1 /* SPDX-License-Identifier: LGPL-2.1+ */
3 This file is part of systemd.
5 Copyright 2010 Lennart Poettering
11 #include <linux/oom.h>
16 #include <stdio_ext.h>
20 #include <sys/personality.h>
21 #include <sys/prctl.h>
22 #include <sys/types.h>
26 #if HAVE_VALGRIND_VALGRIND_H
27 #include <valgrind/valgrind.h>
30 #include "alloc-util.h"
31 #include "architecture.h"
40 #include "process-util.h"
41 #include "raw-clone.h"
42 #include "signal-util.h"
43 #include "stat-util.h"
44 #include "string-table.h"
45 #include "string-util.h"
46 #include "terminal-util.h"
47 #include "user-util.h"
50 int get_process_state(pid_t pid
) {
54 _cleanup_free_
char *line
= NULL
;
58 p
= procfs_file_alloca(pid
, "stat");
60 r
= read_one_line_file(p
, &line
);
66 p
= strrchr(line
, ')');
72 if (sscanf(p
, " %c", &state
) != 1)
75 return (unsigned char) state
;
78 int get_process_comm(pid_t pid
, char **name
) {
85 p
= procfs_file_alloca(pid
, "comm");
87 r
= read_one_line_file(p
, name
);
94 int get_process_cmdline(pid_t pid
, size_t max_length
, bool comm_fallback
, char **line
) {
95 _cleanup_fclose_
FILE *f
= NULL
;
104 /* Retrieves a process' command line. Replaces unprintable characters while doing so by whitespace (coalescing
105 * multiple sequential ones into one). If max_length is != 0 will return a string of the specified size at most
106 * (the trailing NUL byte does count towards the length here!), abbreviated with a "..." ellipsis. If
107 * comm_fallback is true and the process has no command line set (the case for kernel threads), or has a
108 * command line that resolves to the empty string will return the "comm" name of the process instead.
110 * Returns -ESRCH if the process doesn't exist, and -ENOENT if the process has no command line (and
111 * comm_fallback is false). Returns 0 and sets *line otherwise. */
113 p
= procfs_file_alloca(pid
, "cmdline");
122 (void) __fsetlocking(f
, FSETLOCKING_BYCALLER
);
124 if (max_length
== 1) {
126 /* If there's only room for one byte, return the empty string */
134 } else if (max_length
== 0) {
135 size_t len
= 0, allocated
= 0;
137 while ((c
= getc(f
)) != EOF
) {
139 if (!GREEDY_REALLOC(ans
, allocated
, len
+3)) {
161 bool dotdotdot
= false;
164 ans
= new(char, max_length
);
170 while ((c
= getc(f
)) != EOF
) {
197 if (max_length
<= 4) {
201 k
= ans
+ max_length
- 4;
204 /* Eat up final spaces */
205 while (k
> ans
&& isspace(k
[-1])) {
211 strncpy(k
, "...", left
-1);
217 /* Kernel threads have no argv[] */
219 _cleanup_free_
char *t
= NULL
;
227 h
= get_process_comm(pid
, &t
);
232 ans
= strjoin("[", t
, "]");
238 if (l
+ 3 <= max_length
)
239 ans
= strjoin("[", t
, "]");
240 else if (max_length
<= 6) {
242 ans
= new(char, max_length
);
246 memcpy(ans
, "[...]", max_length
-1);
247 ans
[max_length
-1] = 0;
251 t
[max_length
- 6] = 0;
253 /* Chop off final spaces */
255 while (e
> t
&& isspace(e
[-1]))
259 ans
= strjoin("[", t
, "...]");
270 int rename_process(const char name
[]) {
271 static size_t mm_size
= 0;
272 static char *mm
= NULL
;
273 bool truncated
= false;
276 /* This is a like a poor man's setproctitle(). It changes the comm field, argv[0], and also the glibc's
277 * internally used name of the process. For the first one a limit of 16 chars applies; to the second one in
278 * many cases one of 10 (i.e. length of "/sbin/init") — however if we have CAP_SYS_RESOURCES it is unbounded;
279 * to the third one 7 (i.e. the length of "systemd". If you pass a longer string it will likely be
