Confirmation question timed out for %s, assuming positive response.\n", unit_id
);
699 dprintf(fd
, "Couldn't ask confirmation for %s: %m, assuming positive response.\n", unit_id
);
703 static void write_confirm_error(int err
, const char *vc
, const char *unit_id
) {
704 _cleanup_close_
int fd
= -EBADF
;
708 fd
= open_terminal(vc
, O_WRONLY
|O_NOCTTY
|O_CLOEXEC
);
712 write_confirm_error_fd(err
, fd
, unit_id
);
715 static int restore_confirm_stdio(int *saved_stdin
, int *saved_stdout
) {
719 assert(saved_stdout
);
723 if (*saved_stdin
>= 0)
724 if (dup2(*saved_stdin
, STDIN_FILENO
) < 0)
727 if (*saved_stdout
>= 0)
728 if (dup2(*saved_stdout
, STDOUT_FILENO
) < 0)
731 *saved_stdin
= safe_close(*saved_stdin
);
732 *saved_stdout
= safe_close(*saved_stdout
);
738 CONFIRM_PRETEND_FAILURE
= -1,
739 CONFIRM_PRETEND_SUCCESS
= 0,
743 static bool confirm_spawn_disabled(void) {
744 return access("/run/systemd/confirm_spawn_disabled", F_OK
) >= 0;
747 static int ask_for_confirmation(const ExecContext
*context
, const ExecParameters
*params
, const char *cmdline
) {
748 int saved_stdout
= -1, saved_stdin
= -1, r
;
749 _cleanup_free_
char *e
= NULL
;
755 /* For any internal errors, assume a positive response. */
756 r
= setup_confirm_stdio(context
, params
->confirm_spawn
, &saved_stdin
, &saved_stdout
);
758 write_confirm_error(r
, params
->confirm_spawn
, params
->unit_id
);
759 return CONFIRM_EXECUTE
;
762 /* confirm_spawn might have been disabled while we were sleeping. */
763 if (!params
->confirm_spawn
|| confirm_spawn_disabled()) {
768 e
= ellipsize(cmdline
, 60, 100);
776 r
= ask_char(&c
, "yfshiDjcn", "Execute %s? [y, f, s – h for help] ", e
);
778 write_confirm_error_fd(r
, STDOUT_FILENO
, params
->unit_id
);
785 printf("Resuming normal execution.\n");
786 manager_disable_confirm_spawn();
790 printf(" Unit: %s\n",
792 exec_context_dump(context
, stdout
, " ");
793 exec_params_dump(params
, stdout
, " ");
794 continue; /* ask again */
796 printf("Failing execution.\n");
797 r
= CONFIRM_PRETEND_FAILURE
;
800 printf(" c - continue, proceed without asking anymore\n"
801 " D - dump, show the state of the unit\n"
802 " f - fail, don't execute the command and pretend it failed\n"
804 " i - info, show a short summary of the unit\n"
805 " j - jobs, show jobs that are in progress\n"
806 " s - skip, don't execute the command and pretend it succeeded\n"
807 " y - yes, execute the command\n");
808 continue; /* ask again */
812 params
->unit_id
, cmdline
);
813 continue; /* ask again */
815 if (sigqueue(getppid(),
817 (const union sigval
) { .sival_int
= MANAGER_SIGNAL_COMMAND_DUMP_JOBS
}) < 0)
820 continue; /* ask again */
822 /* 'n' was removed in favor of 'f'. */
823 printf("Didn't understand 'n', did you mean 'f'?\n");
824 continue; /* ask again */
826 printf("Skipping execution.\n");
827 r
= CONFIRM_PRETEND_SUCCESS
;
833 assert_not_reached();
839 restore_confirm_stdio(&saved_stdin
, &saved_stdout
);
843 static int get_fixed_user(
844 const char *user_or_uid
,
845 const char **ret_username
,
848 const char **ret_home
,
849 const char **ret_shell
) {
854 assert(ret_username
);
856 /* Note that we don't set $HOME or $SHELL if they are not particularly enlightening anyway
857 * (i.e. are "/" or "/bin/nologin"). */
859 r
= get_user_creds(&user_or_uid
, ret_uid
, ret_gid
, ret_home
, ret_shell
, USER_CREDS_CLEAN
);
863 /* user_or_uid is normalized by get_user_creds to username */
864 *ret_username
= user_or_uid
;
869 static int get_fixed_group(
870 const char *group_or_gid
,
871 const char **ret_groupname
,
876 assert(group_or_gid
);
877 assert(ret_groupname
);
879 r
= get_group_creds(&group_or_gid
, ret_gid
, /* flags = */ 0);
883 /* group_or_gid is normalized by get_group_creds to groupname */
884 *ret_groupname
= group_or_gid
;
889 static int get_supplementary_groups(const ExecContext
*c
, const char *user
,
890 const char *group
, gid_t gid
,
891 gid_t
**supplementary_gids
, int *ngids
) {
894 bool keep_groups
= false;
895 gid_t
*groups
= NULL
;
896 _cleanup_free_ gid_t
*l_gids
= NULL
;
901 * If user is given, then lookup GID and supplementary groups list.
902 * We avoid NSS lookups for gid=0. Also we have to initialize groups
903 * here and as early as possible so we keep the list of supplementary
904 * groups of the caller.
906 if (user
&& gid_is_valid(gid
) && gid
!= 0) {
907 /* First step, initialize groups from /etc/groups */
908 if (initgroups(user
, gid
) < 0)
914 if (strv_isempty(c
->supplementary_groups
))
918 * If SupplementaryGroups= was passed then NGROUPS_MAX has to
919 * be positive, otherwise fail.
922 ngroups_max
= (int) sysconf(_SC_NGROUPS_MAX
);
923 if (ngroups_max
<= 0)
924 return errno_or_else(EOPNOTSUPP
);
926 l_gids
= new(gid_t
, ngroups_max
);
932 * Lookup the list of groups that the user belongs to, we
933 * avoid NSS lookups here too for gid=0.
936 if (getgrouplist(user
, gid
, l_gids
, &k
) < 0)
941 STRV_FOREACH(i
, c
->supplementary_groups
) {
944 if (k
>= ngroups_max
)
948 r
= get_group_creds(&g
, l_gids
+k
, 0);
956 * Sets ngids to zero to drop all supplementary groups, happens
957 * when we are under root and SupplementaryGroups= is empty.
964 /* Otherwise get the final list of supplementary groups */
965 groups
= memdup(l_gids
, sizeof(gid_t
) * k
);
969 *supplementary_gids
= groups
;
977 static int enforce_groups(gid_t gid
, const gid_t
*supplementary_gids
, int ngids
) {
980 /* Handle SupplementaryGroups= if it is not empty */
982 r
= maybe_setgroups(ngids
, supplementary_gids
);
987 if (gid_is_valid(gid
)) {
988 /* Then set our gids */
989 if (setresgid(gid
, gid
, gid
) < 0)
996 static int set_securebits(unsigned bits
, unsigned mask
) {
1000 current
= prctl(PR_GET_SECUREBITS
);
1004 /* Clear all securebits defined in mask and set bits */
1005 applied
= ((unsigned) current
& ~mask
) | bits
;
1006 if ((unsigned) current
== applied
)
1009 if (prctl(PR_SET_SECUREBITS
, applied
) < 0)
1015 static int enforce_user(
1016 const ExecContext
*context
,
1018 uint64_t capability_ambient_set
) {
1022 if (!uid_is_valid(uid
))
1025 /* Sets (but doesn't look up) the UIS and makes sure we keep the capabilities while doing so. For
1026 * setting secure bits the capability CAP_SETPCAP is required, so we also need keep-caps in this
1029 if ((capability_ambient_set
!= 0 || context
->secure_bits
!= 0) && uid
!= 0) {
1031 /* First step: If we need to keep capabilities but drop privileges we need to make sure we
1032 * keep our caps, while we drop privileges. Add KEEP_CAPS to the securebits */
1033 r
= set_securebits(1U << SECURE_KEEP_CAPS
, 0);
1038 /* Second step: actually set the uids */
1039 if (setresuid(uid
, uid
, uid
) < 0)
1042 /* At this point we should have all necessary capabilities but are otherwise a normal user. However,
1043 * the caps might got corrupted due to the setresuid() so we need clean them up later. This is done
1044 * outside of this call. */
1050 static int null_conv(
1052 const struct pam_message
**msg
,
1053 struct pam_response
**resp
,
1054 void *appdata_ptr
) {
1056 /* We don't support conversations */
1058 return PAM_CONV_ERR
;
1061 static int pam_close_session_and_delete_credentials(pam_handle_t
*handle
, int flags
) {
1066 r
= pam_close_session(handle
, flags
);
1067 if (r
!= PAM_SUCCESS
)
1068 log_debug("pam_close_session() failed: %s", pam_strerror(handle
, r
));
1070 s
= pam_setcred(handle
, PAM_DELETE_CRED
| flags
);
1071 if (s
!= PAM_SUCCESS
)
1072 log_debug("pam_setcred(PAM_DELETE_CRED) failed: %s", pam_strerror(handle
, s
));
1074 return r
!= PAM_SUCCESS
? r
: s
;
1079 static int setup_pam(
1085 char ***env
, /* updated on success */
1086 const int fds
[], size_t n_fds
,
1091 static const struct pam_conv conv
= {
1096 _cleanup_(barrier_destroy
) Barrier barrier
= BARRIER_NULL
;
1097 _cleanup_strv_free_
char **e
= NULL
;
1098 pam_handle_t
*handle
= NULL
;
1100 int pam_code
= PAM_SUCCESS
, r
;
1101 bool close_session
= false;
1109 /* We set up PAM in the parent process, then fork. The child
1110 * will then stay around until killed via PR_GET_PDEATHSIG or
1111 * systemd via the cgroup logic. It will then remove the PAM
1112 * session again. The parent process will exec() the actual
1113 * daemon. We do things this way to ensure that the main PID
1114 * of the daemon is the one we initially fork()ed. */
1116 r
= barrier_create(&barrier
);
1120 if (log_get_max_level() < LOG_DEBUG
)
1121 flags
|= PAM_SILENT
;
1123 pam_code
= pam_start(name
, user
, &conv
, &handle
);
1124 if (pam_code
!= PAM_SUCCESS
) {
1130 _cleanup_free_
char *q
= NULL
;
1132 /* Hmm, so no TTY was explicitly passed, but an fd passed to us directly might be a TTY. Let's figure
1133 * out if that's the case, and read the TTY off it. */
1135 if (getttyname_malloc(STDIN_FILENO
, &q
) >= 0)
1136 tty
= strjoina("/dev/", q
);
1140 pam_code
= pam_set_item(handle
, PAM_TTY
, tty
);
1141 if (pam_code
!= PAM_SUCCESS
)
1145 STRV_FOREACH(nv
, *env
) {
1146 pam_code
= pam_putenv(handle
, *nv
);
1147 if (pam_code
!= PAM_SUCCESS
)
1151 pam_code
= pam_acct_mgmt(handle
, flags
);
1152 if (pam_code
!= PAM_SUCCESS
)
1155 pam_code
= pam_setcred(handle
, PAM_ESTABLISH_CRED
| flags
);
1156 if (pam_code
!= PAM_SUCCESS
)
1157 log_debug("pam_setcred(PAM_ESTABLISH_CRED) failed, ignoring: %s", pam_strerror(handle
, pam_code
));
1159 pam_code
= pam_open_session(handle
, flags
);
1160 if (pam_code
!= PAM_SUCCESS
)
1163 close_session
= true;
1165 e
= pam_getenvlist(handle
);
1167 pam_code
= PAM_BUF_ERR
;
1171 /* Block SIGTERM, so that we know that it won't get lost in the child */
1173 assert_se(sigprocmask_many(SIG_BLOCK
, &old_ss
, SIGTERM
) >= 0);
1175 parent_pid
= getpid_cached();
1177 r
= safe_fork("(sd-pam)", 0, NULL
);
1183 /* The child's job is to reset the PAM session on termination */
1184 barrier_set_role(&barrier
, BARRIER_CHILD
);
1186 /* Make sure we don't keep open the passed fds in this child. We assume that otherwise only
1187 * those fds are open here that have been opened by PAM. */
1188 (void) close_many(fds
, n_fds
);
1190 /* Also close the 'exec_fd' in the child, since the service manager waits for the EOF induced
1191 * by the execve() to wait for completion, and if we'd keep the fd open here in the child
1192 * we'd never signal completion. */
1193 exec_fd
= safe_close(exec_fd
);
1195 /* Drop privileges - we don't need any to pam_close_session and this will make
1196 * PR_SET_PDEATHSIG work in most cases. If this fails, ignore the error - but expect sd-pam
1197 * threads to fail to exit normally */
1199 r
= fully_set_uid_gid(uid
, gid
, /* supplementary_gids= */ NULL
, /* n_supplementary_gids= */ 0);
1201 log_warning_errno(r
, "Failed to drop privileges in sd-pam: %m");
1203 (void) ignore_signals(SIGPIPE
);
1205 /* Wait until our parent died. This will only work if the above setresuid() succeeds,
1206 * otherwise the kernel will not allow unprivileged parents kill their privileged children
1207 * this way. We rely on the control groups kill logic to do the rest for us. */
1208 if (prctl(PR_SET_PDEATHSIG
, SIGTERM
) < 0)
1211 /* Tell the parent that our setup is done. This is especially important regarding dropping
1212 * privileges. Otherwise, unit setup might race against our setresuid(2) call.
1214 * If the parent aborted, we'll detect this below, hence ignore return failure here. */
1215 (void) barrier_place(&barrier
);
1217 /* Check if our parent process might already have died? */
1218 if (getppid() == parent_pid
) {
1222 assert_se(sigemptyset(&ss
) >= 0);
1223 assert_se(sigaddset(&ss
, SIGTERM
) >= 0);
1225 assert_se(sigwait(&ss
, &sig
) == 0);
1226 assert(sig
== SIGTERM
);
1229 /* If our parent died we'll end the session */
1230 if (getppid() != parent_pid
) {
1231 pam_code
= pam_close_session_and_delete_credentials(handle
, flags
);
1232 if (pam_code
!= PAM_SUCCESS
)
1239 /* NB: pam_end() when called in child processes should set PAM_DATA_SILENT to let the module
1240 * know about this. See pam_end(3) */
1241 (void) pam_end(handle
, pam_code
| flags
| PAM_DATA_SILENT
);
1245 barrier_set_role(&barrier
, BARRIER_PARENT
);
1247 /* If the child was forked off successfully it will do all the cleanups, so forget about the handle
1251 /* Unblock SIGTERM again in the parent */
1252 assert_se(sigprocmask(SIG_SETMASK
, &old_ss
, NULL
) >= 0);
1254 /* We close the log explicitly here, since the PAM modules might have opened it, but we don't want
1255 * this fd around. */
1258 /* Synchronously wait for the child to initialize. We don't care for errors as we cannot
1259 * recover. However, warn loudly if it happens. */
1260 if (!barrier_place_and_sync(&barrier
))
1261 log_error("PAM initialization failed");
1263 return strv_free_and_replace(*env
, e
);
1266 if (pam_code
!= PAM_SUCCESS
) {
1267 log_error("PAM failed: %s", pam_strerror(handle
, pam_code
));
1268 r
= -EPERM
; /* PAM errors do not map to errno */
1270 log_error_errno(r
, "PAM failed: %m");
1274 pam_code
= pam_close_session_and_delete_credentials(handle
, flags
);
1276 (void) pam_end(handle
, pam_code
| flags
);
1286 static void rename_process_from_path(const char *path
) {
1287 _cleanup_free_
char *buf
= NULL
;
1292 /* This resulting string must fit in 10 chars (i.e. the length of "/sbin/init") to look pretty in
1295 if (path_extract_filename(path
, &buf
) < 0) {
1296 rename_process("(...)");
1300 size_t l
= strlen(buf
);
1302 /* The end of the process name is usually more interesting, since the first bit might just be
1309 char process_name
[11];
1310 process_name
[0] = '(';
1311 memcpy(process_name
+1, p
, l
);
1312 process_name
[1+l
] = ')';
1313 process_name
[1+l
+1] = 0;
1315 (void) rename_process(process_name
);
1318 static bool context_has_address_families(const ExecContext
*c
) {
1321 return c
->address_families_allow_list
||
1322 !set_isempty(c
->address_families
);
1325 static bool context_has_syscall_filters(const ExecContext
*c
) {
1328 return c
->syscall_allow_list
||
1329 !hashmap_isempty(c
->syscall_filter
);
1332 static bool context_has_syscall_logs(const ExecContext
*c
) {
1335 return c
->syscall_log_allow_list
||
1336 !hashmap_isempty(c
->syscall_log
);
1339 static bool context_has_seccomp(const ExecContext
*c
) {
1340 /* We need NNP if we have any form of seccomp and are unprivileged */
1341 return c
->lock_personality
||
1342 c
->memory_deny_write_execute
||
1343 c
->private_devices
||
1345 c
->protect_hostname
||
1346 c
->protect_kernel_tunables
||
1347 c
->protect_kernel_modules
||
1348 c
->protect_kernel_logs
||
1349 context_has_address_families(c
) ||
1350 exec_context_restrict_namespaces_set(c
) ||
1351 c
->restrict_realtime
||
1352 c
->restrict_suid_sgid
||
1353 !set_isempty(c
->syscall_archs
) ||
1354 context_has_syscall_filters(c
) ||
1355 context_has_syscall_logs(c
);
1358 static bool context_has_no_new_privileges(const ExecContext
*c
) {
1361 if (c
->no_new_privileges
)
1364 if (have_effective_cap(CAP_SYS_ADMIN
) > 0) /* if we are privileged, we don't need NNP */
1367 return context_has_seccomp(c
);
1372 static bool seccomp_allows_drop_privileges(const ExecContext
*c
) {
1374 bool has_capget
= false, has_capset
= false, has_prctl
= false;
1378 /* No syscall filter, we are allowed to drop privileges */
1379 if (hashmap_isempty(c
->syscall_filter
))
1382 HASHMAP_FOREACH_KEY(val
, id
, c
->syscall_filter
) {
1383 _cleanup_free_
char *name
= NULL
;
1385 name
= seccomp_syscall_resolve_num_arch(SCMP_ARCH_NATIVE
, PTR_TO_INT(id
) - 1);
1387 if (streq(name
, "capget"))
1389 else if (streq(name
, "capset"))
1391 else if (streq(name
, "prctl"))
1395 if (c
->syscall_allow_list
)
1396 return has_capget
&& has_capset
