]> git.ipfire.org Git - thirdparty/systemd.git/blob - src/core/manager.c
projects / thirdparty / systemd.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
tree-wide: "<n>bit" → "<n>-bit"
[thirdparty/systemd.git] / src / core / manager.c
1 /* SPDX-License-Identifier: LGPL-2.1-or-later */
2
3 #include <errno.h>
4 #include <fcntl.h>
5 #include <linux/kd.h>
6 #include <sys/epoll.h>
7 #include <sys/inotify.h>
8 #include <sys/ioctl.h>
9 #include <sys/mount.h>
10 #include <sys/reboot.h>
11 #include <sys/timerfd.h>
12 #include <sys/utsname.h>
13 #include <sys/wait.h>
14 #include <unistd.h>
15
16 #if HAVE_AUDIT
17 #include <libaudit.h>
18 #endif
19
20 #include "sd-daemon.h"
21 #include "sd-messages.h"
22 #include "sd-path.h"
23
24 #include "all-units.h"
25 #include "alloc-util.h"
26 #include "audit-fd.h"
27 #include "boot-timestamps.h"
28 #include "bus-common-errors.h"
29 #include "bus-error.h"
30 #include "bus-kernel.h"
31 #include "bus-util.h"
32 #include "clean-ipc.h"
33 #include "clock-util.h"
34 #include "common-signal.h"
35 #include "constants.h"
36 #include "core-varlink.h"
37 #include "creds-util.h"
38 #include "dbus-job.h"
39 #include "dbus-manager.h"
40 #include "dbus-unit.h"
41 #include "dbus.h"
42 #include "dirent-util.h"
43 #include "env-util.h"
44 #include "escape.h"
45 #include "event-util.h"
46 #include "exec-util.h"
47 #include "execute.h"
48 #include "exit-status.h"
49 #include "fd-util.h"
50 #include "fileio.h"
51 #include "generator-setup.h"
52 #include "hashmap.h"
53 #include "initrd-util.h"
54 #include "inotify-util.h"
55 #include "install.h"
56 #include "io-util.h"
57 #include "label-util.h"
58 #include "load-fragment.h"
59 #include "locale-setup.h"
60 #include "log.h"
61 #include "macro.h"
62 #include "manager.h"
63 #include "manager-dump.h"
64 #include "manager-serialize.h"
65 #include "memory-util.h"
66 #include "mkdir-label.h"
67 #include "mount-util.h"
68 #include "os-util.h"
69 #include "parse-util.h"
70 #include "path-lookup.h"
71 #include "path-util.h"
72 #include "process-util.h"
73 #include "psi-util.h"
74 #include "ratelimit.h"
75 #include "rlimit-util.h"
76 #include "rm-rf.h"
77 #include "selinux-util.h"
78 #include "signal-util.h"
79 #include "socket-util.h"
80 #include "special.h"
81 #include "stat-util.h"
82 #include "string-table.h"
83 #include "string-util.h"
84 #include "strv.h"
85 #include "strxcpyx.h"
86 #include "sysctl-util.h"
87 #include "syslog-util.h"
88 #include "terminal-util.h"
89 #include "time-util.h"
90 #include "transaction.h"
91 #include "uid-range.h"
92 #include "umask-util.h"
93 #include "unit-name.h"
94 #include "user-util.h"
95 #include "virt.h"
96 #include "watchdog.h"
97
98 #define NOTIFY_RCVBUF_SIZE (8*1024*1024)
99 #define CGROUPS_AGENT_RCVBUF_SIZE (8*1024*1024)
100
101 /* Initial delay and the interval for printing status messages about running jobs */
102 #define JOBS_IN_PROGRESS_WAIT_USEC (2*USEC_PER_SEC)
103 #define JOBS_IN_PROGRESS_QUIET_WAIT_USEC (25*USEC_PER_SEC)
104 #define JOBS_IN_PROGRESS_PERIOD_USEC (USEC_PER_SEC / 3)
105 #define JOBS_IN_PROGRESS_PERIOD_DIVISOR 3
106
107 /* If there are more than 1K bus messages queue across our API and direct buses, then let's not add more on top until
108 * the queue gets more empty. */
109 #define MANAGER_BUS_BUSY_THRESHOLD 1024LU
110
111 /* How many units and jobs to process of the bus queue before returning to the event loop. */
112 #define MANAGER_BUS_MESSAGE_BUDGET 100U
113
114 static int manager_dispatch_notify_fd(sd_event_source *source, int fd, uint32_t revents, void *userdata);
115 static int manager_dispatch_cgroups_agent_fd(sd_event_source *source, int fd, uint32_t revents, void *userdata);
116 static int manager_dispatch_signal_fd(sd_event_source *source, int fd, uint32_t revents, void *userdata);
117 static int manager_dispatch_time_change_fd(sd_event_source *source, int fd, uint32_t revents, void *userdata);
118 static int manager_dispatch_idle_pipe_fd(sd_event_source *source, int fd, uint32_t revents, void *userdata);
119 static int manager_dispatch_user_lookup_fd(sd_event_source *source, int fd, uint32_t revents, void *userdata);
120 static int manager_dispatch_jobs_in_progress(sd_event_source *source, usec_t usec, void *userdata);
121 static int manager_dispatch_run_queue(sd_event_source *source, void *userdata);
122 static int manager_dispatch_sigchld(sd_event_source *source, void *userdata);
123 static int manager_dispatch_timezone_change(sd_event_source *source, const struct inotify_event *event, void *userdata);
124 static int manager_run_environment_generators(Manager *m);
125 static int manager_run_generators(Manager *m);
126 static void manager_vacuum(Manager *m);
127
128 static usec_t manager_watch_jobs_next_time(Manager *m) {
129 usec_t timeout;
130
131 if (MANAGER_IS_USER(m))
132 /* Let the user manager without a timeout show status quickly, so the system manager can make
133 * use of it, if it wants to. */
134 timeout = JOBS_IN_PROGRESS_WAIT_USEC * 2 / 3;
135 else if (show_status_on(m->show_status))
136 /* When status is on, just use the usual timeout. */
137 timeout = JOBS_IN_PROGRESS_WAIT_USEC;
138 else
139 timeout = JOBS_IN_PROGRESS_QUIET_WAIT_USEC;
140
141 return usec_add(now(CLOCK_MONOTONIC), timeout);
142 }
143
144 static void manager_watch_jobs_in_progress(Manager *m) {
145 usec_t next;
146 int r;
147
148 assert(m);
149
150 /* We do not want to show the cylon animation if the user
151 * needs to confirm service executions otherwise confirmation
152 * messages will be screwed by the cylon animation. */
153 if (!manager_is_confirm_spawn_disabled(m))
154 return;
155
156 if (m->jobs_in_progress_event_source)
157 return;
158
159 next = manager_watch_jobs_next_time(m);
160 r = sd_event_add_time(
161 m->event,
162 &m->jobs_in_progress_event_source,
163 CLOCK_MONOTONIC,
164 next, 0,
165 manager_dispatch_jobs_in_progress, m);
166 if (r < 0)
167 return;
168
169 (void) sd_event_source_set_description(m->jobs_in_progress_event_source, "manager-jobs-in-progress");
170 }
171
172 #define CYLON_BUFFER_EXTRA (2*STRLEN(ANSI_RED) + STRLEN(ANSI_HIGHLIGHT_RED) + 2*STRLEN(ANSI_NORMAL))
173
174 static void draw_cylon(char buffer[], size_t buflen, unsigned width, unsigned pos) {
175 char *p = buffer;
176
177 assert(buflen >= CYLON_BUFFER_EXTRA + width + 1);
178 assert(pos <= width+1); /* 0 or width+1 mean that the center light is behind the corner */
179
180 if (pos > 1) {
181 if (pos > 2)
182 p = mempset(p, ' ', pos-2);
183 if (log_get_show_color())
184 p = stpcpy(p, ANSI_RED);
185 *p++ = '*';
186 }
187
188 if (pos > 0 && pos <= width) {
189 if (log_get_show_color())
190 p = stpcpy(p, ANSI_HIGHLIGHT_RED);
191 *p++ = '*';
192 }
193
194 if (log_get_show_color())
195 p = stpcpy(p, ANSI_NORMAL);
196
197 if (pos < width) {
198 if (log_get_show_color())
199 p = stpcpy(p, ANSI_RED);
200 *p++ = '*';
201 if (pos < width-1)
202 p = mempset(p, ' ', width-1-pos);
203 if (log_get_show_color())
204 strcpy(p, ANSI_NORMAL);
205 }
206 }
207
208 static void manager_flip_auto_status(Manager *m, bool enable, const char *reason) {
209 assert(m);
210
211 if (enable) {
212 if (m->show_status == SHOW_STATUS_AUTO)
213 manager_set_show_status(m, SHOW_STATUS_TEMPORARY, reason);
214 } else {
215 if (m->show_status == SHOW_STATUS_TEMPORARY)
216 manager_set_show_status(m, SHOW_STATUS_AUTO, reason);
217 }
218 }
219
220 static void manager_print_jobs_in_progress(Manager *m) {
221 Job *j;
222 unsigned counter = 0, print_nr;
223 char cylon[6 + CYLON_BUFFER_EXTRA + 1];
224 unsigned cylon_pos;
225 uint64_t timeout = 0;
226
227 assert(m);
228 assert(m->n_running_jobs > 0);
229
230 manager_flip_auto_status(m, true, "delay");
231
232 print_nr = (m->jobs_in_progress_iteration / JOBS_IN_PROGRESS_PERIOD_DIVISOR) % m->n_running_jobs;
233
234 HASHMAP_FOREACH(j, m->jobs)
235 if (j->state == JOB_RUNNING && counter++ == print_nr)
236 break;
237
238 /* m->n_running_jobs must be consistent with the contents of m->jobs,
239 * so the above loop must have succeeded in finding j. */
240 assert(counter == print_nr + 1);
241 assert(j);
242
243 cylon_pos = m->jobs_in_progress_iteration % 14;
244 if (cylon_pos >= 8)
245 cylon_pos = 14 - cylon_pos;
246 draw_cylon(cylon, sizeof(cylon), 6, cylon_pos);
247
248 m->jobs_in_progress_iteration++;
249
250 char job_of_n[STRLEN("( of ) ") + DECIMAL_STR_MAX(unsigned)*2] = "";
251 if (m->n_running_jobs > 1)
252 xsprintf(job_of_n, "(%u of %u) ", counter, m->n_running_jobs);
253
254 (void) job_get_timeout(j, &timeout);
255
256 /* We want to use enough information for the user to identify previous lines talking about the same
257 * unit, but keep the message as short as possible. So if 'Starting foo.service' or 'Starting
258 * foo.service - Description' were used, 'foo.service' is enough here. On the other hand, if we used
259 * 'Starting Description' before, then we shall also use 'Description' here. So we pass NULL as the
260 * second argument to unit_status_string(). */
261 const char *ident = unit_status_string(j->unit, NULL);
262
263 const char *time = FORMAT_TIMESPAN(now(CLOCK_MONOTONIC) - j->begin_usec, 1*USEC_PER_SEC);
264 const char *limit = timeout > 0 ? FORMAT_TIMESPAN(timeout - j->begin_usec, 1*USEC_PER_SEC) : "no limit";
265
266 if (m->status_unit_format == STATUS_UNIT_FORMAT_DESCRIPTION)
267 /* When using 'Description', we effectively don't have enough space to show the nested status
268 * without ellipsization, so let's not even try. */
269 manager_status_printf(m, STATUS_TYPE_EPHEMERAL, cylon,
270 "%sA %s job is running for %s (%s / %s)",
271 job_of_n,
272 job_type_to_string(j->type),
273 ident,
274 time, limit);
275 else {
276 const char *status_text = unit_status_text(j->unit);
277
278 manager_status_printf(m, STATUS_TYPE_EPHEMERAL, cylon,
279 "%sJob %s/%s running (%s / %s)%s%s",
280 job_of_n,
281 ident,
282 job_type_to_string(j->type),
283 time, limit,
284 status_text ? ": " : "",
285 strempty(status_text));
286 }
287
288 sd_notifyf(false,
289 "STATUS=%sUser job %s/%s running (%s / %s)...",
290 job_of_n,
291 ident,
292 job_type_to_string(j->type),
293 time, limit);
294 m->status_ready = false;
295 }
296
297 static int have_ask_password(void) {
298 _cleanup_closedir_ DIR *dir = NULL;
299
300 dir = opendir("/run/systemd/ask-password");
301 if (!dir) {
302 if (errno == ENOENT)
303 return false;
304 else
305 return -errno;
306 }
307
308 FOREACH_DIRENT_ALL(de, dir, return -errno)
309 if (startswith(de->d_name, "ask."))
310 return true;
311 return false;
312 }
313
314 static int manager_dispatch_ask_password_fd(sd_event_source *source,
315 int fd, uint32_t revents, void *userdata) {
316 Manager *m = ASSERT_PTR(userdata);
317
318 (void) flush_fd(fd);
319
320 m->have_ask_password = have_ask_password();
321 if (m->have_ask_password < 0)
322 /* Log error but continue. Negative have_ask_password
323 * is treated as unknown status. */
324 log_error_errno(m->have_ask_password, "Failed to list /run/systemd/ask-password: %m");
325
326 return 0;
327 }
328
329 static void manager_close_ask_password(Manager *m) {
330 assert(m);
331
332 m->ask_password_event_source = sd_event_source_disable_unref(m->ask_password_event_source);
333 m->ask_password_inotify_fd = safe_close(m->ask_password_inotify_fd);
334 m->have_ask_password = -EINVAL;
335 }
336
337 static int manager_check_ask_password(Manager *m) {
338 int r;
339
340 assert(m);
341
342 if (!m->ask_password_event_source) {
343 assert(m->ask_password_inotify_fd < 0);
344
345 (void) mkdir_p_label("/run/systemd/ask-password", 0755);
346
347 m->ask_password_inotify_fd = inotify_init1(IN_NONBLOCK|IN_CLOEXEC);
348 if (m->ask_password_inotify_fd < 0)
349 return log_error_errno(errno, "Failed to create inotify object: %m");
350
351 r = inotify_add_watch_and_warn(m->ask_password_inotify_fd,
352 "/run/systemd/ask-password",
353 IN_CREATE|IN_DELETE|IN_MOVE);
354 if (r < 0) {
355 manager_close_ask_password(m);
356 return r;
357 }
358
359 r = sd_event_add_io(m->event, &m->ask_password_event_source,
360 m->ask_password_inotify_fd, EPOLLIN,
361 manager_dispatch_ask_password_fd, m);
362 if (r < 0) {
363 log_error_errno(r, "Failed to add event source for /run/systemd/ask-password: %m");
364 manager_close_ask_password(m);
365 return r;
366 }
367
368 (void) sd_event_source_set_description(m->ask_password_event_source, "manager-ask-password");
369
370 /* Queries might have been added meanwhile... */
371 manager_dispatch_ask_password_fd(m->ask_password_event_source,
372 m->ask_password_inotify_fd, EPOLLIN, m);
373 }
374
375 return m->have_ask_password;
376 }
377
378 static int manager_watch_idle_pipe(Manager *m) {
379 int r;
380
381 assert(m);
382
383 if (m->idle_pipe_event_source)
384 return 0;
385
386 if (m->idle_pipe[2] < 0)
387 return 0;
388
389 r = sd_event_add_io(m->event, &m->idle_pipe_event_source, m->idle_pipe[2], EPOLLIN, manager_dispatch_idle_pipe_fd, m);
390 if (r < 0)
391 return log_error_errno(r, "Failed to watch idle pipe: %m");
392
393 (void) sd_event_source_set_description(m->idle_pipe_event_source, "manager-idle-pipe");
394
395 return 0;
396 }
397
398 static void manager_close_idle_pipe(Manager *m) {
399 assert(m);
400
401 m->idle_pipe_event_source = sd_event_source_disable_unref(m->idle_pipe_event_source);
402
403 safe_close_pair(m->idle_pipe);
404 safe_close_pair(m->idle_pipe + 2);
405 }
406
407 static int manager_setup_time_change(Manager *m) {
408 int r;
409
410 assert(m);
411
412 if (MANAGER_IS_TEST_RUN(m))
413 return 0;
414
415 m->time_change_event_source = sd_event_source_disable_unref(m->time_change_event_source);
416
417 r = event_add_time_change(m->event, &m->time_change_event_source, manager_dispatch_time_change_fd, m);
418 if (r < 0)
419 return log_error_errno(r, "Failed to create time change event source: %m");
420
421 /* Schedule this slightly earlier than the .timer event sources */
422 r = sd_event_source_set_priority(m->time_change_event_source, SD_EVENT_PRIORITY_NORMAL-1);
423 if (r < 0)
424 return log_error_errno(r, "Failed to set priority of time change event sources: %m");
425
426 log_debug("Set up TFD_TIMER_CANCEL_ON_SET timerfd.");
427
428 return 0;
429 }
430
431 static int manager_read_timezone_stat(Manager *m) {
432 struct stat st;
433 bool changed;
434
435 assert(m);
436
437 /* Read the current stat() data of /etc/localtime so that we detect changes */
438 if (lstat("/etc/localtime", &st) < 0) {
439 log_debug_errno(errno, "Failed to stat /etc/localtime, ignoring: %m");
440 changed = m->etc_localtime_accessible;
441 m->etc_localtime_accessible = false;
442 } else {
443 usec_t k;
444
445 k = timespec_load(&st.st_mtim);
446 changed = !m->etc_localtime_accessible || k != m->etc_localtime_mtime;
447
448 m->etc_localtime_mtime = k;
449 m->etc_localtime_accessible = true;
450 }
451
452 return changed;
453 }
454
455 static int manager_setup_timezone_change(Manager *m) {
456 _cleanup_(sd_event_source_unrefp) sd_event_source *new_event = NULL;
457 int r;
458
459 assert(m);
460
461 if (MANAGER_IS_TEST_RUN(m))
462 return 0;
463
464 /* We watch /etc/localtime for three events: change of the link count (which might mean removal from /etc even
465 * though another link might be kept), renames, and file close operations after writing. Note we don't bother
466 * with IN_DELETE_SELF, as that would just report when the inode is removed entirely, i.e. after the link count
467 * went to zero and all fds to it are closed.
