1 /* SPDX-License-Identifier: LGPL-2.1+ */
11 #include <sys/mount.h>
12 #include <sys/personality.h>
13 #include <sys/prctl.h>
14 #include <sys/types.h>
18 #if HAVE_VALGRIND_VALGRIND_H
19 #include <valgrind/valgrind.h>
22 #include "alloc-util.h"
23 #include "architecture.h"
25 #include "errno-util.h"
31 #include "locale-util.h"
34 #include "memory-util.h"
35 #include "missing_sched.h"
36 #include "missing_syscall.h"
37 #include "namespace-util.h"
38 #include "path-util.h"
39 #include "process-util.h"
40 #include "raw-clone.h"
41 #include "rlimit-util.h"
42 #include "signal-util.h"
43 #include "stat-util.h"
44 #include "stdio-util.h"
45 #include "string-table.h"
46 #include "string-util.h"
47 #include "terminal-util.h"
48 #include "user-util.h"
51 /* The kernel limits userspace processes to TASK_COMM_LEN (16 bytes), but allows higher values for its own
52 * workers, e.g. "kworker/u9:3-kcryptd/253:0". Let's pick a fixed smallish limit that will work for the kernel.
54 #define COMM_MAX_LEN 128
56 static int get_process_state(pid_t pid
) {
57 _cleanup_free_
char *line
= NULL
;
64 /* Shortcut: if we are enquired about our own state, we are obviously running */
65 if (pid
== 0 || pid
== getpid_cached())
66 return (unsigned char) 'R';
68 p
= procfs_file_alloca(pid
, "stat");
70 r
= read_one_line_file(p
, &line
);
76 p
= strrchr(line
, ')');
82 if (sscanf(p
, " %c", &state
) != 1)
85 return (unsigned char) state
;
88 int get_process_comm(pid_t pid
, char **ret
) {
89 _cleanup_free_
char *escaped
= NULL
, *comm
= NULL
;
95 if (pid
== 0 || pid
== getpid_cached()) {
96 comm
= new0(char, TASK_COMM_LEN
+ 1); /* Must fit in 16 byte according to prctl(2) */
100 if (prctl(PR_GET_NAME
, comm
) < 0)
105 p
= procfs_file_alloca(pid
, "comm");
107 /* Note that process names of kernel threads can be much longer than TASK_COMM_LEN */
108 r
= read_one_line_file(p
, &comm
);
115 escaped
= new(char, COMM_MAX_LEN
);
119 /* Escape unprintable characters, just in case, but don't grow the string beyond the underlying size */
120 cellescape(escaped
, COMM_MAX_LEN
, comm
);
122 *ret
= TAKE_PTR(escaped
);
126 int get_process_cmdline(pid_t pid
, size_t max_columns
, ProcessCmdlineFlags flags
, char **line
) {
127 _cleanup_fclose_
FILE *f
= NULL
;
128 _cleanup_free_
char *t
= NULL
, *ans
= NULL
;
133 /* This is supposed to be a safety guard against runaway command lines. */
134 size_t max_length
= sc_arg_max();
139 /* Retrieves a process' command line. Replaces non-utf8 bytes by replacement character (�). If
140 * max_columns is != -1 will return a string of the specified console width at most, abbreviated with
141 * an ellipsis. If PROCESS_CMDLINE_COMM_FALLBACK is specified in flags and the process has no command
142 * line set (the case for kernel threads), or has a command line that resolves to the empty string
143 * will return the "comm" name of the process instead. This will use at most _SC_ARG_MAX bytes of
146 * Returns -ESRCH if the process doesn't exist, and -ENOENT if the process has no command line (and
147 * comm_fallback is false). Returns 0 and sets *line otherwise. */
149 p
= procfs_file_alloca(pid
, "cmdline");
150 r
= fopen_unlocked(p
, "re", &f
);
156 /* We assume that each four-byte character uses one or two columns. If we ever check for combining
157 * characters, this assumption will need to be adjusted. */
158 if ((size_t) 4 * max_columns
+ 1 < max_columns
)
159 max_length
= MIN(max_length
, (size_t) 4 * max_columns
+ 1);
161 t
= new(char, max_length
);
165 k
= fread(t
, 1, max_length
, f
);
167 /* Arguments are separated by NULs. Let's replace those with spaces. */
168 for (size_t i
= 0; i
< k
- 1; i
++)
172 t
[k
] = '\0'; /* Normally, t[k] is already NUL, so this is just a guard in case of short read */
174 /* We only treat getting nothing as an error. We *could* also get an error after reading some
175 * data, but we ignore that case, as such an error is rather unlikely and we prefer to get
176 * some data rather than none. */
180 if (!(flags
& PROCESS_CMDLINE_COMM_FALLBACK
))
183 /* Kernel threads have no argv[] */
184 _cleanup_free_
char *t2
= NULL
;
186 r
= get_process_comm(pid
, &t2
);
191 t
= strjoin("[", t2
, "]");
196 delete_trailing_chars(t
, WHITESPACE
);
198 bool eight_bit
= (flags
& PROCESS_CMDLINE_USE_LOCALE
) && !is_locale_utf8();
200 ans
= escape_non_printable_full(t
, max_columns
, eight_bit
);
204 (void) str_realloc(&ans
);