&& has_prctl
;
1398 return !(has_capget
|| has_capset
|| has_prctl
);
1401 static bool skip_seccomp_unavailable(const ExecContext
*c
, const ExecParameters
*p
, const char* msg
) {
1403 if (is_seccomp_available())
1406 log_exec_debug(c
, p
, "SECCOMP features not detected in the kernel, skipping %s", msg
);
1410 static int apply_syscall_filter(const ExecContext
*c
, const ExecParameters
*p
, bool needs_ambient_hack
) {
1411 uint32_t negative_action
, default_action
, action
;
1417 if (!context_has_syscall_filters(c
))
1420 if (skip_seccomp_unavailable(c
, p
, "SystemCallFilter="))
1423 negative_action
= c
->syscall_errno
== SECCOMP_ERROR_NUMBER_KILL
? scmp_act_kill_process() : SCMP_ACT_ERRNO(c
->syscall_errno
);
1425 if (c
->syscall_allow_list
) {
1426 default_action
= negative_action
;
1427 action
= SCMP_ACT_ALLOW
;
1429 default_action
= SCMP_ACT_ALLOW
;
1430 action
= negative_action
;
1433 if (needs_ambient_hack
) {
1434 r
= seccomp_filter_set_add(c
->syscall_filter
, c
->syscall_allow_list
, syscall_filter_sets
+ SYSCALL_FILTER_SET_SETUID
);
1439 return seccomp_load_syscall_filter_set_raw(default_action
, c
->syscall_filter
, action
, false);
1442 static int apply_syscall_log(const ExecContext
*c
, const ExecParameters
*p
) {
1444 uint32_t default_action
, action
;
1450 if (!context_has_syscall_logs(c
))
1454 if (skip_seccomp_unavailable(c
, p
, "SystemCallLog="))
1457 if (c
->syscall_log_allow_list
) {
1458 /* Log nothing but the ones listed */
1459 default_action
= SCMP_ACT_ALLOW
;
1460 action
= SCMP_ACT_LOG
;
1462 /* Log everything but the ones listed */
1463 default_action
= SCMP_ACT_LOG
;
1464 action
= SCMP_ACT_ALLOW
;
1467 return seccomp_load_syscall_filter_set_raw(default_action
, c
->syscall_log
, action
, false);
1469 /* old libseccomp */
1470 log_exec_debug(c
, p
, "SECCOMP feature SCMP_ACT_LOG not available, skipping SystemCallLog=");
1475 static int apply_syscall_archs(const ExecContext
*c
, const ExecParameters
*p
) {
1479 if (set_isempty(c
->syscall_archs
))
1482 if (skip_seccomp_unavailable(c
, p
, "SystemCallArchitectures="))
1485 return seccomp_restrict_archs(c
->syscall_archs
);
1488 static int apply_address_families(const ExecContext
*c
, const ExecParameters
*p
) {
1492 if (!context_has_address_families(c
))
1495 if (skip_seccomp_unavailable(c
, p
, "RestrictAddressFamilies="))
1498 return seccomp_restrict_address_families(c
->address_families
, c
->address_families_allow_list
);
1501 static int apply_memory_deny_write_execute(const ExecContext
*c
, const ExecParameters
*p
) {
1507 if (!c
->memory_deny_write_execute
)
1510 /* use prctl() if kernel supports it (6.3) */
1511 r
= prctl(PR_SET_MDWE
, PR_MDWE_REFUSE_EXEC_GAIN
, 0, 0, 0);
1513 log_exec_debug(c
, p
, "Enabled MemoryDenyWriteExecute= with PR_SET_MDWE");
1516 if (r
< 0 && errno
!= EINVAL
)
1517 return log_exec_debug_errno(c
,
1520 "Failed to enable MemoryDenyWriteExecute= with PR_SET_MDWE: %m");
1521 /* else use seccomp */
1522 log_exec_debug(c
, p
, "Kernel doesn't support PR_SET_MDWE: falling back to seccomp");
1524 if (skip_seccomp_unavailable(c
, p
, "MemoryDenyWriteExecute="))
1527 return seccomp_memory_deny_write_execute();
1530 static int apply_restrict_realtime(const ExecContext
*c
, const ExecParameters
*p
) {
1534 if (!c
->restrict_realtime
)
1537 if (skip_seccomp_unavailable(c
, p
, "RestrictRealtime="))
1540 return seccomp_restrict_realtime();
1543 static int apply_restrict_suid_sgid(const ExecContext
*c
, const ExecParameters
*p
) {
1547 if (!c
->restrict_suid_sgid
)
1550 if (skip_seccomp_unavailable(c
, p
, "RestrictSUIDSGID="))
1553 return seccomp_restrict_suid_sgid();
1556 static int apply_protect_sysctl(const ExecContext
*c
, const ExecParameters
*p
) {
1560 /* Turn off the legacy sysctl() system call. Many distributions turn this off while building the kernel, but
1561 * let's protect even those systems where this is left on in the kernel. */
1563 if (!c
->protect_kernel_tunables
)
1566 if (skip_seccomp_unavailable(c
, p
, "ProtectKernelTunables="))
1569 return seccomp_protect_sysctl();
1572 static int apply_protect_kernel_modules(const ExecContext
*c
, const ExecParameters
*p
) {
1576 /* Turn off module syscalls on ProtectKernelModules=yes */
1578 if (!c
->protect_kernel_modules
)
1581 if (skip_seccomp_unavailable(c
, p
, "ProtectKernelModules="))
1584 return seccomp_load_syscall_filter_set(SCMP_ACT_ALLOW
, syscall_filter_sets
+ SYSCALL_FILTER_SET_MODULE
, SCMP_ACT_ERRNO(EPERM
), false);
1587 static int apply_protect_kernel_logs(const ExecContext
*c
, const ExecParameters
*p
) {
1591 if (!c
->protect_kernel_logs
)
1594 if (skip_seccomp_unavailable(c
, p
, "ProtectKernelLogs="))
1597 return seccomp_protect_syslog();
1600 static int apply_protect_clock(const ExecContext
*c
, const ExecParameters
*p
) {
1604 if (!c
->protect_clock
)
1607 if (skip_seccomp_unavailable(c
, p
, "ProtectClock="))
1610 return seccomp_load_syscall_filter_set(SCMP_ACT_ALLOW
, syscall_filter_sets
+ SYSCALL_FILTER_SET_CLOCK
, SCMP_ACT_ERRNO(EPERM
), false);
1613 static int apply_private_devices(const ExecContext
*c
, const ExecParameters
*p
) {
1617 /* If PrivateDevices= is set, also turn off iopl and all @raw-io syscalls. */
1619 if (!c
->private_devices
)
1622 if (skip_seccomp_unavailable(c
, p
, "PrivateDevices="))
1625 return seccomp_load_syscall_filter_set(SCMP_ACT_ALLOW
, syscall_filter_sets
+ SYSCALL_FILTER_SET_RAW_IO
, SCMP_ACT_ERRNO(EPERM
), false);
1628 static int apply_restrict_namespaces(const ExecContext
*c
, const ExecParameters
*p
) {
1632 if (!exec_context_restrict_namespaces_set(c
))
1635 if (skip_seccomp_unavailable(c
, p
, "RestrictNamespaces="))
1638 return seccomp_restrict_namespaces(c
->restrict_namespaces
);
1641 static int apply_lock_personality(const ExecContext
*c
, const ExecParameters
*p
) {
1642 unsigned long personality
;
1648 if (!c
->lock_personality
)
1651 if (skip_seccomp_unavailable(c
, p
, "LockPersonality="))
1654 personality
= c
->personality
;
1656 /* If personality is not specified, use either PER_LINUX or PER_LINUX32 depending on what is currently set. */
1657 if (personality
== PERSONALITY_INVALID
) {
1659 r
= opinionated_personality(&personality
);
1664 return seccomp_lock_personality(personality
);
1670 static int apply_restrict_filesystems(const ExecContext
*c
, const ExecParameters
*p
) {
1676 if (!exec_context_restrict_filesystems_set(c
))
1679 if (p
->bpf_restrict_fs_map_fd
< 0) {
1680 /* LSM BPF is unsupported or lsm_bpf_setup failed */
1681 log_exec_debug(c
, p
, "LSM BPF not supported, skipping RestrictFileSystems=");
1685 /* We are in a new binary, so dl-open again */
1690 return bpf_restrict_fs_update(c
->restrict_filesystems
, p
->cgroup_id
, p
->bpf_restrict_fs_map_fd
, c
->restrict_filesystems_allow_list
);
1694 static int apply_protect_hostname(const ExecContext
*c
, const ExecParameters
*p
, int *ret_exit_status
) {
1698 if (!c
->protect_hostname
)
1701 if (ns_type_supported(NAMESPACE_UTS
)) {
1702 if (unshare(CLONE_NEWUTS
) < 0) {
1703 if (!ERRNO_IS_NOT_SUPPORTED(errno
) && !ERRNO_IS_PRIVILEGE(errno
)) {
1704 *ret_exit_status
= EXIT_NAMESPACE
;
1705 return log_exec_error_errno(c
,
1708 "Failed to set up UTS namespacing: %m");
1713 "ProtectHostname=yes is configured, but UTS namespace setup is "
1714 "prohibited (container manager?), ignoring namespace setup.");
1719 "ProtectHostname=yes is configured, but the kernel does not "
1720 "support UTS namespaces, ignoring namespace setup.");
1725 if (skip_seccomp_unavailable(c
, p
, "ProtectHostname="))
1728 r
= seccomp_protect_hostname();
1730 *ret_exit_status
= EXIT_SECCOMP
;
1731 return log_exec_error_errno(c
, p
, r
, "Failed to apply hostname restrictions: %m");
1738 static void do_idle_pipe_dance(int idle_pipe
[static 4]) {
1741 idle_pipe
[1] = safe_close(idle_pipe
[1]);
1742 idle_pipe
[2] = safe_close(idle_pipe
[2]);
1744 if (idle_pipe
[0] >= 0) {
1747 r
= fd_wait_for_event(idle_pipe
[0], POLLHUP
, IDLE_TIMEOUT_USEC
);
1749 if (idle_pipe
[3] >= 0 && r
== 0 /* timeout */) {
1752 /* Signal systemd that we are bored and want to continue. */
1753 n
= write(idle_pipe
[3], "x", 1);
1755 /* Wait for systemd to react to the signal above. */
1756 (void) fd_wait_for_event(idle_pipe
[0], POLLHUP
, IDLE_TIMEOUT2_USEC
);
1759 idle_pipe
[0] = safe_close(idle_pipe
[0]);
1763 idle_pipe
[3] = safe_close(idle_pipe
[3]);
1766 static const char *exec_directory_env_name_to_string(ExecDirectoryType t
);
1768 /* And this table also maps ExecDirectoryType, to the environment variable we pass the selected directory to
1769 * the service payload in. */
1770 static const char* const exec_directory_env_name_table
[_EXEC_DIRECTORY_TYPE_MAX
] = {
1771 [EXEC_DIRECTORY_RUNTIME
] = "RUNTIME_DIRECTORY",
1772 [EXEC_DIRECTORY_STATE
] = "STATE_DIRECTORY",
1773 [EXEC_DIRECTORY_CACHE
] = "CACHE_DIRECTORY",
1774 [EXEC_DIRECTORY_LOGS
] = "LOGS_DIRECTORY",
1775 [EXEC_DIRECTORY_CONFIGURATION
] = "CONFIGURATION_DIRECTORY",
1778 DEFINE_PRIVATE_STRING_TABLE_LOOKUP_TO_STRING(exec_directory_env_name
, ExecDirectoryType
);
1780 static int build_environment(
1781 const ExecContext
*c
,
1782 const ExecParameters
*p
,
1783 const CGroupContext
*cgroup_context
,
1786 const char *username
,
1788 dev_t journal_stream_dev
,
1789 ino_t journal_stream_ino
,
1790 const char *memory_pressure_path
,
1793 _cleanup_strv_free_
char **our_env
= NULL
;
1802 #define N_ENV_VARS 19
1803 our_env
= new0(char*, N_ENV_VARS
+ _EXEC_DIRECTORY_TYPE_MAX
);
1808 _cleanup_free_
char *joined
= NULL
;
1810 if (asprintf(&x
, "LISTEN_PID="PID_FMT
, getpid_cached()) < 0)
1812 our_env
[n_env
++] = x
;
1814 if (asprintf(&x
, "LISTEN_FDS=%zu", n_fds
) < 0)
1816 our_env
[n_env
++] = x
;
1818 joined
= strv_join(p
->fd_names
, ":");
1822 x
= strjoin("LISTEN_FDNAMES=", joined
);
1825 our_env
[n_env
++] = x
;
1828 if ((p
->flags
& EXEC_SET_WATCHDOG
) && p
->watchdog_usec
> 0) {
1829 if (asprintf(&x
, "WATCHDOG_PID="PID_FMT
, getpid_cached()) < 0)
1831 our_env
[n_env
++] = x
;
1833 if (asprintf(&x
, "WATCHDOG_USEC="USEC_FMT
, p
->watchdog_usec
) < 0)
1835 our_env
[n_env
++] = x
;
1838 /* If this is D-Bus, tell the nss-systemd module, since it relies on being able to use blocking
1839 * Varlink calls back to us for look up dynamic users in PID 1. Break the deadlock between D-Bus and
1840 * PID 1 by disabling use of PID1' NSS interface for looking up dynamic users. */
1841 if (p
->flags
& EXEC_NSS_DYNAMIC_BYPASS
) {
1842 x
= strdup("SYSTEMD_NSS_DYNAMIC_BYPASS=1");
1845 our_env
[n_env
++] = x
;
1848 /* We query "root" if this is a system unit and User= is not specified. $USER is always set. $HOME
1849 * could cause problem for e.g. getty, since login doesn't override $HOME, and $LOGNAME and $SHELL don't
1850 * really make much sense since we're not logged in. Hence we conditionalize the three based on
1851 * SetLoginEnvironment= switch. */
1852 if (!c
->user
&& !c
->dynamic_user
&& p
->runtime_scope
== RUNTIME_SCOPE_SYSTEM
) {
1853 r
= get_fixed_user("root", &username
, NULL
, NULL
, &home
, &shell
);
1855 return log_exec_debug_errno(c
,
1858 "Failed to determine user credentials for root: %m");
1861 bool set_user_login_env
= exec_context_get_set_login_environment(c
);
1864 x
= strjoin("USER=", username
);
1867 our_env
[n_env
++] = x
;
1869 if (set_user_login_env
) {
1870 x
= strjoin("LOGNAME=", username
);
1873 our_env
[n_env
++] = x
;
1877 if (home
&& set_user_login_env
) {
1878 x
= strjoin("HOME=", home
);
1882 path_simplify(x
+ 5);
1883 our_env
[n_env
++] = x
;
1886 if (shell
&& set_user_login_env
) {
1887 x
= strjoin("SHELL=", shell
);
1891 path_simplify(x
+ 6);
1892 our_env
[n_env
++] = x
;
1895 if (!sd_id128_is_null(p
->invocation_id
)) {
1896 assert(p
->invocation_id_string
);
1898 x
= strjoin("INVOCATION_ID=", p
->invocation_id_string
);
1902 our_env
[n_env
++] = x
;
1905 if (exec_context_needs_term(c
)) {
1906 _cleanup_free_
char *cmdline
= NULL
;
1907 const char *tty_path
, *term
= NULL
;
1909 tty_path
= exec_context_tty_path(c
);
1911 /* If we are forked off PID 1 and we are supposed to operate on /dev/console, then let's try
1912 * to inherit the $TERM set for PID 1. This is useful for containers so that the $TERM the
1913 * container manager passes to PID 1 ends up all the way in the console login shown. */
1915 if (path_equal_ptr(tty_path
, "/dev/console") && getppid() == 1)
1916 term
= getenv("TERM");
1917 else if (tty_path
&& in_charset(skip_dev_prefix(tty_path
), ALPHANUMERICAL
)) {
1918 _cleanup_free_
char *key
= NULL
;
1920 key
= strjoin("systemd.tty.term.", skip_dev_prefix(tty_path
));
1924 r
= proc_cmdline_get_key(key
, 0, &cmdline
);
1926 log_exec_debug_errno(c
,
1929 "Failed to read %s from kernel cmdline, ignoring: %m",
1936 term
= default_term_for_tty(tty_path
);
1938 x
= strjoin("TERM=", term
);
1941 our_env
[n_env
++] = x
;
1944 if (journal_stream_dev
!= 0 && journal_stream_ino
!= 0) {
1945 if (asprintf(&x
, "JOURNAL_STREAM=" DEV_FMT
":" INO_FMT
, journal_stream_dev
, journal_stream_ino
) < 0)
1948 our_env
[n_env
++] = x
;
1951 if (c
->log_namespace
) {
1952 x
= strjoin("LOG_NAMESPACE=", c
->log_namespace
);
1956 our_env
[n_env
++] = x
;
1959 for (ExecDirectoryType t
= 0; t
< _EXEC_DIRECTORY_TYPE_MAX
; t
++) {
1960 _cleanup_free_
char *joined
= NULL
;
1966 if (c
->directories
[t
].n_items
== 0)
1969 n
= exec_directory_env_name_to_string(t
);
1973 for (size_t i
= 0; i
< c
->directories
[t
].n_items
; i
++) {
1974 _cleanup_free_
char *prefixed
= NULL
;
1976 prefixed
= path_join(p
->prefix
[t
], c
->directories
[t
].items
[i
].path
);
1980 if (!strextend_with_separator(&joined
, ":", prefixed
))
1984 x
= strjoin(n
, "=", joined
);
1988 our_env
[n_env
++] = x
;
1991 _cleanup_free_
char *creds_dir
= NULL
;
1992 r
= exec_context_get_credential_directory(c
, p
, p
->unit_id
, &creds_dir
);
1996 x
= strjoin("CREDENTIALS_DIRECTORY=", creds_dir
);
2000 our_env
[n_env
++] = x
;
2003 if (asprintf(&x
, "SYSTEMD_EXEC_PID=" PID_FMT
, getpid_cached()) < 0)
2006 our_env
[n_env
++] = x
;
2008 if (memory_pressure_path
) {
2009 x
= strjoin("MEMORY_PRESSURE_WATCH=", memory_pressure_path
);
2013 our_env
[n_env
++] = x
;
2015 if (cgroup_context
&& !path_equal(memory_pressure_path
, "/dev/null")) {
2016 _cleanup_free_
char *b
= NULL
, *e
= NULL
;
2018 if (asprintf(&b
, "%s " USEC_FMT
" " USEC_FMT
,
2019 MEMORY_PRESSURE_DEFAULT_TYPE
,
2020 cgroup_context
->memory_pressure_threshold_usec
== USEC_INFINITY
? MEMORY_PRESSURE_DEFAULT_THRESHOLD_USEC
:
2021 CLAMP(cgroup_context
->memory_pressure_threshold_usec
, 1U, MEMORY_PRESSURE_DEFAULT_WINDOW_USEC
),
2022 MEMORY_PRESSURE_DEFAULT_WINDOW_USEC
) < 0)
2025 if (base64mem(b
, strlen(b
) + 1, &e
) < 0)
2028 x
= strjoin("MEMORY_PRESSURE_WRITE=", e
);
2032 our_env
[n_env
++] = x
;
2036 assert(n_env
< N_ENV_VARS
+ _EXEC_DIRECTORY_TYPE_MAX
);
2039 *ret
= TAKE_PTR(our_env
);
2044 static int build_pass_environment(const ExecContext
*c
, char ***ret
) {
2045 _cleanup_strv_free_
char **pass_env
= NULL
;
2048 STRV_FOREACH(i
, c
->pass_environment
) {
2049 _cleanup_free_
char *x
= NULL
;
2055 x
= strjoin(*i
, "=", v
);
2059 if (!GREEDY_REALLOC(pass_env
, n_env
+ 2))
2062 pass_env
[n_env
++] = TAKE_PTR(x
);
2063 pass_env
[n_env
] = NULL
;
2066 *ret
= TAKE_PTR(pass_env
);
2071 static int setup_private_users(uid_t ouid
, gid_t ogid
, uid_t uid
, gid_t gid
) {
2072 _cleanup_free_
char *uid_map
= NULL
, *gid_map
= NULL
;
2073 _cleanup_close_pair_
int errno_pipe
[2] = EBADF_PAIR
;
2074 _cleanup_close_
int unshare_ready_fd
= -EBADF
;
2075 _cleanup_(sigkill_waitp
) pid_t pid
= 0;
2080 /* Set up a user namespace and map the original UID/GID (IDs from before any user or group changes, i.e.