282 * Returns 0 if a name was set but truncated, > 0 if it was set but not truncated. */
285 return -EINVAL
; /* let's not confuse users unnecessarily with an empty name */
287 if (!is_main_thread())
288 return -EPERM
; /* Let's not allow setting the process name from other threads than the main one, as we
289 * cache things without locking, and we make assumptions that PR_SET_NAME sets the
290 * process name that isn't correct on any other threads */
294 /* First step, change the comm field. The main thread's comm is identical to the process comm. This means we
295 * can use PR_SET_NAME, which sets the thread name for the calling thread. */
296 if (prctl(PR_SET_NAME
, name
) < 0)
297 log_debug_errno(errno
, "PR_SET_NAME failed: %m");
298 if (l
> 15) /* Linux process names can be 15 chars at max */
301 /* Second step, change glibc's ID of the process name. */
302 if (program_invocation_name
) {
305 k
= strlen(program_invocation_name
);
306 strncpy(program_invocation_name
, name
, k
);
311 /* Third step, completely replace the argv[] array the kernel maintains for us. This requires privileges, but
312 * has the advantage that the argv[] array is exactly what we want it to be, and not filled up with zeros at
313 * the end. This is the best option for changing /proc/self/cmdline. */
315 /* Let's not bother with this if we don't have euid == 0. Strictly speaking we should check for the
316 * CAP_SYS_RESOURCE capability which is independent of the euid. In our own code the capability generally is
317 * present only for euid == 0, hence let's use this as quick bypass check, to avoid calling mmap() if
318 * PR_SET_MM_ARG_{START,END} fails with EPERM later on anyway. After all geteuid() is dead cheap to call, but
321 log_debug("Skipping PR_SET_MM, as we don't have privileges.");
322 else if (mm_size
< l
+1) {
326 nn_size
= PAGE_ALIGN(l
+1);
327 nn
= mmap(NULL
, nn_size
, PROT_READ
|PROT_WRITE
, MAP_PRIVATE
|MAP_ANONYMOUS
, -1, 0);
328 if (nn
== MAP_FAILED
) {
329 log_debug_errno(errno
, "mmap() failed: %m");
333 strncpy(nn
, name
, nn_size
);
335 /* Now, let's tell the kernel about this new memory */
336 if (prctl(PR_SET_MM
, PR_SET_MM_ARG_START
, (unsigned long) nn
, 0, 0) < 0) {
337 log_debug_errno(errno
, "PR_SET_MM_ARG_START failed, proceeding without: %m");
338 (void) munmap(nn
, nn_size
);
342 /* And update the end pointer to the new end, too. If this fails, we don't really know what to do, it's
343 * pretty unlikely that we can rollback, hence we'll just accept the failure, and continue. */
344 if (prctl(PR_SET_MM
, PR_SET_MM_ARG_END
, (unsigned long) nn
+ l
+ 1, 0, 0) < 0)
345 log_debug_errno(errno
, "PR_SET_MM_ARG_END failed, proceeding without: %m");
348 (void) munmap(mm
, mm_size
);
353 strncpy(mm
, name
, mm_size
);
355 /* Update the end pointer, continuing regardless of any failure. */
356 if (prctl(PR_SET_MM
, PR_SET_MM_ARG_END
, (unsigned long) mm
+ l
+ 1, 0, 0) < 0)
357 log_debug_errno(errno
, "PR_SET_MM_ARG_END failed, proceeding without: %m");
361 /* Fourth step: in all cases we'll also update the original argv[], so that our own code gets it right too if
362 * it still looks here */
364 if (saved_argc
> 0) {
370 k
= strlen(saved_argv
[0]);
371 strncpy(saved_argv
[0], name
, k
);
376 for (i
= 1; i
< saved_argc
; i
++) {
380 memzero(saved_argv
[i
], strlen(saved_argv
[i
]));
387 int is_kernel_thread(pid_t pid
) {
388 _cleanup_free_