205 *line
= TAKE_PTR(ans
);
209 int rename_process(const char name
[]) {
210 static size_t mm_size
= 0;
211 static char *mm
= NULL
;
212 bool truncated
= false;
215 /* This is a like a poor man's setproctitle(). It changes the comm field, argv[0], and also the glibc's
216 * internally used name of the process. For the first one a limit of 16 chars applies; to the second one in
217 * many cases one of 10 (i.e. length of "/sbin/init") — however if we have CAP_SYS_RESOURCES it is unbounded;
218 * to the third one 7 (i.e. the length of "systemd". If you pass a longer string it will likely be
221 * Returns 0 if a name was set but truncated, > 0 if it was set but not truncated. */
224 return -EINVAL
; /* let's not confuse users unnecessarily with an empty name */
226 if (!is_main_thread())
227 return -EPERM
; /* Let's not allow setting the process name from other threads than the main one, as we
228 * cache things without locking, and we make assumptions that PR_SET_NAME sets the
229 * process name that isn't correct on any other threads */
233 /* First step, change the comm field. The main thread's comm is identical to the process comm. This means we
234 * can use PR_SET_NAME, which sets the thread name for the calling thread. */
235 if (prctl(PR_SET_NAME
, name
) < 0)
236 log_debug_errno(errno
, "PR_SET_NAME failed: %m");
237 if (l
>= TASK_COMM_LEN
) /* Linux userspace process names can be 15 chars at max */
240 /* Second step, change glibc's ID of the process name. */
241 if (program_invocation_name
) {
244 k
= strlen(program_invocation_name
);
245 strncpy(program_invocation_name
, name
, k
);
250 /* Third step, completely replace the argv[] array the kernel maintains for us. This requires privileges, but
251 * has the advantage that the argv[] array is exactly what we want it to be, and not filled up with zeros at
252 * the end. This is the best option for changing /proc/self/cmdline. */
254 /* Let's not bother with this if we don't have euid == 0. Strictly speaking we should check for the
255 * CAP_SYS_RESOURCE capability which is independent of the euid. In our own code the capability generally is
256 * present only for euid == 0, hence let's use this as quick bypass check, to avoid calling mmap() if
257 * PR_SET_MM_ARG_{START,END} fails with EPERM later on anyway. After all geteuid() is dead cheap to call, but
260 log_debug("Skipping PR_SET_MM, as we don't have privileges.");
261 else if (mm_size
< l
+1) {
265 nn_size
= PAGE_ALIGN(l
+1);
266 nn
= mmap(NULL
, nn_size
, PROT_READ
|PROT_WRITE
, MAP_PRIVATE
|MAP_ANONYMOUS
, -1, 0);
267 if (nn
== MAP_FAILED
) {
268 log_debug_errno(errno
, "mmap() failed: %m");
272 strncpy(nn
, name
, nn_size
);
274 /* Now, let's tell the kernel about this new memory */
275 if (prctl(PR_SET_MM
, PR_SET_MM_ARG_START
, (unsigned long) nn
, 0, 0) < 0) {
276 /* HACK: prctl() API is kind of dumb on this point. The existing end address may already be
277 * below the desired start address, in which case the kernel may have kicked this back due
278 * to a range-check failure (see linux/kernel/sys.c:validate_prctl_map() to see this in
279 * action). The proper solution would be to have a prctl() API that could set both start+end
280 * simultaneously, or at least let us query the existing address to anticipate this condition
281 * and respond accordingly. For now, we can only guess at the cause of this failure and try
282 * a workaround--which will briefly expand the arg space to something potentially huge before
283 * resizing it to what we want. */
284 log_debug_errno(errno
, "PR_SET_MM_ARG_START failed, attempting PR_SET_MM_ARG_END hack: %m");
286 if (prctl(PR_SET_MM
, PR_SET_MM_ARG_END
, (unsigned long) nn
+ l
+ 1, 0, 0) < 0) {
287 log_debug_errno(errno
, "PR_SET_MM_ARG_END hack failed, proceeding without: %m");
288 (void) munmap(nn
, nn_size
);
292 if (prctl(PR_SET_MM
, PR_SET_MM_ARG_START
, (unsigned long) nn
, 0, 0) < 0) {
293 log_debug_errno(errno
, "PR_SET_MM_ARG_START still failed, proceeding without: %m");
297 /* And update the end pointer to the new end, too. If this fails, we don't really know what
298 * to do, it's pretty unlikely that we can rollback, hence we'll just accept the failure,
300 if (prctl(PR_SET_MM
, PR_SET_MM_ARG_END
, (unsigned long) nn
+ l
+ 1, 0, 0) < 0)
301 log_debug_errno(errno
, "PR_SET_MM_ARG_END failed, proceeding without: %m");
305 (void) munmap(mm
, mm_size
);
310 strncpy(mm
, name
