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/mount.h>
21 #include <sys/personality.h>
22 #include <sys/prctl.h>
23 #include <sys/types.h>
27 #if HAVE_VALGRIND_VALGRIND_H
28 #include <valgrind/valgrind.h>
31 #include "alloc-util.h"
32 //#include "architecture.h"
41 #include "process-util.h"
42 #include "raw-clone.h"
43 #include "signal-util.h"
44 //#include "stat-util.h"
45 #include "string-table.h"
46 #include "string-util.h"
47 //#include "terminal-util.h"
48 #include "user-util.h"
51 int get_process_state(pid_t pid) {
55 _cleanup_free_ char *line = NULL;
59 p = procfs_file_alloca(pid, "stat");
61 r = read_one_line_file(p, &line);
67 p = strrchr(line, ')');
73 if (sscanf(p, " %c", &state) != 1)
76 return (unsigned char) state;
79 int get_process_comm(pid_t pid, char **ret) {
80 _cleanup_free_ char *escaped = NULL, *comm = NULL;
87 escaped = new(char, TASK_COMM_LEN);
91 p = procfs_file_alloca(pid, "comm");
93 r = read_one_line_file(p, &comm);
99 /* Escape unprintable characters, just in case, but don't grow the string beyond the underlying size */
100 cellescape(escaped, TASK_COMM_LEN, comm);
102 *ret = TAKE_PTR(escaped);
106 int get_process_cmdline(pid_t pid, size_t max_length, bool comm_fallback, char **line) {
107 _cleanup_fclose_ FILE *f = NULL;
109 char *k, *ans = NULL;
116 /* Retrieves a process' command line. Replaces unprintable characters while doing so by whitespace (coalescing
117 * multiple sequential ones into one). If max_length is != 0 will return a string of the specified size at most
118 * (the trailing NUL byte does count towards the length here!), abbreviated with a "..." ellipsis. If
119 * comm_fallback is true and the process has no command line set (the case for kernel threads), or has a
120 * command line that resolves to the empty string will return the "comm" name of the process instead.
122 * Returns -ESRCH if the process doesn't exist, and -ENOENT if the process has no command line (and
123 * comm_fallback is false). Returns 0 and sets *line otherwise. */
125 p = procfs_file_alloca(pid, "cmdline");
134 (void) __fsetlocking(f, FSETLOCKING_BYCALLER);
136 if (max_length == 1) {
138 /* If there's only room for one byte, return the empty string */
146 } else if (max_length == 0) {
147 size_t len = 0, allocated = 0;
149 while ((c = getc(f)) != EOF) {
151 if (!GREEDY_REALLOC(ans, allocated, len+3)) {
173 bool dotdotdot = false;
176 ans = new(char, max_length);
182 while ((c = getc(f)) != EOF) {
209 if (max_length <= 4) {
213 k = ans + max_length - 4;
216 /* Eat up final spaces */
217 while (k > ans && isspace(k[-1])) {
223 strncpy(k, "...", left-1);
229 /* Kernel threads have no argv[] */
231 _cleanup_free_ char *t = NULL;
239 h = get_process_comm(pid, &t);
244 ans = strjoin("[", t, "]");
250 if (l + 3 <= max_length)
251 ans = strjoin("[", t, "]");
252 else if (max_length <= 6) {
254 ans = new(char, max_length);
258 memcpy(ans, "[...]", max_length-1);
259 ans[max_length-1] = 0;
263 t[max_length - 6] = 0;
265 /* Chop off final spaces */
267 while (e > t && isspace(e[-1]))
271 ans = strjoin("[", t, "...]");
282 int rename_process(const char name[]) {
283 static size_t mm_size = 0;
284 static char *mm = NULL;
285 bool truncated = false;
288 /* This is a like a poor man's setproctitle(). It changes the comm field, argv[0], and also the glibc's
289 * internally used name of the process. For the first one a limit of 16 chars applies; to the second one in
290 * many cases one of 10 (i.e. length of "/sbin/init") — however if we have CAP_SYS_RESOURCES it is unbounded;