char *line
= NULL
;
389 unsigned long long flags
;
395 if (IN_SET(pid
, 0, 1) || pid
== getpid_cached()) /* pid 1, and we ourselves certainly aren't a kernel thread */
397 if (!pid_is_valid(pid
))
400 p
= procfs_file_alloca(pid
, "stat");
401 r
= read_one_line_file(p
, &line
);
407 /* Skip past the comm field */
408 q
= strrchr(line
, ')');
413 /* Skip 6 fields to reach the flags field */
414 for (i
= 0; i
< 6; i
++) {
415 l
= strspn(q
, WHITESPACE
);
420 l
= strcspn(q
, WHITESPACE
);
426 /* Skip preceeding whitespace */
427 l
= strspn(q
, WHITESPACE
);
432 /* Truncate the rest */
433 l
= strcspn(q
, WHITESPACE
);
438 r
= safe_atollu(q
, &flags
);
442 return !!(flags
& PF_KTHREAD
);
445 int get_process_capeff(pid_t pid
, char **capeff
) {
452 p
= procfs_file_alloca(pid
, "status");
454 r
= get_proc_field(p
, "CapEff", WHITESPACE
, capeff
);
461 static int get_process_link_contents(const char *proc_file
, char **name
) {
467 r
= readlink_malloc(proc_file
, name
);
476 int get_process_exe(pid_t pid
, char **name
) {
483 p
= procfs_file_alloca(pid
, "exe");
484 r
= get_process_link_contents(p
, name
);
488 d
= endswith(*name
, " (deleted)");
495 static int get_process_id(pid_t pid
, const char *field
, uid_t
*uid
) {
496 _cleanup_fclose_
FILE *f
= NULL
;
506 p
= procfs_file_alloca(pid
, "status");
514 (void) __fsetlocking(f
, FSETLOCKING_BYCALLER
);
516 FOREACH_LINE(line
, f
, return -errno
) {
521 if (startswith(l
, field
)) {
523 l
+= strspn(l
, WHITESPACE
);
525 l
[strcspn(l
, WHITESPACE
)] = 0;
527 return parse_uid(l
, uid
);
534 int get_process_uid(pid_t pid
, uid_t
*uid
) {
536 if (pid
== 0 || pid
== getpid_cached()) {
541 return get_process_id(pid
, "Uid:", uid
);
544 int get_process_gid(pid_t pid
, gid_t
*gid
) {
546 if (pid
== 0 || pid
== getpid_cached()) {
551 assert_cc(sizeof(uid_t
) == sizeof(gid_t
));
552 return get_process_id(pid
, "Gid:", gid
);
555 int get_process_cwd(pid_t pid
, char **cwd
) {
560 p
= procfs_file_alloca(pid
, "cwd");
562 return get_process_link_contents(p
, cwd
);
565 int get_process_root(pid_t pid
, char **root
) {
570 p
= procfs_file_alloca(pid
, "root");
572 return get_process_link_contents(p
, root
);
575 int get_process_environ(pid_t pid
, char **env
) {
576 _cleanup_fclose_
FILE *f
= NULL
;
577 _cleanup_free_
char *outcome
= NULL
;
580 size_t allocated
= 0, sz
= 0;
585 p
= procfs_file_alloca(pid
, "environ");
594 (void) __fsetlocking(f
, FSETLOCKING_BYCALLER
);
596 while ((c
= fgetc(f
)) != EOF
) {
597 if (!GREEDY_REALLOC(outcome
, allocated
, sz
+ 5))
601 outcome
[sz
++] = '\n';
603 sz
+= cescape_char(c
, outcome
+ sz
);
607 outcome
= strdup("");
613 *env
= TAKE_PTR(outcome
);
618 int get_process_ppid(pid_t pid
, pid_t
*_ppid
) {
620 _cleanup_free_
char *line
= NULL
;
627 if (pid
== 0 || pid
== getpid_cached()) {
632 p
= procfs_file_alloca(pid
, "stat");
633 r
= read_one_line_file(p
, &line
);
639 /* Let's skip the pid and comm fields. The latter is enclosed
640 * in () but does not escape any () in its value, so let's
641 * skip over it manually */
643 p
= strrchr(line
, ')');
655 if ((long unsigned) (pid_t
) ppid
!= ppid
)
658 *_ppid
= (pid_t
) ppid
;
663 int wait_for_terminate(pid_t pid
, siginfo_t
*status
) {
674 if (waitid(P_PID
, pid
, status
, WEXITED
) < 0) {
679 return negative_errno();
688 * < 0 : wait_for_terminate() failed to get the state of the
689 * process, the process was terminated by a signal, or
690 * failed for an unknown reason.
