1 /* SPDX-License-Identifier: LGPL-2.1+ */
11 #include <stdio_ext.h>
15 #include <sys/mount.h>
16 #include <sys/personality.h>
17 #include <sys/prctl.h>
18 #include <sys/types.h>
22 #if HAVE_VALGRIND_VALGRIND_H
23 #include <valgrind/valgrind.h>
26 #include "alloc-util.h"
27 #include "architecture.h"
37 #include "process-util.h"
38 #include "raw-clone.h"
39 #include "signal-util.h"
40 #include "stat-util.h"
41 #include "string-table.h"
42 #include "string-util.h"
43 #include "terminal-util.h"
44 #include "user-util.h"
47 int get_process_state(pid_t pid
) {
51 _cleanup_free_
char *line
= NULL
;
55 p
= procfs_file_alloca(pid
, "stat");
57 r
= read_one_line_file(p
, &line
);
63 p
= strrchr(line
, ')');
69 if (sscanf(p
, " %c", &state
) != 1)
72 return (unsigned char) state
;
75 int get_process_comm(pid_t pid
, char **ret
) {
76 _cleanup_free_
char *escaped
= NULL
, *comm
= NULL
;
83 escaped
= new(char, TASK_COMM_LEN
);
87 p
= procfs_file_alloca(pid
, "comm");
89 r
= read_one_line_file(p
, &comm
);
95 /* Escape unprintable characters, just in case, but don't grow the string beyond the underlying size */
96 cellescape(escaped
, TASK_COMM_LEN
, comm
);
98 *ret
= TAKE_PTR(escaped
);
102 int get_process_cmdline(pid_t pid
, size_t max_length
, bool comm_fallback
, char **line
) {
103 _cleanup_fclose_
FILE *f
= NULL
;
105 char *k
, *ans
= NULL
;
112 /* Retrieves a process' command line. Replaces unprintable characters while doing so by whitespace (coalescing
113 * multiple sequential ones into one). If max_length is != 0 will return a string of the specified size at most
114 * (the trailing NUL byte does count towards the length here!), abbreviated with a "..." ellipsis. If
115 * comm_fallback is true and the process has no command line set (the case for kernel threads), or has a
116 * command line that resolves to the empty string will return the "comm" name of the process instead.
118 * Returns -ESRCH if the process doesn't exist, and -ENOENT if the process has no command line (and
119 * comm_fallback is false). Returns 0 and sets *line otherwise. */
121 p
= procfs_file_alloca(pid
, "cmdline");
130 (void) __fsetlocking(f
, FSETLOCKING_BYCALLER
);
132 if (max_length
== 1) {
134 /* If there's only room for one byte, return the empty string */
142 } else if (max_length
== 0) {
143 size_t len
= 0, allocated
= 0;
145 while ((c
= getc(f
)) != EOF
) {
147 if (!GREEDY_REALLOC(ans
, allocated
, len
+3)) {
169 bool dotdotdot
= false;
172 ans
= new(char, max_length
);
178 while ((c
= getc(f
)) != EOF
) {
205 if (max_length
<= 4) {
209 k
= ans
+ max_length
- 4;
212 /* Eat up final spaces */
213 while (k
> ans
&& isspace(k
[-1])) {
219 strncpy(k
, "...", left
-1);
225 /* Kernel threads have no argv[] */
227 _cleanup_free_
char *t
= NULL
;
235 h
= get_process_comm(pid
, &t
);
240 ans
= strjoin("[", t
, "]");
246 if (l
+ 3 <= max_length
)
247 ans
= strjoin("[", t
, "]");
248 else if (max_length
<= 6) {
250 ans
= new(char, max_length
);
254 memcpy(ans
, "[...]", max_length
-1);
255 ans
[max_length
-1] = 0;
257 t
[max_length
- 6] = 0;
259 /* Chop off final spaces */
260 delete_trailing_chars(t
, WHITESPACE
);
262 ans
= strjoin("[", t
, "...]");
273 int rename_process(const char name
[]) {
274 static size_t mm_size
= 0;
275 static char *mm
= NULL
;
276 bool truncated
= false;
279 /* This is a like a poor man's setproctitle(). It changes the comm field, argv[0], and also the glibc's
280 * internally used name of the process. For the first one a limit of 16 chars applies; to the second one in
281 * many cases one of 10 (i.e. length of "/sbin/init") — however if we have CAP_SYS_RESOURCES it is unbounded;
282 * to the third one 7 (i.e. the length of "systemd". If you pass a longer string it will likely be
285 * Returns 0 if a name was set but truncated, > 0 if it was set but not truncated. */
288 return -EINVAL
; /* let's not confuse users unnecessarily with an empty name */