291 * to the third one 7 (i.e. the length of "systemd". If you pass a longer string it will likely be
294 * Returns 0 if a name was set but truncated, > 0 if it was set but not truncated. */
297 return -EINVAL; /* let's not confuse users unnecessarily with an empty name */
299 if (!is_main_thread())
300 return -EPERM; /* Let's not allow setting the process name from other threads than the main one, as we
301 * cache things without locking, and we make assumptions that PR_SET_NAME sets the
302 * process name that isn't correct on any other threads */
306 /* First step, change the comm field. The main thread's comm is identical to the process comm. This means we
307 * can use PR_SET_NAME, which sets the thread name for the calling thread. */
308 if (prctl(PR_SET_NAME, name) < 0)
309 log_debug_errno(errno, "PR_SET_NAME failed: %m");
310 if (l >= TASK_COMM_LEN) /* Linux process names can be 15 chars at max */
313 /* Second step, change glibc's ID of the process name. */
314 if (program_invocation_name) {
317 k = strlen(program_invocation_name);
318 strncpy(program_invocation_name, name, k);
323 /* Third step, completely replace the argv[] array the kernel maintains for us. This requires privileges, but
324 * has the advantage that the argv[] array is exactly what we want it to be, and not filled up with zeros at
325 * the end. This is the best option for changing /proc/self/cmdline. */
327 /* Let's not bother with this if we don't have euid == 0. Strictly speaking we should check for the
328 * CAP_SYS_RESOURCE capability which is independent of the euid. In our own code the capability generally is
329 * present only for euid == 0, hence let's use this as quick bypass check, to avoid calling mmap() if
330 * PR_SET_MM_ARG_{START,END} fails with EPERM later on anyway. After all geteuid() is dead cheap to call, but
333 log_debug("Skipping PR_SET_MM, as we don't have privileges.");
334 else if (mm_size < l+1) {
338 nn_size = PAGE_ALIGN(l+1);
339 nn = mmap(NULL, nn_size, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
340 if (nn == MAP_FAILED) {
341 log_debug_errno(errno, "mmap() failed: %m");
345 strncpy(nn, name, nn_size);
347 /* Now, let's tell the kernel about this new memory */
348 if (prctl(PR_SET_MM, PR_SET_MM_ARG_START, (unsigned long) nn, 0, 0) < 0) {
349 log_debug_errno(errno, "PR_SET_MM_ARG_START failed, proceeding without: %m");
350 (void) munmap(nn, nn_size);
354 /* And update the end pointer to the new end, too. If this fails, we don't really know what to do, it's
355 * pretty unlikely that we can rollback, hence we'll just accept the failure, and continue. */
356 if (prctl(PR_SET_MM, PR_SET_MM_ARG_END, (unsigned long) nn + l + 1, 0, 0) < 0)
357 log_debug_errno(errno, "PR_SET_MM_ARG_END failed, proceeding without: %m");
360 (void) munmap(mm, mm_size);
365 strncpy(mm, name, mm_size);
367 /* Update the end pointer, continuing regardless of any failure. */
368 if (prctl(PR_SET_MM, PR_SET_MM_ARG_END, (unsigned long) mm + l + 1, 0, 0) < 0)
369 log_debug_errno(errno, "PR_SET_MM_ARG_END failed, proceeding without: %m");
373 /* Fourth step: in all cases we'll also update the original argv[], so that our own code gets it right too if
374 * it still looks here */
376 if (saved_argc > 0) {
382 k = strlen(saved_argv[0]);
383 strncpy(saved_argv[0], name, k);
388 for (i = 1; i < saved_argc; i++) {
392 memzero(saved_argv[i], strlen(saved_argv[i]));
399 int is_kernel_thread(pid_t pid) {
400 _cleanup_free_ char *line = NULL;
401 unsigned long long flags;
407 if (IN_SET(pid, 0, 1) || pid == getpid_cached()) /* pid 1, and we ourselves certainly aren't a kernel thread */