468 *
469 * Note that we never follow symlinks here. This is a simplification, but should cover almost all cases
470 * correctly.
471 *
472 * Note that we create the new event source first here, before releasing the old one. This should optimize
473 * behaviour as this way sd-event can reuse the old watch in case the inode didn't change. */
474
475 r = sd_event_add_inotify(m->event, &new_event, "/etc/localtime",
476 IN_ATTRIB|IN_MOVE_SELF|IN_CLOSE_WRITE|IN_DONT_FOLLOW, manager_dispatch_timezone_change, m);
477 if (r == -ENOENT) {
478 /* If the file doesn't exist yet, subscribe to /etc instead, and wait until it is created either by
479 * O_CREATE or by rename() */
480
481 log_debug_errno(r, "/etc/localtime doesn't exist yet, watching /etc instead.");
482 r = sd_event_add_inotify(m->event, &new_event, "/etc",
483 IN_CREATE|IN_MOVED_TO|IN_ONLYDIR, manager_dispatch_timezone_change, m);
484 }
485 if (r < 0)
486 return log_error_errno(r, "Failed to create timezone change event source: %m");
487
488 /* Schedule this slightly earlier than the .timer event sources */
489 r = sd_event_source_set_priority(new_event, SD_EVENT_PRIORITY_NORMAL-1);
490 if (r < 0)
491 return log_error_errno(r, "Failed to set priority of timezone change event sources: %m");
492
493 sd_event_source_unref(m->timezone_change_event_source);
494 m->timezone_change_event_source = TAKE_PTR(new_event);
495
496 return 0;
497 }
498
499 static int enable_special_signals(Manager *m) {
500 _cleanup_close_ int fd = -EBADF;
501
502 assert(m);
503
504 if (MANAGER_IS_TEST_RUN(m))
505 return 0;
506
507 /* Enable that we get SIGINT on control-alt-del. In containers
508 * this will fail with EPERM (older) or EINVAL (newer), so
509 * ignore that. */
510 if (reboot(RB_DISABLE_CAD) < 0 && !IN_SET(errno, EPERM, EINVAL))
511 log_warning_errno(errno, "Failed to enable ctrl-alt-del handling: %m");
512
513 fd = open_terminal("/dev/tty0", O_RDWR|O_NOCTTY|O_CLOEXEC);
514 if (fd < 0) {
515 /* Support systems without virtual console */
516 if (fd != -ENOENT)
517 log_warning_errno(errno, "Failed to open /dev/tty0: %m");
518 } else {
519 /* Enable that we get SIGWINCH on kbrequest */
520 if (ioctl(fd, KDSIGACCEPT, SIGWINCH) < 0)
521 log_warning_errno(errno, "Failed to enable kbrequest handling: %m");
522 }
523
524 return 0;
525 }
526
527 #define RTSIG_IF_AVAILABLE(signum) (signum <= SIGRTMAX ? signum : -1)
528
529 static int manager_setup_signals(Manager *m) {
530 struct sigaction sa = {
531 .sa_handler = SIG_DFL,
532 .sa_flags = SA_NOCLDSTOP|SA_RESTART,
533 };
534 sigset_t mask;
535 int r;
536
537 assert(m);
538
539 assert_se(sigaction(SIGCHLD, &sa, NULL) == 0);
540
541 /* We make liberal use of realtime signals here. On
542 * Linux/glibc we have 30 of them (with the exception of Linux
543 * on hppa, see below), between SIGRTMIN+0 ... SIGRTMIN+30
544 * (aka SIGRTMAX). */
545
546 assert_se(sigemptyset(&mask) == 0);
547 sigset_add_many(&mask,
548 SIGCHLD, /* Child died */
549 SIGTERM, /* Reexecute daemon */
550 SIGHUP, /* Reload configuration */
551 SIGUSR1, /* systemd: reconnect to D-Bus */
552 SIGUSR2, /* systemd: dump status */
553 SIGINT, /* Kernel sends us this on control-alt-del */
554 SIGWINCH, /* Kernel sends us this on kbrequest (alt-arrowup) */
555 SIGPWR, /* Some kernel drivers and upsd send us this on power failure */
556
557 SIGRTMIN+0, /* systemd: start default.target */
558 SIGRTMIN+1, /* systemd: isolate rescue.target */
559 SIGRTMIN+2, /* systemd: isolate emergency.target */
560 SIGRTMIN+3, /* systemd: start halt.target */
561 SIGRTMIN+4, /* systemd: start poweroff.target */
562 SIGRTMIN+5, /* systemd: start reboot.target */
563 SIGRTMIN+6, /* systemd: start kexec.target */
564 SIGRTMIN+7, /* systemd: start soft-reboot.target */
565
566 /* ... space for more special targets ... */
567
568 SIGRTMIN+13, /* systemd: Immediate halt */
569 SIGRTMIN+14, /* systemd: Immediate poweroff */
570 SIGRTMIN+15, /* systemd: Immediate reboot */
571 SIGRTMIN+16, /* systemd: Immediate kexec */
572 SIGRTMIN+17, /* systemd: Immediate soft-reboot */
573 SIGRTMIN+18, /* systemd: control command */
574
575 /* ... space ... */
576
577 SIGRTMIN+20, /* systemd: enable status messages */
578 SIGRTMIN+21, /* systemd: disable status messages */
579 SIGRTMIN+22, /* systemd: set log level to LOG_DEBUG */
580 SIGRTMIN+23, /* systemd: set log level to LOG_INFO */
581 SIGRTMIN+24, /* systemd: Immediate exit (--user only) */
582 SIGRTMIN+25, /* systemd: reexecute manager */
583
584 /* Apparently Linux on hppa had fewer RT signals until v3.18,
585 * SIGRTMAX was SIGRTMIN+25, and then SIGRTMIN was lowered,
586 * see commit v3.17-7614-g1f25df2eff.
587 *
588 * We cannot unconditionally make use of those signals here,
589 * so let's use a runtime check. Since these commands are
590 * accessible by different means and only really a safety
591 * net, the missing functionality on hppa shouldn't matter.
592 */
593
594 RTSIG_IF_AVAILABLE(SIGRTMIN+26), /* systemd: set log target to journal-or-kmsg */
595 RTSIG_IF_AVAILABLE(SIGRTMIN+27), /* systemd: set log target to console */
596 RTSIG_IF_AVAILABLE(SIGRTMIN+28), /* systemd: set log target to kmsg */
597 RTSIG_IF_AVAILABLE(SIGRTMIN+29), /* systemd: set log target to syslog-or-kmsg (obsolete) */
598
599 /* ... one free signal here SIGRTMIN+30 ... */
600 -1);
601 assert_se(sigprocmask(SIG_SETMASK, &mask, NULL) == 0);
602
603 m->signal_fd = signalfd(-1, &mask, SFD_NONBLOCK|SFD_CLOEXEC);
604 if (m->signal_fd < 0)
605 return -errno;
606
607 r = sd_event_add_io(m->event, &m->signal_event_source, m->signal_fd, EPOLLIN, manager_dispatch_signal_fd, m);
608 if (r < 0)
609 return r;
610
611 (void) sd_event_source_set_description(m->signal_event_source, "manager-signal");
612
613 /* Process signals a bit earlier than the rest of things, but later than notify_fd processing, so that the
614 * notify processing can still figure out to which process/service a message belongs, before we reap the
615 * process. Also, process this before handling cgroup notifications, so that we always collect child exit
616 * status information before detecting that there's no process in a cgroup. */
617 r = sd_event_source_set_priority(m->signal_event_source, SD_EVENT_PRIORITY_NORMAL-6);
618 if (r < 0)
619 return r;
620
621 if (MANAGER_IS_SYSTEM(m))
622 return enable_special_signals(m);
623
624 return 0;
625 }
626
627 static char** sanitize_environment(char **l) {
628
629 /* Let's remove some environment variables that we need ourselves to communicate with our clients */
630 strv_env_unset_many(
631 l,
632 "CACHE_DIRECTORY",
633 "CONFIGURATION_DIRECTORY",
634 "CREDENTIALS_DIRECTORY",
635 "EXIT_CODE",
636 "EXIT_STATUS",
637 "INVOCATION_ID",
638 "JOURNAL_STREAM",
639 "LISTEN_FDNAMES",
640 "LISTEN_FDS",
641 "LISTEN_PID",
642 "LOGS_DIRECTORY",
643 "LOG_NAMESPACE",
644 "MAINPID",
645 "MANAGERPID",
646 "MEMORY_PRESSURE_WATCH",
647 "MEMORY_PRESSURE_WRITE",
648 "MONITOR_EXIT_CODE",
649 "MONITOR_EXIT_STATUS",
650 "MONITOR_INVOCATION_ID",
651 "MONITOR_SERVICE_RESULT",
652 "MONITOR_UNIT",
653 "NOTIFY_SOCKET",
654 "PIDFILE",
655 "REMOTE_ADDR",
656 "REMOTE_PORT",
657 "RUNTIME_DIRECTORY",
658 "SERVICE_RESULT",
659 "STATE_DIRECTORY",
660 "SYSTEMD_EXEC_PID",
661 "TRIGGER_PATH",
662 "TRIGGER_TIMER_MONOTONIC_USEC",
663 "TRIGGER_TIMER_REALTIME_USEC",
664 "TRIGGER_UNIT",
665 "WATCHDOG_PID",
666 "WATCHDOG_USEC",
667 NULL);
668
669 /* Let's order the environment alphabetically, just to make it pretty */
670 return strv_sort(l);
671 }
672
673 int manager_default_environment(Manager *m) {
674 int r;
675
676 assert(m);
677
678 m->transient_environment = strv_free(m->transient_environment);
679
680 if (MANAGER_IS_SYSTEM(m)) {
681 /* The system manager always starts with a clean environment for its children. It does not
682 * import the kernel's or the parents' exported variables.
683 *
684 * The initial passed environment is untouched to keep /proc/self/environ valid; it is used
685 * for tagging the init process inside containers. */
686 m->transient_environment = strv_new("PATH=" DEFAULT_PATH);
687 if (!m->transient_environment)
688 return log_oom();
689
690 /* Import locale variables LC_*= from configuration */
691 (void) locale_setup(&m->transient_environment);
692 } else {
693 /* The user manager passes its own environment along to its children, except for $PATH. */
694 m->transient_environment = strv_copy(environ);
695 if (!m->transient_environment)
696 return log_oom();
697
698 r = strv_env_replace_strdup(&m->transient_environment, "PATH=" DEFAULT_USER_PATH);
699 if (r < 0)
700 return log_oom();
701 }
702
703 sanitize_environment(m->transient_environment);
704 return 0;
705 }
706
707 static int manager_setup_prefix(Manager *m) {
708 struct table_entry {
709 uint64_t type;
710 const char *suffix;
711 };
712
713 static const struct table_entry paths_system[_EXEC_DIRECTORY_TYPE_MAX] = {
714 [EXEC_DIRECTORY_RUNTIME] = { SD_PATH_SYSTEM_RUNTIME, NULL },
715 [EXEC_DIRECTORY_STATE] = { SD_PATH_SYSTEM_STATE_PRIVATE, NULL },
716 [EXEC_DIRECTORY_CACHE] = { SD_PATH_SYSTEM_STATE_CACHE, NULL },
717 [EXEC_DIRECTORY_LOGS] = { SD_PATH_SYSTEM_STATE_LOGS, NULL },
718 [EXEC_DIRECTORY_CONFIGURATION] = { SD_PATH_SYSTEM_CONFIGURATION, NULL },
719 };
720
721 static const struct table_entry paths_user[_EXEC_DIRECTORY_TYPE_MAX] = {
722 [EXEC_DIRECTORY_RUNTIME] = { SD_PATH_USER_RUNTIME, NULL },
723 [EXEC_DIRECTORY_STATE] = { SD_PATH_USER_STATE_PRIVATE, NULL },
724 [EXEC_DIRECTORY_CACHE] = { SD_PATH_USER_STATE_CACHE, NULL },
725 [EXEC_DIRECTORY_LOGS] = { SD_PATH_USER_STATE_PRIVATE, "log" },
726 [EXEC_DIRECTORY_CONFIGURATION] = { SD_PATH_USER_CONFIGURATION, NULL },
727 };
728
729 assert(m);
730
731 const struct table_entry *p = MANAGER_IS_SYSTEM(m) ? paths_system : paths_user;
732 int r;
733
734 for (ExecDirectoryType i = 0; i < _EXEC_DIRECTORY_TYPE_MAX; i++) {
735 r = sd_path_lookup(p[i].type, p[i].suffix, &m->prefix[i]);
736 if (r < 0)
737 return log_warning_errno(r, "Failed to lookup %s path: %m",
738 exec_directory_type_to_string(i));
739 }
740
741 return 0;
742 }
743
744 static void manager_free_unit_name_maps(Manager *m) {
745 m->unit_id_map = hashmap_free(m->unit_id_map);
746 m->unit_name_map = hashmap_free(m->unit_name_map);
747 m->unit_path_cache = set_free(m->unit_path_cache);
748 m->unit_cache_timestamp_hash = 0;
749 }
750
751 static int manager_setup_run_queue(Manager *m) {
752 int r;
753
754 assert(m);
755 assert(!m->run_queue_event_source);
756
757 r = sd_event_add_defer(m->event, &m->run_queue_event_source, manager_dispatch_run_queue, m);
758 if (r < 0)
759 return r;
760
761 r = sd_event_source_set_priority(m->run_queue_event_source, SD_EVENT_PRIORITY_IDLE);
762 if (r < 0)
763 return r;
764
765 r = sd_event_source_set_enabled(m->run_queue_event_source, SD_EVENT_OFF);
766 if (r < 0)
767 return r;
768
769 (void) sd_event_source_set_description(m->run_queue_event_source, "manager-run-queue");
770
771 return 0;
772 }
773
774 static int manager_setup_sigchld_event_source(Manager *m) {
775 int r;
776
777 assert(m);
778 assert(!m->sigchld_event_source);
779
780 r = sd_event_add_defer(m->event, &m->sigchld_event_source, manager_dispatch_sigchld, m);
781 if (r < 0)
782 return r;
783
784 r = sd_event_source_set_priority(m->sigchld_event_source, SD_EVENT_PRIORITY_NORMAL-7);
785 if (r < 0)
786 return r;
787
788 r = sd_event_source_set_enabled(m->sigchld_event_source, SD_EVENT_OFF);
789 if (r < 0)
790 return r;
791
792 (void) sd_event_source_set_description(m->sigchld_event_source, "manager-sigchld");
793
794 return 0;
795 }
796
797 int manager_setup_memory_pressure_event_source(Manager *m) {
798 int r;
799
800 assert(m);
801
802 m->memory_pressure_event_source = sd_event_source_disable_unref(m->memory_pressure_event_source);
803
804 r = sd_event_add_memory_pressure(m->event, &m->memory_pressure_event_source, NULL, NULL);
805 if (r < 0)
806 log_full_errno(ERRNO_IS_NOT_SUPPORTED(r) || ERRNO_IS_PRIVILEGE(r) || (r == -EHOSTDOWN) ? LOG_DEBUG : LOG_NOTICE, r,
807 "Failed to establish memory pressure event source, ignoring: %m");
808 else if (m->default_memory_pressure_threshold_usec != USEC_INFINITY) {
809
810 /* If there's a default memory pressure threshold set, also apply it to the service manager itself */
811 r = sd_event_source_set_memory_pressure_period(
812 m->memory_pressure_event_source,
813 m->default_memory_pressure_threshold_usec,
814 MEMORY_PRESSURE_DEFAULT_WINDOW_USEC);
815 if (r < 0)
816 log_warning_errno(r, "Failed to adjust memory pressure threshold, ignoring: %m");
817 }
818
819 return 0;
820 }
821
822 static int manager_find_credentials_dirs(Manager *m) {
823 const char *e;
824 int r;
825
826 assert(m);
827
828 r = get_credentials_dir(&e);
829 if (r < 0) {
830 if (r != -ENXIO)
831 log_debug_errno(r, "Failed to determine credentials directory, ignoring: %m");
832 } else {
833 m->received_credentials_directory = strdup(e);
834 if (!m->received_credentials_directory)
835 return -ENOMEM;
836 }
837
838 r = get_encrypted_credentials_dir(&e);
839 if (r < 0) {
840 if (r != -ENXIO)