290 if (!is_main_thread())
291 return -EPERM
; /* Let's not allow setting the process name from other threads than the main one, as we
292 * cache things without locking, and we make assumptions that PR_SET_NAME sets the
293 * process name that isn't correct on any other threads */
297 /* First step, change the comm field. The main thread's comm is identical to the process comm. This means we
298 * can use PR_SET_NAME, which sets the thread name for the calling thread. */
299 if (prctl(PR_SET_NAME
, name
) < 0)
300 log_debug_errno(errno
, "PR_SET_NAME failed: %m");
301 if (l
>= TASK_COMM_LEN
) /* Linux process names can be 15 chars at max */
304 /* Second step, change glibc's ID of the process name. */
305 if (program_invocation_name
) {
308 k
= strlen(program_invocation_name
);
309 strncpy(program_invocation_name
, name
, k
);
314 /* Third step, completely replace the argv[] array the kernel maintains for us. This requires privileges, but
315 * has the advantage that the argv[] array is exactly what we want it to be, and not filled up with zeros at
316 * the end. This is the best option for changing /proc/self/cmdline. */
318 /* Let's not bother with this if we don't have euid == 0. Strictly speaking we should check for the
319 * CAP_SYS_RESOURCE capability which is independent of the euid. In our own code the capability generally is
320 * present only for euid == 0, hence let's use this as quick bypass check, to avoid calling mmap() if
321 * PR_SET_MM_ARG_{START,END} fails with EPERM later on anyway. After all geteuid() is dead cheap to call, but
324 log_debug("Skipping PR_SET_MM, as we don't have privileges.");
325 else if (mm_size
< l
+1) {
329 nn_size
= PAGE_ALIGN(l
+1);
330 nn
= mmap(NULL
, nn_size
, PROT_READ
|PROT_WRITE
, MAP_PRIVATE
|MAP_ANONYMOUS
, -1, 0);
331 if (nn
== MAP_FAILED
) {
332 log_debug_errno(errno
, "mmap() failed: %m");
336 strncpy(nn
, name
, nn_size
);
338 /* Now, let's tell the kernel about this new memory */
339 if (prctl(PR_SET_MM
, PR_SET_MM_ARG_START
, (unsigned long) nn
, 0, 0) < 0) {
340 /* HACK: prctl() API is kind of dumb on this point. The existing end address may already be
341 * below the desired start address, in which case the kernel may have kicked this back due
342 * to a range-check failure (see linux/kernel/sys.c:validate_prctl_map() to see this in
343 * action). The proper solution would be to have a prctl() API that could set both start+end
344 * simultaneously, or at least let us query the existing address to anticipate this condition
345 * and respond accordingly. For now, we can only guess at the cause of this failure and try
346 * a workaround--which will briefly expand the arg space to something potentially huge before
347 * resizing it to what we want. */
348 log_debug_errno(errno
, "PR_SET_MM_ARG_START failed, attempting PR_SET_MM_ARG_END hack: %m");
350 if (prctl(PR_SET_MM
, PR_SET_MM_ARG_END
, (unsigned long) nn
+ l
+ 1, 0, 0) < 0) {
351 log_debug_errno(errno
, "PR_SET_MM_ARG_END hack failed, proceeding without: %m");
352 (void) munmap(nn
, nn_size
);
356 if (prctl(PR_SET_MM
, PR_SET_MM_ARG_START
, (unsigned long) nn
, 0, 0) < 0) {
357 log_debug_errno(errno
, "PR_SET_MM_ARG_START still failed, proceeding without: %m");
361 /* And update the end pointer to the new end, too. If this fails, we don't really know what
362 * to do, it's pretty unlikely that we can rollback, hence we'll just accept the failure,
364 if (prctl(PR_SET_MM
, PR_SET_MM_ARG_END
, (unsigned long) nn
+ l
+ 1, 0, 0) < 0)
365 log_debug_errno(errno
, "PR_SET_MM_ARG_END failed, proceeding without: %m");
369 (void) munmap(mm
, mm_size
);
374 strncpy(mm
, name
, mm_size
);
376 /* Update the end pointer, continuing regardless of any failure. */
377 if (prctl(PR_SET_MM
, PR_SET_MM_ARG_END
, (unsigned long) mm
+ l
+ 1, 0, 0) < 0)
378 log_debug_errno(errno
, "PR_SET_MM_ARG_END failed, proceeding without: %m");