409 if (!pid_is_valid(pid))
412 p = procfs_file_alloca(pid, "stat");
413 r = read_one_line_file(p, &line);
419 /* Skip past the comm field */
420 q = strrchr(line, ')');
425 /* Skip 6 fields to reach the flags field */
426 for (i = 0; i < 6; i++) {
427 l = strspn(q, WHITESPACE);
432 l = strcspn(q, WHITESPACE);
438 /* Skip preceeding whitespace */
439 l = strspn(q, WHITESPACE);
444 /* Truncate the rest */
445 l = strcspn(q, WHITESPACE);
450 r = safe_atollu(q, &flags);
454 return !!(flags & PF_KTHREAD);
457 #if 0 /// UNNEEDED by elogind
458 int get_process_capeff(pid_t pid, char **capeff) {
465 p = procfs_file_alloca(pid, "status");
467 r = get_proc_field(p, "CapEff", WHITESPACE, capeff);
475 static int get_process_link_contents(const char *proc_file, char **name) {
481 r = readlink_malloc(proc_file, name);
490 int get_process_exe(pid_t pid, char **name) {
497 p = procfs_file_alloca(pid, "exe");
498 r = get_process_link_contents(p, name);
502 d = endswith(*name, " (deleted)");
509 #if 0 /// UNNEEDED by elogind
510 static int get_process_id(pid_t pid, const char *field, uid_t *uid) {
511 _cleanup_fclose_ FILE *f = NULL;
521 p = procfs_file_alloca(pid, "status");
529 (void) __fsetlocking(f, FSETLOCKING_BYCALLER);
531 FOREACH_LINE(line, f, return -errno) {
536 if (startswith(l, field)) {
538 l += strspn(l, WHITESPACE);
540 l[strcspn(l, WHITESPACE)] = 0;
542 return parse_uid(l, uid);
549 int get_process_uid(pid_t pid, uid_t *uid) {
551 if (pid == 0 || pid == getpid_cached()) {
556 return get_process_id(pid, "Uid:", uid);
559 int get_process_gid(pid_t pid, gid_t *gid) {
561 if (pid == 0 || pid == getpid_cached()) {
566 assert_cc(sizeof(uid_t) == sizeof(gid_t));
567 return get_process_id(pid, "Gid:", gid);
570 int get_process_cwd(pid_t pid, char **cwd) {
575 p = procfs_file_alloca(pid, "cwd");
577 return get_process_link_contents(p, cwd);
580 int get_process_root(pid_t pid, char **root) {
585 p = procfs_file_alloca(pid, "root");
587 return get_process_link_contents(p, root);
590 int get_process_environ(pid_t pid, char **env) {
591 _cleanup_fclose_ FILE *f = NULL;
592 _cleanup_free_ char *outcome = NULL;
595 size_t allocated = 0, sz = 0;
600 p = procfs_file_alloca(pid, "environ");
609 (void) __fsetlocking(f, FSETLOCKING_BYCALLER);
611 while ((c = fgetc(f)) != EOF) {
612 if (!GREEDY_REALLOC(outcome, allocated, sz + 5))
616 outcome[sz++] = '\n';
618 sz += cescape_char(c, outcome + sz);
622 outcome = strdup("");
628 *env = TAKE_PTR(outcome);
633 int get_process_ppid(pid_t pid, pid_t *_ppid) {
635 _cleanup_free_ char *line = NULL;
642 if (pid == 0 || pid == getpid_cached()) {
647 p = procfs_file_alloca(pid, "stat");
648 r = read_one_line_file(p, &line);
654 /* Let's skip the pid and comm fields. The latter is enclosed
655 * in () but does not escape any () in its value, so let's
656 * skip over it manually */
658 p = strrchr(line, ')');
670 if ((long unsigned) (pid_t) ppid != ppid)
673 *_ppid = (pid_t) ppid;
679 int wait_for_terminate(pid_t pid, siginfo_t *status) {
690 if (waitid(P_PID, pid, status, WEXITED) < 0) {
695 return negative_errno();
704 * < 0 : wait_for_terminate() failed to get the state of the
705 * process, the process was terminated by a signal, or
706 * failed for an unknown reason.
707 * >=0 : The process terminated normally, and its exit code is
710 * That is, success is indicated by a return value of zero, and an
711 * error is indicated by a non-zero value.
713 * A warning is emitted if the process terminates abnormally,
714 * and also if it returns non-zero unless check_exit_code is true.