691 * >=0 : The process terminated normally, and its exit code is
694 * That is, success is indicated by a return value of zero, and an
695 * error is indicated by a non-zero value.
697 * A warning is emitted if the process terminates abnormally,
698 * and also if it returns non-zero unless check_exit_code is true.
700 int wait_for_terminate_and_check(const char *name
, pid_t pid
, WaitFlags flags
) {
701 _cleanup_free_
char *buffer
= NULL
;
708 r
= get_process_comm(pid
, &buffer
);
710 log_debug_errno(r
, "Failed to acquire process name of " PID_FMT
", ignoring: %m", pid
);
715 prio
= flags
& WAIT_LOG_ABNORMAL
? LOG_ERR
: LOG_DEBUG
;
717 r
= wait_for_terminate(pid
, &status
);
719 return log_full_errno(prio
, r
, "Failed to wait for %s: %m", strna(name
));
721 if (status
.si_code
== CLD_EXITED
) {
722 if (status
.si_status
!= EXIT_SUCCESS
)
723 log_full(flags
& WAIT_LOG_NON_ZERO_EXIT_STATUS
? LOG_ERR
: LOG_DEBUG
,
724 "%s failed with exit status %i.", strna(name
), status
.si_status
);
726 log_debug("%s succeeded.", name
);
728 return status
.si_status
;
730 } else if (IN_SET(status
.si_code
, CLD_KILLED
, CLD_DUMPED
)) {
732 log_full(prio
, "%s terminated by signal %s.", strna(name
), signal_to_string(status
.si_status
));
736 log_full(prio
, "%s failed due to unknown reason.", strna(name
));
742 * < 0 : wait_for_terminate_with_timeout() failed to get the state of the
743 * process, the process timed out, the process was terminated by a
744 * signal, or failed for an unknown reason.
745 * >=0 : The process terminated normally with no failures.
747 * Success is indicated by a return value of zero, a timeout is indicated
748 * by ETIMEDOUT, and all other child failure states are indicated by error
749 * is indicated by a non-zero value.
751 int wait_for_terminate_with_timeout(pid_t pid
, usec_t timeout
) {
756 assert_se(sigemptyset(&mask
) == 0);
757 assert_se(sigaddset(&mask
, SIGCHLD
) == 0);
759 /* Drop into a sigtimewait-based timeout. Waiting for the
761 until
= now(CLOCK_MONOTONIC
) + timeout
;
764 siginfo_t status
= {};
767 n
= now(CLOCK_MONOTONIC
);
771 r
= sigtimedwait(&mask
, NULL
, timespec_store(&ts
, until
- n
)) < 0 ? -errno
: 0;
772 /* Assuming we woke due to the child exiting. */
773 if (waitid(P_PID
, pid
, &status
, WEXITED
|WNOHANG
) == 0) {
774 if (status
.si_pid
== pid
) {
775 /* This is the correct child.*/
776 if (status
.si_code
== CLD_EXITED
)
777 return (status
.si_status
== 0) ? 0 : -EPROTO
;
782 /* Not the child, check for errors and proceed appropriately */
786 /* Timed out, child is likely hung. */
789 /* Received a different signal and should retry */
792 /* Return any unexpected errors */
801 void sigkill_wait(pid_t pid
) {
804 if (kill(pid
, SIGKILL
) > 0)
805 (void) wait_for_terminate(pid
, NULL
);
808 void sigkill_waitp(pid_t
*pid
) {
819 void sigterm_wait(pid_t pid
) {
822 if (kill_and_sigcont(pid
, SIGTERM
) > 0)
823 (void) wait_for_terminate(pid
, NULL
);
826 int kill_and_sigcont(pid_t pid
, int sig
) {
829 r
= kill(pid
, sig
) < 0 ? -errno
: 0;
831 /* If this worked, also send SIGCONT, unless we already just sent a SIGCONT, or SIGKILL was sent which isn't
832 * affected by a process being suspended anyway. */
833 if (r
>= 0 && !IN_SET(sig
, SIGCONT
, SIGKILL
))
834 (void) kill(pid
, SIGCONT
);
