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"
36 #include "process-util.h"
37 #include "raw-clone.h"
38 #include "signal-util.h"
39 #include "stat-util.h"
40 #include "string-table.h"
41 #include "string-util.h"
42 #include "terminal-util.h"
43 #include "user-util.h"
46 int get_process_state(pid_t pid
) {
50 _cleanup_free_
char *line
= NULL
;
54 p
= procfs_file_alloca(pid
, "stat");
56 r
= read_one_line_file(p
, &line
);
62 p
= strrchr(line
, ')');
68 if (sscanf(p
, " %c", &state
) != 1)
71 return (unsigned char) state
;
74 int get_process_comm(pid_t pid
, char **ret
) {
75 _cleanup_free_
char *escaped
= NULL
, *comm
= NULL
;
82 escaped
= new(char, TASK_COMM_LEN
);
86 p
= procfs_file_alloca(pid
, "comm");
88 r
= read_one_line_file(p
, &comm
);
94 /* Escape unprintable characters, just in case, but don't grow the string beyond the underlying size */
95 cellescape(escaped
, TASK_COMM_LEN
, comm
);
97 *ret
= TAKE_PTR(escaped
);
101 int get_process_cmdline(pid_t pid
, size_t max_length
, bool comm_fallback
, char **line
) {
102 _cleanup_fclose_
FILE *f
= NULL
;
104 char *k
, *ans
= NULL
;
111 /* Retrieves a process' command line. Replaces unprintable characters while doing so by whitespace (coalescing
112 * multiple sequential ones into one). If max_length is != 0 will return a string of the specified size at most
113 * (the trailing NUL byte does count towards the length here!), abbreviated with a "..." ellipsis. If
114 * comm_fallback is true and the process has no command line set (the case for kernel threads), or has a
115 * command line that resolves to the empty string will return the "comm" name of the process instead.
117 * Returns -ESRCH if the process doesn't exist, and -ENOENT if the process has no command line (and
118 * comm_fallback is false). Returns 0 and sets *line otherwise. */
120 p
= procfs_file_alloca(pid
, "cmdline");
129 (void) __fsetlocking(f
, FSETLOCKING_BYCALLER
);
131 if (max_length
== 1) {
133 /* If there's only room for one byte, return the empty string */
141 } else if (max_length
== 0) {
142 size_t len
= 0, allocated
= 0;
144 while ((c
= getc(f
)) != EOF
) {
146 if (!GREEDY_REALLOC(ans
, allocated
, len
+3)) {
168 bool dotdotdot
= false;
171 ans
= new(char, max_length
);
177 while ((c
= getc(f
)) != EOF
) {
204 if (max_length
<= 4) {
208 k
= ans
+ max_length
- 4;
211 /* Eat up final spaces */
212 while (k
> ans
&& isspace(k
[-1])) {
218 strncpy(k
, "...", left
-1);
224 /* Kernel threads have no argv[] */
226 _cleanup_free_
char *t
= NULL
;
234 h
= get_process_comm(pid
, &t
);
239 ans
= strjoin("[", t
, "]");
245 if (l
+ 3 <= max_length
)
246 ans
= strjoin("[", t
, "]");
247 else if (max_length
<= 6) {
249 ans
= new(char, max_length
);
253 memcpy(ans
, "[...]", max_length
-1);
254 ans
[max_length
-1] = 0;
256 t
[max_length
- 6] = 0;
258 /* Chop off final spaces */
259 delete_trailing_chars(t
, WHITESPACE
);
261 ans
= strjoin("[", t
, "...]");
272 int rename_process(const char name
[]) {
273 static size_t mm_size
= 0;
274 static char *mm
= NULL
;
275 bool truncated
= false;
278 /* This is a like a poor man's setproctitle(). It changes the comm field, argv[0], and also the glibc's
279 * internally used name of the process. For the first one a limit of 16 chars applies; to the second one in
280 * many cases one of 10 (i.e. length of "/sbin/init") — however if we have CAP_SYS_RESOURCES it is unbounded;
281 * to the third one 7 (i.e. the length of "systemd". If you pass a longer string it will likely be
284 * Returns 0 if a name was set but truncated, > 0 if it was set but not truncated. */
287 return -EINVAL
; /* let's not confuse users unnecessarily with an empty name */
289 if (!is_main_thread())
290 return -EPERM
; /* Let's not allow setting the process name from other threads than the main one, as we
291 * cache things without locking, and we make assumptions that PR_SET_NAME sets the