, mm_size
);
312 /* Update the end pointer, continuing regardless of any failure. */
313 if (prctl(PR_SET_MM
, PR_SET_MM_ARG_END
, (unsigned long) mm
+ l
+ 1, 0, 0) < 0)
314 log_debug_errno(errno
, "PR_SET_MM_ARG_END failed, proceeding without: %m");
318 /* Fourth step: in all cases we'll also update the original argv[], so that our own code gets it right too if
319 * it still looks here */
321 if (saved_argc
> 0) {
327 k
= strlen(saved_argv
[0]);
328 strncpy(saved_argv
[0], name
, k
);
333 for (i
= 1; i
< saved_argc
; i
++) {
337 memzero(saved_argv
[i
], strlen(saved_argv
[i
]));
344 int is_kernel_thread(pid_t pid
) {
345 _cleanup_free_
char *line
= NULL
;
346 unsigned long long flags
;
352 if (IN_SET(pid
, 0, 1) || pid
== getpid_cached()) /* pid 1, and we ourselves certainly aren't a kernel thread */
354 if (!pid_is_valid(pid
))
357 p
= procfs_file_alloca(pid
, "stat");
358 r
= read_one_line_file(p
, &line
);
364 /* Skip past the comm field */
365 q
= strrchr(line
, ')');
370 /* Skip 6 fields to reach the flags field */
371 for (i
= 0; i
< 6; i
++) {
372 l
= strspn(q
, WHITESPACE
);
377 l
= strcspn(q
, WHITESPACE
);
383 /* Skip preceding whitespace */
384 l
= strspn(q
, WHITESPACE
);
389 /* Truncate the rest */
390 l
= strcspn(q
, WHITESPACE
);
395 r
= safe_atollu(q
, &flags
);
399 return !!(flags
& PF_KTHREAD
);
402 int get_process_capeff(pid_t pid
, char **capeff
) {
409 p
= procfs_file_alloca(pid
, "status");
411 r
= get_proc_field(p
, "CapEff", WHITESPACE
, capeff
);
418 static int get_process_link_contents(const char *proc_file
, char **name
) {
424 r
= readlink_malloc(proc_file
, name
);
433 int get_process_exe(pid_t pid
, char **name
) {
440 p
= procfs_file_alloca(pid
, "exe");
441 r
= get_process_link_contents(p
, name
);
445 d
= endswith(*name
, " (deleted)");
452 static int get_process_id(pid_t pid
, const char *field
, uid_t
*uid
) {
453 _cleanup_fclose_
FILE *f
= NULL
;
463 p
= procfs_file_alloca(pid
, "status");
464 r
= fopen_unlocked(p
, "re", &f
);
471 _cleanup_free_
char *line
= NULL
;
474 r
= read_line(f
, LONG_LINE_MAX
, &line
);
482 if (startswith(l
, field
)) {
484 l
+= strspn(l
, WHITESPACE
);
486 l
[strcspn(l
, WHITESPACE
)] = 0;
488 return parse_uid(l
, uid
);
495 int get_process_uid(pid_t pid
, uid_t
*uid
) {
497 if (pid
== 0 || pid
== getpid_cached()) {
502 return get_process_id(pid
, "Uid:", uid
);
505 int get_process_gid(pid_t pid
, gid_t
*gid
) {
507 if (pid
== 0 || pid
== getpid_cached()) {
512 assert_cc(sizeof(uid_t
) == sizeof(gid_t
));
513 return get_process_id(pid
, "Gid:", gid
);
516 int get_process_cwd(pid_t pid
, char **cwd
) {
521 if (pid
== 0 || pid
== getpid_cached())
522 return safe_getcwd(cwd
);
524 p
= procfs_file_alloca(pid
, "cwd");
526 return get_process_link_contents(p
, cwd
);
529 int get_process_root(pid_t pid
, char **root
) {
534 p
= procfs_file_alloca(pid
, "root");
536 return get_process_link_contents(p
, root
);
539 #define ENVIRONMENT_BLOCK_MAX (5U*1024U*1024U)
541 int get_process_environ(pid_t pid
, char **env
) {
542 _cleanup_fclose_
FILE *f
= NULL
;
543 _cleanup_free_
char *outcome
= NULL
;
544 size_t allocated
= 0, sz
= 0;
551 p
= procfs_file_alloca(pid
, "environ");
553 r
= fopen_unlocked(p
, "re", &f
);
562 if (sz
>= ENVIRONMENT_BLOCK_MAX
)
565 if (!GREEDY_REALLOC(outcome
, allocated
, sz
+ 5))
568 r
= safe_fgetc(f
, &c
);
575 outcome
[sz
++] = '\n';
577 sz
+= cescape_char(c
, outcome
+ sz
);
581 *env
= TAKE_PTR(outcome
);
586 int get_process_ppid(pid_t pid
, pid_t
*_ppid
) {
588 _cleanup_free_
char *line
= NULL
;
595 if (pid
== 0 || pid
== getpid_cached()) {
600 p
= procfs_file_alloca(pid
, "stat");
601 r
= read_one_line_file(p
, &line
);
607 /* Let's skip the pid and comm fields. The latter is enclosed
608 * in () but does not escape any () in its value, so let's
609 * skip over it manually */
611 p
= strrchr(line
, ')');
623 if ((long unsigned) (pid_t
) ppid
!= ppid
)
626 *_ppid
= (pid_t
) ppid
;
631 int wait_for_terminate(pid_t pid
, siginfo_t
*status
) {
642 if (waitid(P_PID
, pid
, status
, WEXITED
) < 0) {
647 return negative_errno();
656 * < 0 : wait_for_terminate() failed to get the state of the
657 * process, the process was terminated by a signal, or
658 * failed for an unknown reason.
659 * >=0 : The process terminated normally, and its exit code is
662 * That is, success is indicated by a return value of zero, and an
663 * error is indicated by a non-zero value.