841 log_debug_errno(r, "Failed to determine encrypted credentials directory, ignoring: %m");
842 } else {
843 m->received_encrypted_credentials_directory = strdup(e);
844 if (!m->received_encrypted_credentials_directory)
845 return -ENOMEM;
846 }
847
848 return 0;
849 }
850
851 void manager_set_switching_root(Manager *m, bool switching_root) {
852 assert(m);
853
854 m->switching_root = MANAGER_IS_SYSTEM(m) && switching_root;
855 }
856
857 int manager_new(RuntimeScope runtime_scope, ManagerTestRunFlags test_run_flags, Manager **_m) {
858 _cleanup_(manager_freep) Manager *m = NULL;
859 int r;
860
861 assert(_m);
862 assert(IN_SET(runtime_scope, RUNTIME_SCOPE_SYSTEM, RUNTIME_SCOPE_USER));
863
864 m = new(Manager, 1);
865 if (!m)
866 return -ENOMEM;
867
868 *m = (Manager) {
869 .runtime_scope = runtime_scope,
870 .objective = _MANAGER_OBJECTIVE_INVALID,
871
872 .status_unit_format = STATUS_UNIT_FORMAT_DEFAULT,
873
874 .default_timer_accuracy_usec = USEC_PER_MINUTE,
875 .default_memory_accounting = MEMORY_ACCOUNTING_DEFAULT,
876 .default_tasks_accounting = true,
877 .default_tasks_max = TASKS_MAX_UNSET,
878 .default_timeout_start_usec = manager_default_timeout(runtime_scope),
879 .default_timeout_stop_usec = manager_default_timeout(runtime_scope),
880 .default_restart_usec = DEFAULT_RESTART_USEC,
881 .default_device_timeout_usec = manager_default_timeout(runtime_scope),
882
883 .original_log_level = -1,
884 .original_log_target = _LOG_TARGET_INVALID,
885
886 .watchdog_overridden[WATCHDOG_RUNTIME] = USEC_INFINITY,
887 .watchdog_overridden[WATCHDOG_REBOOT] = USEC_INFINITY,
888 .watchdog_overridden[WATCHDOG_KEXEC] = USEC_INFINITY,
889 .watchdog_overridden[WATCHDOG_PRETIMEOUT] = USEC_INFINITY,
890
891 .show_status_overridden = _SHOW_STATUS_INVALID,
892
893 .notify_fd = -EBADF,
894 .cgroups_agent_fd = -EBADF,
895 .signal_fd = -EBADF,
896 .user_lookup_fds = PIPE_EBADF,
897 .private_listen_fd = -EBADF,
898 .dev_autofs_fd = -EBADF,
899 .cgroup_inotify_fd = -EBADF,
900 .pin_cgroupfs_fd = -EBADF,
901 .ask_password_inotify_fd = -EBADF,
902 .idle_pipe = { -EBADF, -EBADF, -EBADF, -EBADF},
903
904 /* start as id #1, so that we can leave #0 around as "null-like" value */
905 .current_job_id = 1,
906
907 .have_ask_password = -EINVAL, /* we don't know */
908 .first_boot = -1,
909 .test_run_flags = test_run_flags,
910
911 .default_oom_policy = OOM_STOP,
912
913 .default_memory_pressure_watch = CGROUP_PRESSURE_WATCH_AUTO,
914 .default_memory_pressure_threshold_usec = USEC_INFINITY,
915
916 .dump_ratelimit = {
917 .interval = 10 * USEC_PER_MINUTE,
918 .burst = 10,
919 },
920 };
921
922 #if ENABLE_EFI
923 if (MANAGER_IS_SYSTEM(m) && detect_container() <= 0)
924 boot_timestamps(m->timestamps + MANAGER_TIMESTAMP_USERSPACE,
925 m->timestamps + MANAGER_TIMESTAMP_FIRMWARE,
926 m->timestamps + MANAGER_TIMESTAMP_LOADER);
927 #endif
928
929 /* Prepare log fields we can use for structured logging */
930 if (MANAGER_IS_SYSTEM(m)) {
931 m->unit_log_field = "UNIT=";
932 m->unit_log_format_string = "UNIT=%s";
933
934 m->invocation_log_field = "INVOCATION_ID=";
935 m->invocation_log_format_string = "INVOCATION_ID=%s";
936 } else {
937 m->unit_log_field = "USER_UNIT=";
938 m->unit_log_format_string = "USER_UNIT=%s";
939
940 m->invocation_log_field = "USER_INVOCATION_ID=";
941 m->invocation_log_format_string = "USER_INVOCATION_ID=%s";
942 }
943
944 /* Reboot immediately if the user hits C-A-D more often than 7x per 2s */
945 m->ctrl_alt_del_ratelimit = (const RateLimit) { .interval = 2 * USEC_PER_SEC, .burst = 7 };
946
947 r = manager_default_environment(m);
948 if (r < 0)
949 return r;
950
951 r = hashmap_ensure_allocated(&m->units, &string_hash_ops);
952 if (r < 0)
953 return r;
954
955 r = hashmap_ensure_allocated(&m->cgroup_unit, &path_hash_ops);
956 if (r < 0)
957 return r;
958
959 r = hashmap_ensure_allocated(&m->watch_bus, &string_hash_ops);
960 if (r < 0)
961 return r;
962
963 r = prioq_ensure_allocated(&m->run_queue, compare_job_priority);
964 if (r < 0)
965 return r;
966
967 r = manager_setup_prefix(m);
968 if (r < 0)
969 return r;
970
971 r = manager_find_credentials_dirs(m);
972 if (r < 0)
973 return r;
974
975 r = sd_event_default(&m->event);
976 if (r < 0)
977 return r;
978
979 r = manager_setup_run_queue(m);
980 if (r < 0)
981 return r;
982
983 if (FLAGS_SET(test_run_flags, MANAGER_TEST_RUN_MINIMAL)) {
984 m->cgroup_root = strdup("");
985 if (!m->cgroup_root)
986 return -ENOMEM;
987 } else {
988 r = manager_setup_signals(m);
989 if (r < 0)
990 return r;
991
992 r = manager_setup_cgroup(m);
993 if (r < 0)
994 return r;
995
996 r = manager_setup_time_change(m);
997 if (r < 0)
998 return r;
999
1000 r = manager_read_timezone_stat(m);
1001 if (r < 0)
1002 return r;
1003
1004 (void) manager_setup_timezone_change(m);
1005
1006 r = manager_setup_sigchld_event_source(m);
1007 if (r < 0)
1008 return r;
1009
1010 r = manager_setup_memory_pressure_event_source(m);
1011 if (r < 0)
1012 return r;
1013
1014 #if HAVE_LIBBPF
1015 if (MANAGER_IS_SYSTEM(m) && lsm_bpf_supported(/* initialize = */ true)) {
1016 r = lsm_bpf_setup(m);
1017 if (r < 0)
1018 log_warning_errno(r, "Failed to setup LSM BPF, ignoring: %m");
1019 }
1020 #endif
1021 }
1022
1023 if (test_run_flags == 0) {
1024 if (MANAGER_IS_SYSTEM(m))
1025 r = mkdir_label("/run/systemd/units", 0755);
1026 else {
1027 _cleanup_free_ char *units_path = NULL;
1028 r = xdg_user_runtime_dir(&units_path, "/systemd/units");
1029 if (r < 0)
1030 return r;
1031 r = mkdir_p_label(units_path, 0755);
1032 }
1033
1034 if (r < 0 && r != -EEXIST)
1035 return r;
1036 }
1037
1038 m->taint_usr =
1039 !in_initrd() &&
1040 dir_is_empty("/usr", /* ignore_hidden_or_backup= */ false) > 0;
1041
1042 /* Note that we do not set up the notify fd here. We do that after deserialization,
1043 * since they might have gotten serialized across the reexec. */
1044
1045 *_m = TAKE_PTR(m);
1046
1047 return 0;
1048 }
1049
1050 static int manager_setup_notify(Manager *m) {
1051 int r;
1052
1053 if (MANAGER_IS_TEST_RUN(m))
1054 return 0;
1055
1056 if (m->notify_fd < 0) {
1057 _cleanup_close_ int fd = -EBADF;
1058 union sockaddr_union sa;
1059 socklen_t sa_len;
1060
1061 /* First free all secondary fields */
1062 m->notify_socket = mfree(m->notify_socket);
1063 m->notify_event_source = sd_event_source_disable_unref(m->notify_event_source);
1064
1065 fd = socket(AF_UNIX, SOCK_DGRAM|SOCK_CLOEXEC|SOCK_NONBLOCK, 0);
1066 if (fd < 0)
1067 return log_error_errno(errno, "Failed to allocate notification socket: %m");
1068
1069 fd_increase_rxbuf(fd, NOTIFY_RCVBUF_SIZE);
1070
1071 m->notify_socket = path_join(m->prefix[EXEC_DIRECTORY_RUNTIME], "systemd/notify");
1072 if (!m->notify_socket)
1073 return log_oom();
1074
1075 r = sockaddr_un_set_path(&sa.un, m->notify_socket);
1076 if (r < 0)
1077 return log_error_errno(r, "Notify socket '%s' not valid for AF_UNIX socket address, refusing.",
1078 m->notify_socket);
1079 sa_len = r;
1080
1081 (void) mkdir_parents_label(m->notify_socket, 0755);
1082 (void) sockaddr_un_unlink(&sa.un);
1083
1084 r = mac_selinux_bind(fd, &sa.sa, sa_len);
1085 if (r < 0)
1086 return log_error_errno(r, "bind(%s) failed: %m", m->notify_socket);
1087
1088 r = setsockopt_int(fd, SOL_SOCKET, SO_PASSCRED, true);
1089 if (r < 0)
1090 return log_error_errno(r, "SO_PASSCRED failed: %m");
1091
1092 m->notify_fd = TAKE_FD(fd);
1093
1094 log_debug("Using notification socket %s", m->notify_socket);
1095 }
1096
1097 if (!m->notify_event_source) {
1098 r = sd_event_add_io(m->event, &m->notify_event_source, m->notify_fd, EPOLLIN, manager_dispatch_notify_fd, m);
1099 if (r < 0)
1100 return log_error_errno(r, "Failed to allocate notify event source: %m");
1101
1102 /* Process notification messages a bit earlier than SIGCHLD, so that we can still identify to which
1103 * service an exit message belongs. */
1104 r = sd_event_source_set_priority(m->notify_event_source, SD_EVENT_PRIORITY_NORMAL-8);
1105 if (r < 0)
1106 return log_error_errno(r, "Failed to set priority of notify event source: %m");
1107
1108 (void) sd_event_source_set_description(m->notify_event_source, "manager-notify");
1109 }
1110
1111 return 0;
1112 }
1113
1114 static int manager_setup_cgroups_agent(Manager *m) {
1115
1116 static const union sockaddr_union sa = {
1117 .un.sun_family = AF_UNIX,
1118 .un.sun_path = "/run/systemd/cgroups-agent",
1119 };
1120 int r;
1121
1122 /* This creates a listening socket we receive cgroups agent messages on. We do not use D-Bus for delivering
1123 * these messages from the cgroups agent binary to PID 1, as the cgroups agent binary is very short-living, and
1124 * each instance of it needs a new D-Bus connection. Since D-Bus connections are SOCK_STREAM/AF_UNIX, on
1125 * overloaded systems the backlog of the D-Bus socket becomes relevant, as not more than the configured number
1126 * of D-Bus connections may be queued until the kernel will start dropping further incoming connections,
1127 * possibly resulting in lost cgroups agent messages. To avoid this, we'll use a private SOCK_DGRAM/AF_UNIX
1128 * socket, where no backlog is relevant as communication may take place without an actual connect() cycle, and
1129 * we thus won't lose messages.
1130 *
1131 * Note that PID 1 will forward the agent message to system bus, so that the user systemd instance may listen
1132 * to it. The system instance hence listens on this special socket, but the user instances listen on the system
1133 * bus for these messages. */
1134
1135 if (MANAGER_IS_TEST_RUN(m))
1136 return 0;
1137
1138 if (!MANAGER_IS_SYSTEM(m))
1139 return 0;
1140
1141 r = cg_unified_controller(SYSTEMD_CGROUP_CONTROLLER);
1142 if (r < 0)
1143 return log_error_errno(r, "Failed to determine whether unified cgroups hierarchy is used: %m");
1144 if (r > 0) /* We don't need this anymore on the unified hierarchy */
1145 return 0;
1146
1147 if (m->cgroups_agent_fd < 0) {
1148 _cleanup_close_ int fd = -EBADF;
1149
1150 /* First free all secondary fields */
1151 m->cgroups_agent_event_source = sd_event_source_disable_unref(m->cgroups_agent_event_source);
1152
1153 fd = socket(AF_UNIX, SOCK_DGRAM|SOCK_CLOEXEC|SOCK_NONBLOCK, 0);
1154 if (fd < 0)
1155 return log_error_errno(errno, "Failed to allocate cgroups agent socket: %m");
1156
1157 fd_increase_rxbuf(fd, CGROUPS_AGENT_RCVBUF_SIZE);
1158
1159 (void) sockaddr_un_unlink(&sa.un);
1160
1161 /* Only allow root to connect to this socket */
1162 WITH_UMASK(0077)
1163 r = bind(fd, &sa.sa, SOCKADDR_UN_LEN(sa.un));
1164 if (r < 0)
1165 return log_error_errno(errno, "bind(%s) failed: %m", sa.un.sun_path);
1166
1167 m->cgroups_agent_fd = TAKE_FD(fd);
1168 }
1169
1170 if (!m->cgroups_agent_event_source) {
1171 r = sd_event_add_io(m->event, &m->cgroups_agent_event_source, m->cgroups_agent_fd, EPOLLIN, manager_dispatch_cgroups_agent_fd, m);
1172 if (r < 0)
1173 return log_error_errno(r, "Failed to allocate cgroups agent event source: %m");
1174
1175 /* Process cgroups notifications early. Note that when the agent notification is received
1176 * we'll just enqueue the unit in the cgroup empty queue, hence pick a high priority than
1177 * that. Also see handling of cgroup inotify for the unified cgroup stuff. */
1178 r = sd_event_source_set_priority(m->cgroups_agent_event_source, SD_EVENT_PRIORITY_NORMAL-9);
1179 if (r < 0)
1180 return log_error_errno(r, "Failed to set priority of cgroups agent event source: %m");
1181
1182 (void) sd_event_source_set_description(m->cgroups_agent_event_source, "manager-cgroups-agent");
1183 }
1184
1185 return 0;
1186 }
1187
1188 static int manager_setup_user_lookup_fd(Manager *m) {
1189 int r;
1190
1191 assert(m);
1192
1193 /* Set up the socket pair used for passing UID/GID resolution results from forked off processes to PID
1194 * 1. Background: we can't do name lookups (NSS) from PID 1, since it might involve IPC and thus activation,
1195 * and we might hence deadlock on ourselves. Hence we do all user/group lookups asynchronously from the forked
1196 * off processes right before executing the binaries to start. In order to be able to clean up any IPC objects
1197 * created by a unit (see RemoveIPC=) we need to know in PID 1 the used UID/GID of the executed processes,
1198 * hence we establish this communication channel so that forked off processes can pass their UID/GID
1199 * information back to PID 1. The forked off processes send their resolved UID/GID to PID 1 in a simple
1200 * datagram, along with their unit name, so that we can share one communication socket pair among all units for
1201 * this purpose.
1202 *
1203 * You might wonder why we need a communication channel for this that is independent of the usual notification
1204 * socket scheme (i.e. $NOTIFY_SOCKET). The primary difference is about trust: data sent via the $NOTIFY_SOCKET
1205 * channel is only accepted if it originates from the right unit and if reception was enabled for it. The user
1206 * lookup socket OTOH is only accessible by PID 1 and its children until they exec(), and always available.