292 * process name that isn't correct on any other threads */
296 /* First step, change the comm field. The main thread's comm is identical to the process comm. This means we
297 * can use PR_SET_NAME, which sets the thread name for the calling thread. */
298 if (prctl(PR_SET_NAME
, name
) < 0)
299 log_debug_errno(errno
, "PR_SET_NAME failed: %m");
300 if (l
>= TASK_COMM_LEN
) /* Linux process names can be 15 chars at max */
303 /* Second step, change glibc's ID of the process name. */
304 if (program_invocation_name
) {
307 k
= strlen(program_invocation_name
);
308 strncpy(program_invocation_name
, name
, k
);
313 /* Third step, completely replace the argv[] array the kernel maintains for us. This requires privileges, but
314 * has the advantage that the argv[] array is exactly what we want it to be, and not filled up with zeros at
315 * the end. This is the best option for changing /proc/self/cmdline. */
317 /* Let's not bother with this if we don't have euid == 0. Strictly speaking we should check for the
318 * CAP_SYS_RESOURCE capability which is independent of the euid. In our own code the capability generally is
319 * present only for euid == 0, hence let's use this as quick bypass check, to avoid calling mmap() if
320 * PR_SET_MM_ARG_{START,END} fails with EPERM later on anyway. After all geteuid() is dead cheap to call, but
323 log_debug("Skipping PR_SET_MM, as we don't have privileges.");
324 else if (mm_size
< l
+1) {
328 nn_size
= PAGE_ALIGN(l
+1);
329 nn
= mmap(NULL
, nn_size
, PROT_READ
|PROT_WRITE
, MAP_PRIVATE
|MAP_ANONYMOUS
, -1, 0);
330 if (nn
== MAP_FAILED
) {
331 log_debug_errno(errno
, "mmap() failed: %m");
335 strncpy(nn
, name
, nn_size
);
337 /* Now, let's tell the kernel about this new memory */
338 if (prctl(PR_SET_MM
, PR_SET_MM_ARG_START
, (unsigned long) nn
, 0, 0) < 0) {
339 /* HACK: prctl() API is kind of dumb on this point. The existing end address may already be
340 * below the desired start address, in which case the kernel may have kicked this back due
341 * to a range-check failure (see linux/kernel/sys.c:validate_prctl_map() to see this in
342 * action). The proper solution would be to have a prctl() API that could set both start+end
343 * simultaneously, or at least let us query the existing address to anticipate this condition
344 * and respond accordingly. For now, we can only guess at the cause of this failure and try
345 * a workaround--which will briefly expand the arg space to something potentially huge before
346 * resizing it to what we want. */
347 log_debug_errno(errno
, "PR_SET_MM_ARG_START failed, attempting PR_SET_MM_ARG_END hack: %m");
349 if (prctl(PR_SET_MM
, PR_SET_MM_ARG_END
, (unsigned long) nn
+ l
+ 1, 0, 0) < 0) {
350 log_debug_errno(errno
, "PR_SET_MM_ARG_END hack failed, proceeding without: %m");
351 (void) munmap(nn
, nn_size
);
355 if (prctl(PR_SET_MM
, PR_SET_MM_ARG_START
, (unsigned long) nn
, 0, 0) < 0) {
356 log_debug_errno(errno
, "PR_SET_MM_ARG_START still failed, proceeding without: %m");
360 /* And update the end pointer to the new end, too. If this fails, we don't really know what
361 * to do, it's pretty unlikely that we can rollback, hence we'll just accept the failure,
363 if (prctl(PR_SET_MM
, PR_SET_MM_ARG_END
, (unsigned long) nn
+ l
+ 1, 0, 0) < 0)
364 log_debug_errno(errno
, "PR_SET_MM_ARG_END failed, proceeding without: %m");
368 (void) munmap(mm
, mm_size
);
373 strncpy(mm
, name
, mm_size
);
375 /* Update the end pointer, continuing regardless of any failure. */
376 if (prctl(PR_SET_MM
, PR_SET_MM_ARG_END
, (unsigned long) mm
+ l
+ 1, 0, 0) < 0)
377 log_debug_errno(errno
, "PR_SET_MM_ARG_END failed, proceeding without: %m");
381 /* Fourth step: in all cases we'll also update the original argv[], so that our own code gets it right too if
382 * it still looks here */
384 if (saved_argc
> 0) {
390 k
= strlen(saved_argv
[0]);
391 strncpy(saved_argv
[0], name
, k
);
396 for (i
= 1; i
< saved_argc
; i
++) {