665 * A warning is emitted if the process terminates abnormally,
666 * and also if it returns non-zero unless check_exit_code is true.
668 int wait_for_terminate_and_check(const char *name
, pid_t pid
, WaitFlags flags
) {
669 _cleanup_free_
char *buffer
= NULL
;
676 r
= get_process_comm(pid
, &buffer
);
678 log_debug_errno(r
, "Failed to acquire process name of " PID_FMT
", ignoring: %m", pid
);
683 prio
= flags
& WAIT_LOG_ABNORMAL
? LOG_ERR
: LOG_DEBUG
;
685 r
= wait_for_terminate(pid
, &status
);
687 return log_full_errno(prio
, r
, "Failed to wait for %s: %m", strna(name
));
689 if (status
.si_code
== CLD_EXITED
) {
690 if (status
.si_status
!= EXIT_SUCCESS
)
691 log_full(flags
& WAIT_LOG_NON_ZERO_EXIT_STATUS
? LOG_ERR
: LOG_DEBUG
,
692 "%s failed with exit status %i.", strna(name
), status
.si_status
);
694 log_debug("%s succeeded.", name
);
696 return status
.si_status
;
698 } else if (IN_SET(status
.si_code
, CLD_KILLED
, CLD_DUMPED
)) {
700 log_full(prio
, "%s terminated by signal %s.", strna(name
), signal_to_string(status
.si_status
));
704 log_full(prio
, "%s failed due to unknown reason.", strna(name
));
711 * < 0 : wait_for_terminate_with_timeout() failed to get the state of the process, the process timed out, the process
712 * was terminated by a signal, or failed for an unknown reason.
714 * >=0 : The process terminated normally with no failures.
716 * Success is indicated by a return value of zero, a timeout is indicated by ETIMEDOUT, and all other child failure
717 * states are indicated by error is indicated by a non-zero value.
719 * This call assumes SIGCHLD has been blocked already, in particular before the child to wait for has been forked off
720 * to remain entirely race-free.
722 int wait_for_terminate_with_timeout(pid_t pid
, usec_t timeout
) {
727 assert_se(sigemptyset(&mask
) == 0);
728 assert_se(sigaddset(&mask
, SIGCHLD
) == 0);
730 /* Drop into a sigtimewait-based timeout. Waiting for the
732 until
= now(CLOCK_MONOTONIC
) + timeout
;
735 siginfo_t status
= {};
738 n
= now(CLOCK_MONOTONIC
);
742 r
= sigtimedwait(&mask
, NULL
, timespec_store(&ts
, until
- n
)) < 0 ? -errno
: 0;
743 /* Assuming we woke due to the child exiting. */
744 if (waitid(P_PID
, pid
, &status
, WEXITED
|WNOHANG
) == 0) {
745 if (status
.si_pid
== pid
) {
746 /* This is the correct child.*/
747 if (status
.si_code
== CLD_EXITED
)
748 return (status
.si_status
== 0) ? 0 : -EPROTO
;
753 /* Not the child, check for errors and proceed appropriately */
757 /* Timed out, child is likely hung. */
760 /* Received a different signal and should retry */
763 /* Return any unexpected errors */
772 void sigkill_wait(pid_t pid
) {
775 if (kill(pid
, SIGKILL
) >= 0)
776 (void) wait_for_terminate(pid
, NULL
);
779 void sigkill_waitp(pid_t
*pid
) {
790 void sigterm_wait(pid_t pid
) {
793 if (kill_and_sigcont(pid
, SIGTERM
) >= 0)
794 (void) wait_for_terminate(pid
, NULL
);
797 int kill_and_sigcont(pid_t pid
, int sig
) {
800 r
= kill(pid
, sig
) < 0 ? -errno
: 0;
802 /* If this worked, also send SIGCONT, unless we already just sent a SIGCONT, or SIGKILL was sent which isn't
803 * affected by a process being suspended anyway. */
804 if (r
>= 0 && !IN_SET(sig
, SIGCONT
, SIGKILL
))
805 (void) kill(pid
, SIGCONT
);
810 int getenv_for_pid(pid_t pid
, const char *field
, char **ret
) {
811 _cleanup_fclose_
FILE *f
= NULL
;
821 if (pid
== 0 || pid
== getpid_cached()) {
838 if (!pid_is_valid(pid
))
841 path
= procfs_file_alloca(pid
, "environ");
843 r
= fopen_unlocked(path
, "re", &f
);
851 _cleanup_free_
char *line
= NULL
;
853 if (sum
> ENVIRONMENT_BLOCK_MAX
) /* Give up searching eventually */
856 r
= read_nul_string(f
, LONG_LINE_MAX
, &line
);
859 if (r
== 0) /* EOF */
864 if (strneq(line
, field
, l
) && line
[l
] == '=') {
865 value
= strdup(line
+ l
+ 1);
878 int pid_is_my_child(pid_t pid
) {
885 r
= get_process_ppid(pid
, &ppid
);
889 return ppid
== getpid_cached();
892 bool pid_is_unwaited(pid_t pid
) {
893 /* Checks whether a PID is still valid at all, including a zombie */
898 if (pid
<= 1) /* If we or PID 1 would be dead and have been waited for, this code would not be running */
901 if (pid
== getpid_cached())
904 if (kill(pid
, 0) >= 0)
907 return errno
!= ESRCH
;
910 bool pid_is_alive(pid_t pid
) {
913 /* Checks whether a PID is still valid and not a zombie */
918 if (pid
<= 1) /* If we or PID 1 would be a zombie, this code would not be running */
921 if (pid
== getpid_cached())
924 r
= get_process_state(pid
);
925 if (IN_SET(r
, -ESRCH
, 'Z'))
931 int pid_from_same_root_fs(pid_t pid
) {
937 if (pid
== 0 || pid
== getpid_cached())
940 root
= procfs_file_alloca(pid
, "root");
942 return files_same(root
, "/proc/1/root", 0);