716 int wait_for_terminate_and_check(const char *name, pid_t pid, WaitFlags flags) {
717 _cleanup_free_ char *buffer = NULL;
724 r = get_process_comm(pid, &buffer);
726 log_debug_errno(r, "Failed to acquire process name of " PID_FMT ", ignoring: %m", pid);
731 prio = flags & WAIT_LOG_ABNORMAL ? LOG_ERR : LOG_DEBUG;
733 r = wait_for_terminate(pid, &status);
735 return log_full_errno(prio, r, "Failed to wait for %s: %m", strna(name));
737 if (status.si_code == CLD_EXITED) {
738 if (status.si_status != EXIT_SUCCESS)
739 log_full(flags & WAIT_LOG_NON_ZERO_EXIT_STATUS ? LOG_ERR : LOG_DEBUG,
740 "%s failed with exit status %i.", strna(name), status.si_status);
742 log_debug("%s succeeded.", name);
744 return status.si_status;
746 } else if (IN_SET(status.si_code, CLD_KILLED, CLD_DUMPED)) {
748 log_full(prio, "%s terminated by signal %s.", strna(name), signal_to_string(status.si_status));
752 log_full(prio, "%s failed due to unknown reason.", strna(name));
759 * < 0 : wait_for_terminate_with_timeout() failed to get the state of the process, the process timed out, the process
760 * was terminated by a signal, or failed for an unknown reason.
762 * >=0 : The process terminated normally with no failures.
764 * Success is indicated by a return value of zero, a timeout is indicated by ETIMEDOUT, and all other child failure
765 * states are indicated by error is indicated by a non-zero value.
767 * This call assumes SIGCHLD has been blocked already, in particular before the child to wait for has been forked off
768 * to remain entirely race-free.
770 int wait_for_terminate_with_timeout(pid_t pid, usec_t timeout) {
775 assert_se(sigemptyset(&mask) == 0);
776 assert_se(sigaddset(&mask, SIGCHLD) == 0);
778 /* Drop into a sigtimewait-based timeout. Waiting for the
780 until = now(CLOCK_MONOTONIC) + timeout;
783 siginfo_t status = {};
786 n = now(CLOCK_MONOTONIC);
790 r = sigtimedwait(&mask, NULL, timespec_store(&ts, until - n)) < 0 ? -errno : 0;
791 /* Assuming we woke due to the child exiting. */
792 if (waitid(P_PID, pid, &status, WEXITED|WNOHANG) == 0) {
793 if (status.si_pid == pid) {
794 /* This is the correct child.*/
795 if (status.si_code == CLD_EXITED)
796 return (status.si_status == 0) ? 0 : -EPROTO;
801 /* Not the child, check for errors and proceed appropriately */
805 /* Timed out, child is likely hung. */
808 /* Received a different signal and should retry */
811 /* Return any unexpected errors */
820 #if 0 /// UNNEEDED by elogind
821 void sigkill_wait(pid_t pid) {
824 if (kill(pid, SIGKILL) > 0)
825 (void) wait_for_terminate(pid, NULL);
828 void sigkill_waitp(pid_t *pid) {
840 void sigterm_wait(pid_t pid) {
843 if (kill_and_sigcont(pid, SIGTERM) > 0)
844 (void) wait_for_terminate(pid, NULL);
847 int kill_and_sigcont(pid_t pid, int sig) {
850 r = kill(pid, sig) < 0 ? -errno : 0;
852 /* If this worked, also send SIGCONT, unless we already just sent a SIGCONT, or SIGKILL was sent which isn't
853 * affected by a process being suspended anyway. */
854 if (r >= 0 && !IN_SET(sig, SIGCONT, SIGKILL))
855 (void) kill(pid, SIGCONT);
860 int getenv_for_pid(pid_t pid, const char *field, char **ret) {
861 _cleanup_fclose_ FILE *f = NULL;
871 if (pid == 0 || pid == getpid_cached()) {