1207 *
1208 * Note that this function is called under two circumstances: when we first initialize (in which case we
1209 * allocate both the socket pair and the event source to listen on it), and when we deserialize after a reload
1210 * (in which case the socket pair already exists but we still need to allocate the event source for it). */
1211
1212 if (m->user_lookup_fds[0] < 0) {
1213
1214 /* Free all secondary fields */
1215 safe_close_pair(m->user_lookup_fds);
1216 m->user_lookup_event_source = sd_event_source_disable_unref(m->user_lookup_event_source);
1217
1218 if (socketpair(AF_UNIX, SOCK_DGRAM|SOCK_CLOEXEC, 0, m->user_lookup_fds) < 0)
1219 return log_error_errno(errno, "Failed to allocate user lookup socket: %m");
1220
1221 (void) fd_increase_rxbuf(m->user_lookup_fds[0], NOTIFY_RCVBUF_SIZE);
1222 }
1223
1224 if (!m->user_lookup_event_source) {
1225 r = sd_event_add_io(m->event, &m->user_lookup_event_source, m->user_lookup_fds[0], EPOLLIN, manager_dispatch_user_lookup_fd, m);
1226 if (r < 0)
1227 return log_error_errno(errno, "Failed to allocate user lookup event source: %m");
1228
1229 /* Process even earlier than the notify event source, so that we always know first about valid UID/GID
1230 * resolutions */
1231 r = sd_event_source_set_priority(m->user_lookup_event_source, SD_EVENT_PRIORITY_NORMAL-11);
1232 if (r < 0)
1233 return log_error_errno(errno, "Failed to set priority of user lookup event source: %m");
1234
1235 (void) sd_event_source_set_description(m->user_lookup_event_source, "user-lookup");
1236 }
1237
1238 return 0;
1239 }
1240
1241 static unsigned manager_dispatch_cleanup_queue(Manager *m) {
1242 Unit *u;
1243 unsigned n = 0;
1244
1245 assert(m);
1246
1247 while ((u = m->cleanup_queue)) {
1248 assert(u->in_cleanup_queue);
1249
1250 unit_free(u);
1251 n++;
1252 }
1253
1254 return n;
1255 }
1256
1257 static unsigned manager_dispatch_release_resources_queue(Manager *m) {
1258 unsigned n = 0;
1259 Unit *u;
1260
1261 assert(m);
1262
1263 while ((u = m->release_resources_queue)) {
1264 assert(u->in_release_resources_queue);
1265
1266 LIST_REMOVE(release_resources_queue, m->release_resources_queue, u);
1267 u->in_release_resources_queue = false;
1268
1269 n++;
1270
1271 unit_release_resources(u);
1272 }
1273
1274 return n;
1275 }
1276
1277 enum {
1278 GC_OFFSET_IN_PATH, /* This one is on the path we were traveling */
1279 GC_OFFSET_UNSURE, /* No clue */
1280 GC_OFFSET_GOOD, /* We still need this unit */
1281 GC_OFFSET_BAD, /* We don't need this unit anymore */
1282 _GC_OFFSET_MAX
1283 };
1284
1285 static void unit_gc_mark_good(Unit *u, unsigned gc_marker) {
1286 Unit *other;
1287
1288 u->gc_marker = gc_marker + GC_OFFSET_GOOD;
1289
1290 /* Recursively mark referenced units as GOOD as well */
1291 UNIT_FOREACH_DEPENDENCY(other, u, UNIT_ATOM_REFERENCES)
1292 if (other->gc_marker == gc_marker + GC_OFFSET_UNSURE)
1293 unit_gc_mark_good(other, gc_marker);
1294 }
1295
1296 static void unit_gc_sweep(Unit *u, unsigned gc_marker) {
1297 Unit *other;
1298 bool is_bad;
1299
1300 assert(u);
1301
1302 if (IN_SET(u->gc_marker - gc_marker,
1303 GC_OFFSET_GOOD, GC_OFFSET_BAD, GC_OFFSET_UNSURE, GC_OFFSET_IN_PATH))
1304 return;
1305
1306 if (u->in_cleanup_queue)
1307 goto bad;
1308
1309 if (!unit_may_gc(u))
1310 goto good;
1311
1312 u->gc_marker = gc_marker + GC_OFFSET_IN_PATH;
1313
1314 is_bad = true;
1315
1316 UNIT_FOREACH_DEPENDENCY(other, u, UNIT_ATOM_REFERENCED_BY) {
1317 unit_gc_sweep(other, gc_marker);
1318
1319 if (other->gc_marker == gc_marker + GC_OFFSET_GOOD)
1320 goto good;
1321
1322 if (other->gc_marker != gc_marker + GC_OFFSET_BAD)
1323 is_bad = false;
1324 }
1325
1326 LIST_FOREACH(refs_by_target, ref, u->refs_by_target) {
1327 unit_gc_sweep(ref->source, gc_marker);
1328
1329 if (ref->source->gc_marker == gc_marker + GC_OFFSET_GOOD)
1330 goto good;
1331
1332 if (ref->source->gc_marker != gc_marker + GC_OFFSET_BAD)
1333 is_bad = false;
1334 }
1335
1336 if (is_bad)
1337 goto bad;
1338
1339 /* We were unable to find anything out about this entry, so
1340 * let's investigate it later */
1341 u->gc_marker = gc_marker + GC_OFFSET_UNSURE;
1342 unit_add_to_gc_queue(u);
1343 return;
1344
1345 bad:
1346 /* We definitely know that this one is not useful anymore, so
1347 * let's mark it for deletion */
1348 u->gc_marker = gc_marker + GC_OFFSET_BAD;
1349 unit_add_to_cleanup_queue(u);
1350 return;
1351
1352 good:
1353 unit_gc_mark_good(u, gc_marker);
1354 }
1355
1356 static unsigned manager_dispatch_gc_unit_queue(Manager *m) {
1357 unsigned n = 0, gc_marker;
1358 Unit *u;
1359
1360 assert(m);
1361
1362 /* log_debug("Running GC..."); */
1363
1364 m->gc_marker += _GC_OFFSET_MAX;
1365 if (m->gc_marker + _GC_OFFSET_MAX <= _GC_OFFSET_MAX)
1366 m->gc_marker = 1;
1367
1368 gc_marker = m->gc_marker;
1369
1370 while ((u = m->gc_unit_queue)) {
1371 assert(u->in_gc_queue);
1372
1373 unit_gc_sweep(u, gc_marker);
1374
1375 LIST_REMOVE(gc_queue, m->gc_unit_queue, u);
1376 u->in_gc_queue = false;
1377
1378 n++;
1379
1380 if (IN_SET(u->gc_marker - gc_marker,
1381 GC_OFFSET_BAD, GC_OFFSET_UNSURE)) {
1382 if (u->id)
1383 log_unit_debug(u, "Collecting.");
1384 u->gc_marker = gc_marker + GC_OFFSET_BAD;
1385 unit_add_to_cleanup_queue(u);
1386 }
1387 }
1388
1389 return n;
1390 }
1391
1392 static unsigned manager_dispatch_gc_job_queue(Manager *m) {
1393 unsigned n = 0;
1394 Job *j;
1395
1396 assert(m);
1397
1398 while ((j = m->gc_job_queue)) {
1399 assert(j->in_gc_queue);
1400
1401 LIST_REMOVE(gc_queue, m->gc_job_queue, j);
1402 j->in_gc_queue = false;
1403
1404 n++;
1405
1406 if (!job_may_gc(j))
1407 continue;
1408
1409 log_unit_debug(j->unit, "Collecting job.");
1410 (void) job_finish_and_invalidate(j, JOB_COLLECTED, false, false);
1411 }
1412
1413 return n;
1414 }
1415
1416 static int manager_ratelimit_requeue(sd_event_source *s, uint64_t usec, void *userdata) {
1417 Unit *u = userdata;
1418
1419 assert(u);
1420 assert(s == u->auto_start_stop_event_source);
1421
1422 u->auto_start_stop_event_source = sd_event_source_unref(u->auto_start_stop_event_source);
1423
1424 /* Re-queue to all queues, if the rate limit hit we might have been throttled on any of them. */
1425 unit_submit_to_stop_when_unneeded_queue(u);
1426 unit_submit_to_start_when_upheld_queue(u);
1427 unit_submit_to_stop_when_bound_queue(u);
1428
1429 return 0;
1430 }
1431
1432 static int manager_ratelimit_check_and_queue(Unit *u) {
1433 int r;
1434
1435 assert(u);
1436
1437 if (ratelimit_below(&u->auto_start_stop_ratelimit))
1438 return 1;
1439
1440 /* Already queued, no need to requeue */
1441 if (u->auto_start_stop_event_source)
1442 return 0;
1443
1444 r = sd_event_add_time(
1445 u->manager->event,
1446 &u->auto_start_stop_event_source,
1447 CLOCK_MONOTONIC,
1448 ratelimit_end(&u->auto_start_stop_ratelimit),
1449 0,
1450 manager_ratelimit_requeue,
1451 u);
1452 if (r < 0)
1453 return log_unit_error_errno(u, r, "Failed to queue timer on event loop: %m");
1454
1455 return 0;
1456 }
1457
1458 static unsigned manager_dispatch_stop_when_unneeded_queue(Manager *m) {
1459 unsigned n = 0;
1460 Unit *u;
1461 int r;
1462
1463 assert(m);
1464
1465 while ((u = m->stop_when_unneeded_queue)) {
1466 _cleanup_(sd_bus_error_free) sd_bus_error error = SD_BUS_ERROR_NULL;
1467
1468 assert(u->in_stop_when_unneeded_queue);
1469 LIST_REMOVE(stop_when_unneeded_queue, m->stop_when_unneeded_queue, u);
1470 u->in_stop_when_unneeded_queue = false;
1471
1472 n++;
1473
1474 if (!unit_is_unneeded(u))
1475 continue;
1476
1477 log_unit_debug(u, "Unit is not needed anymore.");
1478
1479 /* If stopping a unit fails continuously we might enter a stop loop here, hence stop acting on the
1480 * service being unnecessary after a while. */
1481
1482 r = manager_ratelimit_check_and_queue(u);
1483 if (r <= 0) {
1484 log_unit_warning(u,
1485 "Unit not needed anymore, but not stopping since we tried this too often recently.%s",
1486 r == 0 ? " Will retry later." : "");
1487 continue;
1488 }
1489
1490 /* Ok, nobody needs us anymore. Sniff. Then let's commit suicide */
1491 r = manager_add_job(u->manager, JOB_STOP, u, JOB_FAIL, NULL, &error, NULL);
1492 if (r < 0)
1493 log_unit_warning_errno(u, r, "Failed to enqueue stop job, ignoring: %s", bus_error_message(&error, r));
1494 }
1495
1496 return n;
1497 }
1498
1499 static unsigned manager_dispatch_start_when_upheld_queue(Manager *m) {
1500 unsigned n = 0;
1501 Unit *u;
1502 int r;
1503
1504 assert(m);
1505
1506 while ((u = m->start_when_upheld_queue)) {
1507 _cleanup_(sd_bus_error_free) sd_bus_error error = SD_BUS_ERROR_NULL;
1508 Unit *culprit = NULL;
1509
1510 assert(u->in_start_when_upheld_queue);
1511 LIST_REMOVE(start_when_upheld_queue, m->start_when_upheld_queue, u);
1512 u->in_start_when_upheld_queue = false;
1513
1514 n++;
1515
1516 if (!unit_is_upheld_by_active(u, &culprit))
1517 continue;
1518
1519 log_unit_debug(u, "Unit is started because upheld by active unit %s.", culprit->id);
1520
1521 /* If stopping a unit fails continuously we might enter a stop loop here, hence stop acting on the
1522 * service being unnecessary after a while. */
1523
1524 r = manager_ratelimit_check_and_queue(u);
1525 if (r <= 0) {
1526 log_unit_warning(u,
1527 "Unit needs to be started because active unit %s upholds it, but not starting since we tried this too often recently.%s",
1528 culprit->id,
1529 r == 0 ? " Will retry later." : "");
1530 continue;
1531 }
1532
1533 r = manager_add_job(u->manager, JOB_START, u, JOB_FAIL, NULL, &error, NULL);
1534 if (r < 0)
1535 log_unit_warning_errno(u, r, "Failed to enqueue start job, ignoring: %s", bus_error_message(&error, r));
1536 }
1537
1538 return n;
1539 }
1540
1541 static unsigned manager_dispatch_stop_when_bound_queue(Manager *m) {
1542 unsigned n = 0;
1543 Unit *u;
1544 int r;
1545
1546 assert(m);
1547
1548 while ((u = m->stop_when_bound_queue)) {
1549 _cleanup_(sd_bus_error_free) sd_bus_error error = SD_BUS_ERROR_NULL;
1550 Unit *culprit = NULL;
1551
1552 assert(u->in_stop_when_bound_queue);
1553 LIST_REMOVE(stop_when_bound_queue, m->stop_when_bound_queue, u);
1554 u->in_stop_when_bound_queue = false;
1555
1556 n++;
1557
1558 if (!unit_is_bound_by_inactive(u, &culprit))
1559 continue;
1560
1561 log_unit_debug(u, "Unit is stopped because bound to inactive unit %s.", culprit->id);
1562
1563 /* If stopping a unit fails continuously we might enter a stop loop here, hence stop acting on the
1564 * service being unnecessary after a while. */
1565
1566 r = manager_ratelimit_check_and_queue(u);
1567 if (r <= 0) {
1568 log_unit_warning(u,
1569 "Unit needs to be stopped because it is bound to inactive unit %s it, but not stopping since we tried this too often recently.%s",
1570 culprit->id,
1571 r == 0 ? " Will retry later." : "");
1572 continue;
1573 }
1574
1575 r = manager_add_job(u->manager, JOB_STOP, u, JOB_REPLACE, NULL, &error, NULL);
1576 if (r < 0)
1577 log_unit_warning_errno(u, r, "Failed to enqueue stop job, ignoring: %s", bus_error_message(&error, r));
1578 }
1579
1580 return n;
1581 }
1582
1583 static void manager_clear_jobs_and_units(Manager *m) {
1584 Unit *u;
1585
1586 assert(m);
1587
1588 while ((u = hashmap_first(m->units)))
1589 unit_free(u);
1590
1591 manager_dispatch_cleanup_queue(m);
1592
1593 assert(!m->load_queue);
1594 assert(prioq_isempty(m->run_queue));
1595 assert(!m->dbus_unit_queue);
1596 assert(!m->dbus_job_queue);
1597 assert(!m->cleanup_queue);
1598 assert(!m->gc_unit_queue);
1599 assert(!m->gc_job_queue);
1600 assert(!m->cgroup_realize_queue);
1601 assert(!m->cgroup_empty_queue);
1602 assert(!m->cgroup_oom_queue);
1603 assert(!m->target_deps_queue);
1604 assert(!m->stop_when_unneeded_queue);
1605 assert(!m->start_when_upheld_queue);
1606 assert(!m->stop_when_bound_queue);
1607 assert(!m->release_resources_queue);
1608
1609 assert(hashmap_isempty(m->jobs));
1610 assert(hashmap_isempty(m->units));
1611
1612 m->n_on_console = 0;
1613 m->n_running_jobs = 0;
1614 m->n_installed_jobs = 0;
1615 m->n_failed_jobs = 0;
1616 }
1617
1618 Manager* manager_free(Manager *m) {
1619 if (!m)
1620 return NULL;
1621
1622 manager_clear_jobs_and_units(m);
1623
1624 for (UnitType c = 0; c < _UNIT_TYPE_MAX; c++)
1625 if (unit_vtable[c]->shutdown)
1626 unit_vtable[c]->shutdown(m);
1627
1628 /* Keep the cgroup hierarchy in place except when we know we are going down for good */
1629 manager_shutdown_cgroup(m, /* delete= */ IN_SET(m->objective, MANAGER_EXIT, MANAGER_REBOOT, MANAGER_POWEROFF, MANAGER_HALT, MANAGER_KEXEC));
1630
1631 lookup_paths_flush_generator(&m->lookup_paths);
1632
1633 bus_done(m);
1634 manager_varlink_done(m);
1635
1636 exec_shared_runtime_vacuum(m);
1637 hashmap_free(m->exec_shared_runtime_by_id);
1638
1639 dynamic_user_vacuum(m, false);
1640 hashmap_free(m->dynamic_users);
1641
1642 hashmap_free(m->units);
1643 hashmap_free(m->units_by_invocation_id);
1644 hashmap_free(m->jobs);
1645 hashmap_free(m->watch_pids);
1646 hashmap_free(m->watch_bus);
1647
1648 prioq_free(m->run_queue);
1649
1650 set_free(m->startup_units);
1651 set_free(m->failed_units);
1652
1653 sd_event_source_unref(m->signal_event_source);
1654 sd_event_source_unref(m->sigchld_event_source);
1655 sd_event_source_unref(m->notify_event_source);
1656 sd_event_source_unref(m->cgroups_agent_event_source);
1657 sd_event_source_unref(m->time_change_event_source);
1658 sd_event_source_unref(m->timezone_change_event_source);
1659 sd_event_source_unref(m->jobs_in_progress_event_source);
1660 sd_event_source_unref(m->run_queue_event_source);
1661 sd_event_source_unref(m->user_lookup_event_source);
1662 sd_event_source_unref(m->memory_pressure_event_source);
1663
1664 safe_close(m->signal_fd);
1665 safe_close(m->notify_fd);
1666 safe_close(m->cgroups_agent_fd);
1667 safe_close_pair(m->user_lookup_fds);
1668
1669 manager_close_ask_password(m);
1670
1671 manager_close_idle_pipe(m);
1672
1673 sd_event_unref(m->event);
1674
1675 free(m->notify_socket);
1676
1677 lookup_paths_free(&m->lookup_paths);
1678 strv_free(m->transient_environment);
1679 strv_free(m->client_environment);
1680
1681 hashmap_free(m->cgroup_unit);
1682 manager_free_unit_name_maps(m);
1683
1684 free(m->switch_root);
1685 free(m->switch_root_init);
1686
1687 free(m->default_smack_process_label);
1688
1689 rlimit_free_all(m->rlimit);
1690
1691 assert(hashmap_isempty(m->units_requiring_mounts_for));
1692 hashmap_free(m->units_requiring_mounts_for);
1693
1694 hashmap_free(m->uid_refs);
1695 hashmap_free(m->gid_refs);
1696
1697 for (ExecDirectoryType dt = 0; dt < _EXEC_DIRECTORY_TYPE_MAX; dt++)
1698 m->prefix[dt] = mfree(m->prefix[dt]);
1699 free(m->received_credentials_directory);
1700 free(m->received_encrypted_credentials_directory);
1701
1702 free(m->watchdog_pretimeout_governor);
1703 free(m->watchdog_pretimeout_governor_overridden);
1704
1705 #if BPF_FRAMEWORK
1706 lsm_bpf_destroy(m->restrict_fs);
1707 #endif
1708
1709 return mfree(m);
1710 }
1711
1712 static void manager_enumerate_perpetual(Manager *m) {
1713 assert(m);
1714
1715 if (FLAGS_SET(m->test_run_flags, MANAGER_TEST_RUN_MINIMAL))