945 bool is_main_thread(void) {
946 static thread_local
int cached
= 0;
948 if (_unlikely_(cached
== 0))
949 cached
= getpid_cached() == gettid() ? 1 : -1;
954 _noreturn_
void freeze(void) {
958 /* Make sure nobody waits for us on a socket anymore */
959 (void) close_all_fds(NULL
, 0);
963 /* Let's not freeze right away, but keep reaping zombies. */
968 r
= waitid(P_ALL
, 0, &si
, WEXITED
);
969 if (r
< 0 && errno
!= EINTR
)
973 /* waitid() failed with an unexpected error, things are really borked. Freeze now! */
978 bool oom_score_adjust_is_valid(int oa
) {
979 return oa
>= OOM_SCORE_ADJ_MIN
&& oa
<= OOM_SCORE_ADJ_MAX
;
982 unsigned long personality_from_string(const char *p
) {
986 return PERSONALITY_INVALID
;
988 /* Parse a personality specifier. We use our own identifiers that indicate specific ABIs, rather than just
989 * hints regarding the register size, since we want to keep things open for multiple locally supported ABIs for
990 * the same register size. */
992 architecture
= architecture_from_string(p
);
993 if (architecture
< 0)
994 return PERSONALITY_INVALID
;
996 if (architecture
== native_architecture())
998 #ifdef SECONDARY_ARCHITECTURE
999 if (architecture
== SECONDARY_ARCHITECTURE
)
1003 return PERSONALITY_INVALID
;
1006 const char* personality_to_string(unsigned long p
) {
1007 int architecture
= _ARCHITECTURE_INVALID
;
1010 architecture
= native_architecture();
1011 #ifdef SECONDARY_ARCHITECTURE
1012 else if (p
== PER_LINUX32
)
1013 architecture
= SECONDARY_ARCHITECTURE
;
1016 if (architecture
< 0)
1019 return architecture_to_string(architecture
);
1022 int safe_personality(unsigned long p
) {
1025 /* So here's the deal, personality() is weirdly defined by glibc. In some cases it returns a failure via errno,
1026 * and in others as negative return value containing an errno-like value. Let's work around this: this is a
1027 * wrapper that uses errno if it is set, and uses the return value otherwise. And then it sets both errno and
1028 * the return value indicating the same issue, so that we are definitely on the safe side.
1030 * See https://github.com/systemd/systemd/issues/6737 */
1033 ret
= personality(p
);
1044 int opinionated_personality(unsigned long *ret
) {
1047 /* Returns the current personality, or PERSONALITY_INVALID if we can't determine it. This function is a bit
1048 * opinionated though, and ignores all the finer-grained bits and exotic personalities, only distinguishing the
1049 * two most relevant personalities: PER_LINUX and PER_LINUX32. */
1051 current
= safe_personality(PERSONALITY_INVALID
);
1055 if (((unsigned long) current
& 0xffff) == PER_LINUX32
)
1063 void valgrind_summary_hack(void) {
1064 #if HAVE_VALGRIND_VALGRIND_H
1065 if (getpid_cached() == 1 && RUNNING_ON_VALGRIND
) {
1067 pid
= raw_clone(SIGCHLD
);
1069 log_emergency_errno(errno
, "Failed to fork off valgrind helper: %m");
1073 log_info("Spawned valgrind helper as PID "PID_FMT
".", pid
);
1074 (void) wait_for_terminate(pid
, NULL
);
1080 int pid_compare_func(const pid_t
*a
, const pid_t
*b
) {
1081 /* Suitable for usage in qsort() */
1085 int ioprio_parse_priority(const char *s
, int *ret
) {
1091 r
= safe_atoi(s
, &i
);
1095 if (!ioprio_priority_is_valid(i
))
1102 /* The cached PID, possible values:
1104 * == UNSET [0] → cache not initialized yet
1105 * == BUSY [-1] → some thread is initializing it at the moment
1106 * any other → the cached PID
1109 #define CACHED_PID_UNSET ((pid_t) 0)
1110 #define CACHED_PID_BUSY ((pid_t) -1)
1112 static pid_t cached_pid
= CACHED_PID_UNSET
;
1114 void reset_cached_pid(void) {
1115 /* Invoked in the child after a fork(), i.e. at the first moment the PID changed */
1116 cached_pid
= CACHED_PID_UNSET
;
1119 /* We use glibc __register_atfork() + __dso_handle directly here, as they are not included in the glibc
1120 * headers. __register_atfork() is mostly equivalent to pthread_atfork(), but doesn't require us to link against
1121 * libpthread, as it is part of glibc anyway. */
1122 extern int __register_atfork(void (*prepare
) (void), void (*parent
) (void), void (*child
) (void), void *dso_handle
);
1123 extern void* __dso_handle _weak_
;
1125 pid_t
getpid_cached(void) {
1126 static bool installed
= false;
1127 pid_t current_value
;
1129 /* getpid_cached() is much like getpid(), but caches the value in local memory, to avoid having to invoke a
1130 * system call each time. This restores glibc behaviour from before 2.24, when getpid() was unconditionally
1131 * cached. Starting with 2.24 getpid() started to become prohibitively expensive when used for detecting when
1132 * objects were used across fork()s. With this caching the old behaviour is somewhat restored.