382 /* Fourth step: in all cases we'll also update the original argv[], so that our own code gets it right too if
383 * it still looks here */
385 if (saved_argc
> 0) {
391 k
= strlen(saved_argv
[0]);
392 strncpy(saved_argv
[0], name
, k
);
397 for (i
= 1; i
< saved_argc
; i
++) {
401 memzero(saved_argv
[i
], strlen(saved_argv
[i
]));
408 int is_kernel_thread(pid_t pid
) {
409 _cleanup_free_
char *line
= NULL
;
410 unsigned long long flags
;
416 if (IN_SET(pid
, 0, 1) || pid
== getpid_cached()) /* pid 1, and we ourselves certainly aren't a kernel thread */
418 if (!pid_is_valid(pid
))
421 p
= procfs_file_alloca(pid
, "stat");
422 r
= read_one_line_file(p
, &line
);
428 /* Skip past the comm field */
429 q
= strrchr(line
, ')');
434 /* Skip 6 fields to reach the flags field */
435 for (i
= 0; i
< 6; i
++) {
436 l
= strspn(q
, WHITESPACE
);
441 l
= strcspn(q
, WHITESPACE
);
447 /* Skip preceding whitespace */
448 l
= strspn(q
, WHITESPACE
);
453 /* Truncate the rest */
454 l
= strcspn(q
, WHITESPACE
);
459 r
= safe_atollu(q
, &flags
);
463 return !!(flags
& PF_KTHREAD
);
466 int get_process_capeff(pid_t pid
, char **capeff
) {
473 p
= procfs_file_alloca(pid
, "status");
475 r
= get_proc_field(p
, "CapEff", WHITESPACE
, capeff
);
482 static int get_process_link_contents(const char *proc_file
, char **name
) {
488 r
= readlink_malloc(proc_file
, name
);
497 int get_process_exe(pid_t pid
, char **name
) {
504 p
= procfs_file_alloca(pid
, "exe");
505 r
= get_process_link_contents(p
, name
);
509 d
= endswith(*name
, " (deleted)");
516 static int get_process_id(pid_t pid
, const char *field
, uid_t
*uid
) {
517 _cleanup_fclose_
FILE *f
= NULL
;
527 p
= procfs_file_alloca(pid
, "status");
535 (void) __fsetlocking(f
, FSETLOCKING_BYCALLER
);
538 _cleanup_free_
char *line
= NULL
;
541 r
= read_line(f
, LONG_LINE_MAX
, &line
);
549 if (startswith(l
, field
)) {
551 l
+= strspn(l
, WHITESPACE
);
553 l
[strcspn(l
, WHITESPACE
)] = 0;
555 return parse_uid(l
, uid
);
562 int get_process_uid(pid_t pid
, uid_t
*uid
) {
564 if (pid
== 0 || pid
== getpid_cached()) {
569 return get_process_id(pid
, "Uid:", uid
);
572 int get_process_gid(pid_t pid
, gid_t
*gid
) {
574 if (pid
== 0 || pid
== getpid_cached()) {
579 assert_cc(sizeof(uid_t
) == sizeof(gid_t
));
580 return get_process_id(pid
, "Gid:", gid
);
583 int get_process_cwd(pid_t pid
, char **cwd
) {
588 p
= procfs_file_alloca(pid
, "cwd");
590 return get_process_link_contents(p
, cwd
);
593 int get_process_root(pid_t pid
, char **root
) {
598 p
= procfs_file_alloca(pid
, "root");
600 return get_process_link_contents(p
, root
);
603 int get_process_environ(pid_t pid
, char **env
) {
604 _cleanup_fclose_
FILE *f
= NULL
;
605 _cleanup_free_
char *outcome
= NULL
;
608 size_t allocated
= 0, sz
= 0;
613 p
= procfs_file_alloca(pid
, "environ");
622 (void) __fsetlocking(f
, FSETLOCKING_BYCALLER
);
624 while ((c
= fgetc(f
)) != EOF
) {
625 if (!GREEDY_REALLOC(outcome
, allocated
, sz
+ 5))
629 outcome
[sz
++] = '\n';
631 sz
+= cescape_char(c
, outcome
+ sz
);
635 outcome
= strdup("");
641 *env
= TAKE_PTR(outcome
);
646 int get_process_ppid(pid_t pid
, pid_t
*_ppid
) {
648 _cleanup_free_
char *line
= NULL
;
655 if (pid
== 0 || pid
== getpid_cached()) {
660 p
= procfs_file_alloca(pid
, "stat");
661 r
= read_one_line_file(p
, &line
);
667 /* Let's skip the pid and comm fields. The latter is enclosed
668 * in () but does not escape any () in its value, so let's
669 * skip over it manually */
671 p
= strrchr(line
, ')');
683 if ((long unsigned) (pid_t
) ppid
!= ppid
)
686 *_ppid
= (pid_t
) ppid
;
691 int wait_for_terminate(pid_t pid
, siginfo_t
*status
) {
702 if (waitid(P_PID
, pid
, status
, WEXITED
) < 0) {
707 return negative_errno();
716 * < 0 : wait_for_terminate() failed to get the state of the
717 * process, the process was terminated by a signal, or
718 * failed for an unknown reason.