1716 return;
1717
1718 /* Let's ask every type to load all units from disk/kernel that it might know */
1719 for (UnitType c = 0; c < _UNIT_TYPE_MAX; c++) {
1720 if (!unit_type_supported(c)) {
1721 log_debug("Unit type .%s is not supported on this system.", unit_type_to_string(c));
1722 continue;
1723 }
1724
1725 if (unit_vtable[c]->enumerate_perpetual)
1726 unit_vtable[c]->enumerate_perpetual(m);
1727 }
1728 }
1729
1730 static void manager_enumerate(Manager *m) {
1731 assert(m);
1732
1733 if (FLAGS_SET(m->test_run_flags, MANAGER_TEST_RUN_MINIMAL))
1734 return;
1735
1736 /* Let's ask every type to load all units from disk/kernel that it might know */
1737 for (UnitType c = 0; c < _UNIT_TYPE_MAX; c++) {
1738 if (!unit_type_supported(c)) {
1739 log_debug("Unit type .%s is not supported on this system.", unit_type_to_string(c));
1740 continue;
1741 }
1742
1743 if (unit_vtable[c]->enumerate)
1744 unit_vtable[c]->enumerate(m);
1745 }
1746
1747 manager_dispatch_load_queue(m);
1748 }
1749
1750 static void manager_coldplug(Manager *m) {
1751 Unit *u;
1752 char *k;
1753 int r;
1754
1755 assert(m);
1756
1757 log_debug("Invoking unit coldplug() handlers%s", special_glyph(SPECIAL_GLYPH_ELLIPSIS));
1758
1759 /* Let's place the units back into their deserialized state */
1760 HASHMAP_FOREACH_KEY(u, k, m->units) {
1761
1762 /* ignore aliases */
1763 if (u->id != k)
1764 continue;
1765
1766 r = unit_coldplug(u);
1767 if (r < 0)
1768 log_warning_errno(r, "We couldn't coldplug %s, proceeding anyway: %m", u->id);
1769 }
1770 }
1771
1772 static void manager_catchup(Manager *m) {
1773 Unit *u;
1774 char *k;
1775
1776 assert(m);
1777
1778 log_debug("Invoking unit catchup() handlers%s", special_glyph(SPECIAL_GLYPH_ELLIPSIS));
1779
1780 /* Let's catch up on any state changes that happened while we were reloading/reexecing */
1781 HASHMAP_FOREACH_KEY(u, k, m->units) {
1782
1783 /* ignore aliases */
1784 if (u->id != k)
1785 continue;
1786
1787 unit_catchup(u);
1788 }
1789 }
1790
1791 static void manager_distribute_fds(Manager *m, FDSet *fds) {
1792 Unit *u;
1793
1794 assert(m);
1795
1796 HASHMAP_FOREACH(u, m->units) {
1797
1798 if (fdset_size(fds) <= 0)
1799 break;
1800
1801 if (!UNIT_VTABLE(u)->distribute_fds)
1802 continue;
1803
1804 UNIT_VTABLE(u)->distribute_fds(u, fds);
1805 }
1806 }
1807
1808 static bool manager_dbus_is_running(Manager *m, bool deserialized) {
1809 Unit *u;
1810
1811 assert(m);
1812
1813 /* This checks whether the dbus instance we are supposed to expose our APIs on is up. We check both the socket
1814 * and the service unit. If the 'deserialized' parameter is true we'll check the deserialized state of the unit
1815 * rather than the current one. */
1816
1817 if (MANAGER_IS_TEST_RUN(m))
1818 return false;
1819
1820 u = manager_get_unit(m, SPECIAL_DBUS_SOCKET);
1821 if (!u)
1822 return false;
1823 if ((deserialized ? SOCKET(u)->deserialized_state : SOCKET(u)->state) != SOCKET_RUNNING)
1824 return false;
1825
1826 u = manager_get_unit(m, SPECIAL_DBUS_SERVICE);
1827 if (!u)
1828 return false;
1829 if (!IN_SET((deserialized ? SERVICE(u)->deserialized_state : SERVICE(u)->state),
1830 SERVICE_RUNNING,
1831 SERVICE_RELOAD,
1832 SERVICE_RELOAD_NOTIFY,
1833 SERVICE_RELOAD_SIGNAL))
1834 return false;
1835
1836 return true;
1837 }
1838
1839 static void manager_setup_bus(Manager *m) {
1840 assert(m);
1841
1842 /* Let's set up our private bus connection now, unconditionally */
1843 (void) bus_init_private(m);
1844
1845 /* If we are in --user mode also connect to the system bus now */
1846 if (MANAGER_IS_USER(m))
1847 (void) bus_init_system(m);
1848
1849 /* Let's connect to the bus now, but only if the unit is supposed to be up */
1850 if (manager_dbus_is_running(m, MANAGER_IS_RELOADING(m))) {
1851 (void) bus_init_api(m);
1852
1853 if (MANAGER_IS_SYSTEM(m))
1854 (void) bus_init_system(m);
1855 }
1856 }
1857
1858 static void manager_preset_all(Manager *m) {
1859 int r;
1860
1861 assert(m);
1862
1863 if (m->first_boot <= 0)
1864 return;
1865
1866 if (!MANAGER_IS_SYSTEM(m))
1867 return;
1868
1869 if (MANAGER_IS_TEST_RUN(m))
1870 return;
1871
1872 /* If this is the first boot, and we are in the host system, then preset everything */
1873 UnitFilePresetMode mode = FIRST_BOOT_FULL_PRESET ? UNIT_FILE_PRESET_FULL : UNIT_FILE_PRESET_ENABLE_ONLY;
1874
1875 r = unit_file_preset_all(RUNTIME_SCOPE_SYSTEM, 0, NULL, mode, NULL, 0);
1876 if (r < 0)
1877 log_full_errno(r == -EEXIST ? LOG_NOTICE : LOG_WARNING, r,
1878 "Failed to populate /etc with preset unit settings, ignoring: %m");
1879 else
1880 log_info("Populated /etc with preset unit settings.");
1881 }
1882
1883 static void manager_ready(Manager *m) {
1884 assert(m);
1885
1886 /* After having loaded everything, do the final round of catching up with what might have changed */
1887
1888 m->objective = MANAGER_OK; /* Tell everyone we are up now */
1889
1890 /* It might be safe to log to the journal now and connect to dbus */
1891 manager_recheck_journal(m);
1892 manager_recheck_dbus(m);
1893
1894 /* Let's finally catch up with any changes that took place while we were reloading/reexecing */
1895 manager_catchup(m);
1896
1897 /* Create a file which will indicate when the manager started loading units the last time. */
1898 if (MANAGER_IS_SYSTEM(m))
1899 (void) touch_file("/run/systemd/systemd-units-load", false,
1900 m->timestamps[MANAGER_TIMESTAMP_UNITS_LOAD].realtime ?: now(CLOCK_REALTIME),
1901 UID_INVALID, GID_INVALID, 0444);
1902 }
1903
1904 Manager* manager_reloading_start(Manager *m) {
1905 m->n_reloading++;
1906 dual_timestamp_get(m->timestamps + MANAGER_TIMESTAMP_UNITS_LOAD);
1907 return m;
1908 }
1909
1910 void manager_reloading_stopp(Manager **m) {
1911 if (*m) {
1912 assert((*m)->n_reloading > 0);
1913 (*m)->n_reloading--;
1914 }
1915 }
1916
1917 int manager_startup(Manager *m, FILE *serialization, FDSet *fds, const char *root) {
1918 int r;
1919
1920 assert(m);
1921
1922 /* If we are running in test mode, we still want to run the generators,
1923 * but we should not touch the real generator directories. */
1924 r = lookup_paths_init_or_warn(&m->lookup_paths, m->runtime_scope,
1925 MANAGER_IS_TEST_RUN(m) ? LOOKUP_PATHS_TEMPORARY_GENERATED : 0,
1926 root);
1927 if (r < 0)
1928 return r;
1929
1930 dual_timestamp_get(m->timestamps + manager_timestamp_initrd_mangle(MANAGER_TIMESTAMP_GENERATORS_START));
1931 r = manager_run_environment_generators(m);
1932 if (r >= 0)
1933 r = manager_run_generators(m);
1934 dual_timestamp_get(m->timestamps + manager_timestamp_initrd_mangle(MANAGER_TIMESTAMP_GENERATORS_FINISH));
1935 if (r < 0)
1936 return r;
1937
1938 manager_preset_all(m);
1939
1940 lookup_paths_log(&m->lookup_paths);
1941
1942 {
1943 /* This block is (optionally) done with the reloading counter bumped */
1944 _unused_ _cleanup_(manager_reloading_stopp) Manager *reloading = NULL;
1945
1946 /* Make sure we don't have a left-over from a previous run */
1947 if (!serialization)
1948 (void) rm_rf(m->lookup_paths.transient, 0);
1949
1950 /* If we will deserialize make sure that during enumeration this is already known, so we increase the
1951 * counter here already */
1952 if (serialization)
1953 reloading = manager_reloading_start(m);
1954
1955 /* First, enumerate what we can from all config files */
1956 dual_timestamp_get(m->timestamps + manager_timestamp_initrd_mangle(MANAGER_TIMESTAMP_UNITS_LOAD_START));
1957 manager_enumerate_perpetual(m);
1958 manager_enumerate(m);
1959 dual_timestamp_get(m->timestamps + manager_timestamp_initrd_mangle(MANAGER_TIMESTAMP_UNITS_LOAD_FINISH));
1960
1961 /* Second, deserialize if there is something to deserialize */
1962 if (serialization) {
1963 r = manager_deserialize(m, serialization, fds);
1964 if (r < 0)
1965 return log_error_errno(r, "Deserialization failed: %m");
1966 }
1967
1968 /* Any fds left? Find some unit which wants them. This is useful to allow container managers to pass
1969 * some file descriptors to us pre-initialized. This enables socket-based activation of entire
1970 * containers. */
1971 manager_distribute_fds(m, fds);
1972
1973 /* We might have deserialized the notify fd, but if we didn't then let's create the bus now */
1974 r = manager_setup_notify(m);
1975 if (r < 0)
1976 /* No sense to continue without notifications, our children would fail anyway. */
1977 return r;
1978
1979 r = manager_setup_cgroups_agent(m);
1980 if (r < 0)
1981 /* Likewise, no sense to continue without empty cgroup notifications. */
1982 return r;
1983
1984 r = manager_setup_user_lookup_fd(m);
1985 if (r < 0)
1986 /* This shouldn't fail, except if things are really broken. */
1987 return r;
1988
1989 /* Connect to the bus if we are good for it */
1990 manager_setup_bus(m);
1991
1992 /* Now that we are connected to all possible buses, let's deserialize who is tracking us. */
1993 r = bus_track_coldplug(m, &m->subscribed, false, m->deserialized_subscribed);
1994 if (r < 0)
1995 log_warning_errno(r, "Failed to deserialized tracked clients, ignoring: %m");
1996 m->deserialized_subscribed = strv_free(m->deserialized_subscribed);
1997
1998 r = manager_varlink_init(m);
1999 if (r < 0)
2000 log_warning_errno(r, "Failed to set up Varlink, ignoring: %m");
2001
2002 /* Third, fire things up! */
2003 manager_coldplug(m);
2004
2005 /* Clean up runtime objects */
2006 manager_vacuum(m);
2007
2008 if (serialization)
2009 /* Let's wait for the UnitNew/JobNew messages being sent, before we notify that the
2010 * reload is finished */
2011 m->send_reloading_done = true;
2012 }
2013
2014 manager_ready(m);
2015
2016 manager_set_switching_root(m, false);
2017
2018 return 0;
2019 }
2020
2021 int manager_add_job(
2022 Manager *m,
2023 JobType type,
2024 Unit *unit,
2025 JobMode mode,
2026 Set *affected_jobs,
2027 sd_bus_error *error,
2028 Job **ret) {
2029
2030 _cleanup_(transaction_abort_and_freep) Transaction *tr = NULL;
2031 int r;
2032
2033 assert(m);
2034 assert(type < _JOB_TYPE_MAX);
2035 assert(unit);
2036 assert(mode < _JOB_MODE_MAX);
2037
2038 if (mode == JOB_ISOLATE && type != JOB_START)
2039 return sd_bus_error_set(error, SD_BUS_ERROR_INVALID_ARGS, "Isolate is only valid for start.");
2040
2041 if (mode == JOB_ISOLATE && !unit->allow_isolate)
2042 return sd_bus_error_set(error, BUS_ERROR_NO_ISOLATION, "Operation refused, unit may not be isolated.");
2043
2044 if (mode == JOB_TRIGGERING && type != JOB_STOP)
2045 return sd_bus_error_set(error, SD_BUS_ERROR_INVALID_ARGS, "--job-mode=triggering is only valid for stop.");
2046
2047 log_unit_debug(unit, "Trying to enqueue job %s/%s/%s", unit->id, job_type_to_string(type), job_mode_to_string(mode));
2048
2049 type = job_type_collapse(type, unit);
2050
2051 tr = transaction_new(mode == JOB_REPLACE_IRREVERSIBLY);
2052 if (!tr)
2053 return -ENOMEM;
2054
2055 r = transaction_add_job_and_dependencies(
2056 tr,
2057 type,
2058 unit,
2059 /* by= */ NULL,
2060 TRANSACTION_MATTERS |
2061 (IN_SET(mode, JOB_IGNORE_DEPENDENCIES, JOB_IGNORE_REQUIREMENTS) ? TRANSACTION_IGNORE_REQUIREMENTS : 0) |
2062 (mode == JOB_IGNORE_DEPENDENCIES ? TRANSACTION_IGNORE_ORDER : 0),
2063 error);
2064 if (r < 0)
2065 return r;
2066
2067 if (mode == JOB_ISOLATE) {
2068 r = transaction_add_isolate_jobs(tr, m);
2069 if (r < 0)
2070 return r;
2071 }
2072
2073 if (mode == JOB_TRIGGERING) {
2074 r = transaction_add_triggering_jobs(tr, unit);
2075 if (r < 0)
2076 return r;
2077 }
2078
2079 r = transaction_activate(tr, m, mode, affected_jobs, error);
2080 if (r < 0)
2081 return r;
2082
2083 log_unit_debug(unit,
2084 "Enqueued job %s/%s as %u", unit->id,
2085 job_type_to_string(type), (unsigned) tr->anchor_job->id);
2086
2087 if (ret)
2088 *ret = tr->anchor_job;
2089
2090 tr = transaction_free(tr);
2091 return 0;
2092 }
2093
2094 int manager_add_job_by_name(Manager *m, JobType type, const char *name, JobMode mode, Set *affected_jobs, sd_bus_error *e, Job **ret) {
2095 Unit *unit = NULL; /* just to appease gcc, initialization is not really necessary */
2096 int r;
2097
2098 assert(m);
2099 assert(type < _JOB_TYPE_MAX);
2100 assert(name);
2101 assert(mode < _JOB_MODE_MAX);
2102
2103 r = manager_load_unit(m, name, NULL, NULL, &unit);
2104 if (r < 0)
2105 return r;
2106 assert(unit);
2107
2108 return manager_add_job(m, type, unit, mode, affected_jobs, e, ret);
2109 }
2110
2111 int manager_add_job_by_name_and_warn(Manager *m, JobType type, const char *name, JobMode mode, Set *affected_jobs, Job **ret) {
2112 _cleanup_(sd_bus_error_free) sd_bus_error error = SD_BUS_ERROR_NULL;
2113 int r;
2114
2115 assert(m);
2116 assert(type < _JOB_TYPE_MAX);
2117 assert(name);
2118 assert(mode < _JOB_MODE_MAX);
2119
2120 r = manager_add_job_by_name(m, type, name, mode, affected_jobs, &error, ret);
2121 if (r < 0)
2122 return log_warning_errno(r, "Failed to enqueue %s job for %s: %s", job_mode_to_string(mode), name, bus_error_message(&error, r));
2123
2124 return r;
2125 }
2126
2127 int manager_propagate_reload(Manager *m, Unit *unit, JobMode mode, sd_bus_error *e) {
2128 int r;
2129 _cleanup_(transaction_abort_and_freep) Transaction *tr = NULL;
2130
2131 assert(m);
2132 assert(unit);
2133 assert(mode < _JOB_MODE_MAX);
2134 assert(mode != JOB_ISOLATE); /* Isolate is only valid for start */
2135
2136 tr = transaction_new(mode == JOB_REPLACE_IRREVERSIBLY);
2137 if (!tr)
2138 return -ENOMEM;
2139
2140 /* We need an anchor job */
2141 r = transaction_add_job_and_dependencies(tr, JOB_NOP, unit, NULL, TRANSACTION_IGNORE_REQUIREMENTS|TRANSACTION_IGNORE_ORDER, e);
2142 if (r < 0)
2143 return r;
2144
2145 /* Failure in adding individual dependencies is ignored, so this always succeeds. */
2146 transaction_add_propagate_reload_jobs(
2147 tr,
2148 unit,
2149 tr->anchor_job,
2150 mode == JOB_IGNORE_DEPENDENCIES ? TRANSACTION_IGNORE_ORDER : 0);
2151
2152 r = transaction_activate(tr, m, mode, NULL, e);
2153 if (r < 0)
2154 return r;
2155
2156 tr = transaction_free(tr);
2157 return 0;
2158 }
2159
2160 Job *manager_get_job(Manager *m, uint32_t id) {
2161 assert(m);
2162
2163 return hashmap_get(m->jobs, UINT32_TO_PTR(id));
2164 }
2165
2166 Unit *manager_get_unit(Manager *m, const char *name) {
2167 assert(m);
2168 assert(name);
2169
2170 return hashmap_get(m->units, name);
2171 }
2172
2173 static int manager_dispatch_target_deps_queue(Manager *m) {
2174 Unit *u;
2175 int r = 0;
2176
2177 assert(m);
2178
2179 while ((u = m->target_deps_queue)) {
2180 _cleanup_free_ Unit **targets = NULL;
2181 int n_targets;
2182
2183 assert(u->in_target_deps_queue);
2184
2185 LIST_REMOVE(target_deps_queue, u->manager->target_deps_queue, u);
2186 u->in_target_deps_queue = false;
2187
2188 /* Take an "atomic" snapshot of dependencies here, as the call below will likely modify the
2189 * dependencies, and we can't have it that hash tables we iterate through are modified while
2190 * we are iterating through them. */
2191 n_targets = unit_get_dependency_array(u, UNIT_ATOM_DEFAULT_TARGET_DEPENDENCIES, &targets);
2192 if (n_targets < 0)
2193 return n_targets;
2194
2195 for (int i = 0; i < n_targets; i++) {
2196 r = unit_add_default_target_dependency(u, targets[i]);