1134 * https://bugzilla.redhat.com/show_bug.cgi?id=1443976
1135 * https://sourceware.org/git/gitweb.cgi?p=glibc.git;h=c579f48edba88380635ab98cb612030e3ed8691e
1138 current_value
= __sync_val_compare_and_swap(&cached_pid
, CACHED_PID_UNSET
, CACHED_PID_BUSY
);
1140 switch (current_value
) {
1142 case CACHED_PID_UNSET
: { /* Not initialized yet, then do so now */
1145 new_pid
= raw_getpid();
1148 /* __register_atfork() either returns 0 or -ENOMEM, in its glibc implementation. Since it's
1149 * only half-documented (glibc doesn't document it but LSB does — though only superficially)
1150 * we'll check for errors only in the most generic fashion possible. */
1152 if (__register_atfork(NULL
, NULL
, reset_cached_pid
, __dso_handle
) != 0) {
1153 /* OOM? Let's try again later */
1154 cached_pid
= CACHED_PID_UNSET
;
1161 cached_pid
= new_pid
;
1165 case CACHED_PID_BUSY
: /* Somebody else is currently initializing */
1166 return raw_getpid();
1168 default: /* Properly initialized */
1169 return current_value
;
1173 int must_be_root(void) {
1178 return log_error_errno(SYNTHETIC_ERRNO(EPERM
), "Need to be root.");
1183 const int except_fds
[],
1184 size_t n_except_fds
,
1188 pid_t original_pid
, pid
;
1189 sigset_t saved_ss
, ss
;
1190 bool block_signals
= false;
1193 /* A wrapper around fork(), that does a couple of important initializations in addition to mere forking. Always
1194 * returns the child's PID in *ret_pid. Returns == 0 in the child, and > 0 in the parent. */
1196 prio
= flags
& FORK_LOG
? LOG_ERR
: LOG_DEBUG
;
1198 original_pid
= getpid_cached();
1200 if (flags
& (FORK_RESET_SIGNALS
|FORK_DEATHSIG
)) {
1201 /* We temporarily block all signals, so that the new child has them blocked initially. This way, we can
1202 * be sure that SIGTERMs are not lost we might send to the child. */
1204 assert_se(sigfillset(&ss
) >= 0);
1205 block_signals
= true;
1207 } else if (flags
& FORK_WAIT
) {
1208 /* Let's block SIGCHLD at least, so that we can safely watch for the child process */
1210 assert_se(sigemptyset(&ss
) >= 0);
1211 assert_se(sigaddset(&ss
, SIGCHLD
) >= 0);
1212 block_signals
= true;
1216 if (sigprocmask(SIG_SETMASK
, &ss
, &saved_ss
) < 0)
1217 return log_full_errno(prio
, errno
, "Failed to set signal mask: %m");
1219 if (flags
& FORK_NEW_MOUNTNS
)
1220 pid
= raw_clone(SIGCHLD
|CLONE_NEWNS
);
1226 if (block_signals
) /* undo what we did above */
1227 (void) sigprocmask(SIG_SETMASK
, &saved_ss
, NULL
);
1229 return log_full_errno(prio
, r
, "Failed to fork: %m");
1232 /* We are in the parent process */
1234 log_debug("Successfully forked off '%s' as PID " PID_FMT
".", strna(name
), pid
);
1236 if (flags
& FORK_WAIT
) {
1237 r
= wait_for_terminate_and_check(name
, pid
, (flags
& FORK_LOG
? WAIT_LOG
: 0));
1240 if (r
!= EXIT_SUCCESS
) /* exit status > 0 should be treated as failure, too */
1244 if (block_signals
) /* undo what we did above */
1245 (void) sigprocmask(SIG_SETMASK
, &saved_ss
, NULL
);
1253 /* We are in the child process */
1255 if (flags
& FORK_REOPEN_LOG
) {
1256 /* Close the logs if requested, before we log anything. And make sure we reopen it if needed. */
1258 log_set_open_when_needed(true);
1262 r
= rename_process(name
);
1264 log_full_errno(flags
& FORK_LOG
? LOG_WARNING
: LOG_DEBUG
,
1265 r
, "Failed to rename process, ignoring: %m");
1268 if (flags
& (FORK_DEATHSIG
|FORK_DEATHSIG_SIGINT
))
1269 if (prctl(PR_SET_PDEATHSIG
, (flags
& FORK_DEATHSIG_SIGINT
) ? SIGINT
: SIGTERM
) < 0) {
1270 log_full_errno(prio
, errno
, "Failed to set death signal: %m");
1271 _exit(EXIT_FAILURE
);
1274 if (flags
& FORK_RESET_SIGNALS
) {
1275 r
= reset_all_signal_handlers();
1277 log_full_errno(prio
, r
, "Failed to reset signal handlers: %m");
1278 _exit(EXIT_FAILURE
);
1281 /* This implicitly undoes the signal mask stuff we did before the fork()ing above */