2081 * the IDs from the user or system manager(s)) to itself, the selected UID/GID to itself, and everything else to
2082 * nobody. In order to be able to write this mapping we need CAP_SETUID in the original user namespace, which
2083 * we however lack after opening the user namespace. To work around this we fork() a temporary child process,
2084 * which waits for the parent to create the new user namespace while staying in the original namespace. The
2085 * child then writes the UID mapping, under full privileges. The parent waits for the child to finish and
2086 * continues execution normally.
2087 * For unprivileged users (i.e. without capabilities), the root to root mapping is excluded. As such, it
2088 * does not need CAP_SETUID to write the single line mapping to itself. */
2090 /* Can only set up multiple mappings with CAP_SETUID. */
2091 if (have_effective_cap(CAP_SETUID
) > 0 && uid
!= ouid
&& uid_is_valid(uid
))
2092 r
= asprintf(&uid_map
,
2093 UID_FMT
" " UID_FMT
" 1\n" /* Map $OUID → $OUID */
2094 UID_FMT
" " UID_FMT
" 1\n", /* Map $UID → $UID */
2095 ouid
, ouid
, uid
, uid
);
2097 r
= asprintf(&uid_map
,
2098 UID_FMT
" " UID_FMT
" 1\n", /* Map $OUID → $OUID */
2104 /* Can only set up multiple mappings with CAP_SETGID. */
2105 if (have_effective_cap(CAP_SETGID
) > 0 && gid
!= ogid
&& gid_is_valid(gid
))
2106 r
= asprintf(&gid_map
,
2107 GID_FMT
" " GID_FMT
" 1\n" /* Map $OGID → $OGID */
2108 GID_FMT
" " GID_FMT
" 1\n", /* Map $GID → $GID */
2109 ogid
, ogid
, gid
, gid
);
2111 r
= asprintf(&gid_map
,
2112 GID_FMT
" " GID_FMT
" 1\n", /* Map $OGID -> $OGID */
2118 /* Create a communication channel so that the parent can tell the child when it finished creating the user
2120 unshare_ready_fd
= eventfd(0, EFD_CLOEXEC
);
2121 if (unshare_ready_fd
< 0)
2124 /* Create a communication channel so that the child can tell the parent a proper error code in case it
2126 if (pipe2(errno_pipe
, O_CLOEXEC
) < 0)
2129 r
= safe_fork("(sd-userns)", FORK_RESET_SIGNALS
|FORK_DEATHSIG_SIGKILL
, &pid
);
2133 _cleanup_close_
int fd
= -EBADF
;
2137 /* Child process, running in the original user namespace. Let's update the parent's UID/GID map from
2138 * here, after the parent opened its own user namespace. */
2141 errno_pipe
[0] = safe_close(errno_pipe
[0]);
2143 /* Wait until the parent unshared the user namespace */
2144 if (read(unshare_ready_fd
, &c
, sizeof(c
)) < 0) {
2149 /* Disable the setgroups() system call in the child user namespace, for good. */
2150 a
= procfs_file_alloca(ppid
, "setgroups");
2151 fd
= open(a
, O_WRONLY
|O_CLOEXEC
);
2153 if (errno
!= ENOENT
) {
2158 /* If the file is missing the kernel is too old, let's continue anyway. */
2160 if (write(fd
, "deny\n", 5) < 0) {
2165 fd
= safe_close(fd
);
2168 /* First write the GID map */
2169 a
= procfs_file_alloca(ppid
, "gid_map");
2170 fd
= open(a
, O_WRONLY
|O_CLOEXEC
);
2175 if (write(fd
, gid_map
, strlen(gid_map
)) < 0) {
2179 fd
= safe_close(fd
);
2181 /* The write the UID map */
2182 a
= procfs_file_alloca(ppid
, "uid_map");
2183 fd
= open(a
, O_WRONLY
|O_CLOEXEC
);
2188 if (write(fd
, uid_map
, strlen(uid_map
)) < 0) {
2193 _exit(EXIT_SUCCESS
);
2196 (void) write(errno_pipe
[1], &r
, sizeof(r
));
2197 _exit(EXIT_FAILURE
);
2200 errno_pipe
[1] = safe_close(errno_pipe
[1]);
2202 if (unshare(CLONE_NEWUSER
) < 0)
2205 /* Let the child know that the namespace is ready now */
2206 if (write(unshare_ready_fd
, &c
, sizeof(c
)) < 0)
2209 /* Try to read an error code from the child */
2210 n
= read(errno_pipe
[0], &r
, sizeof(r
));
2213 if (n
== sizeof(r
)) { /* an error code was sent to us */
2218 if (n
!= 0) /* on success we should have read 0 bytes */
2221 r
= wait_for_terminate_and_check("(sd-userns)", TAKE_PID(pid
), 0);
2224 if (r
!= EXIT_SUCCESS
) /* If something strange happened with the child, let's consider this fatal, too */
2230 static int create_many_symlinks(const char *root
, const char *source
, char **symlinks
) {
2231 _cleanup_free_
char *src_abs
= NULL
;
2236 src_abs
= path_join(root
, source
);
2240 STRV_FOREACH(dst
, symlinks
) {
2241 _cleanup_free_
char *dst_abs
= NULL
;
2243 dst_abs
= path_join(root
, *dst
);
2247 r
= mkdir_parents_label(dst_abs
, 0755);
2251 r
= symlink_idempotent(src_abs
, dst_abs
, true);
2259 static int setup_exec_directory(
2260 const ExecContext
*context
,
2261 const ExecParameters
*params
,
2264 ExecDirectoryType type
,
2265 bool needs_mount_namespace
,
2268 static const int exit_status_table
[_EXEC_DIRECTORY_TYPE_MAX
] = {
2269 [EXEC_DIRECTORY_RUNTIME
] = EXIT_RUNTIME_DIRECTORY
,
2270 [EXEC_DIRECTORY_STATE
] = EXIT_STATE_DIRECTORY
,
2271 [EXEC_DIRECTORY_CACHE
] = EXIT_CACHE_DIRECTORY
,
2272 [EXEC_DIRECTORY_LOGS
] = EXIT_LOGS_DIRECTORY
,
2273 [EXEC_DIRECTORY_CONFIGURATION
] = EXIT_CONFIGURATION_DIRECTORY
,
2279 assert(type
>= 0 && type
< _EXEC_DIRECTORY_TYPE_MAX
);
2280 assert(exit_status
);
2282 if (!params
->prefix
[type
])
2285 if (params
->flags
& EXEC_CHOWN_DIRECTORIES
) {
2286 if (!uid_is_valid(uid
))
2288 if (!gid_is_valid(gid
))
2292 for (size_t i
= 0; i
< context
->directories
[type
].n_items
; i
++) {
2293 _cleanup_free_
char *p
= NULL
, *pp
= NULL
;
2295 p
= path_join(params
->prefix
[type
], context
->directories
[type
].items
[i
].path
);
2301 r
= mkdir_parents_label(p
, 0755);
2305 if (IN_SET(type
, EXEC_DIRECTORY_STATE
, EXEC_DIRECTORY_LOGS
) && params
->runtime_scope
== RUNTIME_SCOPE_USER
) {
2307 /* If we are in user mode, and a configuration directory exists but a state directory
2308 * doesn't exist, then we likely are upgrading from an older systemd version that
2309 * didn't know the more recent addition to the xdg-basedir spec: the $XDG_STATE_HOME
2310 * directory. In older systemd versions EXEC_DIRECTORY_STATE was aliased to
2311 * EXEC_DIRECTORY_CONFIGURATION, with the advent of $XDG_STATE_HOME is is now
2312 * separated. If a service has both dirs configured but only the configuration dir
2313 * exists and the state dir does not, we assume we are looking at an update
2314 * situation. Hence, create a compatibility symlink, so that all expectations are
2317 * (We also do something similar with the log directory, which still doesn't exist in
2318 * the xdg basedir spec. We'll make it a subdir of the state dir.) */
2320 /* this assumes the state dir is always created before the configuration dir */
2321 assert_cc(EXEC_DIRECTORY_STATE
< EXEC_DIRECTORY_LOGS
);
2322 assert_cc(EXEC_DIRECTORY_LOGS
< EXEC_DIRECTORY_CONFIGURATION
);
2324 r
= laccess(p
, F_OK
);
2326 _cleanup_free_
char *q
= NULL
;
2328 /* OK, we know that the state dir does not exist. Let's see if the dir exists
2329 * under the configuration hierarchy. */
2331 if (type
== EXEC_DIRECTORY_STATE
)
2332 q
= path_join(params
->prefix
[EXEC_DIRECTORY_CONFIGURATION
], context
->directories
[type
].items
[i
].path
);
2333 else if (type
== EXEC_DIRECTORY_LOGS
)
2334 q
= path_join(params
->prefix
[EXEC_DIRECTORY_CONFIGURATION
], "log", context
->directories
[type
].items
[i
].path
);
2336 assert_not_reached();
2342 r
= laccess(q
, F_OK
);
2344 /* It does exist! This hence looks like an update. Symlink the
2345 * configuration directory into the state directory. */
2347 r
= symlink_idempotent(q
, p
, /* make_relative= */ true);
2351 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
);
2353 } else if (r
!= -ENOENT
)
2354 log_exec_warning_errno(context
, params
, r
, "Unable to detect whether unit configuration directory '%s' exists, assuming not: %m", q
);
2357 log_exec_warning_errno(context
, params
, r
, "Unable to detect whether unit state directory '%s' is missing, assuming it is: %m", p
);
2360 if (exec_directory_is_private(context
, type
)) {
2361 /* So, here's one extra complication when dealing with DynamicUser=1 units. In that
2362 * case we want to avoid leaving a directory around fully accessible that is owned by
2363 * a dynamic user whose UID is later on reused. To lock this down we use the same
2364 * trick used by container managers to prohibit host users to get access to files of
2365 * the same UID in containers: we place everything inside a directory that has an
2366 * access mode of 0700 and is owned root:root, so that it acts as security boundary
2367 * for unprivileged host code. We then use fs namespacing to make this directory
2368 * permeable for the service itself.
2370 * Specifically: for a service which wants a special directory "foo/" we first create
2371 * a directory "private/" with access mode 0700 owned by root:root. Then we place
2372 * "foo" inside of that directory (i.e. "private/foo/"), and make "foo" a symlink to
2373 * "private/foo". This way, privileged host users can access "foo/" as usual, but
2374 * unprivileged host users can't look into it. Inside of the namespace of the unit
2375 * "private/" is replaced by a more liberally accessible tmpfs, into which the host's
2376 * "private/foo/" is mounted under the same name, thus disabling the access boundary
2377 * for the service and making sure it only gets access to the dirs it needs but no
2378 * others. Tricky? Yes, absolutely, but it works!
2380 * Note that we don't do this for EXEC_DIRECTORY_CONFIGURATION as that's assumed not
2381 * to be owned by the service itself.
2383 * Also, note that we don't do this for EXEC_DIRECTORY_RUNTIME as that's often used
2384 * for sharing files or sockets with other services. */
2386 pp
= path_join(params
->prefix
[type
], "private");
2392 /* First set up private root if it doesn't exist yet, with access mode 0700 and owned by root:root */
2393 r
= mkdir_safe_label(pp
, 0700, 0, 0, MKDIR_WARN_MODE
);
2397 if (!path_extend(&pp
, context
->directories
[type
].items
[i
].path
)) {
2402 /* Create all directories between the configured directory and this private root, and mark them 0755 */
2403 r
= mkdir_parents_label(pp
, 0755);
2407 if (is_dir(p
, false) > 0 &&
2408 (laccess(pp
, F_OK
) == -ENOENT
)) {
2410 /* Hmm, the private directory doesn't exist yet, but the normal one exists? If so, move
2411 * it over. Most likely the service has been upgraded from one that didn't use
2412 * DynamicUser=1, to one that does. */
2414 log_exec_info(context
,
2416 "Found pre-existing public %s= directory %s, migrating to %s.\n"
2417 "Apparently, service previously had DynamicUser= turned off, and has now turned it on.",
2418 exec_directory_type_to_string(type
), p
, pp
);
2420 r
= RET_NERRNO(rename(p
, pp
));
2424 /* Otherwise, create the actual directory for the service */
2426 r
= mkdir_label(pp
, context
->directories
[type
].mode
);
2427 if (r
< 0 && r
!= -EEXIST
)
2431 if (!context
->directories
[type
].items
[i
].only_create
) {
2432 /* And link it up from the original place.
2434 * 1) If a mount namespace is going to be used, then this symlink remains on
2435 * the host, and a new one for the child namespace will be created later.
2436 * 2) It is not necessary to create this symlink when one of its parent
2437 * directories is specified and already created. E.g.
2438 * StateDirectory=foo foo/bar
2439 * In that case, the inode points to pp and p for "foo/bar" are the same:
2440 * pp = "/var/lib/private/foo/bar"
2441 * p = "/var/lib/foo/bar"
2442 * and, /var/lib/foo is a symlink to /var/lib/private/foo. So, not only
2443 * we do not need to create the symlink, but we cannot create the symlink.
2444 * See issue #24783. */
2445 r
= symlink_idempotent(pp
, p
, true);
2451 _cleanup_free_
char *target
= NULL
;
2453 if (type
!= EXEC_DIRECTORY_CONFIGURATION
&&
2454 readlink_and_make_absolute(p
, &target
) >= 0) {
2455 _cleanup_free_
char *q
= NULL
, *q_resolved
= NULL
, *target_resolved
= NULL
;
2457 /* This already exists and is a symlink? Interesting. Maybe it's one created
2458 * by DynamicUser=1 (see above)?