2197 if (r < 0)
2198 return r;
2199 }
2200 }
2201
2202 return r;
2203 }
2204
2205 unsigned manager_dispatch_load_queue(Manager *m) {
2206 Unit *u;
2207 unsigned n = 0;
2208
2209 assert(m);
2210
2211 /* Make sure we are not run recursively */
2212 if (m->dispatching_load_queue)
2213 return 0;
2214
2215 m->dispatching_load_queue = true;
2216
2217 /* Dispatches the load queue. Takes a unit from the queue and
2218 * tries to load its data until the queue is empty */
2219
2220 while ((u = m->load_queue)) {
2221 assert(u->in_load_queue);
2222
2223 unit_load(u);
2224 n++;
2225 }
2226
2227 m->dispatching_load_queue = false;
2228
2229 /* Dispatch the units waiting for their target dependencies to be added now, as all targets that we know about
2230 * should be loaded and have aliases resolved */
2231 (void) manager_dispatch_target_deps_queue(m);
2232
2233 return n;
2234 }
2235
2236 bool manager_unit_cache_should_retry_load(Unit *u) {
2237 assert(u);
2238
2239 /* Automatic reloading from disk only applies to units which were not found sometime in the past, and
2240 * the not-found stub is kept pinned in the unit graph by dependencies. For units that were
2241 * previously loaded, we don't do automatic reloading, and daemon-reload is necessary to update. */
2242 if (u->load_state != UNIT_NOT_FOUND)
2243 return false;
2244
2245 /* The cache has been updated since the last time we tried to load the unit. There might be new
2246 * fragment paths to read. */
2247 if (u->manager->unit_cache_timestamp_hash != u->fragment_not_found_timestamp_hash)
2248 return true;
2249
2250 /* The cache needs to be updated because there are modifications on disk. */
2251 return !lookup_paths_timestamp_hash_same(&u->manager->lookup_paths, u->manager->unit_cache_timestamp_hash, NULL);
2252 }
2253
2254 int manager_load_unit_prepare(
2255 Manager *m,
2256 const char *name,
2257 const char *path,
2258 sd_bus_error *e,
2259 Unit **ret) {
2260
2261 _cleanup_(unit_freep) Unit *cleanup_unit = NULL;
2262 _cleanup_free_ char *nbuf = NULL;
2263 int r;
2264
2265 assert(m);
2266 assert(ret);
2267 assert(name || path);
2268
2269 /* This will prepare the unit for loading, but not actually load anything from disk. */
2270
2271 if (path && !path_is_absolute(path))
2272 return sd_bus_error_setf(e, SD_BUS_ERROR_INVALID_ARGS, "Path %s is not absolute.", path);
2273
2274 if (!name) {
2275 r = path_extract_filename(path, &nbuf);
2276 if (r < 0)
2277 return r;
2278 if (r == O_DIRECTORY)
2279 return sd_bus_error_setf(e, SD_BUS_ERROR_INVALID_ARGS, "Path '%s' refers to directory, refusing.", path);
2280
2281 name = nbuf;
2282 }
2283
2284 UnitType t = unit_name_to_type(name);
2285
2286 if (t == _UNIT_TYPE_INVALID || !unit_name_is_valid(name, UNIT_NAME_PLAIN|UNIT_NAME_INSTANCE)) {
2287 if (unit_name_is_valid(name, UNIT_NAME_TEMPLATE))
2288 return sd_bus_error_setf(e, SD_BUS_ERROR_INVALID_ARGS, "Unit name %s is missing the instance name.", name);
2289
2290 return sd_bus_error_setf(e, SD_BUS_ERROR_INVALID_ARGS, "Unit name %s is not valid.", name);
2291 }
2292
2293 Unit *unit = manager_get_unit(m, name);
2294 if (unit) {
2295 /* The time-based cache allows to start new units without daemon-reload,
2296 * but if they are already referenced (because of dependencies or ordering)
2297 * then we have to force a load of the fragment. As an optimization, check
2298 * first if anything in the usual paths was modified since the last time
2299 * the cache was loaded. Also check if the last time an attempt to load the
2300 * unit was made was before the most recent cache refresh, so that we know
2301 * we need to try again — even if the cache is current, it might have been
2302 * updated in a different context before we had a chance to retry loading
2303 * this particular unit. */
2304 if (manager_unit_cache_should_retry_load(unit))
2305 unit->load_state = UNIT_STUB;
2306 else {
2307 *ret = unit;
2308 return 0; /* The unit was already loaded */
2309 }
2310 } else {
2311 unit = cleanup_unit = unit_new(m, unit_vtable[t]->object_size);
2312 if (!unit)
2313 return -ENOMEM;
2314 }
2315
2316 if (path) {
2317 r = free_and_strdup(&unit->fragment_path, path);
2318 if (r < 0)
2319 return r;
2320 }
2321
2322 r = unit_add_name(unit, name);
2323 if (r < 0)
2324 return r;
2325
2326 unit_add_to_load_queue(unit);
2327 unit_add_to_dbus_queue(unit);
2328 unit_add_to_gc_queue(unit);
2329
2330 *ret = unit;
2331 TAKE_PTR(cleanup_unit);
2332
2333 return 1; /* The unit was added the load queue */
2334 }
2335
2336 int manager_load_unit(
2337 Manager *m,
2338 const char *name,
2339 const char *path,
2340 sd_bus_error *e,
2341 Unit **ret) {
2342 int r;
2343
2344 assert(m);
2345 assert(ret);
2346
2347 /* This will load the unit config, but not actually start any services or anything. */
2348
2349 r = manager_load_unit_prepare(m, name, path, e, ret);
2350 if (r <= 0)
2351 return r;
2352
2353 /* Unit was newly loaded */
2354 manager_dispatch_load_queue(m);
2355 *ret = unit_follow_merge(*ret);
2356 return 0;
2357 }
2358
2359 int manager_load_startable_unit_or_warn(
2360 Manager *m,
2361 const char *name,
2362 const char *path,
2363 Unit **ret) {
2364
2365 /* Load a unit, make sure it loaded fully and is not masked. */
2366
2367 _cleanup_(sd_bus_error_free) sd_bus_error error = SD_BUS_ERROR_NULL;
2368 Unit *unit;
2369 int r;
2370
2371 r = manager_load_unit(m, name, path, &error, &unit);
2372 if (r < 0)
2373 return log_error_errno(r, "Failed to load %s %s: %s",
2374 name ? "unit" : "unit file", name ?: path,
2375 bus_error_message(&error, r));
2376
2377 r = bus_unit_validate_load_state(unit, &error);
2378 if (r < 0)
2379 return log_error_errno(r, "%s", bus_error_message(&error, r));
2380
2381 *ret = unit;
2382 return 0;
2383 }
2384
2385 void manager_clear_jobs(Manager *m) {
2386 Job *j;
2387
2388 assert(m);
2389
2390 while ((j = hashmap_first(m->jobs)))
2391 /* No need to recurse. We're cancelling all jobs. */
2392 job_finish_and_invalidate(j, JOB_CANCELED, false, false);
2393 }
2394
2395 void manager_unwatch_pid(Manager *m, pid_t pid) {
2396 assert(m);
2397
2398 /* First let's drop the unit keyed as "pid". */
2399 (void) hashmap_remove(m->watch_pids, PID_TO_PTR(pid));
2400
2401 /* Then, let's also drop the array keyed by -pid. */
2402 free(hashmap_remove(m->watch_pids, PID_TO_PTR(-pid)));
2403 }
2404
2405 static int manager_dispatch_run_queue(sd_event_source *source, void *userdata) {
2406 Manager *m = ASSERT_PTR(userdata);
2407 Job *j;
2408
2409 assert(source);
2410
2411 while ((j = prioq_peek(m->run_queue))) {
2412 assert(j->installed);
2413 assert(j->in_run_queue);
2414
2415 (void) job_run_and_invalidate(j);
2416 }
2417
2418 if (m->n_running_jobs > 0)
2419 manager_watch_jobs_in_progress(m);
2420
2421 if (m->n_on_console > 0)
2422 manager_watch_idle_pipe(m);
2423
2424 return 1;
2425 }
2426
2427 void manager_trigger_run_queue(Manager *m) {
2428 int r;
2429
2430 assert(m);
2431
2432 r = sd_event_source_set_enabled(
2433 m->run_queue_event_source,
2434 prioq_isempty(m->run_queue) ? SD_EVENT_OFF : SD_EVENT_ONESHOT);
2435 if (r < 0)
2436 log_warning_errno(r, "Failed to enable job run queue event source, ignoring: %m");
2437 }
2438
2439 static unsigned manager_dispatch_dbus_queue(Manager *m) {
2440 unsigned n = 0, budget;
2441 Unit *u;
2442 Job *j;
2443
2444 assert(m);
2445
2446 /* When we are reloading, let's not wait with generating signals, since we need to exit the manager as quickly
2447 * as we can. There's no point in throttling generation of signals in that case. */
2448 if (MANAGER_IS_RELOADING(m) || m->send_reloading_done || m->pending_reload_message)
2449 budget = UINT_MAX; /* infinite budget in this case */
2450 else {
2451 /* Anything to do at all? */
2452 if (!m->dbus_unit_queue && !m->dbus_job_queue)
2453 return 0;
2454
2455 /* Do we have overly many messages queued at the moment? If so, let's not enqueue more on top, let's
2456 * sit this cycle out, and process things in a later cycle when the queues got a bit emptier. */
2457 if (manager_bus_n_queued_write(m) > MANAGER_BUS_BUSY_THRESHOLD)
2458 return 0;
2459
2460 /* Only process a certain number of units/jobs per event loop iteration. Even if the bus queue wasn't
2461 * overly full before this call we shouldn't increase it in size too wildly in one step, and we
2462 * shouldn't monopolize CPU time with generating these messages. Note the difference in counting of
2463 * this "budget" and the "threshold" above: the "budget" is decreased only once per generated message,
2464 * regardless how many buses/direct connections it is enqueued on, while the "threshold" is applied to
2465 * each queued instance of bus message, i.e. if the same message is enqueued to five buses/direct
2466 * connections it will be counted five times. This difference in counting ("references"
2467 * vs. "instances") is primarily a result of the fact that it's easier to implement it this way,
2468 * however it also reflects the thinking that the "threshold" should put a limit on used queue memory,
2469 * i.e. space, while the "budget" should put a limit on time. Also note that the "threshold" is
2470 * currently chosen much higher than the "budget". */
2471 budget = MANAGER_BUS_MESSAGE_BUDGET;
2472 }
2473
2474 while (budget != 0 && (u = m->dbus_unit_queue)) {
2475
2476 assert(u->in_dbus_queue);
2477
2478 bus_unit_send_change_signal(u);
2479 n++;
2480
2481 if (budget != UINT_MAX)
2482 budget--;
2483 }
2484
2485 while (budget != 0 && (j = m->dbus_job_queue)) {
2486 assert(j->in_dbus_queue);
2487
2488 bus_job_send_change_signal(j);
2489 n++;
2490
2491 if (budget != UINT_MAX)
2492 budget--;
2493 }
2494
2495 if (m->send_reloading_done) {
2496 m->send_reloading_done = false;
2497 bus_manager_send_reloading(m, false);
2498 n++;
2499 }
2500
2501 if (m->pending_reload_message) {
2502 bus_send_pending_reload_message(m);
2503 n++;
2504 }
2505
2506 return n;
2507 }
2508
2509 static int manager_dispatch_cgroups_agent_fd(sd_event_source *source, int fd, uint32_t revents, void *userdata) {
2510 Manager *m = userdata;
2511 char buf[PATH_MAX];
2512 ssize_t n;
2513
2514 n = recv(fd, buf, sizeof(buf), 0);
2515 if (n < 0)
2516 return log_error_errno(errno, "Failed to read cgroups agent message: %m");
2517 if (n == 0) {
2518 log_error("Got zero-length cgroups agent message, ignoring.");
2519 return 0;
2520 }
2521 if ((size_t) n >= sizeof(buf)) {
2522 log_error("Got overly long cgroups agent message, ignoring.");
2523 return 0;
2524 }
2525
2526 if (memchr(buf, 0, n)) {
2527 log_error("Got cgroups agent message with embedded NUL byte, ignoring.");
2528 return 0;
2529 }
2530 buf[n] = 0;
2531
2532 manager_notify_cgroup_empty(m, buf);
2533 (void) bus_forward_agent_released(m, buf);
2534
2535 return 0;
2536 }
2537
2538 static bool manager_process_barrier_fd(char * const *tags, FDSet *fds) {
2539
2540 /* nothing else must be sent when using BARRIER=1 */
2541 if (strv_contains(tags, "BARRIER=1")) {
2542 if (strv_length(tags) != 1)
2543 log_warning("Extra notification messages sent with BARRIER=1, ignoring everything.");
2544 else if (fdset_size(fds) != 1)
2545 log_warning("Got incorrect number of fds with BARRIER=1, closing them.");
2546
2547 /* Drop the message if BARRIER=1 was found */
2548 return true;
2549 }
2550
2551 return false;
2552 }
2553
2554 static void manager_invoke_notify_message(
2555 Manager *m,
2556 Unit *u,
2557 const struct ucred *ucred,
2558 char * const *tags,
2559 FDSet *fds) {
2560
2561 assert(m);
2562 assert(u);
2563 assert(ucred);
2564 assert(tags);
2565
2566 if (u->notifygen == m->notifygen) /* Already invoked on this same unit in this same iteration? */
2567 return;
2568 u->notifygen = m->notifygen;
2569
2570 if (UNIT_VTABLE(u)->notify_message)
2571 UNIT_VTABLE(u)->notify_message(u, ucred, tags, fds);
2572
2573 else if (DEBUG_LOGGING) {
2574 _cleanup_free_ char *buf = NULL, *x = NULL, *y = NULL;
2575
2576 buf = strv_join(tags, ", ");
2577 if (buf)
2578 x = ellipsize(buf, 20, 90);
2579 if (x)
2580 y = cescape(x);
2581
2582 log_unit_debug(u, "Got notification message \"%s\", ignoring.", strnull(y));
2583 }
2584 }
2585
2586 static int manager_dispatch_notify_fd(sd_event_source *source, int fd, uint32_t revents, void *userdata) {
2587
2588 _cleanup_fdset_free_ FDSet *fds = NULL;
2589 Manager *m = ASSERT_PTR(userdata);
2590 char buf[NOTIFY_BUFFER_MAX+1];
2591 struct iovec iovec = {
2592 .iov_base = buf,
2593 .iov_len = sizeof(buf)-1,
2594 };
2595 CMSG_BUFFER_TYPE(CMSG_SPACE(sizeof(struct ucred)) +
2596 CMSG_SPACE(sizeof(int) * NOTIFY_FD_MAX)) control;
2597 struct msghdr msghdr = {
2598 .msg_iov = &iovec,
2599 .msg_iovlen = 1,
2600 .msg_control = &control,
2601 .msg_controllen = sizeof(control),
2602 };
2603
2604 struct cmsghdr *cmsg;
2605 struct ucred *ucred = NULL;
2606 _cleanup_free_ Unit **array_copy = NULL;
2607 _cleanup_strv_free_ char **tags = NULL;
2608 Unit *u1, *u2, **array;
2609 int r, *fd_array = NULL;
2610 size_t n_fds = 0;
2611 bool found = false;
2612 ssize_t n;
2613
2614 assert(m->notify_fd == fd);
2615
2616 if (revents != EPOLLIN) {
2617 log_warning("Got unexpected poll event for notify fd.");
2618 return 0;
2619 }
2620
2621 n = recvmsg_safe(m->notify_fd, &msghdr, MSG_DONTWAIT|MSG_CMSG_CLOEXEC|MSG_TRUNC);
2622 if (n < 0) {
2623 if (ERRNO_IS_TRANSIENT(n))
2624 return 0; /* Spurious wakeup, try again */
2625 if (n == -EXFULL) {
2626 log_warning("Got message with truncated control data (too many fds sent?), ignoring.");
2627 return 0;
2628 }
2629 /* If this is any other, real error, then let's stop processing this socket. This of course
2630 * means we won't take notification messages anymore, but that's still better than busy
2631 * looping around this: being woken up over and over again but being unable to actually read
2632 * the message off the socket. */
2633 return log_error_errno(n, "Failed to receive notification message: %m");
2634 }
2635
2636 CMSG_FOREACH(cmsg, &msghdr) {
2637 if (cmsg->cmsg_level == SOL_SOCKET && cmsg->cmsg_type == SCM_RIGHTS) {
2638
2639 assert(!fd_array);
2640 fd_array = CMSG_TYPED_DATA(cmsg, int);
2641 n_fds = (cmsg->cmsg_len - CMSG_LEN(0)) / sizeof(int);
2642
2643 } else if (cmsg->cmsg_level == SOL_SOCKET &&
2644 cmsg->cmsg_type == SCM_CREDENTIALS &&
2645 cmsg->cmsg_len == CMSG_LEN(sizeof(struct ucred))) {
2646
2647 assert(!ucred);
2648 ucred = CMSG_TYPED_DATA(cmsg, struct ucred);
2649 }
2650 }
2651
2652 if (n_fds > 0) {
2653 assert(fd_array);
2654
2655 r = fdset_new_array(&fds, fd_array, n_fds);
2656 if (r < 0) {
2657 close_many(fd_array, n_fds);
2658 log_oom();
2659 return 0;
2660 }
2661 }
2662
2663 if (!ucred || !pid_is_valid(ucred->pid)) {
2664 log_warning("Received notify message without valid credentials. Ignoring.");
2665 return 0;
2666 }
2667
2668 if ((size_t) n >= sizeof(buf) || (msghdr.msg_flags & MSG_TRUNC)) {
2669 log_warning("Received notify message exceeded maximum size. Ignoring.");
2670 return 0;
2671 }
2672
2673 /* As extra safety check, let's make sure the string we get doesn't contain embedded NUL bytes.