839 int getenv_for_pid(pid_t pid
, const char *field
, char **ret
) {
840 _cleanup_fclose_
FILE *f
= NULL
;
850 if (pid
== 0 || pid
== getpid_cached()) {
867 path
= procfs_file_alloca(pid
, "environ");
869 f
= fopen(path
, "re");
877 (void) __fsetlocking(f
, FSETLOCKING_BYCALLER
);
885 for (i
= 0; i
< sizeof(line
)-1; i
++) {
889 if (_unlikely_(c
== EOF
)) {
899 if (strneq(line
, field
, l
) && line
[l
] == '=') {
900 value
= strdup(line
+ l
+ 1);
914 bool pid_is_unwaited(pid_t pid
) {
915 /* Checks whether a PID is still valid at all, including a zombie */
920 if (pid
<= 1) /* If we or PID 1 would be dead and have been waited for, this code would not be running */
923 if (pid
== getpid_cached())
926 if (kill(pid
, 0) >= 0)
929 return errno
!= ESRCH
;
932 bool pid_is_alive(pid_t pid
) {
935 /* Checks whether a PID is still valid and not a zombie */
940 if (pid
<= 1) /* If we or PID 1 would be a zombie, this code would not be running */
943 if (pid
== getpid_cached())
946 r
= get_process_state(pid
);
947 if (IN_SET(r
, -ESRCH
, 'Z'))
953 int pid_from_same_root_fs(pid_t pid
) {
959 if (pid
== 0 || pid
== getpid_cached())
962 root
= procfs_file_alloca(pid
, "root");
964 return files_same(root
, "/proc/1/root", 0);
967 bool is_main_thread(void) {
968 static thread_local
int cached
= 0;
970 if (_unlikely_(cached
== 0))
971 cached
= getpid_cached() == gettid() ? 1 : -1;
976 _noreturn_
void freeze(void) {
980 /* Make sure nobody waits for us on a socket anymore */
981 close_all_fds(NULL
, 0);
985 /* Let's not freeze right away, but keep reaping zombies. */
990 r
= waitid(P_ALL
, 0, &si
, WEXITED
);
991 if (r
< 0 && errno
!= EINTR
)
995 /* waitid() failed with an unexpected error, things are really borked. Freeze now! */
1000 bool oom_score_adjust_is_valid(int oa
) {
1001 return oa
>= OOM_SCORE_ADJ_MIN
&& oa
<= OOM_SCORE_ADJ_MAX
;
1004 unsigned long personality_from_string(const char *p
) {
1008 return PERSONALITY_INVALID
;
1010 /* Parse a personality specifier. We use our own identifiers that indicate specific ABIs, rather than just
1011 * hints regarding the register size, since we want to keep things open for multiple locally supported ABIs for
1012 * the same register size. */
1014 architecture
= architecture_from_string(p
);
1015 if (architecture
< 0)
1016 return PERSONALITY_INVALID
;
1018 if (architecture
== native_architecture())
1020 #ifdef SECONDARY_ARCHITECTURE
1021 if (architecture
== SECONDARY_ARCHITECTURE
)
1025 return PERSONALITY_INVALID
;
1028 const char* personality_to_string(unsigned long p
) {
1029 int architecture
= _ARCHITECTURE_INVALID
;
1032 architecture
= native_architecture();
1033 #ifdef SECONDARY_ARCHITECTURE
1034 else if (p
== PER_LINUX32
)
1035 architecture
= SECONDARY_ARCHITECTURE
;
1038 if (architecture
< 0)
1041 return architecture_to_string(architecture
);
1044 int safe_personality(unsigned long p
) {
1047 /* So here's the deal, personality() is weirdly defined by glibc. In some cases it returns a failure via errno,
1048 * and in others as negative return value containing an errno-like value. Let's work around this: this is a
1049 * wrapper that uses errno if it is set, and uses the return value otherwise. And then it sets both errno and
1050 * the return value indicating the same issue, so that we are definitely on the safe side.