719 * >=0 : The process terminated normally, and its exit code is
722 * That is, success is indicated by a return value of zero, and an
723 * error is indicated by a non-zero value.
725 * A warning is emitted if the process terminates abnormally,
726 * and also if it returns non-zero unless check_exit_code is true.
728 int wait_for_terminate_and_check(const char *name
, pid_t pid
, WaitFlags flags
) {
729 _cleanup_free_
char *buffer
= NULL
;
736 r
= get_process_comm(pid
, &buffer
);
738 log_debug_errno(r
, "Failed to acquire process name of " PID_FMT
", ignoring: %m", pid
);
743 prio
= flags
& WAIT_LOG_ABNORMAL
? LOG_ERR
: LOG_DEBUG
;
745 r
= wait_for_terminate(pid
, &status
);
747 return log_full_errno(prio
, r
, "Failed to wait for %s: %m", strna(name
));
749 if (status
.si_code
== CLD_EXITED
) {
750 if (status
.si_status
!= EXIT_SUCCESS
)
751 log_full(flags
& WAIT_LOG_NON_ZERO_EXIT_STATUS
? LOG_ERR
: LOG_DEBUG
,
752 "%s failed with exit status %i.", strna(name
), status
.si_status
);
754 log_debug("%s succeeded.", name
);
756 return status
.si_status
;
758 } else if (IN_SET(status
.si_code
, CLD_KILLED
, CLD_DUMPED
)) {
760 log_full(prio
, "%s terminated by signal %s.", strna(name
), signal_to_string(status
.si_status
));
764 log_full(prio
, "%s failed due to unknown reason.", strna(name
));
771 * < 0 : wait_for_terminate_with_timeout() failed to get the state of the process, the process timed out, the process
772 * was terminated by a signal, or failed for an unknown reason.
774 * >=0 : The process terminated normally with no failures.
776 * Success is indicated by a return value of zero, a timeout is indicated by ETIMEDOUT, and all other child failure
777 * states are indicated by error is indicated by a non-zero value.
779 * This call assumes SIGCHLD has been blocked already, in particular before the child to wait for has been forked off
780 * to remain entirely race-free.
782 int wait_for_terminate_with_timeout(pid_t pid
, usec_t timeout
) {
787 assert_se(sigemptyset(&mask
) == 0);
788 assert_se(sigaddset(&mask
, SIGCHLD
) == 0);
790 /* Drop into a sigtimewait-based timeout. Waiting for the
792 until
= now(CLOCK_MONOTONIC
) + timeout
;
795 siginfo_t status
= {};
798 n
= now(CLOCK_MONOTONIC
);
802 r
= sigtimedwait(&mask
, NULL
, timespec_store(&ts
, until
- n
)) < 0 ? -errno
: 0;
803 /* Assuming we woke due to the child exiting. */
804 if (waitid(P_PID
, pid
, &status
, WEXITED
|WNOHANG
) == 0) {
805 if (status
.si_pid
== pid
) {
806 /* This is the correct child.*/
807 if (status
.si_code
== CLD_EXITED
)
808 return (status
.si_status
== 0) ? 0 : -EPROTO
;
813 /* Not the child, check for errors and proceed appropriately */
817 /* Timed out, child is likely hung. */
820 /* Received a different signal and should retry */
823 /* Return any unexpected errors */
832 void sigkill_wait(pid_t pid
) {
835 if (kill(pid
, SIGKILL
) > 0)
836 (void) wait_for_terminate(pid
, NULL
);
839 void sigkill_waitp(pid_t
*pid
) {
850 void sigterm_wait(pid_t pid
) {
853 if (kill_and_sigcont(pid
, SIGTERM
) > 0)
854 (void) wait_for_terminate(pid
, NULL
);
857 int kill_and_sigcont(pid_t pid
, int sig
) {
860 r
= kill(pid
, sig
) < 0 ? -errno
: 0;
862 /* If this worked, also send SIGCONT, unless we already just sent a SIGCONT, or SIGKILL was sent which isn't
863 * affected by a process being suspended anyway. */
864 if (r
>= 0 && !IN_SET(sig