2674 * We permit one trailing NUL byte in the message, but don't expect it. */
2675 if (n > 1 && memchr(buf, 0, n-1)) {
2676 log_warning("Received notify message with embedded NUL bytes. Ignoring.");
2677 return 0;
2678 }
2679
2680 /* Make sure it's NUL-terminated, then parse it to obtain the tags list. */
2681 buf[n] = 0;
2682 tags = strv_split_newlines(buf);
2683 if (!tags) {
2684 log_oom();
2685 return 0;
2686 }
2687
2688 /* Possibly a barrier fd, let's see. */
2689 if (manager_process_barrier_fd(tags, fds)) {
2690 log_debug("Received barrier notification message from PID " PID_FMT ".", ucred->pid);
2691 return 0;
2692 }
2693
2694 /* Increase the generation counter used for filtering out duplicate unit invocations. */
2695 m->notifygen++;
2696
2697 /* Notify every unit that might be interested, which might be multiple. */
2698 u1 = manager_get_unit_by_pid_cgroup(m, ucred->pid);
2699 u2 = hashmap_get(m->watch_pids, PID_TO_PTR(ucred->pid));
2700 array = hashmap_get(m->watch_pids, PID_TO_PTR(-ucred->pid));
2701 if (array) {
2702 size_t k = 0;
2703
2704 while (array[k])
2705 k++;
2706
2707 array_copy = newdup(Unit*, array, k+1);
2708 if (!array_copy)
2709 log_oom();
2710 }
2711 /* And now invoke the per-unit callbacks. Note that manager_invoke_notify_message() will handle
2712 * duplicate units make sure we only invoke each unit's handler once. */
2713 if (u1) {
2714 manager_invoke_notify_message(m, u1, ucred, tags, fds);
2715 found = true;
2716 }
2717 if (u2) {
2718 manager_invoke_notify_message(m, u2, ucred, tags, fds);
2719 found = true;
2720 }
2721 if (array_copy)
2722 for (size_t i = 0; array_copy[i]; i++) {
2723 manager_invoke_notify_message(m, array_copy[i], ucred, tags, fds);
2724 found = true;
2725 }
2726
2727 if (!found)
2728 log_warning("Cannot find unit for notify message of PID "PID_FMT", ignoring.", ucred->pid);
2729
2730 if (fdset_size(fds) > 0)
2731 log_warning("Got extra auxiliary fds with notification message, closing them.");
2732
2733 return 0;
2734 }
2735
2736 static void manager_invoke_sigchld_event(
2737 Manager *m,
2738 Unit *u,
2739 const siginfo_t *si) {
2740
2741 assert(m);
2742 assert(u);
2743 assert(si);
2744
2745 /* Already invoked the handler of this unit in this iteration? Then don't process this again */
2746 if (u->sigchldgen == m->sigchldgen)
2747 return;
2748 u->sigchldgen = m->sigchldgen;
2749
2750 log_unit_debug(u, "Child "PID_FMT" belongs to %s.", si->si_pid, u->id);
2751 unit_unwatch_pid(u, si->si_pid);
2752
2753 if (UNIT_VTABLE(u)->sigchld_event)
2754 UNIT_VTABLE(u)->sigchld_event(u, si->si_pid, si->si_code, si->si_status);
2755 }
2756
2757 static int manager_dispatch_sigchld(sd_event_source *source, void *userdata) {
2758 Manager *m = ASSERT_PTR(userdata);
2759 siginfo_t si = {};
2760 int r;
2761
2762 assert(source);
2763
2764 /* First we call waitid() for a PID and do not reap the zombie. That way we can still access
2765 * /proc/$PID for it while it is a zombie. */
2766
2767 if (waitid(P_ALL, 0, &si, WEXITED|WNOHANG|WNOWAIT) < 0) {
2768
2769 if (errno != ECHILD)
2770 log_error_errno(errno, "Failed to peek for child with waitid(), ignoring: %m");
2771
2772 goto turn_off;
2773 }
2774
2775 if (si.si_pid <= 0)
2776 goto turn_off;
2777
2778 if (IN_SET(si.si_code, CLD_EXITED, CLD_KILLED, CLD_DUMPED)) {
2779 _cleanup_free_ Unit **array_copy = NULL;
2780 _cleanup_free_ char *name = NULL;
2781 Unit *u1, *u2, **array;
2782
2783 (void) get_process_comm(si.si_pid, &name);
2784
2785 log_debug("Child "PID_FMT" (%s) died (code=%s, status=%i/%s)",
2786 si.si_pid, strna(name),
2787 sigchld_code_to_string(si.si_code),
2788 si.si_status,
2789 strna(si.si_code == CLD_EXITED
2790 ? exit_status_to_string(si.si_status, EXIT_STATUS_FULL)
2791 : signal_to_string(si.si_status)));
2792
2793 /* Increase the generation counter used for filtering out duplicate unit invocations */
2794 m->sigchldgen++;
2795
2796 /* And now figure out the unit this belongs to, it might be multiple... */
2797 u1 = manager_get_unit_by_pid_cgroup(m, si.si_pid);
2798 u2 = hashmap_get(m->watch_pids, PID_TO_PTR(si.si_pid));
2799 array = hashmap_get(m->watch_pids, PID_TO_PTR(-si.si_pid));
2800 if (array) {
2801 size_t n = 0;
2802
2803 /* Count how many entries the array has */
2804 while (array[n])
2805 n++;
2806
2807 /* Make a copy of the array so that we don't trip up on the array changing beneath us */
2808 array_copy = newdup(Unit*, array, n+1);
2809 if (!array_copy)
2810 log_oom();
2811 }
2812
2813 /* Finally, execute them all. Note that u1, u2 and the array might contain duplicates, but
2814 * that's fine, manager_invoke_sigchld_event() will ensure we only invoke the handlers once for
2815 * each iteration. */
2816 if (u1) {
2817 /* We check for oom condition, in case we got SIGCHLD before the oom notification.
2818 * We only do this for the cgroup the PID belonged to. */
2819 (void) unit_check_oom(u1);
2820
2821 /* We check if systemd-oomd performed a kill so that we log and notify appropriately */
2822 (void) unit_check_oomd_kill(u1);
2823
2824 manager_invoke_sigchld_event(m, u1, &si);
2825 }
2826 if (u2)
2827 manager_invoke_sigchld_event(m, u2, &si);
2828 if (array_copy)
2829 for (size_t i = 0; array_copy[i]; i++)
2830 manager_invoke_sigchld_event(m, array_copy[i], &si);
2831 }
2832
2833 /* And now, we actually reap the zombie. */
2834 if (waitid(P_PID, si.si_pid, &si, WEXITED) < 0) {
2835 log_error_errno(errno, "Failed to dequeue child, ignoring: %m");
2836 return 0;
2837 }
2838
2839 return 0;
2840
2841 turn_off:
2842 /* All children processed for now, turn off event source */
2843
2844 r = sd_event_source_set_enabled(m->sigchld_event_source, SD_EVENT_OFF);
2845 if (r < 0)
2846 return log_error_errno(r, "Failed to disable SIGCHLD event source: %m");
2847
2848 return 0;
2849 }
2850
2851 static void manager_start_special(Manager *m, const char *name, JobMode mode) {
2852 Job *job;
2853
2854 if (manager_add_job_by_name_and_warn(m, JOB_START, name, mode, NULL, &job) < 0)
2855 return;
2856
2857 const char *s = unit_status_string(job->unit, NULL);
2858
2859 log_info("Activating special unit %s...", s);
2860
2861 sd_notifyf(false,
2862 "STATUS=Activating special unit %s...", s);
2863 m->status_ready = false;
2864 }
2865
2866 static void manager_handle_ctrl_alt_del(Manager *m) {
2867 /* If the user presses C-A-D more than
2868 * 7 times within 2s, we reboot/shutdown immediately,
2869 * unless it was disabled in system.conf */
2870
2871 if (ratelimit_below(&m->ctrl_alt_del_ratelimit) || m->cad_burst_action == EMERGENCY_ACTION_NONE)
2872 manager_start_special(m, SPECIAL_CTRL_ALT_DEL_TARGET, JOB_REPLACE_IRREVERSIBLY);
2873 else
2874 emergency_action(m, m->cad_burst_action, EMERGENCY_ACTION_WARN, NULL, -1,
2875 "Ctrl-Alt-Del was pressed more than 7 times within 2s");
2876 }
2877
2878 static int manager_dispatch_signal_fd(sd_event_source *source, int fd, uint32_t revents, void *userdata) {
2879 Manager *m = ASSERT_PTR(userdata);
2880 ssize_t n;
2881 struct signalfd_siginfo sfsi;
2882 int r;
2883
2884 assert(m->signal_fd == fd);
2885
2886 if (revents != EPOLLIN) {
2887 log_warning("Got unexpected events from signal file descriptor.");
2888 return 0;
2889 }
2890
2891 n = read(m->signal_fd, &sfsi, sizeof(sfsi));
2892 if (n < 0) {
2893 if (ERRNO_IS_TRANSIENT(errno))
2894 return 0;
2895
2896 /* We return an error here, which will kill this handler,
2897 * to avoid a busy loop on read error. */
2898 return log_error_errno(errno, "Reading from signal fd failed: %m");
2899 }
2900 if (n != sizeof(sfsi)) {
2901 log_warning("Truncated read from signal fd (%zi bytes), ignoring!", n);
2902 return 0;
2903 }
2904
2905 log_received_signal(sfsi.ssi_signo == SIGCHLD ||
2906 (sfsi.ssi_signo == SIGTERM && MANAGER_IS_USER(m))
2907 ? LOG_DEBUG : LOG_INFO,
2908 &sfsi);
2909
2910 switch (sfsi.ssi_signo) {
2911
2912 case SIGCHLD:
2913 r = sd_event_source_set_enabled(m->sigchld_event_source, SD_EVENT_ON);
2914 if (r < 0)
2915 log_warning_errno(r, "Failed to enable SIGCHLD event source, ignoring: %m");
2916
2917 break;
2918
2919 case SIGTERM:
2920 if (MANAGER_IS_SYSTEM(m)) {
2921 /* This is for compatibility with the original sysvinit */
2922 if (verify_run_space_and_log("Refusing to reexecute") < 0)
2923 break;
2924
2925 m->objective = MANAGER_REEXECUTE;
2926 break;
2927 }
2928
2929 _fallthrough_;
2930 case SIGINT:
2931 if (MANAGER_IS_SYSTEM(m))
2932 manager_handle_ctrl_alt_del(m);
2933 else
2934 manager_start_special(m, SPECIAL_EXIT_TARGET, JOB_REPLACE_IRREVERSIBLY);
2935 break;
2936
2937 case SIGWINCH:
2938 /* This is a nop on non-init */
2939 if (MANAGER_IS_SYSTEM(m))
2940 manager_start_special(m, SPECIAL_KBREQUEST_TARGET, JOB_REPLACE);
2941
2942 break;
2943
2944 case SIGPWR:
2945 /* This is a nop on non-init */
2946 if (MANAGER_IS_SYSTEM(m))
2947 manager_start_special(m, SPECIAL_SIGPWR_TARGET, JOB_REPLACE);
2948
2949 break;
2950
2951 case SIGUSR1:
2952 if (manager_dbus_is_running(m, false)) {
2953 log_info("Trying to reconnect to bus...");
2954
2955 (void) bus_init_api(m);
2956
2957 if (MANAGER_IS_SYSTEM(m))
2958 (void) bus_init_system(m);
2959 } else
2960 manager_start_special(m, SPECIAL_DBUS_SERVICE, JOB_REPLACE);
2961
2962 break;
2963
2964 case SIGUSR2: {
2965 _cleanup_free_ char *dump = NULL;
2966
2967 r = manager_get_dump_string(m, /* patterns= */ NULL, &dump);
2968 if (r < 0) {
2969 log_warning_errno(errno, "Failed to acquire manager dump: %m");
2970 break;
2971 }
2972
2973 log_dump(LOG_INFO, dump);
2974 break;
2975 }
2976
2977 case SIGHUP:
2978 if (verify_run_space_and_log("Refusing to reload") < 0)
2979 break;
2980
2981 m->objective = MANAGER_RELOAD;
2982 break;
2983
2984 default: {
2985
2986 /* Starting SIGRTMIN+0 */
2987 static const struct {
2988 const char *target;
2989 JobMode mode;
2990 } target_table[] = {
2991 [0] = { SPECIAL_DEFAULT_TARGET, JOB_ISOLATE },
2992 [1] = { SPECIAL_RESCUE_TARGET, JOB_ISOLATE },
2993 [2] = { SPECIAL_EMERGENCY_TARGET, JOB_ISOLATE },
2994 [3] = { SPECIAL_HALT_TARGET, JOB_REPLACE_IRREVERSIBLY },
2995 [4] = { SPECIAL_POWEROFF_TARGET, JOB_REPLACE_IRREVERSIBLY },
2996 [5] = { SPECIAL_REBOOT_TARGET, JOB_REPLACE_IRREVERSIBLY },
2997 [6] = { SPECIAL_KEXEC_TARGET, JOB_REPLACE_IRREVERSIBLY },
2998 [7] = { SPECIAL_SOFT_REBOOT_TARGET, JOB_REPLACE_IRREVERSIBLY },
2999 };
3000
3001 /* Starting SIGRTMIN+13, so that target halt and system halt are 10 apart */
3002 static const ManagerObjective objective_table[] = {
3003 [0] = MANAGER_HALT,
3004 [1] = MANAGER_POWEROFF,
3005 [2] = MANAGER_REBOOT,
3006 [3] = MANAGER_KEXEC,
3007 [4] = MANAGER_SOFT_REBOOT,
3008 };
3009
3010 if ((int) sfsi.ssi_signo >= SIGRTMIN+0 &&
3011 (int) sfsi.ssi_signo < SIGRTMIN+(int) ELEMENTSOF(target_table)) {
3012 int idx = (int) sfsi.ssi_signo - SIGRTMIN;
3013 manager_start_special(m, target_table[idx].target, target_table[idx].mode);
3014 break;
3015 }
3016
3017 if ((int) sfsi.ssi_signo >= SIGRTMIN+13 &&
3018 (int) sfsi.ssi_signo < SIGRTMIN+13+(int) ELEMENTSOF(objective_table)) {
3019 m->objective = objective_table[sfsi.ssi_signo - SIGRTMIN - 13];
3020 break;
3021 }
3022
3023 switch (sfsi.ssi_signo - SIGRTMIN) {
3024
3025 case 18: {
3026 bool generic = false;
3027
3028 if (sfsi.ssi_code != SI_QUEUE)
3029 generic = true;
3030 else {
3031 /* Override a few select commands by our own PID1-specific logic */
3032
3033 switch (sfsi.ssi_int) {
3034
3035 case _COMMON_SIGNAL_COMMAND_LOG_LEVEL_BASE..._COMMON_SIGNAL_COMMAND_LOG_LEVEL_END:
3036 manager_override_log_level(m, sfsi.ssi_int - _COMMON_SIGNAL_COMMAND_LOG_LEVEL_BASE);
3037 break;
3038
3039 case COMMON_SIGNAL_COMMAND_CONSOLE:
3040 manager_override_log_target(m, LOG_TARGET_CONSOLE);
3041 break;
3042
3043 case COMMON_SIGNAL_COMMAND_JOURNAL:
3044 manager_override_log_target(m, LOG_TARGET_JOURNAL);
3045 break;
3046
3047 case COMMON_SIGNAL_COMMAND_KMSG:
3048 manager_override_log_target(m, LOG_TARGET_KMSG);
3049 break;
3050
3051 case COMMON_SIGNAL_COMMAND_NULL:
3052 manager_override_log_target(m, LOG_TARGET_NULL);
3053 break;
3054
3055 default:
3056 generic = true;
3057 }
3058 }
3059
3060 if (generic)
3061 return sigrtmin18_handler(source, &sfsi, NULL);
3062
3063 break;
3064 }
3065
3066 case 20:
3067 manager_override_show_status(m, SHOW_STATUS_YES, "signal");
3068 break;
3069
3070 case 21:
3071 manager_override_show_status(m, SHOW_STATUS_NO, "signal");
3072 break;
3073
3074 case 22:
3075 manager_override_log_level(m, LOG_DEBUG);
3076 break;
3077
3078 case 23:
3079 manager_restore_original_log_level(m);
3080 break;
3081
3082 case 24:
3083 if (MANAGER_IS_USER(m)) {
3084 m->objective = MANAGER_EXIT;
3085 return 0;
3086 }
3087
3088 /* This is a nop on init */
3089 break;
3090
3091 case 25:
3092 m->objective = MANAGER_REEXECUTE;
3093 break;
3094
3095 case 26:
3096 case 29: /* compatibility: used to be mapped to LOG_TARGET_SYSLOG_OR_KMSG */
3097 manager_restore_original_log_target(m);
3098 break;
3099
3100 case 27:
3101 manager_override_log_target(m, LOG_TARGET_CONSOLE);
3102 break;
3103
3104 case 28:
3105 manager_override_log_target(m, LOG_TARGET_KMSG);
3106 break;
3107
3108 default:
3109 log_warning("Got unhandled signal <%s>.", signal_to_string(sfsi.ssi_signo));
3110 }
3111 }}
3112
3113 return 0;
3114 }
3115
3116 static int manager_dispatch_time_change_fd(sd_event_source *source, int fd, uint32_t revents, void *userdata) {
3117 Manager *m = ASSERT_PTR(userdata);
3118 Unit *u;
3119
3120 log_struct(LOG_DEBUG,
3121 "MESSAGE_ID=" SD_MESSAGE_TIME_CHANGE_STR,
3122 LOG_MESSAGE("Time has been changed"));
3123
3124 /* Restart the watch */
3125 (void) manager_setup_time_change(m);
3126
3127 HASHMAP_FOREACH(u, m->units)
3128 if (UNIT_VTABLE(u)->time_change)
3129 UNIT_VTABLE(u)->time_change(u);
3130
3131 return 0;
3132 }
3133
3134 static int manager_dispatch_timezone_change(
3135 sd_event_source *source,
3136 const struct inotify_event *e,
3137 void *userdata) {
3138
3139 Manager *m = ASSERT_PTR(userdata);
3140 int changed;
3141 Unit *u;
3142
3143 log_debug("inotify event for /etc/localtime");
3144
3145 changed = manager_read_timezone_stat(m);
3146 if (changed <= 0)
3147 return changed;
3148
3149 /* Something changed, restart the watch, to ensure we watch the new /etc/localtime if it changed */
3150 (void) manager_setup_timezone_change(m);
3151
3152 /* Read the new timezone */
3153 tzset();
3154
3155 log_debug("Timezone has been changed (now: %s).", tzname[daylight]);