1052 * See https://github.com/systemd/systemd/issues/6737 */
1055 ret
= personality(p
);
1066 int opinionated_personality(unsigned long *ret
) {
1069 /* Returns the current personality, or PERSONALITY_INVALID if we can't determine it. This function is a bit
1070 * opinionated though, and ignores all the finer-grained bits and exotic personalities, only distinguishing the
1071 * two most relevant personalities: PER_LINUX and PER_LINUX32. */
1073 current
= safe_personality(PERSONALITY_INVALID
);
1077 if (((unsigned long) current
& 0xffff) == PER_LINUX32
)
1085 void valgrind_summary_hack(void) {
1086 #if HAVE_VALGRIND_VALGRIND_H
1087 if (getpid_cached() == 1 && RUNNING_ON_VALGRIND
) {
1089 pid
= raw_clone(SIGCHLD
);
1091 log_emergency_errno(errno
, "Failed to fork off valgrind helper: %m");
1095 log_info("Spawned valgrind helper as PID "PID_FMT
".", pid
);
1096 (void) wait_for_terminate(pid
, NULL
);
1102 int pid_compare_func(const void *a
, const void *b
) {
1103 const pid_t
*p
= a
, *q
= b
;
1105 /* Suitable for usage in qsort() */
1114 int ioprio_parse_priority(const char *s
, int *ret
) {
1120 r
= safe_atoi(s
, &i
);
1124 if (!ioprio_priority_is_valid(i
))
1131 /* The cached PID, possible values:
1133 * == UNSET [0] → cache not initialized yet
1134 * == BUSY [-1] → some thread is initializing it at the moment
1135 * any other → the cached PID
1138 #define CACHED_PID_UNSET ((pid_t) 0)
1139 #define CACHED_PID_BUSY ((pid_t) -1)
1141 static pid_t cached_pid
= CACHED_PID_UNSET
;
1143 void reset_cached_pid(void) {
1144 /* Invoked in the child after a fork(), i.e. at the first moment the PID changed */
1145 cached_pid
= CACHED_PID_UNSET
;
1148 /* We use glibc __register_atfork() + __dso_handle directly here, as they are not included in the glibc
1149 * headers. __register_atfork() is mostly equivalent to pthread_atfork(), but doesn't require us to link against
1150 * libpthread, as it is part of glibc anyway. */
1151 extern int __register_atfork(void (*prepare
) (void), void (*parent
) (void), void (*child
) (void), void * __dso_handle
);
1152 extern void* __dso_handle
__attribute__ ((__weak__
));
1154 pid_t
getpid_cached(void) {
1155 static bool installed
= false;
1156 pid_t current_value
;
1158 /* getpid_cached() is much like getpid(), but caches the value in local memory, to avoid having to invoke a
1159 * system call each time. This restores glibc behaviour from before 2.24, when getpid() was unconditionally
1160 * cached. Starting with 2.24 getpid() started to become prohibitively expensive when used for detecting when
1161 * objects were used across fork()s. With this caching the old behaviour is somewhat restored.
1163 * https://bugzilla.redhat.com/show_bug.cgi?id=1443976
1164 * https://sourceware.org/git/gitweb.cgi?p=glibc.git;h=c579f48edba88380635ab98cb612030e3ed8691e
1167 current_value
= __sync_val_compare_and_swap(&cached_pid
, CACHED_PID_UNSET
, CACHED_PID_BUSY
);
1169 switch (current_value
) {
1171 case CACHED_PID_UNSET
: { /* Not initialized yet, then do so now */
1174 new_pid
= raw_getpid();
1177 /* __register_atfork() either returns 0 or -ENOMEM, in its glibc implementation. Since it's
1178 * only half-documented (glibc doesn't document it but LSB does — though only superficially)
1179 * we'll check for errors only in the most generic fashion possible. */
1181 if (__register_atfork(NULL
, NULL
, reset_cached_pid
, __dso_handle
) != 0) {
1182 /* OOM? Let's try again later */
1183 cached_pid