, SIGCONT
, SIGKILL
))
865 (void) kill(pid
, SIGCONT
);
870 int getenv_for_pid(pid_t pid
, const char *field
, char **ret
) {
871 _cleanup_fclose_
FILE *f
= NULL
;
881 if (pid
== 0 || pid
== getpid_cached()) {
898 path
= procfs_file_alloca(pid
, "environ");
900 f
= fopen(path
, "re");
908 (void) __fsetlocking(f
, FSETLOCKING_BYCALLER
);
916 for (i
= 0; i
< sizeof(line
)-1; i
++) {
920 if (_unlikely_(c
== EOF
)) {
930 if (strneq(line
, field
, l
) && line
[l
] == '=') {
931 value
= strdup(line
+ l
+ 1);
945 bool pid_is_unwaited(pid_t pid
) {
946 /* Checks whether a PID is still valid at all, including a zombie */
951 if (pid
<= 1) /* If we or PID 1 would be dead and have been waited for, this code would not be running */
954 if (pid
== getpid_cached())
957 if (kill(pid
, 0) >= 0)
960 return errno
!= ESRCH
;
963 bool pid_is_alive(pid_t pid
) {
966 /* Checks whether a PID is still valid and not a zombie */
971 if (pid
<= 1) /* If we or PID 1 would be a zombie, this code would not be running */
974 if (pid
== getpid_cached())
977 r
= get_process_state(pid
);
978 if (IN_SET(r
, -ESRCH
, 'Z'))
984 int pid_from_same_root_fs(pid_t pid
) {
990 if (pid
== 0 || pid
== getpid_cached())
993 root
= procfs_file_alloca(pid
, "root");
995 return files_same(root
, "/proc/1/root", 0);
998 bool is_main_thread(void) {
999 static thread_local
int cached
= 0;
1001 if (_unlikely_(cached
== 0))
1002 cached
= getpid_cached() == gettid() ? 1 : -1;
1007 _noreturn_
void freeze(void) {
1011 /* Make sure nobody waits for us on a socket anymore */
1012 close_all_fds(NULL
, 0);
1016 /* Let's not freeze right away, but keep reaping zombies. */
1021 r
= waitid(P_ALL
, 0, &si
, WEXITED
);
1022 if (r
< 0 && errno
!= EINTR
)
1026 /* waitid() failed with an unexpected error, things are really borked. Freeze now! */
1031 bool oom_score_adjust_is_valid(int oa
) {
1032 return oa
>= OOM_SCORE_ADJ_MIN
&& oa
<= OOM_SCORE_ADJ_MAX
;
1035 unsigned long personality_from_string(const char *p
) {
1039 return PERSONALITY_INVALID
;
1041 /* Parse a personality specifier. We use our own identifiers that indicate specific ABIs, rather than just
1042 * hints regarding the register size, since we want to keep things open for multiple locally supported ABIs for
1043 * the same register size. */
1045 architecture
= architecture_from_string(p
);
1046 if (architecture
< 0)
1047 return PERSONALITY_INVALID
;
1049 if (architecture
== native_architecture())
1051 #ifdef SECONDARY_ARCHITECTURE
1052 if (architecture
== SECONDARY_ARCHITECTURE
)
1056 return PERSONALITY_INVALID
;
1059 const char* personality_to_string(unsigned long p
) {
1060 int architecture
= _ARCHITECTURE_INVALID
;
1063 architecture
= native_architecture();
1064 #ifdef SECONDARY_ARCHITECTURE
1065 else if (p
== PER_LINUX32
)
1066 architecture
= SECONDARY_ARCHITECTURE
;
1069 if (architecture
< 0)
1072 return architecture_to_string(architecture
);
1075 int safe_personality(unsigned long p
) {
1078 /* So here's the deal, personality() is weirdly defined by glibc. In some cases it returns a failure via errno,
1079 * and in others as negative return value containing an errno-like value. Let's work around this: this is a
1080 * wrapper that uses errno if it is set, and uses the return value otherwise. And then it sets both errno and
1081 * the return value indicating the same issue, so that we are definitely on the safe side.