2460 * We do this for all directory types except for ConfigurationDirectory=,
2461 * since they all support the private/ symlink logic at least in some
2462 * configurations, see above. */
2464 r
= chase(target
, NULL
, 0, &target_resolved
, NULL
);
2468 q
= path_join(params
->prefix
[type
], "private", context
->directories
[type
].items
[i
].path
);
2474 /* /var/lib or friends may be symlinks. So, let's chase them also. */
2475 r
= chase(q
, NULL
, CHASE_NONEXISTENT
, &q_resolved
, NULL
);
2479 if (path_equal(q_resolved
, target_resolved
)) {
2481 /* Hmm, apparently DynamicUser= was once turned on for this service,
2482 * but is no longer. Let's move the directory back up. */
2484 log_exec_info(context
,
2486 "Found pre-existing private %s= directory %s, migrating to %s.\n"
2487 "Apparently, service previously had DynamicUser= turned on, and has now turned it off.",
2488 exec_directory_type_to_string(type
), q
, p
);
2490 r
= RET_NERRNO(unlink(p
));
2494 r
= RET_NERRNO(rename(q
, p
));
2500 r
= mkdir_label(p
, context
->directories
[type
].mode
);
2505 if (type
== EXEC_DIRECTORY_CONFIGURATION
) {
2508 /* Don't change the owner/access mode of the configuration directory,
2509 * as in the common case it is not written to by a service, and shall
2510 * not be writable. */
2512 r
= RET_NERRNO(stat(p
, &st
));
2516 /* Still complain if the access mode doesn't match */
2517 if (((st
.st_mode
^ context
->directories
[type
].mode
) & 07777) != 0)
2518 log_exec_warning(context
,
2520 "%s \'%s\' already exists but the mode is different. "
2521 "(File system: %o %sMode: %o)",
2522 exec_directory_type_to_string(type
), context
->directories
[type
].items
[i
].path
,
2523 st
.st_mode
& 07777, exec_directory_type_to_string(type
), context
->directories
[type
].mode
& 07777);
2530 /* Lock down the access mode (we use chmod_and_chown() to make this idempotent. We don't
2531 * specify UID/GID here, so that path_chown_recursive() can optimize things depending on the
2532 * current UID/GID ownership.) */
2533 r
= chmod_and_chown(pp
?: p
, context
->directories
[type
].mode
, UID_INVALID
, GID_INVALID
);
2537 /* Skip the rest (which deals with ownership) in user mode, since ownership changes are not
2538 * available to user code anyway */
2539 if (params
->runtime_scope
!= RUNTIME_SCOPE_SYSTEM
)
2542 /* Then, change the ownership of the whole tree, if necessary. When dynamic users are used we
2543 * drop the suid/sgid bits, since we really don't want SUID/SGID files for dynamic UID/GID
2544 * assignments to exist. */
2545 r
= path_chown_recursive(pp
?: p
, uid
, gid
, context
->dynamic_user
? 01777 : 07777, AT_SYMLINK_FOLLOW
);
2550 /* If we are not going to run in a namespace, set up the symlinks - otherwise
2551 * they are set up later, to allow configuring empty var/run/etc. */
2552 if (!needs_mount_namespace
)
2553 for (size_t i
= 0; i
< context
->directories
[type
].n_items
; i
++) {
2554 r
= create_many_symlinks(params
->prefix
[type
],
2555 context
->directories
[type
].items
[i
].path
,
2556 context
->directories
[type
].items
[i
].symlinks
);
2564 *exit_status
= exit_status_table
[type
];
2569 static int setup_smack(
2570 const ExecParameters
*params
,
2571 const ExecContext
*context
,
2572 int executable_fd
) {
2576 assert(executable_fd
>= 0);
2578 if (context
->smack_process_label
) {
2579 r
= mac_smack_apply_pid(0, context
->smack_process_label
);
2582 } else if (params
->fallback_smack_process_label
) {
2583 _cleanup_free_
char *exec_label
= NULL
;
2585 r
= mac_smack_read_fd(executable_fd
, SMACK_ATTR_EXEC
, &exec_label
);
2586 if (r
< 0 && !ERRNO_IS_XATTR_ABSENT(r
))
2589 r
= mac_smack_apply_pid(0, exec_label
?: params
->fallback_smack_process_label
);
2598 static int compile_bind_mounts(
2599 const ExecContext
*context
,
2600 const ExecParameters
*params
,
2601 BindMount
**ret_bind_mounts
,
2602 size_t *ret_n_bind_mounts
,
2603 char ***ret_empty_directories
) {
2605 _cleanup_strv_free_
char **empty_directories
= NULL
;
2606 BindMount
*bind_mounts
= NULL
;
2612 assert(ret_bind_mounts
);
2613 assert(ret_n_bind_mounts
);
2614 assert(ret_empty_directories
);
2616 CLEANUP_ARRAY(bind_mounts
, h
, bind_mount_free_many
);
2618 n
= context
->n_bind_mounts
;
2619 for (ExecDirectoryType t
= 0; t
< _EXEC_DIRECTORY_TYPE_MAX
; t
++) {
2620 if (!params
->prefix
[t
])
2623 for (size_t i
= 0; i
< context
->directories
[t
].n_items
; i
++)
2624 n
+= !context
->directories
[t
].items
[i
].only_create
;
2628 *ret_bind_mounts
= NULL
;
2629 *ret_n_bind_mounts
= 0;
2630 *ret_empty_directories
= NULL
;
2634 bind_mounts
= new(BindMount
, n
);
2638 for (size_t i
= 0; i
< context
->n_bind_mounts
; i
++) {
2639 BindMount
*item
= context
->bind_mounts
+ i
;
2640 _cleanup_free_
char *s
= NULL
, *d
= NULL
;
2642 s
= strdup(item
->source
);
2646 d
= strdup(item
->destination
);
2650 bind_mounts
[h
++] = (BindMount
) {
2651 .source
= TAKE_PTR(s
),
2652 .destination
= TAKE_PTR(d
),
2653 .read_only
= item
->read_only
,
2654 .recursive
= item
->recursive
,
2655 .ignore_enoent
= item
->ignore_enoent
,
2659 for (ExecDirectoryType t
= 0; t
< _EXEC_DIRECTORY_TYPE_MAX
; t
++) {
2660 if (!params
->prefix
[t
])
2663 if (context
->directories
[t
].n_items
== 0)
2666 if (exec_directory_is_private(context
, t
) &&
2667 !exec_context_with_rootfs(context
)) {
2670 /* So this is for a dynamic user, and we need to make sure the process can access its own
2671 * directory. For that we overmount the usually inaccessible "private" subdirectory with a
2672 * tmpfs that makes it accessible and is empty except for the submounts we do this for. */
2674 private_root
= path_join(params
->prefix
[t
], "private");
2678 r
= strv_consume(&empty_directories
, private_root
);
2683 for (size_t i
= 0; i
< context
->directories
[t
].n_items
; i
++) {
2684 _cleanup_free_
char *s
= NULL
, *d
= NULL
;
2686 /* When one of the parent directories is in the list, we cannot create the symlink
2687 * for the child directory. See also the comments in setup_exec_directory(). */
2688 if (context
->directories
[t
].items
[i
].only_create
)
2691 if (exec_directory_is_private(context
, t
))
2692 s
= path_join(params
->prefix
[t
], "private", context
->directories
[t
].items
[i
].path
);
2694 s
= path_join(params
->prefix
[t
], context
->directories
[t
].items
[i
].path
);
2698 if (exec_directory_is_private(context
, t
) &&
2699 exec_context_with_rootfs(context
))
2700 /* When RootDirectory= or RootImage= are set, then the symbolic link to the private
2701 * directory is not created on the root directory. So, let's bind-mount the directory
2702 * on the 'non-private' place. */
2703 d
= path_join(params
->prefix
[t
], context
->directories
[t
].items
[i
].path
);
2709 bind_mounts
[h
++] = (BindMount
) {
2710 .source
= TAKE_PTR(s
),
2711 .destination
= TAKE_PTR(d
),
2713 .nosuid
= context
->dynamic_user
, /* don't allow suid/sgid when DynamicUser= is on */
2715 .ignore_enoent
= false,
2722 *ret_bind_mounts
= TAKE_PTR(bind_mounts
);
2723 *ret_n_bind_mounts
= n
;
2724 *ret_empty_directories
= TAKE_PTR(empty_directories
);
2729 /* ret_symlinks will contain a list of pairs src:dest that describes
2730 * the symlinks to create later on. For example, the symlinks needed
2731 * to safely give private directories to DynamicUser=1 users. */
2732 static int compile_symlinks(
2733 const ExecContext
*context
,
2734 const ExecParameters
*params
,
2735 bool setup_os_release_symlink
,
2736 char ***ret_symlinks
) {
2738 _cleanup_strv_free_
char **symlinks
= NULL
;
2743 assert(ret_symlinks
);
2745 for (ExecDirectoryType dt
= 0; dt
< _EXEC_DIRECTORY_TYPE_MAX
; dt
++) {
2746 for (size_t i
= 0; i
< context
->directories
[dt
].n_items
; i
++) {
2747 _cleanup_free_
char *private_path
= NULL
, *path
= NULL
;
2749 STRV_FOREACH(symlink
, context
->directories
[dt
].items
[i
].symlinks
) {
2750 _cleanup_free_
char *src_abs
= NULL
, *dst_abs
= NULL
;
2752 src_abs
= path_join(params
->prefix
[dt
], context
->directories
[dt
].items
[i
].path
);
2753 dst_abs
= path_join(params
->prefix
[dt
], *symlink
);
2754 if (!src_abs
|| !dst_abs
)
2757 r
= strv_consume_pair(&symlinks
, TAKE_PTR(src_abs
), TAKE_PTR(dst_abs
));
2762 if (!exec_directory_is_private(context
, dt
) ||
2763 exec_context_with_rootfs(context
) ||
2764 context
->directories
[dt
].items
[i
].only_create
)
2767 private_path
= path_join(params
->prefix
[dt
], "private", context
->directories
[dt
].items
[i
].path
);
2771 path
= path_join(params
->prefix
[dt
], context
->directories
[dt
].items
[i
].path
);
2775 r
= strv_consume_pair(&symlinks
, TAKE_PTR(private_path
), TAKE_PTR(path
));
2781 /* We make the host's os-release available via a symlink, so that we can copy it atomically
2782 * and readers will never get a half-written version. Note that, while the paths specified here are
2783 * absolute, when they are processed in namespace.c they will be made relative automatically, i.e.:
2784 * 'os-release -> .os-release-stage/os-release' is what will be created. */
2785 if (setup_os_release_symlink
) {
2786 r
= strv_extend_many(
2788 "/run/host/.os-release-stage/os-release",
2789 "/run/host/os-release");
2794 *ret_symlinks
= TAKE_PTR(symlinks
);
2799 static bool insist_on_sandboxing(
2800 const ExecContext
*context
,
2801 const char *root_dir
,
2802 const char *root_image
,
2803 const BindMount
*bind_mounts
,
2804 size_t n_bind_mounts
) {
2807 assert(n_bind_mounts
== 0 || bind_mounts
);
2809 /* Checks whether we need to insist on fs namespacing. i.e. whether we have settings configured that
2810 * would alter the view on the file system beyond making things read-only or invisible, i.e. would
2811 * rearrange stuff in a way we cannot ignore gracefully. */
2813 if (context
->n_temporary_filesystems
> 0)
2816 if (root_dir
|| root_image
)
2819 if (context
->n_mount_images
> 0)
2822 if (context
->dynamic_user
)
2825 if (context
->n_extension_images
> 0 || !strv_isempty(context
->extension_directories
))
2828 /* If there are any bind mounts set that don't map back onto themselves, fs namespacing becomes
2830 for (size_t i
= 0; i
< n_bind_mounts
; i
++)
2831 if (!path_equal(bind_mounts
[i
].source
, bind_mounts
[i
].destination
))
2834 if (context
->log_namespace
)
2840 static int setup_ephemeral(
2841 const ExecContext
*context
,
2842 ExecRuntime
*runtime
,
2843 char **root_image
, /* both input and output! modified if ephemeral logic enabled */
2844 char **root_directory
) { /* ditto */
2846 _cleanup_close_
int fd
= -EBADF
;
2847 _cleanup_free_
char *new_root
= NULL
;
2852 assert(root_directory
);
2854 if (!*root_image
&& !*root_directory
)
2857 if (!runtime
|| !runtime
->ephemeral_copy
)
2860 assert(runtime
->ephemeral_storage_socket
[0] >= 0);
2861 assert(runtime
->ephemeral_storage_socket
[1] >= 0);
2863 new_root
= strdup(runtime
->ephemeral_copy
);
2865 return log_oom_debug();
2867 r
= posix_lock(runtime
->ephemeral_storage_socket
[0], LOCK_EX
);
2869 return log_debug_errno(r
, "Failed to lock ephemeral storage socket: %m");
2871 CLEANUP_POSIX_UNLOCK(runtime
->ephemeral_storage_socket
[0]);
2873 fd
= receive_one_fd(runtime
->ephemeral_storage_socket
[0], MSG_PEEK
|MSG_DONTWAIT
);
2875 /* We got an fd! That means ephemeral has already been set up, so nothing to do here. */
2878 return log_debug_errno(fd
, "Failed to receive file descriptor queued on ephemeral storage socket: %m");
2881 log_debug("Making ephemeral copy of %s to %s", *root_image
, new_root
);
2883 fd
= copy_file(*root_image
,
2891 return log_debug_errno(fd
, "Failed to copy image %s to %s: %m",
2892 *root_image
, new_root
);
2894 /* A root image might be subject to lots of random writes so let's try to disable COW on it
2895 * which tends to not perform well in combination with lots of random writes.
2897 * Note: btrfs actually isn't impressed by us setting the flag after making the reflink'ed
2898 * copy, but we at least want to make the intention clear.
2900 r
= chattr_fd(fd
, FS_NOCOW_FL
, FS_NOCOW_FL
, NULL
);
2902 log_debug_errno(fd
, "Failed to disable copy-on-write for %s, ignoring: %m", new_root
);
2904 assert(*root_directory
);
2906 log_debug("Making ephemeral snapshot of %s to %s", *root_directory
, new_root
);
2908 fd
= btrfs_subvol_snapshot_at(
2909 AT_FDCWD
, *root_directory
,
2911 BTRFS_SNAPSHOT_FALLBACK_COPY
|
2912 BTRFS_SNAPSHOT_FALLBACK_DIRECTORY
|
2913 BTRFS_SNAPSHOT_RECURSIVE
|
2914 BTRFS_SNAPSHOT_LOCK_BSD
);
2916 return log_debug_errno(fd
, "Failed to snapshot directory %s to %s: %m",
2917 *root_directory
, new_root
);
2920 r
= send_one_fd(runtime
->ephemeral_storage_socket
[1], fd
, MSG_DONTWAIT
);
2922 return log_debug_errno(r
, "Failed to queue file descriptor on ephemeral storage socket: %m");
2925 free_and_replace(*root_image
, new_root
);
2927 assert(*root_directory
);
2928 free_and_replace(*root_directory
, new_root
);
2934 static int verity_settings_prepare(
2935 VeritySettings
*verity
,
2936 const char *root_image
,
2937 const void *root_hash
,
2938 size_t root_hash_size
,
2939 const char *root_hash_path
,
2940 const void *root_hash_sig
,
2941 size_t root_hash_sig_size
,
2942 const char *root_hash_sig_path
,
2943 const char *verity_data_path
) {
2952 d
= memdup(root_hash
, root_hash_size
);
2956 free_and_replace(verity
->root_hash
, d
);
2957 verity
->root_hash_size
= root_hash_size
;
2958 verity
->designator
= PARTITION_ROOT
;
2961 if (root_hash_sig
) {
2964 d
= memdup(root_hash_sig
, root_hash_sig_size
);
2968 free_and_replace(verity
->root_hash_sig
, d
);
2969 verity
->root_hash_sig_size
= root_hash_sig_size
;
2970 verity
->designator
= PARTITION_ROOT
;
2973 if (verity_data_path
) {
2974 r
= free_and_strdup(&verity
->data_path
, verity_data_path
);
2979 r
= verity_settings_load(
2983 root_hash_sig_path
);
2985 return log_debug_errno(r
, "Failed to load root hash: %m");
2990 static int pick_versions(
2991 const ExecContext
*context
,
2992 const ExecParameters
*params
,
2993 char **ret_root_image
,
2994 char **ret_root_directory
) {
3000 assert(ret_root_image
);
3001 assert(ret_root_directory
);
3003 if (context
->root_image
) {
3004 _cleanup_(pick_result_done
) PickResult result
= PICK_RESULT_NULL
;
3006 r
= path_pick(/* toplevel_path= */ NULL
,
3007 /* toplevel_fd= */ AT_FDCWD
,
3008 context
->root_image
,
3009 &pick_filter_image_raw
,
3010 PICK_ARCHITECTURE
|PICK_TRIES
|PICK_RESOLVE
,
3016 return log_exec_debug_errno(context
, params
, SYNTHETIC_ERRNO(ENOENT
), "No matching entry in .v/ directory %s found.", context
->root_image
);
3018 *ret_root_image
= TAKE_PTR(result
.path
);
3019 *ret_root_directory
= NULL
;
3023 if (context
->root_directory
) {
3024 _cleanup_(pick_result_done
) PickResult result
= PICK_RESULT_NULL
;
3026 r
= path_pick(/* toplevel_path= */ NULL
,
3027 /* toplevel_fd= */ AT_FDCWD
,
3028 context
->root_directory
,
3029 &pick_filter_image_dir
,
3030 PICK_ARCHITECTURE
|PICK_TRIES
|PICK_RESOLVE
,
3036 return log_exec_debug_errno(context
, params
, SYNTHETIC_ERRNO(ENOENT
), "No matching entry in .v/ directory %s found.", context
->root_directory
);
3038 *ret_root_image
= NULL
;
3039 *ret_root_directory
= TAKE_PTR(result
.path
);
3043 *ret_root_image
= *ret_root_directory
= NULL
;
3047 static int apply_mount_namespace(
3048 ExecCommandFlags command_flags
,
3049 const ExecContext
*context
,
3050 const ExecParameters
*params
,
3051 ExecRuntime
*runtime
,
3052 const char *memory_pressure_path
,
3053 bool needs_sandboxing
,
3054 char **error_path
) {
3056 _cleanup_(verity_settings_done
) VeritySettings verity
= VERITY_SETTINGS_DEFAULT
;
3057 _cleanup_strv_free_
char **empty_directories
= NULL
, **symlinks
= NULL
,
3058 **read_write_paths_cleanup
= NULL
;
3059 _cleanup_free_
char *creds_path
= NULL
, *incoming_dir
= NULL
, *propagate_dir
= NULL
,
3060 *extension_dir
= NULL
, *host_os_release_stage
= NULL
, *root_image
= NULL
, *root_dir
= NULL
;
3061 const char *tmp_dir
= NULL
, *var_tmp_dir
= NULL
;
3062 char **read_write_paths
;
3063 bool setup_os_release_symlink
;
3064 BindMount
*bind_mounts
= NULL
;
3065 size_t n_bind_mounts
= 0;
3070 CLEANUP_ARRAY(bind_mounts
, n_bind_mounts
, bind_mount_free_many
);
3072 if (params
->flags
& EXEC_APPLY_CHROOT
) {
3081 r
= setup_ephemeral(
3090 r
= compile_bind_mounts(context
, params
, &bind_mounts
, &n_bind_mounts
, &empty_directories
);
3094 /* We need to make the pressure path writable even if /sys/fs/cgroups is made read-only, as the
3095 * service will need to write to it in order to start the notifications. */
3096 if (context
->protect_control_groups
&& memory_pressure_path
&& !streq(memory_pressure_path
, "/dev/null")) {
3097 read_write_paths_cleanup
= strv_copy(context
->read_write_paths
);
3098 if (!read_write_paths_cleanup
)
3101 r
= strv_extend(&read_write_paths_cleanup
, memory_pressure_path
);
3105 read_write_paths
= read_write_paths_cleanup
;
3107 read_write_paths
= context
->read_write_paths
;
3109 if (needs_sandboxing
) {
3110 /* The runtime struct only contains the parent of the private /tmp, which is non-accessible
3111 * to world users. Inside of it there's a /tmp that is sticky, and that's the one we want to
3112 * use here. This does not apply when we are using /run/systemd/empty as fallback. */
3114 if (context
->private_tmp
&& runtime
&& runtime
->shared
) {
3115 if (streq_ptr(runtime
->shared
->tmp_dir
, RUN_SYSTEMD_EMPTY
))
3116 tmp_dir
= runtime
->shared
->tmp_dir
;
3117 else if (runtime
->shared
->tmp_dir
)
3118 tmp_dir
= strjoina(runtime
->shared
->tmp_dir
, "/tmp");
3120 if (streq_ptr(runtime
->shared
->var_tmp_dir
, RUN_SYSTEMD_EMPTY
))
3121 var_tmp_dir
= runtime
->shared
->var_tmp_dir
;
3122 else if (runtime
->shared
->var_tmp_dir
)
3123 var_tmp_dir
= strjoina(runtime
->shared
->var_tmp_dir
, "/tmp");
3127 /* Symlinks (exec dirs, os-release) are set up after other mounts, before they are made read-only. */
3128 setup_os_release_symlink
= needs_sandboxing
&& exec_context_get_effective_mount_apivfs(context
) && (root_dir