888 path = procfs_file_alloca(pid, "environ");
890 f = fopen(path, "re");
898 (void) __fsetlocking(f, FSETLOCKING_BYCALLER);
906 for (i = 0; i < sizeof(line)-1; i++) {
910 if (_unlikely_(c == EOF)) {
920 if (strneq(line, field, l) && line[l] == '=') {
921 value = strdup(line + l + 1);
935 bool pid_is_unwaited(pid_t pid) {
936 /* Checks whether a PID is still valid at all, including a zombie */
941 if (pid <= 1) /* If we or PID 1 would be dead and have been waited for, this code would not be running */
944 if (pid == getpid_cached())
947 if (kill(pid, 0) >= 0)
950 return errno != ESRCH;
953 bool pid_is_alive(pid_t pid) {
956 /* Checks whether a PID is still valid and not a zombie */
961 if (pid <= 1) /* If we or PID 1 would be a zombie, this code would not be running */
964 if (pid == getpid_cached())
967 r = get_process_state(pid);
968 if (IN_SET(r, -ESRCH, 'Z'))
974 #if 0 /// UNNEEDED by elogind
975 int pid_from_same_root_fs(pid_t pid) {
981 if (pid == 0 || pid == getpid_cached())
984 root = procfs_file_alloca(pid, "root");
986 return files_same(root, "/proc/1/root", 0);
990 bool is_main_thread(void) {
991 static thread_local int cached = 0;
993 if (_unlikely_(cached == 0))
994 cached = getpid_cached() == gettid() ? 1 : -1;
999 #if 0 /// UNNEEDED by elogind
1000 _noreturn_ void freeze(void) {
1004 /* Make sure nobody waits for us on a socket anymore */
1005 close_all_fds(NULL, 0);
1009 /* Let's not freeze right away, but keep reaping zombies. */
1014 r = waitid(P_ALL, 0, &si, WEXITED);
1015 if (r < 0 && errno != EINTR)
1019 /* waitid() failed with an unexpected error, things are really borked. Freeze now! */
1024 bool oom_score_adjust_is_valid(int oa) {
1025 return oa >= OOM_SCORE_ADJ_MIN && oa <= OOM_SCORE_ADJ_MAX;
1028 unsigned long personality_from_string(const char *p) {
1032 return PERSONALITY_INVALID;
1034 /* Parse a personality specifier. We use our own identifiers that indicate specific ABIs, rather than just
1035 * hints regarding the register size, since we want to keep things open for multiple locally supported ABIs for
1036 * the same register size. */
1038 architecture = architecture_from_string(p);
1039 if (architecture < 0)
1040 return PERSONALITY_INVALID;
1042 if (architecture == native_architecture())
1044 #ifdef SECONDARY_ARCHITECTURE
1045 if (architecture == SECONDARY_ARCHITECTURE)
1049 return PERSONALITY_INVALID;
1052 const char* personality_to_string(unsigned long p) {
1053 int architecture = _ARCHITECTURE_INVALID;
1056 architecture = native_architecture();
1057 #ifdef SECONDARY_ARCHITECTURE
1058 else if (p == PER_LINUX32)
1059 architecture = SECONDARY_ARCHITECTURE;
1062 if (architecture < 0)
1065 return architecture_to_string(architecture);
1068 int safe_personality(unsigned long p) {
1071 /* So here's the deal, personality() is weirdly defined by glibc. In some cases it returns a failure via errno,
1072 * and in others as negative return value containing an errno-like value. Let's work around this: this is a
1073 * wrapper that uses errno if it is set, and uses the return value otherwise. And then it sets both errno and
1074 * the return value indicating the same issue, so that we are definitely on the safe side.