400 memzero(saved_argv
[i
], strlen(saved_argv
[i
]));
407 int is_kernel_thread(pid_t pid
) {
408 _cleanup_free_
char *line
= NULL
;
409 unsigned long long flags
;
415 if (IN_SET(pid
, 0, 1) || pid
== getpid_cached()) /* pid 1, and we ourselves certainly aren't a kernel thread */
417 if (!pid_is_valid(pid
))
420 p
= procfs_file_alloca(pid
, "stat");
421 r
= read_one_line_file(p
, &line
);
427 /* Skip past the comm field */
428 q
= strrchr(line
, ')');
433 /* Skip 6 fields to reach the flags field */
434 for (i
= 0; i
< 6; i
++) {
435 l
= strspn(q
, WHITESPACE
);
440 l
= strcspn(q
, WHITESPACE
);
446 /* Skip preceding whitespace */
447 l
= strspn(q
, WHITESPACE
);
452 /* Truncate the rest */
453 l
= strcspn(q
, WHITESPACE
);
458 r
= safe_atollu(q
, &flags
);
462 return !!(flags
& PF_KTHREAD
);
465 int get_process_capeff(pid_t pid
, char **capeff
) {
472 p
= procfs_file_alloca(pid
, "status");
474 r
= get_proc_field(p
, "CapEff", WHITESPACE
, capeff
);
481 static int get_process_link_contents(const char *proc_file
, char **name
) {
487 r
= readlink_malloc(proc_file
, name
);
496 int get_process_exe(pid_t pid
, char **name
) {
503 p
= procfs_file_alloca(pid
, "exe");
504 r
= get_process_link_contents(p
, name
);
508 d
= endswith(*name
, " (deleted)");
515 static int get_process_id(pid_t pid
, const char *field
, uid_t
*uid
) {
516 _cleanup_fclose_
FILE *f
= NULL
;
526 p
= procfs_file_alloca(pid
, "status");
534 (void) __fsetlocking(f
, FSETLOCKING_BYCALLER
);
537 _cleanup_free_
char *line
= NULL
;
540 r
= read_line(f
, LONG_LINE_MAX
, &line
);
548 if (startswith(l
, field
)) {
550 l
+= strspn(l
, WHITESPACE
);
552 l
[strcspn(l
, WHITESPACE
)] = 0;
554 return parse_uid(l
, uid
);
561 int get_process_uid(pid_t pid
, uid_t
*uid
) {
563 if (pid
== 0 || pid
== getpid_cached()) {
568 return get_process_id(pid
, "Uid:", uid
);
571 int get_process_gid(pid_t pid
, gid_t
*gid
) {
573 if (pid
== 0 || pid
== getpid_cached()) {
578 assert_cc(sizeof(uid_t
) == sizeof(gid_t
));
579 return get_process_id(pid
, "Gid:", gid
);
582 int get_process_cwd(pid_t pid
, char **cwd
) {
587 p
= procfs_file_alloca(pid
, "cwd");
589 return get_process_link_contents(p
, cwd
);
592 int get_process_root(pid_t pid
, char **root
) {
597 p
= procfs_file_alloca(pid
, "root");
599 return get_process_link_contents(p
, root
);
602 int get_process_environ(pid_t pid
, char **env
) {
603 _cleanup_fclose_
FILE *f
= NULL
;
604 _cleanup_free_
char *outcome
= NULL
;
607 size_t allocated
= 0, sz
= 0;
612 p
= procfs_file_alloca(pid
, "environ");
621 (void) __fsetlocking(f
, FSETLOCKING_BYCALLER
);
623 while ((c
= fgetc(f
)) != EOF
) {
624 if (!GREEDY_REALLOC(outcome
, allocated
, sz
+ 5))
628 outcome
[sz
++] = '\n';
630 sz
+= cescape_char(c
, outcome
+ sz
);
634 outcome
= strdup("");
640 *env
= TAKE_PTR(outcome
);
645 int get_process_ppid(pid_t pid
, pid_t
*_ppid
) {
647 _cleanup_free_
char *line
= NULL
;
654 if (pid
== 0 || pid
== getpid_cached()) {
659 p
= procfs_file_alloca(pid
, "stat");
660 r
= read_one_line_file(p
, &line
);
666 /* Let's skip the pid and comm fields. The latter is enclosed
667 * in () but does not escape any () in its value, so let's
668 * skip over it manually */
670 p
= strrchr(line
, ')');
682 if ((long unsigned) (pid_t
) ppid
!= ppid
)
685 *_ppid
= (pid_t
) ppid
;
690 int wait_for_terminate(pid_t pid
, siginfo_t
*status
) {
701 if (waitid(P_PID
, pid
, status
, WEXITED
) < 0) {
706 return negative_errno();
715 * < 0 : wait_for_terminate() failed to get the state of the
716 * process, the process was terminated by a signal, or
717 * failed for an unknown reason.
718 * >=0 : The process terminated normally, and its exit code is
721 * That is, success is indicated by a return value of zero, and an
722 * error is indicated by a non-zero value.
724 * A warning is emitted if the process terminates abnormally,
725 * and also if it returns non-zero unless check_exit_code is true.