1282 r
= reset_signal_mask();
1284 log_full_errno(prio
, r
, "Failed to reset signal mask: %m");
1285 _exit(EXIT_FAILURE
);
1287 } else if (block_signals
) { /* undo what we did above */
1288 if (sigprocmask(SIG_SETMASK
, &saved_ss
, NULL
) < 0) {
1289 log_full_errno(prio
, errno
, "Failed to restore signal mask: %m");
1290 _exit(EXIT_FAILURE
);
1294 if (flags
& FORK_DEATHSIG
) {
1296 /* Let's see if the parent PID is still the one we started from? If not, then the parent
1297 * already died by the time we set PR_SET_PDEATHSIG, hence let's emulate the effect */
1301 /* Parent is in a differn't PID namespace. */;
1302 else if (ppid
!= original_pid
) {
1303 log_debug("Parent died early, raising SIGTERM.");
1304 (void) raise(SIGTERM
);
1305 _exit(EXIT_FAILURE
);
1309 if (FLAGS_SET(flags
, FORK_NEW_MOUNTNS
| FORK_MOUNTNS_SLAVE
)) {
1311 /* Optionally, make sure we never propagate mounts to the host. */
1313 if (mount(NULL
, "/", NULL
, MS_SLAVE
| MS_REC
, NULL
) < 0) {
1314 log_full_errno(prio
, errno
, "Failed to remount root directory as MS_SLAVE: %m");
1315 _exit(EXIT_FAILURE
);
1319 if (flags
& FORK_CLOSE_ALL_FDS
) {
1320 /* Close the logs here in case it got reopened above, as close_all_fds() would close them for us */
1323 r
= close_all_fds(except_fds
, n_except_fds
);
1325 log_full_errno(prio
, r
, "Failed to close all file descriptors: %m");
1326 _exit(EXIT_FAILURE
);
1330 /* When we were asked to reopen the logs, do so again now */
1331 if (flags
& FORK_REOPEN_LOG
) {
1333 log_set_open_when_needed(false);
1336 if (flags
& FORK_NULL_STDIO
) {
1337 r
= make_null_stdio();
1339 log_full_errno(prio
, r
, "Failed to connect stdin/stdout to /dev/null: %m");
1340 _exit(EXIT_FAILURE
);
1343 } else if (flags
& FORK_STDOUT_TO_STDERR
) {
1344 if (dup2(STDERR_FILENO
, STDOUT_FILENO
) < 0) {
1345 log_full_errno(prio
, errno
, "Failed to connect stdout to stderr: %m");
1346 _exit(EXIT_FAILURE
);
1350 if (flags
& FORK_RLIMIT_NOFILE_SAFE
) {
1351 r
= rlimit_nofile_safe();
1353 log_full_errno(prio
, r
, "Failed to lower RLIMIT_NOFILE's soft limit to 1K: %m");
1354 _exit(EXIT_FAILURE
);
1359 *ret_pid
= getpid_cached();
1365 const char *outer_name
,
1366 const char *inner_name
,
1367 const int except_fds
[],
1368 size_t n_except_fds
,
1379 /* This is much like safe_fork(), but forks twice, and joins the specified namespaces in the middle
1380 * process. This ensures that we are fully a member of the destination namespace, with pidns an all, so that
1381 * /proc/self/fd works correctly. */
1383 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
);
1391 r
= namespace_enter(pidns_fd
, mntns_fd
, netns_fd
, userns_fd
, root_fd
);
1393 log_full_errno(FLAGS_SET(flags
, FORK_LOG
) ? LOG_ERR
: LOG_DEBUG
, r
, "Failed to join namespace: %m");
1394 _exit(EXIT_FAILURE
);
1397 /* We mask a few flags here that either make no sense for the grandchild, or that we don't have to do again */
1398 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
);
1400 _exit(EXIT_FAILURE
);
1408 r
= wait_for_terminate_and_check(inner_name
, pid
, FLAGS_SET(flags
, FORK_LOG
) ? WAIT_LOG
: 0);
1410 _exit(EXIT_FAILURE
);
1418 int fork_agent(const char *name
, const int except
[], size_t n_except
, pid_t
*ret_pid
, const char *path
, ...) {
1419 bool stdout_is_tty
, stderr_is_tty
;
1427 /* Spawns a temporary TTY agent, making sure it goes away when we go away */
1429 r
= safe_fork_full(name
, except
, n_except
, FORK_RESET_SIGNALS
|FORK_DEATHSIG
|FORK_CLOSE_ALL_FDS
, ret_pid
);
1437 stdout_is_tty
= isatty(STDOUT_FILENO
);
1438 stderr_is_tty
= isatty(STDERR_FILENO
);
1440 if (!stdout_is_tty
|| !stderr_is_tty
) {
1443 /* Detach from stdout/stderr. and reopen
1444 * /dev/tty for them. This is important to
1445 * ensure that when systemctl is started via
1446 * popen() or a similar call that expects to