|| root_image
);
3129 r
= compile_symlinks(context
, params
, setup_os_release_symlink
, &symlinks
);
3133 if (context
->mount_propagation_flag
== MS_SHARED
)
3134 log_exec_debug(context
,
3136 "shared mount propagation hidden by other fs namespacing unit settings: ignoring");
3138 r
= exec_context_get_credential_directory(context
, params
, params
->unit_id
, &creds_path
);
3142 if (params
->runtime_scope
== RUNTIME_SCOPE_SYSTEM
) {
3143 propagate_dir
= path_join("/run/systemd/propagate/", params
->unit_id
);
3147 incoming_dir
= strdup("/run/systemd/incoming");
3151 extension_dir
= strdup("/run/systemd/unit-extensions");
3155 /* If running under a different root filesystem, propagate the host's os-release. We make a
3156 * copy rather than just bind mounting it, so that it can be updated on soft-reboot. */
3157 if (setup_os_release_symlink
) {
3158 host_os_release_stage
= strdup("/run/systemd/propagate/.os-release-stage");
3159 if (!host_os_release_stage
)
3163 assert(params
->runtime_scope
== RUNTIME_SCOPE_USER
);
3165 if (asprintf(&extension_dir
, "/run/user/" UID_FMT
"/systemd/unit-extensions", geteuid()) < 0)
3168 if (setup_os_release_symlink
) {
3169 if (asprintf(&host_os_release_stage
,
3170 "/run/user/" UID_FMT
"/systemd/propagate/.os-release-stage",
3177 r
= verity_settings_prepare(
3180 context
->root_hash
, context
->root_hash_size
, context
->root_hash_path
,
3181 context
->root_hash_sig
, context
->root_hash_sig_size
, context
->root_hash_sig_path
,
3182 context
->root_verity
);
3187 NamespaceParameters parameters
= {
3188 .runtime_scope
= params
->runtime_scope
,
3190 .root_directory
= root_dir
,
3191 .root_image
= root_image
,
3192 .root_image_options
= context
->root_image_options
,
3193 .root_image_policy
= context
->root_image_policy
?: &image_policy_service
,
3195 .read_write_paths
= read_write_paths
,
3196 .read_only_paths
= needs_sandboxing
? context
->read_only_paths
: NULL
,
3197 .inaccessible_paths
= needs_sandboxing
? context
->inaccessible_paths
: NULL
,
3199 .exec_paths
= needs_sandboxing
? context
->exec_paths
: NULL
,
3200 .no_exec_paths
= needs_sandboxing
? context
->no_exec_paths
: NULL
,
3202 .empty_directories
= empty_directories
,
3203 .symlinks
= symlinks
,
3205 .bind_mounts
= bind_mounts
,
3206 .n_bind_mounts
= n_bind_mounts
,
3208 .temporary_filesystems
= context
->temporary_filesystems
,
3209 .n_temporary_filesystems
= context
->n_temporary_filesystems
,
3211 .mount_images
= context
->mount_images
,
3212 .n_mount_images
= context
->n_mount_images
,
3213 .mount_image_policy
= context
->mount_image_policy
?: &image_policy_service
,
3216 .var_tmp_dir
= var_tmp_dir
,
3218 .creds_path
= creds_path
,
3219 .log_namespace
= context
->log_namespace
,
3220 .mount_propagation_flag
= context
->mount_propagation_flag
,
3224 .extension_images
= context
->extension_images
,
3225 .n_extension_images
= context
->n_extension_images
,
3226 .extension_image_policy
= context
->extension_image_policy
?: &image_policy_sysext
,
3227 .extension_directories
= context
->extension_directories
,
3229 .propagate_dir
= propagate_dir
,
3230 .incoming_dir
= incoming_dir
,
3231 .extension_dir
= extension_dir
,
3232 .notify_socket
= root_dir
|| root_image
? params
->notify_socket
: NULL
,
3233 .host_os_release_stage
= host_os_release_stage
,
3235 /* If DynamicUser=no and RootDirectory= is set then lets pass a relaxed sandbox info,
3236 * otherwise enforce it, don't ignore protected paths and fail if we are enable to apply the
3237 * sandbox inside the mount namespace. */
3238 .ignore_protect_paths
= !needs_sandboxing
&& !context
->dynamic_user
&& root_dir
,
3240 .protect_control_groups
= needs_sandboxing
&& context
->protect_control_groups
,
3241 .protect_kernel_tunables
= needs_sandboxing
&& context
->protect_kernel_tunables
,
3242 .protect_kernel_modules
= needs_sandboxing
&& context
->protect_kernel_modules
,
3243 .protect_kernel_logs
= needs_sandboxing
&& context
->protect_kernel_logs
,
3244 .protect_hostname
= needs_sandboxing
&& context
->protect_hostname
,
3246 .private_dev
= needs_sandboxing
&& context
->private_devices
,
3247 .private_network
= needs_sandboxing
&& exec_needs_network_namespace(context
),
3248 .private_ipc
= needs_sandboxing
&& exec_needs_ipc_namespace(context
),
3250 .mount_apivfs
= needs_sandboxing
&& exec_context_get_effective_mount_apivfs(context
),
3252 /* If NNP is on, we can turn on MS_NOSUID, since it won't have any effect anymore. */
3253 .mount_nosuid
= needs_sandboxing
&& context
->no_new_privileges
&& !mac_selinux_use(),
3255 .protect_home
= needs_sandboxing
? context
->protect_home
: false,
3256 .protect_system
= needs_sandboxing
? context
->protect_system
: false,
3257 .protect_proc
= needs_sandboxing
? context
->protect_proc
: false,
3258 .proc_subset
= needs_sandboxing
? context
->proc_subset
: false,
3261 r
= setup_namespace(¶meters
, error_path
);
3262 /* If we couldn't set up the namespace this is probably due to a missing capability. setup_namespace() reports
3263 * that with a special, recognizable error ENOANO. In this case, silently proceed, but only if exclusively
3264 * sandboxing options were used, i.e. nothing such as RootDirectory= or BindMount= that would result in a
3265 * completely different execution environment. */
3267 if (insist_on_sandboxing(
3269 root_dir
, root_image
,
3272 return log_exec_debug_errno(context
,
3274 SYNTHETIC_ERRNO(EOPNOTSUPP
),
3275 "Failed to set up namespace, and refusing to continue since "
3276 "the selected namespacing options alter mount environment non-trivially.\n"
3277 "Bind mounts: %zu, temporary filesystems: %zu, root directory: %s, root image: %s, dynamic user: %s",
3279 context
->n_temporary_filesystems
,
3282 yes_no(context
->dynamic_user
));
3284 log_exec_debug(context
, params
, "Failed to set up namespace, assuming containerized execution and ignoring.");
3291 static int apply_working_directory(
3292 const ExecContext
*context
,
3293 const ExecParameters
*params
,
3294 ExecRuntime
*runtime
,
3302 assert(exit_status
);
3304 if (context
->working_directory_home
) {
3306 *exit_status
= EXIT_CHDIR
;
3312 wd
= empty_to_root(context
->working_directory
);
3314 if (params
->flags
& EXEC_APPLY_CHROOT
)
3315 r
= RET_NERRNO(chdir(wd
));
3317 _cleanup_close_
int dfd
= -EBADF
;
3320 (runtime
? runtime
->ephemeral_copy
: NULL
) ?: context
->root_directory
,
3321 CHASE_PREFIX_ROOT
|CHASE_AT_RESOLVE_IN_ROOT
,
3322 /* ret_path= */ NULL
,
3325 r
= RET_NERRNO(fchdir(dfd
));
3328 if (r
< 0 && !context
->working_directory_missing_ok
) {
3329 *exit_status
= EXIT_CHDIR
;
3336 static int apply_root_directory(
3337 const ExecContext
*context
,
3338 const ExecParameters
*params
,
3339 ExecRuntime
*runtime
,
3340 const bool needs_mount_ns
,
3344 assert(exit_status
);
3346 if (params
->flags
& EXEC_APPLY_CHROOT
)
3347 if (!needs_mount_ns
&& context
->root_directory
)
3348 if (chroot((runtime
? runtime
->ephemeral_copy
: NULL
) ?: context
->root_directory
) < 0) {
3349 *exit_status
= EXIT_CHROOT
;
3356 static int setup_keyring(
3357 const ExecContext
*context
,
3358 const ExecParameters
*p
,
3359 uid_t uid
, gid_t gid
) {
3361 key_serial_t keyring
;
3369 /* Let's set up a new per-service "session" kernel keyring for each system service. This has the benefit that
3370 * each service runs with its own keyring shared among all processes of the service, but with no hook-up beyond
3371 * that scope, and in particular no link to the per-UID keyring. If we don't do this the keyring will be
3372 * automatically created on-demand and then linked to the per-UID keyring, by the kernel. The kernel's built-in
3373 * on-demand behaviour is very appropriate for login users, but probably not so much for system services, where
3374 * UIDs are not necessarily specific to a service but reused (at least in the case of UID 0). */
3376 if (context
->keyring_mode
== EXEC_KEYRING_INHERIT
)
3379 /* Acquiring a reference to the user keyring is nasty. We briefly change identity in order to get things set up
3380 * properly by the kernel. If we don't do that then we can't create it atomically, and that sucks for parallel
3381 * execution. This mimics what pam_keyinit does, too. Setting up session keyring, to be owned by the right user
3382 * & group is just as nasty as acquiring a reference to the user keyring. */
3384 saved_uid
= getuid();
3385 saved_gid
= getgid();
3387 if (gid_is_valid(gid
) && gid
!= saved_gid
) {
3388 if (setregid(gid
, -1) < 0)
3389 return log_exec_error_errno(context
,
3392 "Failed to change GID for user keyring: %m");
3395 if (uid_is_valid(uid
) && uid
!= saved_uid
) {
3396 if (setreuid(uid
, -1) < 0) {
3397 r
= log_exec_error_errno(context
,
3400 "Failed to change UID for user keyring: %m");
3405 keyring
= keyctl(KEYCTL_JOIN_SESSION_KEYRING
, 0, 0, 0, 0);
3406 if (keyring
== -1) {
3407 if (errno
== ENOSYS
)
3408 log_exec_debug_errno(context
,
3411 "Kernel keyring not supported, ignoring.");
3412 else if (ERRNO_IS_PRIVILEGE(errno
))
3413 log_exec_debug_errno(context
,
3416 "Kernel keyring access prohibited, ignoring.");
3417 else if (errno
== EDQUOT
)
3418 log_exec_debug_errno(context
,
3421 "Out of kernel keyrings to allocate, ignoring.");
3423 r
= log_exec_error_errno(context
,
3426 "Setting up kernel keyring failed: %m");
3431 /* When requested link the user keyring into the session keyring. */
3432 if (context
->keyring_mode
== EXEC_KEYRING_SHARED
) {
3434 if (keyctl(KEYCTL_LINK
,
3435 KEY_SPEC_USER_KEYRING
,
3436 KEY_SPEC_SESSION_KEYRING
, 0, 0) < 0) {
3437 r
= log_exec_error_errno(context
,
3440 "Failed to link user keyring into session keyring: %m");
3445 /* Restore uid/gid back */
3446 if (uid_is_valid(uid
) && uid
!= saved_uid
) {
3447 if (setreuid(saved_uid
, -1) < 0) {
3448 r
= log_exec_error_errno(context
,
3451 "Failed to change UID back for user keyring: %m");
3456 if (gid_is_valid(gid
) && gid
!= saved_gid
) {
3457 if (setregid(saved_gid
, -1) < 0)
3458 return log_exec_error_errno(context
,
3461 "Failed to change GID back for user keyring: %m");
3464 /* Populate they keyring with the invocation ID by default, as original saved_uid. */
3465 if (!sd_id128_is_null(p
->invocation_id
)) {
3468 key
= add_key("user",
3471 sizeof(p
->invocation_id
),
3472 KEY_SPEC_SESSION_KEYRING
);
3474 log_exec_debug_errno(context
,
3477 "Failed to add invocation ID to keyring, ignoring: %m");
3479 if (keyctl(KEYCTL_SETPERM
, key
,
3480 KEY_POS_VIEW
|KEY_POS_READ
|KEY_POS_SEARCH
|
3481 KEY_USR_VIEW
|KEY_USR_READ
|KEY_USR_SEARCH
, 0, 0) < 0)
3482 r
= log_exec_error_errno(context
,
3485 "Failed to restrict invocation ID permission: %m");
3490 /* Revert back uid & gid for the last time, and exit */
3491 /* no extra logging, as only the first already reported error matters */
3492 if (getuid() != saved_uid
)
3493 (void) setreuid(saved_uid
, -1);
3495 if (getgid() != saved_gid
)
3496 (void) setregid(saved_gid
, -1);
3501 static void append_socket_pair(int *array
, size_t *n
, const int pair
[static 2]) {
3507 array
[(*n
)++] = pair
[0];
3509 array
[(*n
)++] = pair
[1];
3512 static int close_remaining_fds(
3513 const ExecParameters
*params
,
3514 const ExecRuntime
*runtime
,
3516 const int *fds
, size_t n_fds
) {
3518 size_t n_dont_close
= 0;
3519 int dont_close
[n_fds
+ 14];
3523 if (params
->stdin_fd
>= 0)
3524 dont_close
[n_dont_close
++] = params
->stdin_fd
;
3525 if (params
->stdout_fd
>= 0)
3526 dont_close
[n_dont_close
++] = params
->stdout_fd
;
3527 if (params
->stderr_fd
>= 0)
3528 dont_close
[n_dont_close
++] = params
->stderr_fd
;
3531 dont_close
[n_dont_close
++] = socket_fd
;
3533 memcpy(dont_close
+ n_dont_close
, fds
, sizeof(int) * n_fds
);
3534 n_dont_close
+= n_fds
;
3538 append_socket_pair(dont_close
, &n_dont_close
, runtime
->ephemeral_storage_socket
);
3540 if (runtime
&& runtime
->shared
) {
3541 append_socket_pair(dont_close
, &n_dont_close
, runtime
->shared
->netns_storage_socket
);
3542 append_socket_pair(dont_close
, &n_dont_close
, runtime
->shared
->ipcns_storage_socket
);
3545 if (runtime
&& runtime
->dynamic_creds
) {
3546 if (runtime
->dynamic_creds
->user
)
3547 append_socket_pair(dont_close
, &n_dont_close
, runtime
->dynamic_creds
->user
->storage_socket
);
3548 if (runtime
->dynamic_creds
->group
)
3549 append_socket_pair(dont_close
, &n_dont_close
, runtime
->dynamic_creds
->group
->storage_socket
);
3552 if (params
->user_lookup_fd
>= 0)
3553 dont_close
[n_dont_close
++] = params
->user_lookup_fd
;
3555 return close_all_fds(dont_close
, n_dont_close
);
3558 static int send_user_lookup(
3559 const char *unit_id
,
3566 /* Send the resolved UID/GID to PID 1 after we learnt it. We send a single datagram, containing the UID/GID
3567 * data as well as the unit name. Note that we suppress sending this if no user/group to resolve was
3570 if (user_lookup_fd
< 0)
3573 if (!uid_is_valid(uid
) && !gid_is_valid(gid
))
3576 if (writev(user_lookup_fd
,
3578 IOVEC_MAKE(&uid
, sizeof(uid
)),
3579 IOVEC_MAKE(&gid
, sizeof(gid
)),
3580 IOVEC_MAKE_STRING(unit_id
) }, 3) < 0)
3586 static int acquire_home(const ExecContext
*c
, uid_t uid
, const char** home
, char **buf
) {
3593 /* If WorkingDirectory=~ is set, try to acquire a usable home directory. */
3598 if (!c
->working_directory_home
)
3601 r
= get_home_dir(buf
);
3609 static int compile_suggested_paths(const ExecContext
*c
, const ExecParameters
*p
, char ***ret
) {
3610 _cleanup_strv_free_
char ** list
= NULL
;
3617 assert(c
->dynamic_user
);
3619 /* Compile a list of paths that it might make sense to read the owning UID from to use as initial candidate for
3620 * dynamic UID allocation, in order to save us from doing costly recursive chown()s of the special
3623 for (ExecDirectoryType t
= 0; t
< _EXEC_DIRECTORY_TYPE_MAX
; t
++) {
3624 if (t
== EXEC_DIRECTORY_CONFIGURATION
)
3630 for (size_t i
= 0; i
< c
->directories
[t
].n_items
; i
++) {
3633 if (exec_directory_is_private(c
, t
))
3634 e
= path_join(p
->prefix
[t
], "private", c
->directories
[t
].items
[i
].path
);
3636 e
= path_join(p
->prefix
[t
], c
->directories
[t
].items
[i
].path
);
3640 r
= strv_consume(&list
, e
);
3646 *ret
= TAKE_PTR(list
);
3651 static int exec_context_cpu_affinity_from_numa(const ExecContext
*c
, CPUSet
*ret
) {
3652 _cleanup_(cpu_set_reset
) CPUSet s
= {};
3658 if (!c
->numa_policy
.nodes
.set
) {
3659 log_debug("Can't derive CPU affinity mask from NUMA mask because NUMA mask is not set, ignoring");
3663 r
= numa_to_cpu_set(&c
->numa_policy
, &s
);
3669 return cpu_set_add_all(ret
, &s
);
3672 static int add_shifted_fd(int *fds
, size_t fds_size
, size_t *n_fds
, int *fd
) {
3677 assert(*n_fds
< fds_size
);
3683 if (*fd
< 3 + (int) *n_fds
) {
3684 /* Let's move the fd up, so that it's outside of the fd range we will use to store
3685 * the fds we pass to the process (or which are closed only during execve). */
3687 r
= fcntl(*fd
, F_DUPFD_CLOEXEC
, 3 + (int) *n_fds
);
3691 close_and_replace(*fd
, r
);
3694 fds
[(*n_fds
)++] = *fd
;
3698 static int connect_unix_harder(const ExecContext
*c
, const ExecParameters
*p
, const OpenFile
*of
, int ofd
) {
3699 union sockaddr_union addr
= {
3700 .un
.sun_family
= AF_UNIX
,
3703 static const int socket_types
[] = { SOCK_DGRAM
, SOCK_STREAM
, SOCK_SEQPACKET
};
3711 r
= sockaddr_un_set_path(&addr
.un
, FORMAT_PROC_FD_PATH(ofd
));
3713 return log_exec_error_errno(c
, p
, r
, "Failed to set sockaddr for %s: %m", of
->path
);
3717 for (size_t i
= 0; i
< ELEMENTSOF(socket_types
); i
++) {
3718 _cleanup_close_
int fd
= -EBADF
;
3720 fd
= socket(AF_UNIX
, socket_types
[i
] | SOCK_CLOEXEC
, 0);
3722 return log_exec_error_errno(c
,
3725 "Failed to create socket for %s: %m",
3728 r
= RET_NERRNO(connect(fd
, &addr
.sa
, sa_len
));
3729 if (r
== -EPROTOTYPE
)
3732 return log_exec_error_errno(c
,
3735 "Failed to connect socket for %s: %m",
3741 return log_exec_error_errno(c
,
3743 SYNTHETIC_ERRNO(EPROTOTYPE
), "Failed to connect socket for \"%s\".",
3747 static int get_open_file_fd(const ExecContext
*c
, const ExecParameters
*p
, const OpenFile
*of
) {
3749 _cleanup_close_
int fd
= -EBADF
, ofd
= -EBADF
;
3755 ofd
= open(of
->path
, O_PATH
| O_CLOEXEC
);
3757 return log_exec_error_errno(c
, p
, errno
, "Could not open \"%s\": %m", of
->path
);
3759 if (fstat(ofd
, &st
) < 0)
3760 return log_exec_error_errno(c
, p
, errno
, "Failed to stat %s: %m", of
->path
);
3762 if (S_ISSOCK(st
.st_mode
)) {
3763 fd
= connect_unix_harder(c
, p
, of
, ofd
);
3767 if (FLAGS_SET(of
->flags
, OPENFILE_READ_ONLY
) && shutdown(fd
, SHUT_WR
) < 0)
3768 return log_exec_error_errno(c
, p
, errno
, "Failed to shutdown send for socket %s: %m",
3771 log_exec_debug(c
, p
, "socket %s opened (fd=%d)", of
->path
, fd
);
3773 int flags
= FLAGS_SET(of
->flags
, OPENFILE_READ_ONLY
) ? O_RDONLY
: O_RDWR
;
3774 if (FLAGS_SET(of
->flags
, OPENFILE_APPEND
))
3776 else if (FLAGS_SET(of
->flags
, OPENFILE_TRUNCATE
))
3779 fd
= fd_reopen(ofd
, flags
| O_CLOEXEC
);
3781 return log_exec_error_errno(c
, p
, fd
, "Failed to open file %s: %m", of
->path
);
3783 log_exec_debug(c
, p
, "file %s opened (fd=%d)", of
->path
, fd
);
3789 static int collect_open_file_fds(const ExecContext
*c
, ExecParameters
*p
, size_t *n_fds
) {
3796 LIST_FOREACH(open_files
, of
, p
->open_files
) {
3797 _cleanup_close_
int fd
= -EBADF
;
3799 fd
= get_open_file_fd(c
, p
, of
);
3801 if (FLAGS_SET(of
->flags
, OPENFILE_GRACEFUL
)) {
3802 log_exec_debug_errno(c
, p
, fd
, "Failed to get OpenFile= file descriptor for %s, ignoring: %m", of
->path
);
3809 if (!GREEDY_REALLOC(p
->fds
, *n_fds
+ 1))
3812 r
= strv_extend(&p
->fd_names
, of
->fdname
);
3816 p
->fds
[*n_fds
] = TAKE_FD(fd
);
3824 static void log_command_line(
3825 const ExecContext
*context
,
3826 const ExecParameters
*params
,
3828 const char *executable
,
3839 _cleanup_free_
char *cmdline
= quote_command_line(argv
, SHELL_ESCAPE_EMPTY
);
3841 log_exec_struct(context
, params
, LOG_DEBUG
,
3842 "EXECUTABLE=%s", executable