= CACHED_PID_UNSET
;
1190 cached_pid
= new_pid
;
1194 case CACHED_PID_BUSY
: /* Somebody else is currently initializing */
1195 return raw_getpid();
1197 default: /* Properly initialized */
1198 return current_value
;
1202 int must_be_root(void) {
1207 log_error("Need to be root.");
1213 const int except_fds
[],
1214 size_t n_except_fds
,
1218 pid_t original_pid
, pid
;
1219 sigset_t saved_ss
, ss
;
1220 bool block_signals
= false;
1223 /* A wrapper around fork(), that does a couple of important initializations in addition to mere forking. Always
1224 * returns the child's PID in *ret_pid. Returns == 0 in the child, and > 0 in the parent. */
1226 prio
= flags
& FORK_LOG
? LOG_ERR
: LOG_DEBUG
;
1228 original_pid
= getpid_cached();
1230 if (flags
& (FORK_RESET_SIGNALS
|FORK_DEATHSIG
)) {
1232 /* We temporarily block all signals, so that the new child has them blocked initially. This way, we can
1233 * be sure that SIGTERMs are not lost we might send to the child. */
1235 if (sigfillset(&ss
) < 0)
1236 return log_full_errno(prio
, errno
, "Failed to reset signal set: %m");
1238 block_signals
= true;
1240 } else if (flags
& FORK_WAIT
) {
1242 /* Let's block SIGCHLD at least, so that we can safely watch for the child process */
1244 if (sigemptyset(&ss
) < 0)
1245 return log_full_errno(prio
, errno
, "Failed to clear signal set: %m");
1247 if (sigaddset(&ss
, SIGCHLD
) < 0)
1248 return log_full_errno(prio
, errno
, "Failed to add SIGCHLD to signal set: %m");
1250 block_signals
= true;
1254 if (sigprocmask(SIG_SETMASK
, &ss
, &saved_ss
) < 0)
1255 return log_full_errno(prio
, errno
, "Failed to set signal mask: %m");
1257 if (flags
& FORK_NEW_MOUNTNS
)
1258 pid
= raw_clone(SIGCHLD
|CLONE_NEWNS
);
1264 if (block_signals
) /* undo what we did above */
1265 (void) sigprocmask(SIG_SETMASK
, &saved_ss
, NULL
);
1267 return log_full_errno(prio
, r
, "Failed to fork: %m");
1270 /* We are in the parent process */
1272 log_debug("Successfully forked off '%s' as PID " PID_FMT
".", strna(name
), pid
);
1274 if (flags
& FORK_WAIT
) {
1275 r
= wait_for_terminate_and_check(name
, pid
, (flags
& FORK_LOG
? WAIT_LOG
: 0));
1278 if (r
!= EXIT_SUCCESS
) /* exit status > 0 should be treated as failure, too */
1282 if (block_signals
) /* undo what we did above */
1283 (void) sigprocmask(SIG_SETMASK
, &saved_ss
, NULL
);
1291 /* We are in the child process */
1293 if (flags
& FORK_REOPEN_LOG
) {
1294 /* Close the logs if requested, before we log anything. And make sure we reopen it if needed. */
1296 log_set_open_when_needed(true);
1300 r
= rename_process(name
);
1302 log_full_errno(flags
& FORK_LOG
? LOG_WARNING
: LOG_DEBUG
,
1303 r
, "Failed to rename process, ignoring: %m");
1306 if (flags
& FORK_DEATHSIG
)
1307 if (prctl(PR_SET_PDEATHSIG
, SIGTERM
) < 0) {
1308 log_full_errno(prio
, errno
, "Failed to set death signal: %m");
1309 _exit(EXIT_FAILURE
);
1312 if (flags
& FORK_RESET_SIGNALS
) {
1313 r
= reset_all_signal_handlers();
1315 log_full_errno(prio
, r
, "Failed to reset signal handlers: %m");
1316 _exit(EXIT_FAILURE
);
1319 /* This implicitly undoes the signal mask stuff we did before the fork()ing above */
1320 r
= reset_signal_mask();
1322 log_full_errno(prio
, r
, "Failed to reset signal mask: %m");
1323 _exit(EXIT_FAILURE
);
1325 } else if (block_signals
) { /* undo what we did above */
1326 if (sigprocmask(SIG_SETMASK
, &saved_ss
, NULL
) < 0) {
1327 log_full_errno(prio
, errno
, "Failed to restore signal mask: %m");