1076 * See https://github.com/systemd/systemd/issues/6737 */
1079 ret = personality(p);
1090 int opinionated_personality(unsigned long *ret) {
1093 /* Returns the current personality, or PERSONALITY_INVALID if we can't determine it. This function is a bit
1094 * opinionated though, and ignores all the finer-grained bits and exotic personalities, only distinguishing the
1095 * two most relevant personalities: PER_LINUX and PER_LINUX32. */
1097 current = safe_personality(PERSONALITY_INVALID);
1101 if (((unsigned long) current & 0xffff) == PER_LINUX32)
1109 void valgrind_summary_hack(void) {
1110 #if HAVE_VALGRIND_VALGRIND_H
1111 if (getpid_cached() == 1 && RUNNING_ON_VALGRIND) {
1113 pid = raw_clone(SIGCHLD);
1115 log_emergency_errno(errno, "Failed to fork off valgrind helper: %m");
1119 log_info("Spawned valgrind helper as PID "PID_FMT".", pid);
1120 (void) wait_for_terminate(pid, NULL);
1126 int pid_compare_func(const void *a, const void *b) {
1127 const pid_t *p = a, *q = b;
1129 /* 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))
1156 /* The cached PID, possible values:
1158 * == UNSET [0] → cache not initialized yet
1159 * == BUSY [-1] → some thread is initializing it at the moment
1160 * any other → the cached PID
1163 #define CACHED_PID_UNSET ((pid_t) 0)
1164 #define CACHED_PID_BUSY ((pid_t) -1)
1166 static pid_t cached_pid = CACHED_PID_UNSET;
1168 void reset_cached_pid(void) {
1169 /* Invoked in the child after a fork(), i.e. at the first moment the PID changed */
1170 cached_pid = CACHED_PID_UNSET;
1173 /* We use glibc __register_atfork() + __dso_handle directly here, as they are not included in the glibc
1174 * headers. __register_atfork() is mostly equivalent to pthread_atfork(), but doesn't require us to link against
1175 * libpthread, as it is part of glibc anyway. */
1177 extern int __register_atfork(void (*prepare) (void), void (*parent) (void), void (*child) (void), void * __dso_handle);
1178 extern void* __dso_handle __attribute__ ((__weak__));
1179 #endif // ifdef __GLIBC__
1181 pid_t getpid_cached(void) {
1182 static bool installed = false;
1183 pid_t current_value;
1185 /* getpid_cached() is much like getpid(), but caches the value in local memory, to avoid having to invoke a
1186 * system call each time. This restores glibc behaviour from before 2.24, when getpid() was unconditionally
1187 * cached. Starting with 2.24 getpid() started to become prohibitively expensive when used for detecting when
1188 * objects were used across fork()s. With this caching the old behaviour is somewhat restored.
1190 * https://bugzilla.redhat.com/show_bug.cgi?id=1443976
1191 * https://sourceware.org/git/gitweb.cgi?p=glibc.git;h=c579f48edba88380635ab98cb612030e3ed8691e
1194 current_value = __sync_val_compare_and_swap(&cached_pid, CACHED_PID_UNSET, CACHED_PID_BUSY);
1196 switch (current_value) {
1198 case CACHED_PID_UNSET: { /* Not initialized yet, then do so now */
1201 new_pid = raw_getpid();
1204 /* __register_atfork() either returns 0 or -ENOMEM, in its glibc implementation. Since it's
1205 * only half-documented (glibc doesn't document it but LSB does — though only superficially)
1206 * we'll check for errors only in the most generic fashion possible. */
1208 if (__register_atfork(NULL, NULL, reset_cached_pid, __dso_handle) != 0) {
1209 /* OOM? Let's try again later */
1210 cached_pid = CACHED_PID_UNSET;
1217 cached_pid = new_pid;
1221 case CACHED_PID_BUSY: /* Somebody else is currently initializing */
1222 return raw_getpid();
1224 default: /* Properly initialized */
1225 return current_value;
1229 int must_be_root(void) {