727 int wait_for_terminate_and_check(const char *name
, pid_t pid
, WaitFlags flags
) {
728 _cleanup_free_
char *buffer
= NULL
;
735 r
= get_process_comm(pid
, &buffer
);
737 log_debug_errno(r
, "Failed to acquire process name of " PID_FMT
", ignoring: %m", pid
);
742 prio
= flags
& WAIT_LOG_ABNORMAL
? LOG_ERR
: LOG_DEBUG
;
744 r
= wait_for_terminate(pid
, &status
);
746 return log_full_errno(prio
, r
, "Failed to wait for %s: %m", strna(name
));
748 if (status
.si_code
== CLD_EXITED
) {
749 if (status
.si_status
!= EXIT_SUCCESS
)
750 log_full(flags
& WAIT_LOG_NON_ZERO_EXIT_STATUS
? LOG_ERR
: LOG_DEBUG
,
751 "%s failed with exit status %i.", strna(name
), status
.si_status
);
753 log_debug("%s succeeded.", name
);
755 return status
.si_status
;
757 } else if (IN_SET(status
.si_code
, CLD_KILLED
, CLD_DUMPED
)) {
759 log_full(prio
, "%s terminated by signal %s.", strna(name
), signal_to_string(status
.si_status
));
763 log_full(prio
, "%s failed due to unknown reason.", strna(name
));
770 * < 0 : wait_for_terminate_with_timeout() failed to get the state of the process, the process timed out, the process
771 * was terminated by a signal, or failed for an unknown reason.
773 * >=0 : The process terminated normally with no failures.
775 * Success is indicated by a return value of zero, a timeout is indicated by ETIMEDOUT, and all other child failure
776 * states are indicated by error is indicated by a non-zero value.
778 * This call assumes SIGCHLD has been blocked already, in particular before the child to wait for has been forked off
779 * to remain entirely race-free.
781 int wait_for_terminate_with_timeout(pid_t pid
, usec_t timeout
) {
786 assert_se(sigemptyset(&mask
) == 0);
787 assert_se(sigaddset(&mask
, SIGCHLD
) == 0);
789 /* Drop into a sigtimewait-based timeout. Waiting for the
791 until
= now(CLOCK_MONOTONIC
) + timeout
;
794 siginfo_t status
= {};
797 n
= now(CLOCK_MONOTONIC
);
801 r
= sigtimedwait(&mask
, NULL
, timespec_store(&ts
, until
- n
)) < 0 ? -errno
: 0;
802 /* Assuming we woke due to the child exiting. */
803 if (waitid(P_PID
, pid
, &status
, WEXITED
|WNOHANG
) == 0) {
804 if (status
.si_pid
== pid
) {
805 /* This is the correct child.*/
806 if (status
.si_code
== CLD_EXITED
)
807 return (status
.si_status
== 0) ? 0 : -EPROTO
;
812 /* Not the child, check for errors and proceed appropriately */
816 /* Timed out, child is likely hung. */
819 /* Received a different signal and should retry */
822 /* Return any unexpected errors */
831 void sigkill_wait(pid_t pid
) {
834 if (kill(pid
, SIGKILL
) >= 0)
835 (void) wait_for_terminate(pid
, NULL
);
838 void sigkill_waitp(pid_t
*pid
) {
849 void sigterm_wait(pid_t pid
) {
852 if (kill_and_sigcont(pid
, SIGTERM
) >= 0)
853 (void) wait_for_terminate(pid
, NULL
);
856 int kill_and_sigcont(pid_t pid
, int sig
) {
859 r
= kill(pid
, sig
) < 0 ? -errno
: 0;
861 /* If this worked, also send SIGCONT, unless we already just sent a SIGCONT, or SIGKILL was sent which isn't
862 * affected by a process being suspended anyway. */
863 if (r
>= 0 && !IN_SET(sig
, SIGCONT
, SIGKILL
))
864 (void) kill(pid
, SIGCONT
);
869 int getenv_for_pid(pid_t pid
, const char *field
, char **ret
) {
870 _cleanup_fclose_
FILE *f
= NULL
;
880 if (pid
== 0 || pid
== getpid_cached()) {
897 path
= procfs_file_alloca(pid
, "environ");
899 f
= fopen(path
, "re");
907 (void) __fsetlocking(f
, FSETLOCKING_BYCALLER
);
915 for (i
= 0; i
< sizeof(line
)-1; i
++) {
919 if (_unlikely_(c
== EOF
)) {
929 if (strneq(line
, field
, l
) && line
[l
] == '=') {
930 value
= strdup(line
+ l
+ 1);
944 bool pid_is_unwaited(pid_t pid
) {
945 /* Checks whether a PID is still valid at all, including a zombie */
950 if (pid
<= 1) /* If we or PID 1 would be dead and have been waited for, this code would not be running */
953 if (pid
== getpid_cached())
956 if (kill(pid
, 0) >= 0)
959 return errno
!= ESRCH
;
962 bool pid_is_alive(pid_t pid
) {
965 /* Checks whether a PID is still valid and not a zombie */
970 if (pid
<= 1) /* If we or PID 1 would be a zombie, this code would not be running */
973 if (pid
== getpid_cached())
976 r
= get_process_state(pid
);
977 if (IN_SET(r
, -ESRCH
, 'Z'))
983 int pid_from_same_root_fs(pid_t pid
) {
989 if (pid
== 0 || pid
== getpid_cached())
992 root
= procfs_file_alloca(pid
, "root");
994 return files_same(root
, "/proc/1/root", 0);
997 bool is_main_thread(void) {
998 static thread_local
int cached
= 0;
1000 if (_unlikely_(cached
== 0))
1001 cached
= getpid_cached() == gettid() ? 1 : -1;
1006 _noreturn_