1083 * See https://github.com/systemd/systemd/issues/6737 */
1086 ret
= personality(p
);
1097 int opinionated_personality(unsigned long *ret
) {
1100 /* Returns the current personality, or PERSONALITY_INVALID if we can't determine it. This function is a bit
1101 * opinionated though, and ignores all the finer-grained bits and exotic personalities, only distinguishing the
1102 * two most relevant personalities: PER_LINUX and PER_LINUX32. */
1104 current
= safe_personality(PERSONALITY_INVALID
);
1108 if (((unsigned long) current
& 0xffff) == PER_LINUX32
)
1116 void valgrind_summary_hack(void) {
1117 #if HAVE_VALGRIND_VALGRIND_H
1118 if (getpid_cached() == 1 && RUNNING_ON_VALGRIND
) {
1120 pid
= raw_clone(SIGCHLD
);
1122 log_emergency_errno(errno
, "Failed to fork off valgrind helper: %m");
1126 log_info("Spawned valgrind helper as PID "PID_FMT
".", pid
);
1127 (void) wait_for_terminate(pid
, NULL
);
1133 int pid_compare_func(const pid_t
*a
, const pid_t
*b
) {
1134 /* Suitable for usage in qsort() */
1138 int ioprio_parse_priority(const char *s
, int *ret
) {
1144 r
= safe_atoi(s
, &i
);
1148 if (!ioprio_priority_is_valid(i
))
1155 /* The cached PID, possible values:
1157 * == UNSET [0] → cache not initialized yet
1158 * == BUSY [-1] → some thread is initializing it at the moment
1159 * any other → the cached PID
1162 #define CACHED_PID_UNSET ((pid_t) 0)
1163 #define CACHED_PID_BUSY ((pid_t) -1)
1165 static pid_t cached_pid
= CACHED_PID_UNSET
;
1167 void reset_cached_pid(void) {
1168 /* Invoked in the child after a fork(), i.e. at the first moment the PID changed */
1169 cached_pid
= CACHED_PID_UNSET
;
1172 /* We use glibc __register_atfork() + __dso_handle directly here, as they are not included in the glibc
1173 * headers. __register_atfork() is mostly equivalent to pthread_atfork(), but doesn't require us to link against
1174 * libpthread, as it is part of glibc anyway. */
1175 extern int __register_atfork(void (*prepare
) (void), void (*parent
) (void), void (*child
) (void), void *dso_handle
);
1176 extern void* __dso_handle
__attribute__ ((__weak__
));
1178 pid_t
getpid_cached(void) {
1179 static bool installed
= false;
1180 pid_t current_value
;
1182 /* getpid_cached() is much like getpid(), but caches the value in local memory, to avoid having to invoke a
1183 * system call each time. This restores glibc behaviour from before 2.24, when getpid() was unconditionally
1184 * cached. Starting with 2.24 getpid() started to become prohibitively expensive when used for detecting when
1185 * objects were used across fork()s. With this caching the old behaviour is somewhat restored.
1187 * https://bugzilla.redhat.com/show_bug.cgi?id=1443976
1188 * https://sourceware.org/git/gitweb.cgi?p=glibc.git;h=c579f48edba88380635ab98cb612030e3ed8691e
1191 current_value
= __sync_val_compare_and_swap(&cached_pid
, CACHED_PID_UNSET
, CACHED_PID_BUSY
);
1193 switch (current_value
) {
1195 case CACHED_PID_UNSET
: { /* Not initialized yet, then do so now */
1198 new_pid
= raw_getpid();
1201 /* __register_atfork() either returns 0 or -ENOMEM, in its glibc implementation. Since it's
1202 * only half-documented (glibc doesn't document it but LSB does — though only superficially)
1203 * we'll check for errors only in the most generic fashion possible. */
1205 if (__register_atfork(NULL
, NULL
, reset_cached_pid
, __dso_handle
) != 0) {
1206 /* OOM? Let's try again later */
1207 cached_pid
= CACHED_PID_UNSET
;
1214 cached_pid
= new_pid
;
1218 case CACHED_PID_BUSY
: /* Somebody else is currently initializing */
1219 return raw_getpid();
1221 default: /* Properly initialized */
1222 return current_value