1234 log_error("Need to be root.");
1240 const int except_fds[],
1241 size_t n_except_fds,
1245 pid_t original_pid, pid;
1246 sigset_t saved_ss, ss;
1247 bool block_signals = false;
1250 /* A wrapper around fork(), that does a couple of important initializations in addition to mere forking. Always
1251 * returns the child's PID in *ret_pid. Returns == 0 in the child, and > 0 in the parent. */
1253 prio = flags & FORK_LOG ? LOG_ERR : LOG_DEBUG;
1255 original_pid = getpid_cached();
1257 if (flags & (FORK_RESET_SIGNALS|FORK_DEATHSIG)) {
1259 /* We temporarily block all signals, so that the new child has them blocked initially. This way, we can
1260 * be sure that SIGTERMs are not lost we might send to the child. */
1262 if (sigfillset(&ss) < 0)
1263 return log_full_errno(prio, errno, "Failed to reset signal set: %m");
1265 block_signals = true;
1267 } else if (flags & FORK_WAIT) {
1269 /* Let's block SIGCHLD at least, so that we can safely watch for the child process */
1271 if (sigemptyset(&ss) < 0)
1272 return log_full_errno(prio, errno, "Failed to clear signal set: %m");
1274 if (sigaddset(&ss, SIGCHLD) < 0)
1275 return log_full_errno(prio, errno, "Failed to add SIGCHLD to signal set: %m");
1277 block_signals = true;
1281 if (sigprocmask(SIG_SETMASK, &ss, &saved_ss) < 0)
1282 return log_full_errno(prio, errno, "Failed to set signal mask: %m");
1284 if (flags & FORK_NEW_MOUNTNS)
1285 pid = raw_clone(SIGCHLD|CLONE_NEWNS);
1291 if (block_signals) /* undo what we did above */
1292 (void) sigprocmask(SIG_SETMASK, &saved_ss, NULL);
1294 return log_full_errno(prio, r, "Failed to fork: %m");
1297 /* We are in the parent process */
1299 log_debug("Successfully forked off '%s' as PID " PID_FMT ".", strna(name), pid);
1301 if (flags & FORK_WAIT) {
1302 r = wait_for_terminate_and_check(name, pid, (flags & FORK_LOG ? WAIT_LOG : 0));
1305 if (r != EXIT_SUCCESS) /* exit status > 0 should be treated as failure, too */
1309 if (block_signals) /* undo what we did above */
1310 (void) sigprocmask(SIG_SETMASK, &saved_ss, NULL);
1318 /* We are in the child process */
1320 if (flags & FORK_REOPEN_LOG) {
1321 /* Close the logs if requested, before we log anything. And make sure we reopen it if needed. */
1323 log_set_open_when_needed(true);
1327 r = rename_process(name);
1329 log_full_errno(flags & FORK_LOG ? LOG_WARNING : LOG_DEBUG,
1330 r, "Failed to rename process, ignoring: %m");
1333 if (flags & FORK_DEATHSIG)
1334 if (prctl(PR_SET_PDEATHSIG, SIGTERM) < 0) {
1335 log_full_errno(prio, errno, "Failed to set death signal: %m");
1336 _exit(EXIT_FAILURE);
1339 if (flags & FORK_RESET_SIGNALS) {
1340 r = reset_all_signal_handlers();
1342 log_full_errno(prio, r, "Failed to reset signal handlers: %m");
1343 _exit(EXIT_FAILURE);
1346 /* This implicitly undoes the signal mask stuff we did before the fork()ing above */
1347 r = reset_signal_mask();
1349 log_full_errno(prio, r, "Failed to reset signal mask: %m");
1350 _exit(EXIT_FAILURE);
1352 } else if (block_signals) { /* undo what we did above */
1353 if (sigprocmask(SIG_SETMASK, &saved_ss, NULL) < 0) {
1354 log_full_errno(prio, errno, "Failed to restore signal mask: %m");
1355 _exit(EXIT_FAILURE);
1359 if (flags & FORK_DEATHSIG) {
1361 /* Let's see if the parent PID is still the one we started from? If not, then the parent
1362 * already died by the time we set PR_SET_PDEATHSIG, hence let's emulate the effect */
1366 /* Parent is in a differn't PID namespace. */;
1367 else if (ppid != original_pid) {
1368 log_debug("Parent died early, raising SIGTERM.");
1369 (void) raise(SIGTERM);