void freeze(void) {
1010 /* Make sure nobody waits for us on a socket anymore */
1011 close_all_fds(NULL
, 0);
1015 /* Let's not freeze right away, but keep reaping zombies. */
1020 r
= waitid(P_ALL
, 0, &si
, WEXITED
);
1021 if (r
< 0 && errno
!= EINTR
)
1025 /* waitid() failed with an unexpected error, things are really borked. Freeze now! */
1030 bool oom_score_adjust_is_valid(int oa
) {
1031 return oa
>= OOM_SCORE_ADJ_MIN
&& oa
<= OOM_SCORE_ADJ_MAX
;
1034 unsigned long personality_from_string(const char *p
) {
1038 return PERSONALITY_INVALID
;
1040 /* Parse a personality specifier. We use our own identifiers that indicate specific ABIs, rather than just
1041 * hints regarding the register size, since we want to keep things open for multiple locally supported ABIs for
1042 * the same register size. */
1044 architecture
= architecture_from_string(p
);
1045 if (architecture
< 0)
1046 return PERSONALITY_INVALID
;
1048 if (architecture
== native_architecture())
1050 #ifdef SECONDARY_ARCHITECTURE
1051 if (architecture
== SECONDARY_ARCHITECTURE
)
1055 return PERSONALITY_INVALID
;
1058 const char* personality_to_string(unsigned long p
) {
1059 int architecture
= _ARCHITECTURE_INVALID
;
1062 architecture
= native_architecture();
1063 #ifdef SECONDARY_ARCHITECTURE
1064 else if (p
== PER_LINUX32
)
1065 architecture
= SECONDARY_ARCHITECTURE
;
1068 if (architecture
< 0)
1071 return architecture_to_string(architecture
);
1074 int safe_personality(unsigned long p
) {
1077 /* So here's the deal, personality() is weirdly defined by glibc. In some cases it returns a failure via errno,
1078 * and in others as negative return value containing an errno-like value. Let's work around this: this is a
1079 * wrapper that uses errno if it is set, and uses the return value otherwise. And then it sets both errno and
1080 * the return value indicating the same issue, so that we are definitely on the safe side.
1082 * See https://github.com/systemd/systemd/issues/6737 */
1085 ret
= personality(p
);
1096 int opinionated_personality(unsigned long *ret
) {
1099 /* Returns the current personality, or PERSONALITY_INVALID if we can't determine it. This function is a bit
1100 * opinionated though, and ignores all the finer-grained bits and exotic personalities, only distinguishing the
1101 * two most relevant personalities: PER_LINUX and PER_LINUX32. */
1103 current
= safe_personality(PERSONALITY_INVALID
);
1107 if (((unsigned long) current
& 0xffff) == PER_LINUX32
)
1115 void valgrind_summary_hack(void) {
1116 #if HAVE_VALGRIND_VALGRIND_H
1117 if (getpid_cached() == 1 && RUNNING_ON_VALGRIND
) {
1119 pid
= raw_clone(SIGCHLD
);
1121 log_emergency_errno(errno
, "Failed to fork off valgrind helper: %m");
1125 log_info("Spawned valgrind helper as PID "PID_FMT
".", pid
);
1126 (void) wait_for_terminate(pid
, NULL
);
1132 int pid_compare_func(const pid_t
*a
, const pid_t
*b
) {
1133 /* Suitable for usage in qsort() */
1137 int ioprio_parse_priority(const char *s
, int *ret
) {
1143 r
= safe_atoi(s
, &i
);
1147 if (!ioprio_priority_is_valid(i
))
1154 /* The cached PID, possible values:
1156 * == UNSET [0] → cache not initialized yet
1157 * == BUSY [-1] → some thread is initializing it at the moment
1158 * any other → the cached PID
1161 #define CACHED_PID_UNSET ((pid_t) 0)
1162 #define CACHED_PID_BUSY ((pid_t) -1)
1164 static pid_t cached_pid
= CACHED_PID_UNSET
;
1166 void reset_cached_pid(void) {
1167 /* Invoked in the child after a fork(), i.e. at the first moment the PID changed */
1168 cached_pid
= CACHED_PID_UNSET
;
1171 /* We use glibc __register_atfork() + __dso_handle directly here, as they are not included in the glibc
1172 * headers. __register_atfork() is mostly equivalent to pthread_atfork(), but doesn't require us to link against
1173 * libpthread, as it is part of glibc anyway. */
1174 extern int __register_atfork(void (*prepare
) (void), void (*parent
) (void), void (*child
) (void), void *dso_handle
);
1175 extern void* __dso_handle
__attribute__ ((__weak__
));
1177 pid_t
getpid_cached(void) {
1178 static bool installed
= false;
1179 pid_t current_value
;
1181 /* getpid_cached() is much like getpid(), but caches the value in local memory, to avoid having to invoke a
1182 * system call each time. This restores glibc behaviour from before 2.24, when getpid() was unconditionally
1183 * cached. Starting with 2.24 getpid() started to become prohibitively expensive when used for detecting when
1184 * objects were used across fork()s. With this caching the old behaviour is somewhat restored.