;
1226 int must_be_root(void) {
1231 log_error("Need to be root.");
1237 const int except_fds
[],
1238 size_t n_except_fds
,
1242 pid_t original_pid
, pid
;
1243 sigset_t saved_ss
, ss
;
1244 bool block_signals
= false;
1247 /* A wrapper around fork(), that does a couple of important initializations in addition to mere forking. Always
1248 * returns the child's PID in *ret_pid. Returns == 0 in the child, and > 0 in the parent. */
1250 prio
= flags
& FORK_LOG
? LOG_ERR
: LOG_DEBUG
;
1252 original_pid
= getpid_cached();
1254 if (flags
& (FORK_RESET_SIGNALS
|FORK_DEATHSIG
)) {
1256 /* We temporarily block all signals, so that the new child has them blocked initially. This way, we can
1257 * be sure that SIGTERMs are not lost we might send to the child. */
1259 if (sigfillset(&ss
) < 0)
1260 return log_full_errno(prio
, errno
, "Failed to reset signal set: %m");
1262 block_signals
= true;
1264 } else if (flags
& FORK_WAIT
) {
1266 /* Let's block SIGCHLD at least, so that we can safely watch for the child process */
1268 if (sigemptyset(&ss
) < 0)
1269 return log_full_errno(prio
, errno
, "Failed to clear signal set: %m");
1271 if (sigaddset(&ss
, SIGCHLD
) < 0)
1272 return log_full_errno(prio
, errno
, "Failed to add SIGCHLD to signal set: %m");
1274 block_signals
= true;
1278 if (sigprocmask(SIG_SETMASK
, &ss
, &saved_ss
) < 0)
1279 return log_full_errno(prio
, errno
, "Failed to set signal mask: %m");
1281 if (flags
& FORK_NEW_MOUNTNS
)
1282 pid
= raw_clone(SIGCHLD
|CLONE_NEWNS
);
1288 if (block_signals
) /* undo what we did above */
1289 (void) sigprocmask(SIG_SETMASK
, &saved_ss
, NULL
);
1291 return log_full_errno(prio
, r
, "Failed to fork: %m");
1294 /* We are in the parent process */
1296 log_debug("Successfully forked off '%s' as PID " PID_FMT
".", strna(name
), pid
);
1298 if (flags
& FORK_WAIT
) {
1299 r
= wait_for_terminate_and_check(name
, pid
, (flags
& FORK_LOG
? WAIT_LOG
: 0));
1302 if (r
!= EXIT_SUCCESS
) /* exit status > 0 should be treated as failure, too */
1306 if (block_signals
) /* undo what we did above */
1307 (void) sigprocmask(SIG_SETMASK
, &saved_ss
, NULL
);
1315 /* We are in the child process */
1317 if (flags
& FORK_REOPEN_LOG
) {
1318 /* Close the logs if requested, before we log anything. And make sure we reopen it if needed. */
1320 log_set_open_when_needed(true);
1324 r
= rename_process(name
);
1326 log_full_errno(flags
& FORK_LOG
? LOG_WARNING
: LOG_DEBUG
,
1327 r
, "Failed to rename process, ignoring: %m");
1330 if (flags
& FORK_DEATHSIG
)
1331 if (prctl(PR_SET_PDEATHSIG
, SIGTERM
) < 0) {
1332 log_full_errno(prio
, errno
, "Failed to set death signal: %m");
1333 _exit(EXIT_FAILURE
);
1336 if (flags
& FORK_RESET_SIGNALS
) {
1337 r
= reset_all_signal_handlers();
1339 log_full_errno(prio
, r
, "Failed to reset signal handlers: %m");
1340 _exit(EXIT_FAILURE
);
1343 /* This implicitly undoes the signal mask stuff we did before the fork()ing above */
1344 r
= reset_signal_mask();
1346 log_full_errno(prio
, r
, "Failed to reset signal mask: %m");
1347 _exit(EXIT_FAILURE
);
1349 } else if (block_signals
) { /* undo what we did above */
1350 if (sigprocmask(SIG_SETMASK
, &saved_ss
, NULL
) < 0) {
1351 log_full_errno(prio
, errno
, "Failed to restore signal mask: %m");
1352 _exit(EXIT_FAILURE
);
1356 if (flags
& FORK_DEATHSIG
) {
1358 /* Let's see if the parent PID is still the one we started from? If not, then the parent
1359 * already died by the time we set PR_SET_PDEATHSIG, hence let's emulate the effect */
1363 /* Parent is in a differn't PID namespace. */;
1364 else if (ppid
!= original_pid
) {
1365 log_debug("Parent died early, raising SIGTERM.");