,
3843 LOG_EXEC_MESSAGE(params
, "%s: %s", msg
, strnull(cmdline
)),
3844 LOG_EXEC_INVOCATION_ID(params
));
3847 static bool exec_context_need_unprivileged_private_users(
3848 const ExecContext
*context
,
3849 const ExecParameters
*params
) {
3854 /* These options require PrivateUsers= when used in user units, as we need to be in a user namespace
3855 * to have permission to enable them when not running as root. If we have effective CAP_SYS_ADMIN
3856 * (system manager) then we have privileges and don't need this. */
3857 if (params
->runtime_scope
!= RUNTIME_SCOPE_USER
)
3860 return context
->private_users
||
3861 context
->private_tmp
||
3862 context
->private_devices
||
3863 context
->private_network
||
3864 context
->network_namespace_path
||
3865 context
->private_ipc
||
3866 context
->ipc_namespace_path
||
3867 context
->private_mounts
> 0 ||
3868 context
->mount_apivfs
||
3869 context
->n_bind_mounts
> 0 ||
3870 context
->n_temporary_filesystems
> 0 ||
3871 context
->root_directory
||
3872 !strv_isempty(context
->extension_directories
) ||
3873 context
->protect_system
!= PROTECT_SYSTEM_NO
||
3874 context
->protect_home
!= PROTECT_HOME_NO
||
3875 context
->protect_kernel_tunables
||
3876 context
->protect_kernel_modules
||
3877 context
->protect_kernel_logs
||
3878 context
->protect_control_groups
||
3879 context
->protect_clock
||
3880 context
->protect_hostname
||
3881 !strv_isempty(context
->read_write_paths
) ||
3882 !strv_isempty(context
->read_only_paths
) ||
3883 !strv_isempty(context
->inaccessible_paths
) ||
3884 !strv_isempty(context
->exec_paths
) ||
3885 !strv_isempty(context
->no_exec_paths
);
3888 static bool exec_context_shall_confirm_spawn(const ExecContext
*context
) {
3891 if (confirm_spawn_disabled())
3894 /* For some reasons units remaining in the same process group
3895 * as PID 1 fail to acquire the console even if it's not used
3896 * by any process. So skip the confirmation question for them. */
3897 return !context
->same_pgrp
;
3900 static int exec_context_named_iofds(
3901 const ExecContext
*c
,
3902 const ExecParameters
*p
,
3903 int named_iofds
[static 3]) {
3906 const char* stdio_fdname
[3];
3911 assert(named_iofds
);
3913 targets
= (c
->std_input
== EXEC_INPUT_NAMED_FD
) +
3914 (c
->std_output
== EXEC_OUTPUT_NAMED_FD
) +
3915 (c
->std_error
== EXEC_OUTPUT_NAMED_FD
);
3917 for (size_t i
= 0; i
< 3; i
++)
3918 stdio_fdname
[i
] = exec_context_fdname(c
, i
);
3920 n_fds
= p
->n_storage_fds
+ p
->n_socket_fds
;
3922 for (size_t i
= 0; i
< n_fds
&& targets
> 0; i
++)
3923 if (named_iofds
[STDIN_FILENO
] < 0 &&
3924 c
->std_input
== EXEC_INPUT_NAMED_FD
&&
3925 stdio_fdname
[STDIN_FILENO
] &&
3926 streq(p
->fd_names
[i
], stdio_fdname
[STDIN_FILENO
])) {
3928 named_iofds
[STDIN_FILENO
] = p
->fds
[i
];
3931 } else if (named_iofds
[STDOUT_FILENO
] < 0 &&
3932 c
->std_output
== EXEC_OUTPUT_NAMED_FD
&&
3933 stdio_fdname
[STDOUT_FILENO
] &&
3934 streq(p
->fd_names
[i
], stdio_fdname
[STDOUT_FILENO
])) {
3936 named_iofds
[STDOUT_FILENO
] = p
->fds
[i
];
3939 } else if (named_iofds
[STDERR_FILENO
] < 0 &&
3940 c
->std_error
== EXEC_OUTPUT_NAMED_FD
&&
3941 stdio_fdname
[STDERR_FILENO
] &&
3942 streq(p
->fd_names
[i
], stdio_fdname
[STDERR_FILENO
])) {
3944 named_iofds
[STDERR_FILENO
] = p
->fds
[i
];
3948 return targets
== 0 ? 0 : -ENOENT
;
3951 static void exec_shared_runtime_close(ExecSharedRuntime
*shared
) {
3955 safe_close_pair(shared
->netns_storage_socket
);
3956 safe_close_pair(shared
->ipcns_storage_socket
);
3959 static void exec_runtime_close(ExecRuntime
*rt
) {
3963 safe_close_pair(rt
->ephemeral_storage_socket
);
3965 exec_shared_runtime_close(rt
->shared
);
3966 dynamic_creds_close(rt
->dynamic_creds
);
3969 static void exec_params_close(ExecParameters
*p
) {
3973 p
->stdin_fd
= safe_close(p
->stdin_fd
);
3974 p
->stdout_fd
= safe_close(p
->stdout_fd
);
3975 p
->stderr_fd
= safe_close(p
->stderr_fd
);
3979 const ExecCommand
*command
,
3980 const ExecContext
*context
,
3981 ExecParameters
*params
,
3982 ExecRuntime
*runtime
,
3983 const CGroupContext
*cgroup_context
,
3986 _cleanup_strv_free_
char **our_env
= NULL
, **pass_env
= NULL
, **joined_exec_search_path
= NULL
, **accum_env
= NULL
, **replaced_argv
= NULL
;
3988 _cleanup_free_ gid_t
*supplementary_gids
= NULL
;
3989 const char *username
= NULL
, *groupname
= NULL
;
3990 _cleanup_free_
char *home_buffer
= NULL
, *memory_pressure_path
= NULL
;
3991 const char *home
= NULL
, *shell
= NULL
;
3992 char **final_argv
= NULL
;
3993 dev_t journal_stream_dev
= 0;
3994 ino_t journal_stream_ino
= 0;
3995 bool userns_set_up
= false;
3996 bool needs_sandboxing
, /* Do we need to set up full sandboxing? (i.e. all namespacing, all MAC stuff, caps, yadda yadda */
3997 needs_setuid
, /* Do we need to do the actual setresuid()/setresgid() calls? */
3998 needs_mount_namespace
, /* Do we need to set up a mount namespace for this kernel? */
3999 needs_ambient_hack
; /* Do we need to apply the ambient capabilities hack? */
4000 bool keep_seccomp_privileges
= false;
4002 _cleanup_free_
char *mac_selinux_context_net
= NULL
;
4003 bool use_selinux
= false;
4006 bool use_smack
= false;
4009 bool use_apparmor
= false;
4012 uint64_t saved_bset
= 0;
4014 uid_t saved_uid
= getuid();
4015 gid_t saved_gid
= getgid();
4016 uid_t uid
= UID_INVALID
;
4017 gid_t gid
= GID_INVALID
;
4018 size_t n_fds
, /* fds to pass to the child */
4019 n_keep_fds
; /* total number of fds not to close */
4021 _cleanup_free_ gid_t
*gids_after_pam
= NULL
;
4022 int ngids_after_pam
= 0;
4024 int socket_fd
= -EBADF
, named_iofds
[3] = EBADF_TRIPLET
;
4025 size_t n_storage_fds
, n_socket_fds
;
4030 assert(exit_status
);
4032 /* This should be mostly redundant, as the log level is also passed as an argument of the executor,
4033 * and is already applied earlier. Just for safety. */
4034 if (context
->log_level_max
>= 0)
4035 log_set_max_level(context
->log_level_max
);
4037 /* Explicitly test for CVE-2021-4034 inspired invocations */
4038 if (!command
->path
|| strv_isempty(command
->argv
)) {
4039 *exit_status
= EXIT_EXEC
;
4040 return log_exec_error_errno(
4043 SYNTHETIC_ERRNO(EINVAL
),
4044 "Invalid command line arguments.");
4047 LOG_CONTEXT_PUSH_EXEC(context
, params
);
4049 if (context
->std_input
== EXEC_INPUT_SOCKET
||
4050 context
->std_output
== EXEC_OUTPUT_SOCKET
||
4051 context
->std_error
== EXEC_OUTPUT_SOCKET
) {
4053 if (params
->n_socket_fds
> 1)
4054 return log_exec_error_errno(context
, params
, SYNTHETIC_ERRNO(EINVAL
), "Got more than one socket.");
4056 if (params
->n_socket_fds
== 0)
4057 return log_exec_error_errno(context
, params
, SYNTHETIC_ERRNO(EINVAL
), "Got no socket.");
4059 socket_fd
= params
->fds
[0];
4060 n_storage_fds
= n_socket_fds
= 0;
4062 n_socket_fds
= params
->n_socket_fds
;
4063 n_storage_fds
= params
->n_storage_fds
;
4065 n_fds
= n_socket_fds
+ n_storage_fds
;
4067 r
= exec_context_named_iofds(context
, params
, named_iofds
);
4069 return log_exec_error_errno(context
, params
, r
, "Failed to load a named file descriptor: %m");
4071 rename_process_from_path(command
->path
);
4073 /* We reset exactly these signals, since they are the only ones we set to SIG_IGN in the main
4074 * daemon. All others we leave untouched because we set them to SIG_DFL or a valid handler initially,
4075 * both of which will be demoted to SIG_DFL. */
4076 (void) default_signals(SIGNALS_CRASH_HANDLER
,
4079 if (context
->ignore_sigpipe
)
4080 (void) ignore_signals(SIGPIPE
);
4082 r
= reset_signal_mask();
4084 *exit_status
= EXIT_SIGNAL_MASK
;
4085 return log_exec_error_errno(context
, params
, r
, "Failed to set process signal mask: %m");
4088 if (params
->idle_pipe
)
4089 do_idle_pipe_dance(params
->idle_pipe
);
4091 /* Close fds we don't need very early to make sure we don't block init reexecution because it cannot bind its
4092 * sockets. Among the fds we close are the logging fds, and we want to keep them closed, so that we don't have
4093 * any fds open we don't really want open during the transition. In order to make logging work, we switch the
4094 * log subsystem into open_when_needed mode, so that it reopens the logs on every single log call. */
4097 log_set_open_when_needed(true);
4098 log_settle_target();
4100 /* In case anything used libc syslog(), close this here, too */
4103 r
= collect_open_file_fds(context
, params
, &n_fds
);
4105 *exit_status
= EXIT_FDS
;
4106 return log_exec_error_errno(context
, params
, r
, "Failed to get OpenFile= file descriptors: %m");
4109 int keep_fds
[n_fds
+ 3];
4110 memcpy_safe(keep_fds
, params
->fds
, n_fds
* sizeof(int));
4113 r
= add_shifted_fd(keep_fds
, ELEMENTSOF(keep_fds
), &n_keep_fds
, ¶ms
->exec_fd
);
4115 *exit_status
= EXIT_FDS
;
4116 return log_exec_error_errno(context
, params
, r
, "Failed to collect shifted fd: %m");
4120 r
= add_shifted_fd(keep_fds
, ELEMENTSOF(keep_fds
), &n_keep_fds
, ¶ms
->bpf_restrict_fs_map_fd
);
4122 *exit_status
= EXIT_FDS
;
4123 return log_exec_error_errno(context
, params
, r
, "Failed to collect shifted fd: %m");
4127 r
= close_remaining_fds(params
, runtime
, socket_fd
, keep_fds
, n_keep_fds
);
4129 *exit_status
= EXIT_FDS
;
4130 return log_exec_error_errno(context
, params
, r
, "Failed to close unwanted file descriptors: %m");
4133 if (!context
->same_pgrp
&&
4135 *exit_status
= EXIT_SETSID
;
4136 return log_exec_error_errno(context
, params
, errno
, "Failed to create new process session: %m");
4139 exec_context_tty_reset(context
, params
);
4141 if (params
->shall_confirm_spawn
&& exec_context_shall_confirm_spawn(context
)) {
4142 _cleanup_free_
char *cmdline
= NULL
;
4144 cmdline
= quote_command_line(command
->argv
, SHELL_ESCAPE_EMPTY
);
4146 *exit_status
= EXIT_MEMORY
;
4150 r
= ask_for_confirmation(context
, params
, cmdline
);
4151 if (r
!= CONFIRM_EXECUTE
) {
4152 if (r
== CONFIRM_PRETEND_SUCCESS
) {
4153 *exit_status
= EXIT_SUCCESS
;
4157 *exit_status
= EXIT_CONFIRM
;
4158 return log_exec_error_errno(context
, params
, SYNTHETIC_ERRNO(ECANCELED
),
4159 "Execution cancelled by the user");
4163 /* We are about to invoke NSS and PAM modules. Let's tell them what we are doing here, maybe they care. This is
4164 * used by nss-resolve to disable itself when we are about to start systemd-resolved, to avoid deadlocks. Note
4165 * that these env vars do not survive the execve(), which means they really only apply to the PAM and NSS
4166 * invocations themselves. Also note that while we'll only invoke NSS modules involved in user management they
4167 * might internally call into other NSS modules that are involved in hostname resolution, we never know. */
4168 if (setenv("SYSTEMD_ACTIVATION_UNIT", params
->unit_id
, true) != 0 ||
4169 setenv("SYSTEMD_ACTIVATION_SCOPE", runtime_scope_to_string(params
->runtime_scope
), true) != 0) {
4170 *exit_status
= EXIT_MEMORY
;
4171 return log_exec_error_errno(context
, params
, errno
, "Failed to update environment: %m");
4174 if (context
->dynamic_user
&& runtime
&& runtime
->dynamic_creds
) {
4175 _cleanup_strv_free_
char **suggested_paths
= NULL
;
4177 /* On top of that, make sure we bypass our own NSS module nss-systemd comprehensively for any NSS
4178 * checks, if DynamicUser=1 is used, as we shouldn't create a feedback loop with ourselves here. */
4179 if (putenv((char*) "SYSTEMD_NSS_DYNAMIC_BYPASS=1") != 0) {
4180 *exit_status
= EXIT_USER
;
4181 return log_exec_error_errno(context
, params
, errno
, "Failed to update environment: %m");
4184 r
= compile_suggested_paths(context
, params
, &suggested_paths
);
4186 *exit_status
= EXIT_MEMORY
;
4190 r
= dynamic_creds_realize(runtime
->dynamic_creds
, suggested_paths
, &uid
, &gid
);
4192 *exit_status
= EXIT_USER
;
4194 return log_exec_error_errno(context
, params
, SYNTHETIC_ERRNO(EOPNOTSUPP
),
4195 "Failed to update dynamic user credentials: User or group with specified name already exists.");
4196 return log_exec_error_errno(context
, params
, r
, "Failed to update dynamic user credentials: %m");
4199 if (!uid_is_valid(uid
)) {
4200 *exit_status
= EXIT_USER
;
4201 return log_exec_error_errno(context
, params
, SYNTHETIC_ERRNO(ESRCH
), "UID validation failed for \""UID_FMT
"\"", uid
);
4204 if (!gid_is_valid(gid
)) {
4205 *exit_status
= EXIT_USER
;
4206 return log_exec_error_errno(context
, params
, SYNTHETIC_ERRNO(ESRCH
), "GID validation failed for \""GID_FMT
"\"", gid
);
4209 if (runtime
->dynamic_creds
->user
)
4210 username
= runtime
->dynamic_creds
->user
->name
;
4213 if (context
->user
) {
4214 r
= get_fixed_user(context
->user
, &username
, &uid
, &gid
, &home
, &shell
);
4216 *exit_status
= EXIT_USER
;
4217 return log_exec_error_errno(context
, params
, r
, "Failed to determine user credentials: %m");
4221 if (context
->group
) {
4222 r
= get_fixed_group(context
->group
, &groupname
, &gid
);
4224 *exit_status
= EXIT_GROUP
;
4225 return log_exec_error_errno(context
, params
, r
, "Failed to determine group credentials: %m");
4230 /* Initialize user supplementary groups and get SupplementaryGroups= ones */
4231 r
= get_supplementary_groups(context
, username
, groupname
, gid
,
4232 &supplementary_gids
, &ngids
);
4234 *exit_status
= EXIT_GROUP
;
4235 return log_exec_error_errno(context
, params
, r
, "Failed to determine supplementary groups: %m");
4238 r
= send_user_lookup(params
->unit_id
, params
->user_lookup_fd
, uid
, gid
);
4240 *exit_status
= EXIT_USER
;
4241 return log_exec_error_errno(context
, params
, r
, "Failed to send user credentials to PID1: %m");
4244 params
->user_lookup_fd
= safe_close(params
->user_lookup_fd
);
4246 r
= acquire_home(context
, uid
, &home
, &home_buffer
);
4248 *exit_status
= EXIT_CHDIR
;
4249 return log_exec_error_errno(context
, params
, r
, "Failed to determine $HOME for user: %m");
4252 /* If a socket is connected to STDIN/STDOUT/STDERR, we must drop O_NONBLOCK */
4254 (void) fd_nonblock(socket_fd
, false);
4256 /* Journald will try to look-up our cgroup in order to populate _SYSTEMD_CGROUP and _SYSTEMD_UNIT fields.
4257 * Hence we need to migrate to the target cgroup from init.scope before connecting to journald */
4258 if (params
->cgroup_path
) {
4259 _cleanup_free_
char *p
= NULL
;
4261 r
= exec_params_get_cgroup_path(params
, cgroup_context
, &p
);
4263 *exit_status
= EXIT_CGROUP
;
4264 return log_exec_error_errno(context
, params
, r
, "Failed to acquire cgroup path: %m");
4267 r
= cg_attach_everywhere(params
->cgroup_supported
, p
, 0, NULL
, NULL
);
4268 if (r
== -EUCLEAN
) {
4269 *exit_status
= EXIT_CGROUP
;
4270 return log_exec_error_errno(context
, params
, r
, "Failed to attach process to cgroup %s "
4271 "because the cgroup or one of its parents or "
4272 "siblings is in the threaded mode: %m", p
);
4275 *exit_status
= EXIT_CGROUP
;
4276 return log_exec_error_errno(context
, params
, r
, "Failed to attach to cgroup %s: %m", p
);
4280 if (context
->network_namespace_path
&& runtime
&& runtime
->shared
&& runtime
->shared
->netns_storage_socket
[0] >= 0) {
4281 r
= open_shareable_ns_path(runtime
->shared
->netns_storage_socket
, context
->network_namespace_path
, CLONE_NEWNET
);
4283 *exit_status
= EXIT_NETWORK
;
4284 return log_exec_error_errno(context
, params
, r
, "Failed to open network namespace path %s: %m", context
->network_namespace_path
);
4288 if (context
->ipc_namespace_path
&& runtime
&& runtime
->shared
&& runtime
->shared
->ipcns_storage_socket
[0] >= 0) {
4289 r
= open_shareable_ns_path(runtime
->shared
->ipcns_storage_socket
, context
->ipc_namespace_path
, CLONE_NEWIPC
);
4291 *exit_status
= EXIT_NAMESPACE
;
4292 return log_exec_error_errno(context
, params
, r
, "Failed to open IPC namespace path %s: %m", context
->ipc_namespace_path
);
4296 r
= setup_input(context
, params
, socket_fd
, named_iofds
);
4298 *exit_status
= EXIT_STDIN
;
4299 return log_exec_error_errno(context
, params
, r
, "Failed to set up standard input: %m");
4302 r
= setup_output(context
, params
, STDOUT_FILENO
, socket_fd
, named_iofds
, basename(command
->path
), uid
, gid
, &journal_stream_dev
, &journal_stream_ino
);
4304 *exit_status
= EXIT_STDOUT
;
4305 return log_exec_error_errno(context
, params
, r
, "Failed to set up standard output: %m");
4308 r
= setup_output(context
, params
, STDERR_FILENO
, socket_fd
, named_iofds
, basename(command
->path
), uid
, gid
, &journal_stream_dev
, &journal_stream_ino
);
4310 *exit_status
= EXIT_STDERR
;
4311 return log_exec_error_errno(context
, params
, r
, "Failed to set up standard error output: %m");
4314 if (context
->oom_score_adjust_set
) {
4315 /* When we can't make this change due to EPERM, then let's silently skip over it. User
4316 * namespaces prohibit write access to this file, and we shouldn't trip up over that. */
4317 r
= set_oom_score_adjust(context
->oom_score_adjust
);
4318 if (ERRNO_IS_NEG_PRIVILEGE(r
))
4319 log_exec_debug_errno(context
, params
, r
,
4320 "Failed to adjust OOM setting, assuming containerized execution, ignoring: %m");
4322 *exit_status
= EXIT_OOM_ADJUST
;
4323 return log_exec_error_errno(context
, params
, r
, "Failed to adjust OOM setting: %m");
4327 if (context
->coredump_filter_set
) {
4328 r
= set_coredump_filter(context
->coredump_filter
);
4329 if (ERRNO_IS_NEG_PRIVILEGE(r
))
4330 log_exec_debug_errno(context
, params
, r
, "Failed to adjust coredump_filter, ignoring: %m");
4332 *exit_status
= EXIT_LIMITS
;
4333 return log_exec_error_errno(context
, params
, r
, "Failed to adjust coredump_filter: %m");
4337 if (context
->nice_set
) {
4338 r
= setpriority_closest(context
->nice
);
4340 *exit_status
= EXIT_NICE
;
4341 return log_exec_error_errno(context
, params
, r
, "Failed to set up process scheduling priority (nice level): %m");
4345 if (context
->cpu_sched_set
) {
4346 struct sched_param param
= {
4347 .sched_priority
= context
->cpu_sched_priority
,
4350 r
= sched_setscheduler(0,
4351 context
->cpu_sched_policy
|
4352 (context
->cpu_sched_reset_on_fork
?