1186 * https://bugzilla.redhat.com/show_bug.cgi?id=1443976
1187 * https://sourceware.org/git/gitweb.cgi?p=glibc.git;h=c579f48edba88380635ab98cb612030e3ed8691e
1190 current_value
= __sync_val_compare_and_swap(&cached_pid
, CACHED_PID_UNSET
, CACHED_PID_BUSY
);
1192 switch (current_value
) {
1194 case CACHED_PID_UNSET
: { /* Not initialized yet, then do so now */
1197 new_pid
= raw_getpid();
1200 /* __register_atfork() either returns 0 or -ENOMEM, in its glibc implementation. Since it's
1201 * only half-documented (glibc doesn't document it but LSB does — though only superficially)
1202 * we'll check for errors only in the most generic fashion possible. */
1204 if (__register_atfork(NULL
, NULL
, reset_cached_pid
, __dso_handle
) != 0) {
1205 /* OOM? Let's try again later */
1206 cached_pid
= CACHED_PID_UNSET
;
1213 cached_pid
= new_pid
;
1217 case CACHED_PID_BUSY
: /* Somebody else is currently initializing */
1218 return raw_getpid();
1220 default: /* Properly initialized */
1221 return current_value
;
1225 int must_be_root(void) {
1230 return log_error_errno(SYNTHETIC_ERRNO(EPERM
), "Need to be root.");
1235 const int except_fds
[],
1236 size_t n_except_fds
,
1240 pid_t original_pid
, pid
;
1241 sigset_t saved_ss
, ss
;
1242 bool block_signals
= false;
1245 /* A wrapper around fork(), that does a couple of important initializations in addition to mere forking. Always
1246 * returns the child's PID in *ret_pid. Returns == 0 in the child, and > 0 in the parent. */
1248 prio
= flags
& FORK_LOG
? LOG_ERR
: LOG_DEBUG
;
1250 original_pid
= getpid_cached();
1252 if (flags
& (FORK_RESET_SIGNALS
|FORK_DEATHSIG
)) {
1254 /* We temporarily block all signals, so that the new child has them blocked initially. This way, we can
1255 * be sure that SIGTERMs are not lost we might send to the child. */
1257 if (sigfillset(&ss
) < 0)
1258 return log_full_errno(prio
, errno
, "Failed to reset signal set: %m");
1260 block_signals
= true;
1262 } else if (flags
& FORK_WAIT
) {
1264 /* Let's block SIGCHLD at least, so that we can safely watch for the child process */
1266 if (sigemptyset(&ss
) < 0)
1267 return log_full_errno(prio
, errno
, "Failed to clear signal set: %m");
1269 if (sigaddset(&ss
, SIGCHLD
) < 0)
1270 return log_full_errno(prio
, errno
, "Failed to add SIGCHLD to signal set: %m");
1272 block_signals
= true;
1276 if (sigprocmask(SIG_SETMASK
, &ss
, &saved_ss
) < 0)
1277 return log_full_errno(prio
, errno
, "Failed to set signal mask: %m");
1279 if (flags
& FORK_NEW_MOUNTNS
)
1280 pid
= raw_clone(SIGCHLD
|CLONE_NEWNS
);
1286 if (block_signals
) /* undo what we did above */
1287 (void) sigprocmask(SIG_SETMASK
, &saved_ss
, NULL
);
1289 return log_full_errno(prio
, r
, "Failed to fork: %m");
1292 /* We are in the parent process */
1294 log_debug("Successfully forked off '%s' as PID " PID_FMT
".", strna(name
), pid
);
1296 if (flags
& FORK_WAIT
) {
1297 r
= wait_for_terminate_and_check(name
, pid
, (flags
& FORK_LOG
? WAIT_LOG
: 0));
1300 if (r
!= EXIT_SUCCESS
) /* exit status > 0 should be treated as failure, too */
1304 if (block_signals
) /* undo what we did above */
1305 (void) sigprocmask(SIG_SETMASK
, &saved_ss
, NULL
);