1366 (void) raise(SIGTERM
);
1367 _exit(EXIT_FAILURE
);
1371 if (FLAGS_SET(flags
, FORK_NEW_MOUNTNS
| FORK_MOUNTNS_SLAVE
)) {
1373 /* Optionally, make sure we never propagate mounts to the host. */
1375 if (mount(NULL
, "/", NULL
, MS_SLAVE
| MS_REC
, NULL
) < 0) {
1376 log_full_errno(prio
, errno
, "Failed to remount root directory as MS_SLAVE: %m");
1377 _exit(EXIT_FAILURE
);
1381 if (flags
& FORK_CLOSE_ALL_FDS
) {
1382 /* Close the logs here in case it got reopened above, as close_all_fds() would close them for us */
1385 r
= close_all_fds(except_fds
, n_except_fds
);
1387 log_full_errno(prio
, r
, "Failed to close all file descriptors: %m");
1388 _exit(EXIT_FAILURE
);
1392 /* When we were asked to reopen the logs, do so again now */
1393 if (flags
& FORK_REOPEN_LOG
) {
1395 log_set_open_when_needed(false);
1398 if (flags
& FORK_NULL_STDIO
) {
1399 r
= make_null_stdio();
1401 log_full_errno(prio
, r
, "Failed to connect stdin/stdout to /dev/null: %m");
1402 _exit(EXIT_FAILURE
);
1407 *ret_pid
= getpid_cached();
1412 int fork_agent(const char *name
, const int except
[], size_t n_except
, pid_t
*ret_pid
, const char *path
, ...) {
1413 bool stdout_is_tty
, stderr_is_tty
;
1421 /* Spawns a temporary TTY agent, making sure it goes away when we go away */
1423 r
= safe_fork_full(name
, except
, n_except
, FORK_RESET_SIGNALS
|FORK_DEATHSIG
|FORK_CLOSE_ALL_FDS
, ret_pid
);
1431 stdout_is_tty
= isatty(STDOUT_FILENO
);
1432 stderr_is_tty
= isatty(STDERR_FILENO
);
1434 if (!stdout_is_tty
|| !stderr_is_tty
) {
1437 /* Detach from stdout/stderr. and reopen
1438 * /dev/tty for them. This is important to
1439 * ensure that when systemctl is started via
1440 * popen() or a similar call that expects to
1441 * read EOF we actually do generate EOF and
1442 * not delay this indefinitely by because we
1443 * keep an unused copy of stdin around. */
1444 fd
= open("/dev/tty", O_WRONLY
);
1446 log_error_errno(errno
, "Failed to open /dev/tty: %m");
1447 _exit(EXIT_FAILURE
);
1450 if (!stdout_is_tty
&& dup2(fd
, STDOUT_FILENO
) < 0) {
1451 log_error_errno(errno
, "Failed to dup2 /dev/tty: %m");
1452 _exit(EXIT_FAILURE
);
1455 if (!stderr_is_tty
&& dup2(fd
, STDERR_FILENO
) < 0) {
1456 log_error_errno(errno
, "Failed to dup2 /dev/tty: %m");
1457 _exit(EXIT_FAILURE
);
1460 safe_close_above_stdio(fd
);
1463 /* Count arguments */
1465 for (n
= 0; va_arg(ap
, char*); n
++)
1470 l
= newa(char*, n
+ 1);
1472 /* Fill in arguments */
1474 for (i
= 0; i
<= n
; i
++)
1475 l
[i
] = va_arg(ap
, char*);
1479 _exit(EXIT_FAILURE
);
1482 int set_oom_score_adjust(int value
) {
1483 char t
[DECIMAL_STR_MAX(int)];
1485 sprintf(t
, "%i", value
);
1487 return write_string_file("/proc/self/oom_score_adj", t
,
1488 WRITE_STRING_FILE_VERIFY_ON_FAILURE
|WRITE_STRING_FILE_DISABLE_BUFFER
);
1491 static const char *const ioprio_class_table
[] = {
1492 [IOPRIO_CLASS_NONE
] = "none",
1493 [IOPRIO_CLASS_RT
] = "realtime",
1494 [IOPRIO_CLASS_BE
] = "best-effort",
1495 [IOPRIO_CLASS_IDLE
] = "idle"
1498 DEFINE_STRING_TABLE_LOOKUP_WITH_FALLBACK(ioprio_class
, int, IOPRIO_N_CLASSES
);
1500 static const char *const sigchld_code_table
[] = {
1501 [CLD_EXITED
] = "exited",
1502 [CLD_KILLED
] = "killed",
1503 [CLD_DUMPED
] = "dumped",
1504 [CLD_TRAPPED
] = "trapped",
1505 [CLD_STOPPED
] = "stopped",
1506 [CLD_CONTINUED
] = "continued",
1509 DEFINE_STRING_TABLE_LOOKUP(sigchld_code
, int);
1511 static const char* const sched_policy_table
[] = {
1512 [SCHED_OTHER
] = "other",
1513 [SCHED_BATCH
] = "batch",
1514 [SCHED_IDLE
] = "idle",
1515 [SCHED_FIFO
] = "fifo",
1519 DEFINE_STRING_TABLE_LOOKUP_WITH_FALLBACK(sched_policy
, int, INT_MAX
);