1370 _exit(EXIT_FAILURE);
1374 if ((flags & (FORK_NEW_MOUNTNS|FORK_MOUNTNS_SLAVE)) == (FORK_NEW_MOUNTNS|FORK_MOUNTNS_SLAVE)) {
1376 /* Optionally, make sure we never propagate mounts to the host. */
1378 if (mount(NULL, "/", NULL, MS_SLAVE | MS_REC, NULL) < 0) {
1379 log_full_errno(prio, errno, "Failed to remount root directory as MS_SLAVE: %m");
1380 _exit(EXIT_FAILURE);
1384 if (flags & FORK_CLOSE_ALL_FDS) {
1385 /* Close the logs here in case it got reopened above, as close_all_fds() would close them for us */
1388 r = close_all_fds(except_fds, n_except_fds);
1390 log_full_errno(prio, r, "Failed to close all file descriptors: %m");
1391 _exit(EXIT_FAILURE);
1395 /* When we were asked to reopen the logs, do so again now */
1396 if (flags & FORK_REOPEN_LOG) {
1398 log_set_open_when_needed(false);
1401 if (flags & FORK_NULL_STDIO) {
1402 r = make_null_stdio();
1404 log_full_errno(prio, r, "Failed to connect stdin/stdout to /dev/null: %m");
1405 _exit(EXIT_FAILURE);
1410 *ret_pid = getpid_cached();
1415 int fork_agent(const char *name, const int except[], size_t n_except, pid_t *ret_pid, const char *path, ...) {
1416 bool stdout_is_tty, stderr_is_tty;
1424 /* Spawns a temporary TTY agent, making sure it goes away when we go away */
1426 r = safe_fork_full(name, except, n_except, FORK_RESET_SIGNALS|FORK_DEATHSIG|FORK_CLOSE_ALL_FDS, ret_pid);
1434 stdout_is_tty = isatty(STDOUT_FILENO);
1435 stderr_is_tty = isatty(STDERR_FILENO);
1437 if (!stdout_is_tty || !stderr_is_tty) {
1440 /* Detach from stdout/stderr. and reopen
1441 * /dev/tty for them. This is important to
1442 * ensure that when systemctl is started via
1443 * popen() or a similar call that expects to
1444 * read EOF we actually do generate EOF and
1445 * not delay this indefinitely by because we
1446 * keep an unused copy of stdin around. */
1447 fd = open("/dev/tty", O_WRONLY);
1449 log_error_errno(errno, "Failed to open /dev/tty: %m");
1450 _exit(EXIT_FAILURE);
1453 if (!stdout_is_tty && dup2(fd, STDOUT_FILENO) < 0) {
1454 log_error_errno(errno, "Failed to dup2 /dev/tty: %m");
1455 _exit(EXIT_FAILURE);
1458 if (!stderr_is_tty && dup2(fd, STDERR_FILENO) < 0) {
1459 log_error_errno(errno, "Failed to dup2 /dev/tty: %m");
1460 _exit(EXIT_FAILURE);
1463 safe_close_above_stdio(fd);
1466 /* Count arguments */
1468 for (n = 0; va_arg(ap, char*); n++)
1473 l = newa(char*, n + 1);
1475 /* Fill in arguments */
1477 for (i = 0; i <= n; i++)
1478 l[i] = va_arg(ap, char*);
1482 _exit(EXIT_FAILURE);
1485 int set_oom_score_adjust(int value) {
1486 char t[DECIMAL_STR_MAX(int)];
1488 sprintf(t, "%i", value);
1490 return write_string_file("/proc/self/oom_score_adj", t,
1491 WRITE_STRING_FILE_VERIFY_ON_FAILURE|WRITE_STRING_FILE_DISABLE_BUFFER);
1494 #if 0 /// UNNEEDED by elogind
1495 static const char *const ioprio_class_table[] = {
1496 [IOPRIO_CLASS_NONE] = "none",
1497 [IOPRIO_CLASS_RT] = "realtime",
1498 [IOPRIO_CLASS_BE] = "best-effort",
1499 [IOPRIO_CLASS_IDLE] = "idle"
1502 DEFINE_STRING_TABLE_LOOKUP_WITH_FALLBACK(ioprio_class, int, IOPRIO_N_CLASSES);
1504 static const char *const sigchld_code_table[] = {
1505 [CLD_EXITED] = "exited",
1506 [CLD_KILLED] = "killed",
1507 [CLD_DUMPED] = "dumped",
1508 [CLD_TRAPPED] = "trapped",
1509 [CLD_STOPPED] = "stopped",
1510 [CLD_CONTINUED] = "continued",
1513 DEFINE_STRING_TABLE_LOOKUP(sigchld_code, int);
1515 static const char* const sched_policy_table[] = {
1516 [SCHED_OTHER] = "other",
1517 [SCHED_BATCH] = "batch",
1518 [SCHED_IDLE] = "idle",
1519 [SCHED_FIFO] = "fifo",
1523 DEFINE_STRING_TABLE_LOOKUP_WITH_FALLBACK(sched_policy, int, INT_MAX);