1313 /* We are in the child process */
1315 if (flags
& FORK_REOPEN_LOG
) {
1316 /* Close the logs if requested, before we log anything. And make sure we reopen it if needed. */
1318 log_set_open_when_needed(true);
1322 r
= rename_process(name
);
1324 log_full_errno(flags
& FORK_LOG
? LOG_WARNING
: LOG_DEBUG
,
1325 r
, "Failed to rename process, ignoring: %m");
1328 if (flags
& FORK_DEATHSIG
)
1329 if (prctl(PR_SET_PDEATHSIG
, SIGTERM
) < 0) {
1330 log_full_errno(prio
, errno
, "Failed to set death signal: %m");
1331 _exit(EXIT_FAILURE
);
1334 if (flags
& FORK_RESET_SIGNALS
) {
1335 r
= reset_all_signal_handlers();
1337 log_full_errno(prio
, r
, "Failed to reset signal handlers: %m");
1338 _exit(EXIT_FAILURE
);
1341 /* This implicitly undoes the signal mask stuff we did before the fork()ing above */
1342 r
= reset_signal_mask();
1344 log_full_errno(prio
, r
, "Failed to reset signal mask: %m");
1345 _exit(EXIT_FAILURE
);
1347 } else if (block_signals
) { /* undo what we did above */
1348 if (sigprocmask(SIG_SETMASK
, &saved_ss
, NULL
) < 0) {
1349 log_full_errno(prio
, errno
, "Failed to restore signal mask: %m");
1350 _exit(EXIT_FAILURE
);
1354 if (flags
& FORK_DEATHSIG
) {
1356 /* Let's see if the parent PID is still the one we started from? If not, then the parent
1357 * already died by the time we set PR_SET_PDEATHSIG, hence let's emulate the effect */
1361 /* Parent is in a differn't PID namespace. */;
1362 else if (ppid
!= original_pid
) {
1363 log_debug("Parent died early, raising SIGTERM.");
1364 (void) raise(SIGTERM
);
1365 _exit(EXIT_FAILURE
);
1369 if (FLAGS_SET(flags
, FORK_NEW_MOUNTNS
| FORK_MOUNTNS_SLAVE
)) {
1371 /* Optionally, make sure we never propagate mounts to the host. */
1373 if (mount(NULL
, "/", NULL
, MS_SLAVE
| MS_REC
, NULL
) < 0) {
1374 log_full_errno(prio
, errno
, "Failed to remount root directory as MS_SLAVE: %m");
1375 _exit(EXIT_FAILURE
);
1379 if (flags
& FORK_CLOSE_ALL_FDS
) {
1380 /* Close the logs here in case it got reopened above, as close_all_fds() would close them for us */
1383 r
= close_all_fds(except_fds
, n_except_fds
);
1385 log_full_errno(prio
, r
, "Failed to close all file descriptors: %m");
1386 _exit(EXIT_FAILURE
);
1390 /* When we were asked to reopen the logs, do so again now */
1391 if (flags
& FORK_REOPEN_LOG
) {
1393 log_set_open_when_needed(false);
1396 if (flags
& FORK_NULL_STDIO
) {
1397 r
= make_null_stdio();
1399 log_full_errno(prio
, r
, "Failed to connect stdin/stdout to /dev/null: %m");
1400 _exit(EXIT_FAILURE
);
1405 *ret_pid
= getpid_cached();
1411 const char *outer_name
,
1412 const char *inner_name
,
1413 const int except_fds
[],
1414 size_t n_except_fds
,
1425 /* This is much like safe_fork(), but forks twice, and joins the specified namespaces in the middle
1426 * process. This ensures that we are fully a member of the destination namespace, with pidns an all, so that
1427 * /proc/self/fd works correctly. */
1429 r
= safe_fork_full(outer_name
, except_fds
, n_except_fds
, (flags
|FORK_DEATHSIG
) & ~(FORK_REOPEN_LOG
|FORK_NEW_MOUNTNS
|FORK_MOUNTNS_SLAVE
), ret_pid
);
1437 r
= namespace_enter(pidns_fd
, mntns_fd
, netns_fd
, userns_fd
, root_fd
);