4353 SCHED_RESET_ON_FORK
: 0),
4356 *exit_status
= EXIT_SETSCHEDULER
;
4357 return log_exec_error_errno(context
, params
, errno
, "Failed to set up CPU scheduling: %m");
4361 if (context
->cpu_affinity_from_numa
|| context
->cpu_set
.set
) {
4362 _cleanup_(cpu_set_reset
) CPUSet converted_cpu_set
= {};
4363 const CPUSet
*cpu_set
;
4365 if (context
->cpu_affinity_from_numa
) {
4366 r
= exec_context_cpu_affinity_from_numa(context
, &converted_cpu_set
);
4368 *exit_status
= EXIT_CPUAFFINITY
;
4369 return log_exec_error_errno(context
, params
, r
, "Failed to derive CPU affinity mask from NUMA mask: %m");
4372 cpu_set
= &converted_cpu_set
;
4374 cpu_set
= &context
->cpu_set
;
4376 if (sched_setaffinity(0, cpu_set
->allocated
, cpu_set
->set
) < 0) {
4377 *exit_status
= EXIT_CPUAFFINITY
;
4378 return log_exec_error_errno(context
, params
, errno
, "Failed to set up CPU affinity: %m");
4382 if (mpol_is_valid(numa_policy_get_type(&context
->numa_policy
))) {
4383 r
= apply_numa_policy(&context
->numa_policy
);
4384 if (ERRNO_IS_NEG_NOT_SUPPORTED(r
))
4385 log_exec_debug_errno(context
, params
, r
, "NUMA support not available, ignoring.");
4387 *exit_status
= EXIT_NUMA_POLICY
;
4388 return log_exec_error_errno(context
, params
, r
, "Failed to set NUMA memory policy: %m");
4392 if (context
->ioprio_set
)
4393 if (ioprio_set(IOPRIO_WHO_PROCESS
, 0, context
->ioprio
) < 0) {
4394 *exit_status
= EXIT_IOPRIO
;
4395 return log_exec_error_errno(context
, params
, errno
, "Failed to set up IO scheduling priority: %m");
4398 if (context
->timer_slack_nsec
!= NSEC_INFINITY
)
4399 if (prctl(PR_SET_TIMERSLACK
, context
->timer_slack_nsec
) < 0) {
4400 *exit_status
= EXIT_TIMERSLACK
;
4401 return log_exec_error_errno(context
, params
, errno
, "Failed to set up timer slack: %m");
4404 if (context
->personality
!= PERSONALITY_INVALID
) {
4405 r
= safe_personality(context
->personality
);
4407 *exit_status
= EXIT_PERSONALITY
;
4408 return log_exec_error_errno(context
, params
, r
, "Failed to set up execution domain (personality): %m");
4413 if (context
->utmp_id
) {
4414 _cleanup_free_
char *username_alloc
= NULL
;
4416 if (!username
&& context
->utmp_mode
== EXEC_UTMP_USER
) {
4417 username_alloc
= uid_to_name(uid_is_valid(uid
) ? uid
: saved_uid
);
4418 if (!username_alloc
) {
4419 *exit_status
= EXIT_USER
;
4424 const char *line
= context
->tty_path
?
4425 (path_startswith(context
->tty_path
, "/dev/") ?: context
->tty_path
) :
4427 utmp_put_init_process(context
->utmp_id
, getpid_cached(), getsid(0),
4429 context
->utmp_mode
== EXEC_UTMP_INIT
? INIT_PROCESS
:
4430 context
->utmp_mode
== EXEC_UTMP_LOGIN
? LOGIN_PROCESS
:
4432 username
?: username_alloc
);
4436 if (uid_is_valid(uid
)) {
4437 r
= chown_terminal(STDIN_FILENO
, uid
);
4439 *exit_status
= EXIT_STDIN
;
4440 return log_exec_error_errno(context
, params
, r
, "Failed to change ownership of terminal: %m");
4444 if (params
->cgroup_path
) {
4445 /* If delegation is enabled we'll pass ownership of the cgroup to the user of the new process. On cgroup v1
4446 * this is only about systemd's own hierarchy, i.e. not the controller hierarchies, simply because that's not
4447 * safe. On cgroup v2 there's only one hierarchy anyway, and delegation is safe there, hence in that case only
4448 * touch a single hierarchy too. */
4450 if (params
->flags
& EXEC_CGROUP_DELEGATE
) {
4451 _cleanup_free_
char *p
= NULL
;
4453 r
= cg_set_access(SYSTEMD_CGROUP_CONTROLLER
, params
->cgroup_path
, uid
, gid
);
4455 *exit_status
= EXIT_CGROUP
;
4456 return log_exec_error_errno(context
, params
, r
, "Failed to adjust control group access: %m");
4459 r
= exec_params_get_cgroup_path(params
, cgroup_context
, &p
);
4461 *exit_status
= EXIT_CGROUP
;
4462 return log_exec_error_errno(context
, params
, r
, "Failed to acquire cgroup path: %m");
4465 r
= cg_set_access_recursive(SYSTEMD_CGROUP_CONTROLLER
, p
, uid
, gid
);
4467 *exit_status
= EXIT_CGROUP
;
4468 return log_exec_error_errno(context
, params
, r
, "Failed to adjust control subgroup access: %m");
4473 if (cgroup_context
&& cg_unified() > 0 && is_pressure_supported() > 0) {
4474 if (cgroup_context_want_memory_pressure(cgroup_context
)) {
4475 r
= cg_get_path("memory", params
->cgroup_path
, "memory.pressure", &memory_pressure_path
);
4477 *exit_status
= EXIT_MEMORY
;
4481 r
= chmod_and_chown(memory_pressure_path
, 0644, uid
, gid
);
4483 log_exec_full_errno(context
, params
, r
== -ENOENT
|| ERRNO_IS_PRIVILEGE(r
) ? LOG_DEBUG
: LOG_WARNING
, r
,
4484 "Failed to adjust ownership of '%s', ignoring: %m", memory_pressure_path
);
4485 memory_pressure_path
= mfree(memory_pressure_path
);
4487 } else if (cgroup_context
->memory_pressure_watch
== CGROUP_PRESSURE_WATCH_OFF
) {
4488 memory_pressure_path
= strdup("/dev/null"); /* /dev/null is explicit indicator for turning of memory pressure watch */
4489 if (!memory_pressure_path
) {
4490 *exit_status
= EXIT_MEMORY
;
4497 needs_mount_namespace
= exec_needs_mount_namespace(context
, params
, runtime
);
4499 for (ExecDirectoryType dt
= 0; dt
< _EXEC_DIRECTORY_TYPE_MAX
; dt
++) {
4500 r
= setup_exec_directory(context
, params
, uid
, gid
, dt
, needs_mount_namespace
, exit_status
);
4502 return log_exec_error_errno(context
, params
, r
, "Failed to set up special execution directory in %s: %m", params
->prefix
[dt
]);
4505 r
= exec_setup_credentials(context
, params
, params
->unit_id
, uid
, gid
);
4507 *exit_status
= EXIT_CREDENTIALS
;
4508 return log_exec_error_errno(context
, params
, r
, "Failed to set up credentials: %m");
4511 r
= build_environment(
4521 memory_pressure_path
,
4524 *exit_status
= EXIT_MEMORY
;
4528 r
= build_pass_environment(context
, &pass_env
);
4530 *exit_status
= EXIT_MEMORY
;
4534 /* The $PATH variable is set to the default path in params->environment. However, this is overridden
4535 * if user-specified fields have $PATH set. The intention is to also override $PATH if the unit does
4536 * not specify PATH but the unit has ExecSearchPath. */
4537 if (!strv_isempty(context
->exec_search_path
)) {
4538 _cleanup_free_
char *joined
= NULL
;
4540 joined
= strv_join(context
->exec_search_path
, ":");
4542 *exit_status
= EXIT_MEMORY
;
4546 r
= strv_env_assign(&joined_exec_search_path
, "PATH", joined
);
4548 *exit_status
= EXIT_MEMORY
;
4553 accum_env
= strv_env_merge(params
->environment
,
4555 joined_exec_search_path
,
4557 context
->environment
,
4560 *exit_status
= EXIT_MEMORY
;
4563 accum_env
= strv_env_clean(accum_env
);
4565 (void) umask(context
->umask
);
4567 r
= setup_keyring(context
, params
, uid
, gid
);
4569 *exit_status
= EXIT_KEYRING
;
4570 return log_exec_error_errno(context
, params
, r
, "Failed to set up kernel keyring: %m");
4573 /* We need sandboxing if the caller asked us to apply it and the command isn't explicitly excepted
4575 needs_sandboxing
= (params
->flags
& EXEC_APPLY_SANDBOXING
) && !(command
->flags
& EXEC_COMMAND_FULLY_PRIVILEGED
);
4577 /* We need the ambient capability hack, if the caller asked us to apply it and the command is marked
4578 * for it, and the kernel doesn't actually support ambient caps. */
4579 needs_ambient_hack
= (params
->flags
& EXEC_APPLY_SANDBOXING
) && (command
->flags
& EXEC_COMMAND_AMBIENT_MAGIC
) && !ambient_capabilities_supported();
4581 /* We need setresuid() if the caller asked us to apply sandboxing and the command isn't explicitly
4582 * excepted from either whole sandboxing or just setresuid() itself, and the ambient hack is not
4584 if (needs_ambient_hack
)
4585 needs_setuid
= false;
4587 needs_setuid
= (params
->flags
& EXEC_APPLY_SANDBOXING
) && !(command
->flags
& (EXEC_COMMAND_FULLY_PRIVILEGED
|EXEC_COMMAND_NO_SETUID
));
4589 uint64_t capability_ambient_set
= context
->capability_ambient_set
;
4591 if (needs_sandboxing
) {
4592 /* MAC enablement checks need to be done before a new mount ns is created, as they rely on
4593 * /sys being present. The actual MAC context application will happen later, as late as
4594 * possible, to avoid impacting our own code paths. */
4597 use_selinux
= mac_selinux_use();
4600 use_smack
= mac_smack_use();
4603 use_apparmor
= mac_apparmor_use();
4607 if (needs_sandboxing
) {
4610 /* Let's set the resource limits before we call into PAM, so that pam_limits wins over what
4611 * is set here. (See below.) */
4613 r
= setrlimit_closest_all((const struct rlimit
* const *) context
->rlimit
, &which_failed
);
4615 *exit_status
= EXIT_LIMITS
;
4616 return log_exec_error_errno(context
, params
, r
, "Failed to adjust resource limit RLIMIT_%s: %m", rlimit_to_string(which_failed
));
4620 if (needs_setuid
&& context
->pam_name
&& username
) {
4621 /* Let's call into PAM after we set up our own idea of resource limits so that pam_limits
4622 * wins here. (See above.) */
4624 /* All fds passed in the fds array will be closed in the pam child process. */
4625 r
= setup_pam(context
->pam_name
, username
, uid
, gid
, context
->tty_path
, &accum_env
, params
->fds
, n_fds
, params
->exec_fd
);
4627 *exit_status
= EXIT_PAM
;
4628 return log_exec_error_errno(context
, params
, r
, "Failed to set up PAM session: %m");
4631 if (ambient_capabilities_supported()) {
4632 uint64_t ambient_after_pam
;
4634 /* PAM modules might have set some ambient caps. Query them here and merge them into
4635 * the caps we want to set in the end, so that we don't end up unsetting them. */
4636 r
= capability_get_ambient(&ambient_after_pam
);
4638 *exit_status
= EXIT_CAPABILITIES
;
4639 return log_exec_error_errno(context
, params
, r
, "Failed to query ambient caps: %m");
4642 capability_ambient_set
|= ambient_after_pam
;
4645 ngids_after_pam
= getgroups_alloc(&gids_after_pam
);
4646 if (ngids_after_pam
< 0) {
4647 *exit_status
= EXIT_GROUP
;
4648 return log_exec_error_errno(context
, params
, ngids_after_pam
, "Failed to obtain groups after setting up PAM: %m");
4652 if (needs_sandboxing
&& exec_context_need_unprivileged_private_users(context
, params
)) {
4653 /* If we're unprivileged, set up the user namespace first to enable use of the other namespaces.
4654 * Users with CAP_SYS_ADMIN can set up user namespaces last because they will be able to
4655 * set up all of the other namespaces (i.e. network, mount, UTS) without a user namespace. */
4657 r
= setup_private_users(saved_uid
, saved_gid
, uid
, gid
);
4658 /* If it was requested explicitly and we can't set it up, fail early. Otherwise, continue and let
4659 * the actual requested operations fail (or silently continue). */
4660 if (r
< 0 && context
->private_users
) {
4661 *exit_status
= EXIT_USER
;
4662 return log_exec_error_errno(context
, params
, r
, "Failed to set up user namespacing for unprivileged user: %m");
4665 log_exec_info_errno(context
, params
, r
, "Failed to set up user namespacing for unprivileged user, ignoring: %m");
4667 userns_set_up
= true;
4670 if (exec_needs_network_namespace(context
) && runtime
&& runtime
->shared
&& runtime
->shared
->netns_storage_socket
[0] >= 0) {
4672 /* Try to enable network namespacing if network namespacing is available and we have
4673 * CAP_NET_ADMIN. We need CAP_NET_ADMIN to be able to configure the loopback device in the
4674 * new network namespace. And if we don't have that, then we could only create a network
4675 * namespace without the ability to set up "lo". Hence gracefully skip things then. */
4676 if (ns_type_supported(NAMESPACE_NET
) && have_effective_cap(CAP_NET_ADMIN
) > 0) {
4677 r
= setup_shareable_ns(runtime
->shared
->netns_storage_socket
, CLONE_NEWNET
);
4678 if (ERRNO_IS_NEG_PRIVILEGE(r
))
4679 log_exec_notice_errno(context
, params
, r
,
4680 "PrivateNetwork=yes is configured, but network namespace setup not permitted, proceeding without: %m");
4682 *exit_status
= EXIT_NETWORK
;
4683 return log_exec_error_errno(context
, params
, r
, "Failed to set up network namespacing: %m");
4685 } else if (context
->network_namespace_path
) {
4686 *exit_status
= EXIT_NETWORK
;
4687 return log_exec_error_errno(context
, params
, SYNTHETIC_ERRNO(EOPNOTSUPP
),
4688 "NetworkNamespacePath= is not supported, refusing.");
4690 log_exec_notice(context
, params
, "PrivateNetwork=yes is configured, but the kernel does not support or we lack privileges for network namespace, proceeding without.");
4693 if (exec_needs_ipc_namespace(context
) && runtime
&& runtime
->shared
&& runtime
->shared
->ipcns_storage_socket
[0] >= 0) {
4695 if (ns_type_supported(NAMESPACE_IPC
)) {
4696 r
= setup_shareable_ns(runtime
->shared
->ipcns_storage_socket
, CLONE_NEWIPC
);
4698 log_exec_warning_errno(context
, params
, r
,
4699 "PrivateIPC=yes is configured, but IPC namespace setup failed, ignoring: %m");
4701 *exit_status
= EXIT_NAMESPACE
;
4702 return log_exec_error_errno(context
, params
, r
, "Failed to set up IPC namespacing: %m");
4704 } else if (context
->ipc_namespace_path
) {
4705 *exit_status
= EXIT_NAMESPACE
;
4706 return log_exec_error_errno(context
, params
, SYNTHETIC_ERRNO(EOPNOTSUPP
),
4707 "IPCNamespacePath= is not supported, refusing.");
4709 log_exec_warning(context
, params
, "PrivateIPC=yes is configured, but the kernel does not support IPC namespaces, ignoring.");
4712 if (needs_mount_namespace
) {
4713 _cleanup_free_
char *error_path
= NULL
;
4715 r
= apply_mount_namespace(command
->flags
,
4719 memory_pressure_path
,
4723 *exit_status
= EXIT_NAMESPACE
;
4724 return log_exec_error_errno(context
, params
, r
, "Failed to set up mount namespacing%s%s: %m",
4725 error_path
? ": " : "", strempty(error_path
));
4729 if (needs_sandboxing
) {
4730 r
= apply_protect_hostname(context
, params
, exit_status
);
4735 if (context
->memory_ksm
>= 0)
4736 if (prctl(PR_SET_MEMORY_MERGE
, context
->memory_ksm
) < 0) {
4737 if (ERRNO_IS_NOT_SUPPORTED(errno
))
4738 log_exec_debug_errno(context
,
4741 "KSM support not available, ignoring.");
4743 *exit_status
= EXIT_KSM
;
4744 return log_exec_error_errno(context
, params
, errno
, "Failed to set KSM: %m");
4748 /* Drop groups as early as possible.
4749 * This needs to be done after PrivateDevices=yes setup as device nodes should be owned by the host's root.
4750 * For non-root in a userns, devices will be owned by the user/group before the group change, and nobody. */
4752 _cleanup_free_ gid_t
*gids_to_enforce
= NULL
;
4753 int ngids_to_enforce
= 0;
4755 ngids_to_enforce
= merge_gid_lists(supplementary_gids
,
4760 if (ngids_to_enforce
< 0) {
4761 *exit_status
= EXIT_GROUP
;
4762 return log_exec_error_errno(context
, params
,
4764 "Failed to merge group lists. Group membership might be incorrect: %m");
4767 r
= enforce_groups(gid
, gids_to_enforce
, ngids_to_enforce
);
4769 *exit_status
= EXIT_GROUP
;
4770 return log_exec_error_errno(context
, params
, r
, "Changing group credentials failed: %m");
4774 /* If the user namespace was not set up above, try to do it now.