1447 * read EOF we actually do generate EOF and
1448 * not delay this indefinitely by because we
1449 * keep an unused copy of stdin around. */
1450 fd
= open("/dev/tty", O_WRONLY
);
1452 log_error_errno(errno
, "Failed to open /dev/tty: %m");
1453 _exit(EXIT_FAILURE
);
1456 if (!stdout_is_tty
&& dup2(fd
, STDOUT_FILENO
) < 0) {
1457 log_error_errno(errno
, "Failed to dup2 /dev/tty: %m");
1458 _exit(EXIT_FAILURE
);
1461 if (!stderr_is_tty
&& dup2(fd
, STDERR_FILENO
) < 0) {
1462 log_error_errno(errno
, "Failed to dup2 /dev/tty: %m");
1463 _exit(EXIT_FAILURE
);
1466 safe_close_above_stdio(fd
);
1469 (void) rlimit_nofile_safe();
1471 /* Count arguments */
1473 for (n
= 0; va_arg(ap
, char*); n
++)
1478 l
= newa(char*, n
+ 1);
1480 /* Fill in arguments */
1482 for (i
= 0; i
<= n
; i
++)
1483 l
[i
] = va_arg(ap
, char*);
1487 _exit(EXIT_FAILURE
);
1490 int set_oom_score_adjust(int value
) {
1491 char t
[DECIMAL_STR_MAX(int)];
1493 sprintf(t
, "%i", value
);
1495 return write_string_file("/proc/self/oom_score_adj", t
,
1496 WRITE_STRING_FILE_VERIFY_ON_FAILURE
|WRITE_STRING_FILE_DISABLE_BUFFER
);
1499 int pidfd_get_pid(int fd
, pid_t
*ret
) {
1500 char path
[STRLEN("/proc/self/fdinfo/") + DECIMAL_STR_MAX(int)];
1501 _cleanup_free_
char *fdinfo
= NULL
;
1508 xsprintf(path
, "/proc/self/fdinfo/%i", fd
);
1510 r
= read_full_file(path
, &fdinfo
, NULL
);
1511 if (r
== -ENOENT
) /* if fdinfo doesn't exist we assume the process does not exist */
1516 p
= startswith(fdinfo
, "Pid:");
1518 p
= strstr(fdinfo
, "\nPid:");
1520 return -ENOTTY
; /* not a pidfd? */
1525 p
+= strspn(p
, WHITESPACE
);
1526 p
[strcspn(p
, WHITESPACE
)] = 0;
1528 return parse_pid(p
, ret
);
1531 static int rlimit_to_nice(rlim_t limit
) {
1533 return PRIO_MAX
-1; /* i.e. 19 */
1535 if (limit
>= -PRIO_MIN
+ PRIO_MAX
)
1536 return PRIO_MIN
; /* i.e. -20 */
1538 return PRIO_MAX
- (int) limit
;
1541 int setpriority_closest(int priority
) {
1542 int current
, limit
, saved_errno
;
1543 struct rlimit highest
;
1545 /* Try to set requested nice level */
1546 if (setpriority(PRIO_PROCESS
, 0, priority
) >= 0)
1549 /* Permission failed */
1550 saved_errno
= -errno
;
1551 if (!ERRNO_IS_PRIVILEGE(saved_errno
))
1555 current
= getpriority(PRIO_PROCESS
, 0);
1559 if (priority
== current
)
1562 /* Hmm, we'd expect that raising the nice level from our status quo would always work. If it doesn't,
1563 * then the whole setpriority() system call is blocked to us, hence let's propagate the error
1565 if (priority
> current
)
1568 if (getrlimit(RLIMIT_NICE
, &highest
) < 0)
1571 limit
= rlimit_to_nice(highest
.rlim_cur
);
1573 /* We are already less nice than limit allows us */
1574 if (current
< limit
) {
1575 log_debug("Cannot raise nice level, permissions and the resource limit do not allow it.");
1579 /* Push to the allowed limit */
1580 if (setpriority(PRIO_PROCESS
, 0, limit
) < 0)
1583 log_debug("Cannot set requested nice level (%i), used next best (%i).", priority
, limit
);
1587 static const char *const ioprio_class_table
[] = {
1588 [IOPRIO_CLASS_NONE
] = "none",
1589 [IOPRIO_CLASS_RT
] = "realtime",
1590 [IOPRIO_CLASS_BE
] = "best-effort",
1591 [IOPRIO_CLASS_IDLE
] = "idle",
1594 DEFINE_STRING_TABLE_LOOKUP_WITH_FALLBACK(ioprio_class
, int, IOPRIO_N_CLASSES
);
1596 static const char *const sigchld_code_table
[] = {
1597 [CLD_EXITED
] = "exited",
1598 [CLD_KILLED
] = "killed",
1599 [CLD_DUMPED
] = "dumped",
1600 [CLD_TRAPPED
] = "trapped",
1601 [CLD_STOPPED
] = "stopped",
1602 [CLD_CONTINUED
] = "continued",
1605 DEFINE_STRING_TABLE_LOOKUP(sigchld_code
, int);
1607 static const char* const sched_policy_table
[] = {
1608 [SCHED_OTHER
] = "other",
1609 [SCHED_BATCH
] = "batch",
1610 [SCHED_IDLE
] = "idle",
1611 [SCHED_FIFO
] = "fifo",
1615 DEFINE_STRING_TABLE_LOOKUP_WITH_FALLBACK(sched_policy
, int, INT_MAX
);