1439 log_full_errno(FLAGS_SET(flags
, FORK_LOG
) ? LOG_ERR
: LOG_DEBUG
, r
, "Failed to join namespace: %m");
1440 _exit(EXIT_FAILURE
);
1443 /* We mask a few flags here that either make no sense for the grandchild, or that we don't have to do again */
1444 r
= safe_fork_full(inner_name
, except_fds
, n_except_fds
, flags
& ~(FORK_WAIT
|FORK_RESET_SIGNALS
|FORK_CLOSE_ALL_FDS
|FORK_NULL_STDIO
), &pid
);
1446 _exit(EXIT_FAILURE
);
1454 r
= wait_for_terminate_and_check(inner_name
, pid
, FLAGS_SET(flags
, FORK_LOG
) ? WAIT_LOG
: 0);
1456 _exit(EXIT_FAILURE
);
1464 int fork_agent(const char *name
, const int except
[], size_t n_except
, pid_t
*ret_pid
, const char *path
, ...) {
1465 bool stdout_is_tty
, stderr_is_tty
;
1473 /* Spawns a temporary TTY agent, making sure it goes away when we go away */
1475 r
= safe_fork_full(name
, except
, n_except
, FORK_RESET_SIGNALS
|FORK_DEATHSIG
|FORK_CLOSE_ALL_FDS
, ret_pid
);
1483 stdout_is_tty
= isatty(STDOUT_FILENO
);
1484 stderr_is_tty
= isatty(STDERR_FILENO
);
1486 if (!stdout_is_tty
|| !stderr_is_tty
) {
1489 /* Detach from stdout/stderr. and reopen
1490 * /dev/tty for them. This is important to
1491 * ensure that when systemctl is started via
1492 * popen() or a similar call that expects to
1493 * read EOF we actually do generate EOF and
1494 * not delay this indefinitely by because we
1495 * keep an unused copy of stdin around. */
1496 fd
= open("/dev/tty", O_WRONLY
);
1498 log_error_errno(errno
, "Failed to open /dev/tty: %m");
1499 _exit(EXIT_FAILURE
);
1502 if (!stdout_is_tty
&& dup2(fd
, STDOUT_FILENO
) < 0) {
1503 log_error_errno(errno
, "Failed to dup2 /dev/tty: %m");
1504 _exit(EXIT_FAILURE
);
1507 if (!stderr_is_tty
&& dup2(fd
, STDERR_FILENO
) < 0) {
1508 log_error_errno(errno
, "Failed to dup2 /dev/tty: %m");
1509 _exit(EXIT_FAILURE
);
1512 safe_close_above_stdio(fd
);
1515 /* Count arguments */
1517 for (n
= 0; va_arg(ap
, char*); n
++)
1522 l
= newa(char*, n
+ 1);
1524 /* Fill in arguments */
1526 for (i
= 0; i
<= n
; i
++)
1527 l
[i
] = va_arg(ap
, char*);
1531 _exit(EXIT_FAILURE
);
1534 int set_oom_score_adjust(int value
) {
1535 char t
[DECIMAL_STR_MAX(int)];
1537 sprintf(t
, "%i", value
);
1539 return write_string_file("/proc/self/oom_score_adj", t
,
1540 WRITE_STRING_FILE_VERIFY_ON_FAILURE
|WRITE_STRING_FILE_DISABLE_BUFFER
);
1543 static const char *const ioprio_class_table
[] = {
1544 [IOPRIO_CLASS_NONE
] = "none",
1545 [IOPRIO_CLASS_RT
] = "realtime",
1546 [IOPRIO_CLASS_BE
] = "best-effort",
1547 [IOPRIO_CLASS_IDLE
] = "idle"
1550 DEFINE_STRING_TABLE_LOOKUP_WITH_FALLBACK(ioprio_class
, int, IOPRIO_N_CLASSES
);
1552 static const char *const sigchld_code_table
[] = {
1553 [CLD_EXITED
] = "exited",
1554 [CLD_KILLED
] = "killed",
1555 [CLD_DUMPED
] = "dumped",
1556 [CLD_TRAPPED
] = "trapped",
1557 [CLD_STOPPED
] = "stopped",
1558 [CLD_CONTINUED
] = "continued",
1561 DEFINE_STRING_TABLE_LOOKUP(sigchld_code
, int);
1563 static const char* const sched_policy_table
[] = {
1564 [SCHED_OTHER
] = "other",
1565 [SCHED_BATCH
] = "batch",
1566 [SCHED_IDLE
] = "idle",
1567 [SCHED_FIFO
] = "fifo",
1571 DEFINE_STRING_TABLE_LOOKUP_WITH_FALLBACK(sched_policy
, int, INT_MAX
);