4775 * It's preferred to set up the user namespace later (after all other namespaces) so as not to be
4776 * restricted by rules pertaining to combining user namespaces with other namespaces (e.g. in the
4777 * case of mount namespaces being less privileged when the mount point list is copied from a
4778 * different user namespace). */
4780 if (needs_sandboxing
&& context
->private_users
&& !userns_set_up
) {
4781 r
= setup_private_users(saved_uid
, saved_gid
, uid
, gid
);
4783 *exit_status
= EXIT_USER
;
4784 return log_exec_error_errno(context
, params
, r
, "Failed to set up user namespacing: %m");
4788 /* Now that the mount namespace has been set up and privileges adjusted, let's look for the thing we
4791 _cleanup_free_
char *executable
= NULL
;
4792 _cleanup_close_
int executable_fd
= -EBADF
;
4793 r
= find_executable_full(command
->path
, /* root= */ NULL
, context
->exec_search_path
, false, &executable
, &executable_fd
);
4795 *exit_status
= EXIT_EXEC
;
4796 log_exec_struct_errno(context
, params
, LOG_NOTICE
, r
,
4797 "MESSAGE_ID=" SD_MESSAGE_SPAWN_FAILED_STR
,
4798 LOG_EXEC_MESSAGE(params
,
4799 "Unable to locate executable '%s': %m",
4801 "EXECUTABLE=%s", command
->path
);
4802 /* If the error will be ignored by manager, tune down the log level here. Missing executable
4803 * is very much expected in this case. */
4804 return r
!= -ENOMEM
&& FLAGS_SET(command
->flags
, EXEC_COMMAND_IGNORE_FAILURE
) ? 1 : r
;
4807 r
= add_shifted_fd(keep_fds
, ELEMENTSOF(keep_fds
), &n_keep_fds
, &executable_fd
);
4809 *exit_status
= EXIT_FDS
;
4810 return log_exec_error_errno(context
, params
, r
, "Failed to collect shifted fd: %m");
4814 if (needs_sandboxing
&& use_selinux
&& params
->selinux_context_net
) {
4819 else if (params
->n_socket_fds
== 1)
4820 /* If stdin is not connected to a socket but we are triggered by exactly one socket unit then we
4821 * use context from that fd to compute the label. */
4822 fd
= params
->fds
[0];
4825 r
= mac_selinux_get_child_mls_label(fd
, executable
, context
->selinux_context
, &mac_selinux_context_net
);
4827 if (!context
->selinux_context_ignore
) {
4828 *exit_status
= EXIT_SELINUX_CONTEXT
;
4829 return log_exec_error_errno(context
,
4832 "Failed to determine SELinux context: %m");
4834 log_exec_debug_errno(context
,
4837 "Failed to determine SELinux context, ignoring: %m");
4843 /* We repeat the fd closing here, to make sure that nothing is leaked from the PAM modules. Note that
4844 * we are more aggressive this time, since we don't need socket_fd and the netns and ipcns fds any
4845 * more. We do keep exec_fd however, if we have it, since we need to keep it open until the final
4846 * execve(). But first, close the remaining sockets in the context objects. */
4848 exec_runtime_close(runtime
);
4849 exec_params_close(params
);
4851 r
= close_all_fds(keep_fds
, n_keep_fds
);
4853 r
= pack_fds(params
->fds
, n_fds
);
4855 r
= flag_fds(params
->fds
, n_socket_fds
, n_fds
, context
->non_blocking
);
4857 *exit_status
= EXIT_FDS
;
4858 return log_exec_error_errno(context
, params
, r
, "Failed to adjust passed file descriptors: %m");
4861 /* At this point, the fds we want to pass to the program are all ready and set up, with O_CLOEXEC turned off
4862 * and at the right fd numbers. The are no other fds open, with one exception: the exec_fd if it is defined,
4863 * and it has O_CLOEXEC set, after all we want it to be closed by the execve(), so that our parent knows we
4866 secure_bits
= context
->secure_bits
;
4868 if (needs_sandboxing
) {
4871 /* Set the RTPRIO resource limit to 0, but only if nothing else was explicitly requested.
4872 * (Note this is placed after the general resource limit initialization, see above, in order
4873 * to take precedence.) */
4874 if (context
->restrict_realtime
&& !context
->rlimit
[RLIMIT_RTPRIO
]) {
4875 if (setrlimit(RLIMIT_RTPRIO
, &RLIMIT_MAKE_CONST(0)) < 0) {
4876 *exit_status
= EXIT_LIMITS
;
4877 return log_exec_error_errno(context
, params
, errno
, "Failed to adjust RLIMIT_RTPRIO resource limit: %m");
4882 /* LSM Smack needs the capability CAP_MAC_ADMIN to change the current execution security context of the
4883 * process. This is the latest place before dropping capabilities. Other MAC context are set later. */
4884 if (use_smack
&& context
->smack_process_label
) {
4885 r
= setup_smack(params
, context
, executable_fd
);
4886 if (r
< 0 && !context
->smack_process_label_ignore
) {
4887 *exit_status
= EXIT_SMACK_PROCESS_LABEL
;
4888 return log_exec_error_errno(context
, params
, r
, "Failed to set SMACK process label: %m");
4893 bset
= context
->capability_bounding_set
;
4894 /* If the ambient caps hack is enabled (which means the kernel can't do them, and the user asked for
4895 * our magic fallback), then let's add some extra caps, so that the service can drop privs of its own,
4896 * instead of us doing that */
4897 if (needs_ambient_hack
)
4898 bset
|= (UINT64_C(1) << CAP_SETPCAP
) |
4899 (UINT64_C(1) << CAP_SETUID
) |
4900 (UINT64_C(1) << CAP_SETGID
);
4903 /* If the service has any form of a seccomp filter and it allows dropping privileges, we'll
4904 * keep the needed privileges to apply it even if we're not root. */
4906 uid_is_valid(uid
) &&
4907 context_has_seccomp(context
) &&
4908 seccomp_allows_drop_privileges(context
)) {
4909 keep_seccomp_privileges
= true;
4911 if (prctl(PR_SET_KEEPCAPS
, 1) < 0) {
4912 *exit_status
= EXIT_USER
;
4913 return log_exec_error_errno(context
, params
, errno
, "Failed to enable keep capabilities flag: %m");
4916 /* Save the current bounding set so we can restore it after applying the seccomp
4919 bset
|= (UINT64_C(1) << CAP_SYS_ADMIN
) |
4920 (UINT64_C(1) << CAP_SETPCAP
);
4924 if (!cap_test_all(bset
)) {
4925 r
= capability_bounding_set_drop(bset
, /* right_now= */ false);
4927 *exit_status
= EXIT_CAPABILITIES
;
4928 return log_exec_error_errno(context
, params
, r
, "Failed to drop capabilities: %m");
4932 /* Ambient capabilities are cleared during setresuid() (in enforce_user()) even with
4935 * To be able to raise the ambient capabilities after setresuid() they have to be added to
4936 * the inherited set and keep caps has to be set (done in enforce_user()). After setresuid()
4937 * the ambient capabilities can be raised as they are present in the permitted and
4938 * inhertiable set. However it is possible that someone wants to set ambient capabilities
4939 * without changing the user, so we also set the ambient capabilities here.
4941 * The requested ambient capabilities are raised in the inheritable set if the second
4942 * argument is true. */
4943 if (!needs_ambient_hack
) {
4944 r
= capability_ambient_set_apply(capability_ambient_set
, /* also_inherit= */ true);
4946 *exit_status
= EXIT_CAPABILITIES
;
4947 return log_exec_error_errno(context
, params
, r
, "Failed to apply ambient capabilities (before UID change): %m");
4952 /* chroot to root directory first, before we lose the ability to chroot */
4953 r
= apply_root_directory(context
, params
, runtime
, needs_mount_namespace
, exit_status
);
4955 return log_exec_error_errno(context
, params
, r
, "Chrooting to the requested root directory failed: %m");
4958 if (uid_is_valid(uid
)) {
4959 r
= enforce_user(context
, uid
, capability_ambient_set
);
4961 *exit_status
= EXIT_USER
;
4962 return log_exec_error_errno(context
, params
, r
, "Failed to change UID to " UID_FMT
": %m", uid
);
4965 if (keep_seccomp_privileges
) {
4966 if (!FLAGS_SET(capability_ambient_set
, (UINT64_C(1) << CAP_SETUID
))) {
4967 r
= drop_capability(CAP_SETUID
);
4969 *exit_status
= EXIT_USER
;
4970 return log_exec_error_errno(context
, params
, r
, "Failed to drop CAP_SETUID: %m");
4974 r
= keep_capability(CAP_SYS_ADMIN
);
4976 *exit_status
= EXIT_USER
;
4977 return log_exec_error_errno(context
, params
, r
, "Failed to keep CAP_SYS_ADMIN: %m");
4980 r
= keep_capability(CAP_SETPCAP
);
4982 *exit_status
= EXIT_USER
;
4983 return log_exec_error_errno(context
, params
, r
, "Failed to keep CAP_SETPCAP: %m");
4987 if (!needs_ambient_hack
&& capability_ambient_set
!= 0) {
4989 /* Raise the ambient capabilities after user change. */
4990 r
= capability_ambient_set_apply(capability_ambient_set
, /* also_inherit= */ false);
4992 *exit_status
= EXIT_CAPABILITIES
;
4993 return log_exec_error_errno(context
, params
, r
, "Failed to apply ambient capabilities (after UID change): %m");
4999 /* Apply working directory here, because the working directory might be on NFS and only the user
5000 * running this service might have the correct privilege to change to the working directory. Also, it
5001 * is absolutely 💣 crucial 💣 we applied all mount namespacing rearrangements before this, so that
5002 * the cwd cannot be used to pin directories outside of the sandbox. */
5003 r
= apply_working_directory(context
, params
, runtime
, home
, exit_status
);
5005 return log_exec_error_errno(context
, params
, r
, "Changing to the requested working directory failed: %m");
5007 if (needs_sandboxing
) {
5008 /* Apply other MAC contexts late, but before seccomp syscall filtering, as those should really be last to
5009 * influence our own codepaths as little as possible. Moreover, applying MAC contexts usually requires
5010 * syscalls that are subject to seccomp filtering, hence should probably be applied before the syscalls
5011 * are restricted. */
5015 char *exec_context
= mac_selinux_context_net
?: context
->selinux_context
;
5018 r
= setexeccon(exec_context
);
5020 if (!context
->selinux_context_ignore
) {
5021 *exit_status
= EXIT_SELINUX_CONTEXT
;
5022 return log_exec_error_errno(context
, params
, r
, "Failed to change SELinux context to %s: %m", exec_context
);
5024 log_exec_debug_errno(context
,
5027 "Failed to change SELinux context to %s, ignoring: %m",
5035 if (use_apparmor
&& context
->apparmor_profile
) {
5036 r
= aa_change_onexec(context
->apparmor_profile
);
5037 if (r
< 0 && !context
->apparmor_profile_ignore
) {
5038 *exit_status
= EXIT_APPARMOR_PROFILE
;
5039 return log_exec_error_errno(context
,
5042 "Failed to prepare AppArmor profile change to %s: %m",
5043 context
->apparmor_profile
);
5048 /* PR_GET_SECUREBITS is not privileged, while PR_SET_SECUREBITS is. So to suppress potential
5049 * EPERMs we'll try not to call PR_SET_SECUREBITS unless necessary. Setting securebits
5050 * requires CAP_SETPCAP. */
5051 if (prctl(PR_GET_SECUREBITS
) != secure_bits
) {
5052 /* CAP_SETPCAP is required to set securebits. This capability is raised into the
5053 * effective set here.
5055 * The effective set is overwritten during execve() with the following values:
5057 * - ambient set (for non-root processes)
5059 * - (inheritable | bounding) set for root processes)
5061 * Hence there is no security impact to raise it in the effective set before execve
5063 r
= capability_gain_cap_setpcap(/* return_caps= */ NULL
);
5065 *exit_status
= EXIT_CAPABILITIES
;
5066 return log_exec_error_errno(context
, params
, r
, "Failed to gain CAP_SETPCAP for setting secure bits");
5068 if (prctl(PR_SET_SECUREBITS
, secure_bits
) < 0) {
5069 *exit_status
= EXIT_SECUREBITS
;
5070 return log_exec_error_errno(context
, params
, errno
, "Failed to set process secure bits: %m");
5074 if (context_has_no_new_privileges(context
))
5075 if (prctl(PR_SET_NO_NEW_PRIVS
, 1, 0, 0, 0) < 0) {
5076 *exit_status
= EXIT_NO_NEW_PRIVILEGES
;
5077 return log_exec_error_errno(context
, params
, errno
, "Failed to disable new privileges: %m");
5081 r
= apply_address_families(context
, params
);
5083 *exit_status
= EXIT_ADDRESS_FAMILIES
;
5084 return log_exec_error_errno(context
, params
, r
, "Failed to restrict address families: %m");
5087 r
= apply_memory_deny_write_execute(context
, params
);
5089 *exit_status
= EXIT_SECCOMP
;
5090 return log_exec_error_errno(context
, params
, r
, "Failed to disable writing to executable memory: %m");
5093 r
= apply_restrict_realtime(context
, params
);
5095 *exit_status
= EXIT_SECCOMP
;
5096 return log_exec_error_errno(context
, params
, r
, "Failed to apply realtime restrictions: %m");
5099 r
= apply_restrict_suid_sgid(context
, params
);
5101 *exit_status
= EXIT_SECCOMP
;
5102 return log_exec_error_errno(context
, params
, r
, "Failed to apply SUID/SGID restrictions: %m");
5105 r
= apply_restrict_namespaces(context
, params
);
5107 *exit_status
= EXIT_SECCOMP
;
5108 return log_exec_error_errno(context
, params
, r
, "Failed to apply namespace restrictions: %m");
5111 r
= apply_protect_sysctl(context
, params
);
5113 *exit_status
= EXIT_SECCOMP
;
5114 return log_exec_error_errno(context
, params
, r
, "Failed to apply sysctl restrictions: %m");
5117 r
= apply_protect_kernel_modules(context
, params
);
5119 *exit_status
= EXIT_SECCOMP
;
5120 return log_exec_error_errno(context
, params
, r
, "Failed to apply module loading restrictions: %m");
5123 r
= apply_protect_kernel_logs(context
, params
);
5125 *exit_status
= EXIT_SECCOMP
;
5126 return log_exec_error_errno(context
, params
, r
, "Failed to apply kernel log restrictions: %m");
5129 r
= apply_protect_clock(context
, params
);
5131 *exit_status
= EXIT_SECCOMP
;
5132 return log_exec_error_errno(context
, params
, r
, "Failed to apply clock restrictions: %m");
5135 r
= apply_private_devices(context
, params
);
5137 *exit_status
= EXIT_SECCOMP
;
5138 return log_exec_error_errno(context
, params
, r
, "Failed to set up private devices: %m");
5141 r
= apply_syscall_archs(context
, params
);
5143 *exit_status
= EXIT_SECCOMP
;
5144 return log_exec_error_errno(context
, params
, r
, "Failed to apply syscall architecture restrictions: %m");
5147 r
= apply_lock_personality(context
, params
);
5149 *exit_status
= EXIT_SECCOMP
;
5150 return log_exec_error_errno(context
, params
, r
, "Failed to lock personalities: %m");
5153 r
= apply_syscall_log(context
, params
);
5155 *exit_status
= EXIT_SECCOMP
;
5156 return log_exec_error_errno(context
, params
, r
, "Failed to apply system call log filters: %m");
5161 r
= apply_restrict_filesystems(context
, params
);
5163 *exit_status
= EXIT_BPF
;
5164 return log_exec_error_errno(context
, params
, r
, "Failed to restrict filesystems: %m");
5169 /* This really should remain as close to the execve() as possible, to make sure our own code is affected
5170 * by the filter as little as possible. */
5171 r
= apply_syscall_filter(context
, params
, needs_ambient_hack
);
5173 *exit_status
= EXIT_SECCOMP
;
5174 return log_exec_error_errno(context
, params
, r
, "Failed to apply system call filters: %m");
5177 if (keep_seccomp_privileges
) {
5178 /* Restore the capability bounding set with what's expected from the service + the
5179 * ambient capabilities hack */
5180 if (!cap_test_all(saved_bset
)) {
5181 r
= capability_bounding_set_drop(saved_bset
, /* right_now= */ false);
5183 *exit_status
= EXIT_CAPABILITIES
;
5184 return log_exec_error_errno(context
, params
, r
, "Failed to drop bset capabilities: %m");
5188 /* Only drop CAP_SYS_ADMIN if it's not in the bounding set, otherwise we'll break
5189 * applications that use it. */
5190 if (!FLAGS_SET(saved_bset
, (UINT64_C(1) << CAP_SYS_ADMIN
))) {
5191 r
= drop_capability(CAP_SYS_ADMIN
);
5193 *exit_status
= EXIT_USER
;
5194 return log_exec_error_errno(context
, params
, r
, "Failed to drop CAP_SYS_ADMIN: %m");
5198 /* Only drop CAP_SETPCAP if it's not in the bounding set, otherwise we'll break
5199 * applications that use it. */
5200 if (!FLAGS_SET(saved_bset
, (UINT64_C(1) << CAP_SETPCAP
))) {
5201 r
= drop_capability(CAP_SETPCAP
);
5203 *exit_status
= EXIT_USER
;
5204 return log_exec_error_errno(context
, params
, r
, "Failed to drop CAP_SETPCAP: %m");
5208 if (prctl(PR_SET_KEEPCAPS
, 0) < 0) {
5209 *exit_status
= EXIT_USER
;
5210 return log_exec_error_errno(context
, params
, errno
, "Failed to drop keep capabilities flag: %m");
5217 if (!strv_isempty(context
->unset_environment
)) {
5220 ee
= strv_env_delete(accum_env
, 1, context
->unset_environment
);
5222 *exit_status
= EXIT_MEMORY
;
5226 strv_free_and_replace(accum_env
, ee
);
5229 if (!FLAGS_SET(command
->flags
, EXEC_COMMAND_NO_ENV_EXPAND
)) {
5230 _cleanup_strv_free_
char **unset_variables
= NULL
, **bad_variables
= NULL
;
5232 r
= replace_env_argv(command
->argv
, accum_env
, &replaced_argv
, &unset_variables
, &bad_variables
);
5234 *exit_status
= EXIT_MEMORY
;
5235 return log_exec_error_errno(context
,
5238 "Failed to replace environment variables: %m");
5240 final_argv
= replaced_argv
;
5242 if (!strv_isempty(unset_variables
)) {
5243 _cleanup_free_
char *ju
= strv_join(unset_variables
, ", ");
5244 log_exec_warning(context
,
5246 "Referenced but unset environment variable evaluates to an empty string: %s",
5250 if (!strv_isempty(bad_variables
)) {
5251 _cleanup_free_
char *jb
= strv_join(bad_variables
, ", ");
5252 log_exec_warning(context
,
5254 "Invalid environment variable name evaluates to an empty string: %s",
5258 final_argv
= command
->argv
;
5260 log_command_line(context
, params
, "Executing", executable
, final_argv
);
5262 if (params
->exec_fd
>= 0) {
5265 /* We have finished with all our initializations. Let's now let the manager know that. From this point
5266 * on, if the manager sees POLLHUP on the exec_fd, then execve() was successful. */
5268 if (write(params
->exec_fd
, &hot
, sizeof(hot
)) < 0) {
5269 *exit_status
= EXIT_EXEC
;
5270 return log_exec_error_errno(context
, params
, errno
, "Failed to enable exec_fd: %m");
5274 r
= fexecve_or_execve(executable_fd
, executable
, final_argv
, accum_env
);
5276 if (params
->exec_fd
>= 0) {
5279 /* The execve() failed. This means the exec_fd is still open. Which means we need to tell the manager
5280 * that POLLHUP on it no longer means execve() succeeded. */
5282 if (write(params
->exec_fd
, &hot
, sizeof(hot
)) < 0) {
5283 *exit_status
= EXIT_EXEC
;
5284 return log_exec_error_errno(context
, params
, errno
, "Failed to disable exec_fd: %m");
5288 *exit_status
= EXIT_EXEC
;
5289 return log_exec_error_errno(context
, params
, r
, "Failed to execute %s: %m", executable
);