1328 _exit(EXIT_FAILURE
);
1332 if (flags
& FORK_DEATHSIG
) {
1334 /* Let's see if the parent PID is still the one we started from? If not, then the parent
1335 * already died by the time we set PR_SET_PDEATHSIG, hence let's emulate the effect */
1339 /* Parent is in a differn't PID namespace. */;
1340 else if (ppid
!= original_pid
) {
1341 log_debug("Parent died early, raising SIGTERM.");
1342 (void) raise(SIGTERM
);
1343 _exit(EXIT_FAILURE
);
1347 if (flags
& FORK_CLOSE_ALL_FDS
) {
1348 /* Close the logs here in case it got reopened above, as close_all_fds() would close them for us */
1351 r
= close_all_fds(except_fds
, n_except_fds
);
1353 log_full_errno(prio
, r
, "Failed to close all file descriptors: %m");
1354 _exit(EXIT_FAILURE
);
1358 /* When we were asked to reopen the logs, do so again now */
1359 if (flags
& FORK_REOPEN_LOG
) {
1361 log_set_open_when_needed(false);
1364 if (flags
& FORK_NULL_STDIO
) {
1365 r
= make_null_stdio();
1367 log_full_errno(prio
, r
, "Failed to connect stdin/stdout to /dev/null: %m");
1368 _exit(EXIT_FAILURE
);
1373 *ret_pid
= getpid_cached();
1378 int fork_agent(const char *name
, const int except
[], size_t n_except
, pid_t
*ret_pid
, const char *path
, ...) {
1379 bool stdout_is_tty
, stderr_is_tty
;
1387 /* Spawns a temporary TTY agent, making sure it goes away when we go away */
1389 r
= safe_fork_full(name
, except
, n_except
, FORK_RESET_SIGNALS
|FORK_DEATHSIG
|FORK_CLOSE_ALL_FDS
, ret_pid
);
1397 stdout_is_tty
= isatty(STDOUT_FILENO
);
1398 stderr_is_tty
= isatty(STDERR_FILENO
);
1400 if (!stdout_is_tty
|| !stderr_is_tty
) {
1403 /* Detach from stdout/stderr. and reopen
1404 * /dev/tty for them. This is important to
1405 * ensure that when systemctl is started via
1406 * popen() or a similar call that expects to
1407 * read EOF we actually do generate EOF and
1408 * not delay this indefinitely by because we
1409 * keep an unused copy of stdin around. */
1410 fd
= open("/dev/tty", O_WRONLY
);
1412 log_error_errno(errno
, "Failed to open /dev/tty: %m");
1413 _exit(EXIT_FAILURE
);
1416 if (!stdout_is_tty
&& dup2(fd
, STDOUT_FILENO
) < 0) {
1417 log_error_errno(errno
, "Failed to dup2 /dev/tty: %m");
1418 _exit(EXIT_FAILURE
);
1421 if (!stderr_is_tty
&& dup2(fd
, STDERR_FILENO
) < 0) {
1422 log_error_errno(errno
, "Failed to dup2 /dev/tty: %m");
1423 _exit(EXIT_FAILURE
);
1426 safe_close_above_stdio(fd
);
1429 /* Count arguments */
1431 for (n
= 0; va_arg(ap
, char*); n
++)
1436 l
= newa(char*, n
+ 1);
1438 /* Fill in arguments */
1440 for (i
= 0; i
<= n
; i
++)
1441 l
[i
] = va_arg(ap
, char*);
1445 _exit(EXIT_FAILURE
);
1448 static const char *const ioprio_class_table
[] = {
1449 [IOPRIO_CLASS_NONE
] = "none",
1450 [IOPRIO_CLASS_RT
] = "realtime",
1451 [IOPRIO_CLASS_BE
] = "best-effort",
1452 [IOPRIO_CLASS_IDLE
] = "idle"
1455 DEFINE_STRING_TABLE_LOOKUP_WITH_FALLBACK(ioprio_class
, int, IOPRIO_N_CLASSES
);
1457 static const char *const sigchld_code_table
[] = {
1458 [CLD_EXITED
] = "exited",
1459 [CLD_KILLED
] = "killed",
1460 [CLD_DUMPED
] = "dumped",
1461 [CLD_TRAPPED
] = "trapped",
1462 [CLD_STOPPED
] = "stopped",
1463 [CLD_CONTINUED
] = "continued",
1466 DEFINE_STRING_TABLE_LOOKUP(sigchld_code
, int);
1468 static const char* const sched_policy_table
[] = {
1469 [SCHED_OTHER
] = "other",
1470 [SCHED_BATCH
] = "batch",
1471 [SCHED_IDLE
] = "idle",
1472 [SCHED_FIFO
] = "fifo",
1476 DEFINE_STRING_TABLE_LOOKUP_WITH_FALLBACK(sched_policy
, int, INT_MAX
);