3156
3157 HASHMAP_FOREACH(u, m->units)
3158 if (UNIT_VTABLE(u)->timezone_change)
3159 UNIT_VTABLE(u)->timezone_change(u);
3160
3161 return 0;
3162 }
3163
3164 static int manager_dispatch_idle_pipe_fd(sd_event_source *source, int fd, uint32_t revents, void *userdata) {
3165 Manager *m = ASSERT_PTR(userdata);
3166
3167 assert(m->idle_pipe[2] == fd);
3168
3169 /* There's at least one Type=idle child that just gave up on us waiting for the boot process to
3170 * complete. Let's now turn off any further console output if there's at least one service that needs
3171 * console access, so that from now on our own output should not spill into that service's output
3172 * anymore. After all, we support Type=idle only to beautify console output and it generally is set
3173 * on services that want to own the console exclusively without our interference. */
3174 m->no_console_output = m->n_on_console > 0;
3175
3176 /* Acknowledge the child's request, and let all other children know too that they shouldn't wait
3177 * any longer by closing the pipes towards them, which is what they are waiting for. */
3178 manager_close_idle_pipe(m);
3179
3180 return 0;
3181 }
3182
3183 static int manager_dispatch_jobs_in_progress(sd_event_source *source, usec_t usec, void *userdata) {
3184 Manager *m = ASSERT_PTR(userdata);
3185 int r;
3186
3187 assert(source);
3188
3189 manager_print_jobs_in_progress(m);
3190
3191 r = sd_event_source_set_time_relative(source, JOBS_IN_PROGRESS_PERIOD_USEC);
3192 if (r < 0)
3193 return r;
3194
3195 return sd_event_source_set_enabled(source, SD_EVENT_ONESHOT);
3196 }
3197
3198 int manager_loop(Manager *m) {
3199 RateLimit rl = { .interval = 1*USEC_PER_SEC, .burst = 50000 };
3200 int r;
3201
3202 assert(m);
3203 assert(m->objective == MANAGER_OK); /* Ensure manager_startup() has been called */
3204
3205 manager_check_finished(m);
3206
3207 /* There might still be some zombies hanging around from before we were exec()'ed. Let's reap them. */
3208 r = sd_event_source_set_enabled(m->sigchld_event_source, SD_EVENT_ON);
3209 if (r < 0)
3210 return log_error_errno(r, "Failed to enable SIGCHLD event source: %m");
3211
3212 while (m->objective == MANAGER_OK) {
3213
3214 (void) watchdog_ping();
3215
3216 if (!ratelimit_below(&rl)) {
3217 /* Yay, something is going seriously wrong, pause a little */
3218 log_warning("Looping too fast. Throttling execution a little.");
3219 sleep(1);
3220 }
3221
3222 if (manager_dispatch_load_queue(m) > 0)
3223 continue;
3224
3225 if (manager_dispatch_gc_job_queue(m) > 0)
3226 continue;
3227
3228 if (manager_dispatch_gc_unit_queue(m) > 0)
3229 continue;
3230
3231 if (manager_dispatch_cleanup_queue(m) > 0)
3232 continue;
3233
3234 if (manager_dispatch_cgroup_realize_queue(m) > 0)
3235 continue;
3236
3237 if (manager_dispatch_start_when_upheld_queue(m) > 0)
3238 continue;
3239
3240 if (manager_dispatch_stop_when_bound_queue(m) > 0)
3241 continue;
3242
3243 if (manager_dispatch_stop_when_unneeded_queue(m) > 0)
3244 continue;
3245
3246 if (manager_dispatch_release_resources_queue(m) > 0)
3247 continue;
3248
3249 if (manager_dispatch_dbus_queue(m) > 0)
3250 continue;
3251
3252 /* Sleep for watchdog runtime wait time */
3253 r = sd_event_run(m->event, watchdog_runtime_wait());
3254 if (r < 0)
3255 return log_error_errno(r, "Failed to run event loop: %m");
3256 }
3257
3258 return m->objective;
3259 }
3260
3261 int manager_load_unit_from_dbus_path(Manager *m, const char *s, sd_bus_error *e, Unit **_u) {
3262 _cleanup_free_ char *n = NULL;
3263 sd_id128_t invocation_id;
3264 Unit *u;
3265 int r;
3266
3267 assert(m);
3268 assert(s);
3269 assert(_u);
3270
3271 r = unit_name_from_dbus_path(s, &n);
3272 if (r < 0)
3273 return r;
3274
3275 /* Permit addressing units by invocation ID: if the passed bus path is suffixed by a 128-bit ID then
3276 * we use it as invocation ID. */
3277 r = sd_id128_from_string(n, &invocation_id);
3278 if (r >= 0) {
3279 u = hashmap_get(m->units_by_invocation_id, &invocation_id);
3280 if (u) {
3281 *_u = u;
3282 return 0;
3283 }
3284
3285 return sd_bus_error_setf(e, BUS_ERROR_NO_UNIT_FOR_INVOCATION_ID,
3286 "No unit with the specified invocation ID " SD_ID128_FORMAT_STR " known.",
3287 SD_ID128_FORMAT_VAL(invocation_id));
3288 }
3289
3290 /* If this didn't work, we check if this is a unit name */
3291 if (!unit_name_is_valid(n, UNIT_NAME_PLAIN|UNIT_NAME_INSTANCE)) {
3292 _cleanup_free_ char *nn = NULL;
3293
3294 nn = cescape(n);
3295 return sd_bus_error_setf(e, SD_BUS_ERROR_INVALID_ARGS,
3296 "Unit name %s is neither a valid invocation ID nor unit name.", strnull(nn));
3297 }
3298
3299 r = manager_load_unit(m, n, NULL, e, &u);
3300 if (r < 0)
3301 return r;
3302
3303 *_u = u;
3304 return 0;
3305 }
3306
3307 int manager_get_job_from_dbus_path(Manager *m, const char *s, Job **_j) {
3308 const char *p;
3309 unsigned id;
3310 Job *j;
3311 int r;
3312
3313 assert(m);
3314 assert(s);
3315 assert(_j);
3316
3317 p = startswith(s, "/org/freedesktop/systemd1/job/");
3318 if (!p)
3319 return -EINVAL;
3320
3321 r = safe_atou(p, &id);
3322 if (r < 0)
3323 return r;
3324
3325 j = manager_get_job(m, id);
3326 if (!j)
3327 return -ENOENT;
3328
3329 *_j = j;
3330
3331 return 0;
3332 }
3333
3334 void manager_send_unit_audit(Manager *m, Unit *u, int type, bool success) {
3335
3336 #if HAVE_AUDIT
3337 _cleanup_free_ char *p = NULL;
3338 const char *msg;
3339 int audit_fd, r;
3340
3341 if (!MANAGER_IS_SYSTEM(m))
3342 return;
3343
3344 audit_fd = get_audit_fd();
3345 if (audit_fd < 0)
3346 return;
3347
3348 /* Don't generate audit events if the service was already
3349 * started and we're just deserializing */
3350 if (MANAGER_IS_RELOADING(m))
3351 return;
3352
3353 r = unit_name_to_prefix_and_instance(u->id, &p);
3354 if (r < 0) {
3355 log_warning_errno(r, "Failed to extract prefix and instance of unit name, ignoring: %m");
3356 return;
3357 }
3358
3359 msg = strjoina("unit=", p);
3360 if (audit_log_user_comm_message(audit_fd, type, msg, "systemd", NULL, NULL, NULL, success) < 0) {
3361 if (ERRNO_IS_PRIVILEGE(errno)) {
3362 /* We aren't allowed to send audit messages? Then let's not retry again. */
3363 log_debug_errno(errno, "Failed to send audit message, closing audit socket: %m");
3364 close_audit_fd();
3365 } else
3366 log_warning_errno(errno, "Failed to send audit message, ignoring: %m");
3367 }
3368 #endif
3369
3370 }
3371
3372 void manager_send_unit_plymouth(Manager *m, Unit *u) {
3373 static const union sockaddr_union sa = PLYMOUTH_SOCKET;
3374 _cleanup_free_ char *message = NULL;
3375 _cleanup_close_ int fd = -EBADF;
3376 int n = 0;
3377
3378 /* Don't generate plymouth events if the service was already
3379 * started and we're just deserializing */
3380 if (MANAGER_IS_RELOADING(m))
3381 return;
3382
3383 if (!MANAGER_IS_SYSTEM(m))
3384 return;
3385
3386 if (detect_container() > 0)
3387 return;
3388
3389 if (!UNIT_VTABLE(u)->notify_plymouth)
3390 return;
3391
3392 /* We set SOCK_NONBLOCK here so that we rather drop the
3393 * message then wait for plymouth */
3394 fd = socket(AF_UNIX, SOCK_STREAM|SOCK_CLOEXEC|SOCK_NONBLOCK, 0);
3395 if (fd < 0) {
3396 log_error_errno(errno, "socket() failed: %m");
3397 return;
3398 }
3399
3400 if (connect(fd, &sa.sa, SOCKADDR_UN_LEN(sa.un)) < 0) {
3401 if (!IN_SET(errno, EAGAIN, ENOENT) && !ERRNO_IS_DISCONNECT(errno))
3402 log_error_errno(errno, "connect() failed: %m");
3403 return;
3404 }
3405
3406 if (asprintf(&message, "U\002%c%s%n", (int) (strlen(u->id) + 1), u->id, &n) < 0)
3407 return (void) log_oom();
3408
3409 errno = 0;
3410 if (write(fd, message, n + 1) != n + 1)
3411 if (!IN_SET(errno, EAGAIN, ENOENT) && !ERRNO_IS_DISCONNECT(errno))
3412 log_error_errno(errno, "Failed to write Plymouth message: %m");
3413 }
3414
3415 usec_t manager_get_watchdog(Manager *m, WatchdogType t) {
3416 assert(m);
3417
3418 if (MANAGER_IS_USER(m))
3419 return USEC_INFINITY;
3420
3421 if (timestamp_is_set(m->watchdog_overridden[t]))
3422 return m->watchdog_overridden[t];
3423
3424 return m->watchdog[t];
3425 }
3426
3427 void manager_set_watchdog(Manager *m, WatchdogType t, usec_t timeout) {
3428
3429 assert(m);
3430
3431 if (MANAGER_IS_USER(m))
3432 return;
3433
3434 if (m->watchdog[t] == timeout)
3435 return;
3436
3437 if (t == WATCHDOG_RUNTIME) {
3438 if (!timestamp_is_set(m->watchdog_overridden[WATCHDOG_RUNTIME]))
3439 (void) watchdog_setup(timeout);
3440 } else if (t == WATCHDOG_PRETIMEOUT)
3441 if (m->watchdog_overridden[WATCHDOG_PRETIMEOUT] == USEC_INFINITY)
3442 (void) watchdog_setup_pretimeout(timeout);
3443
3444 m->watchdog[t] = timeout;
3445 }
3446
3447 void manager_override_watchdog(Manager *m, WatchdogType t, usec_t timeout) {
3448
3449 assert(m);
3450
3451 if (MANAGER_IS_USER(m))
3452 return;
3453
3454 if (m->watchdog_overridden[t] == timeout)
3455 return;
3456
3457 if (t == WATCHDOG_RUNTIME) {
3458 usec_t usec = timestamp_is_set(timeout) ? timeout : m->watchdog[t];
3459
3460 (void) watchdog_setup(usec);
3461 } else if (t == WATCHDOG_PRETIMEOUT)
3462 (void) watchdog_setup_pretimeout(timeout);
3463
3464 m->watchdog_overridden[t] = timeout;
3465 }
3466
3467 int manager_set_watchdog_pretimeout_governor(Manager *m, const char *governor) {
3468 _cleanup_free_ char *p = NULL;
3469 int r;
3470
3471 assert(m);
3472
3473 if (MANAGER_IS_USER(m))
3474 return 0;
3475
3476 if (streq_ptr(m->watchdog_pretimeout_governor, governor))
3477 return 0;
3478
3479 p = strdup(governor);
3480 if (!p)
3481 return -ENOMEM;
3482
3483 r = watchdog_setup_pretimeout_governor(governor);
3484 if (r < 0)
3485 return r;
3486
3487 return free_and_replace(m->watchdog_pretimeout_governor, p);
3488 }
3489
3490 int manager_override_watchdog_pretimeout_governor(Manager *m, const char *governor) {
3491 _cleanup_free_ char *p = NULL;
3492 int r;
3493
3494 assert(m);
3495
3496 if (MANAGER_IS_USER(m))
3497 return 0;
3498
3499 if (streq_ptr(m->watchdog_pretimeout_governor_overridden, governor))
3500 return 0;
3501
3502 p = strdup(governor);
3503 if (!p)
3504 return -ENOMEM;
3505
3506 r = watchdog_setup_pretimeout_governor(governor);
3507 if (r < 0)
3508 return r;
3509
3510 return free_and_replace(m->watchdog_pretimeout_governor_overridden, p);
3511 }
3512
3513 int manager_reload(Manager *m) {
3514 _unused_ _cleanup_(manager_reloading_stopp) Manager *reloading = NULL;
3515 _cleanup_fdset_free_ FDSet *fds = NULL;
3516 _cleanup_fclose_ FILE *f = NULL;
3517 int r;
3518
3519 assert(m);
3520
3521 r = manager_open_serialization(m, &f);
3522 if (r < 0)
3523 return log_error_errno(r, "Failed to create serialization file: %m");
3524
3525 fds = fdset_new();
3526 if (!fds)
3527 return log_oom();
3528
3529 /* We are officially in reload mode from here on. */
3530 reloading = manager_reloading_start(m);
3531
3532 r = manager_serialize(m, f, fds, false);
3533 if (r < 0)
3534 return r;
3535
3536 if (fseeko(f, 0, SEEK_SET) < 0)
3537 return log_error_errno(errno, "Failed to seek to beginning of serialization: %m");
3538
3539 /* 💀 This is the point of no return, from here on there is no way back. 💀 */
3540 reloading = NULL;
3541
3542 bus_manager_send_reloading(m, true);
3543
3544 /* Start by flushing out all jobs and units, all generated units, all runtime environments, all dynamic users
3545 * and everything else that is worth flushing out. We'll get it all back from the serialization — if we need
3546 * it. */
3547
3548 manager_clear_jobs_and_units(m);
3549 lookup_paths_flush_generator(&m->lookup_paths);
3550 lookup_paths_free(&m->lookup_paths);
3551 exec_shared_runtime_vacuum(m);
3552 dynamic_user_vacuum(m, false);
3553 m->uid_refs = hashmap_free(m->uid_refs);
3554 m->gid_refs = hashmap_free(m->gid_refs);
3555
3556 r = lookup_paths_init_or_warn(&m->lookup_paths, m->runtime_scope, 0, NULL);
3557 if (r < 0)
3558 return r;
3559
3560 (void) manager_run_environment_generators(m);
3561 (void) manager_run_generators(m);
3562
3563 lookup_paths_log(&m->lookup_paths);
3564
3565 /* We flushed out generated files, for which we don't watch mtime, so we should flush the old map. */
3566 manager_free_unit_name_maps(m);
3567
3568 /* First, enumerate what we can from kernel and suchlike */
3569 manager_enumerate_perpetual(m);
3570 manager_enumerate(m);
3571
3572 /* Second, deserialize our stored data */
3573 r = manager_deserialize(m, f, fds);
3574 if (r < 0)
3575 log_warning_errno(r, "Deserialization failed, proceeding anyway: %m");
3576
3577 /* We don't need the serialization anymore */
3578 f = safe_fclose(f);
3579
3580 /* Re-register notify_fd as event source, and set up other sockets/communication channels we might need */
3581 (void) manager_setup_notify(m);
3582 (void) manager_setup_cgroups_agent(m);
3583 (void) manager_setup_user_lookup_fd(m);
3584
3585 /* Third, fire things up! */
3586 manager_coldplug(m);
3587
3588 /* Clean up runtime objects no longer referenced */
3589 manager_vacuum(m);
3590
3591 /* Clean up deserialized tracked clients */
3592 m->deserialized_subscribed = strv_free(m->deserialized_subscribed);
3593
3594 /* Consider the reload process complete now. */
3595 assert(m->n_reloading > 0);
3596 m->n_reloading--;
3597
3598 manager_ready(m);
3599
3600 m->send_reloading_done = true;
3601 return 0;
3602 }
3603
3604 void manager_reset_failed(Manager *m) {
3605 Unit *u;
3606
3607 assert(m);
3608
3609 HASHMAP_FOREACH(u, m->units)
3610 unit_reset_failed(u);
3611 }
3612
3613 bool manager_unit_inactive_or_pending(Manager *m, const char *name) {
3614 Unit *u;
3615
3616 assert(m);
3617 assert(name);
3618
3619 /* Returns true if the unit is inactive or going down */
3620 u = manager_get_unit(m, name);
3621 if (!u)
3622 return true;
3623
3624 return unit_inactive_or_pending(u);
3625 }
3626
3627 static void log_taint_string(Manager *m) {
3628 _cleanup_free_ char *taint = NULL;
3629
3630 assert(m);
3631
3632 if (MANAGER_IS_USER(m) || m->taint_logged)
3633 return;
3634
3635 m->taint_logged = true; /* only check for taint once */
3636
3637 taint = manager_taint_string(m);
3638 if (isempty(taint))
3639 return;
3640
3641 log_struct(LOG_NOTICE,
3642 LOG_MESSAGE("System is tainted: %s", taint),
3643 "TAINT=%s", taint,
3644 "MESSAGE_ID=" SD_MESSAGE_TAINTED_STR);
3645 }
3646
3647 static void manager_notify_finished(Manager *m) {
3648 usec_t firmware_usec, loader_usec, kernel_usec, initrd_usec, userspace_usec, total_usec;
3649
3650 if (MANAGER_IS_TEST_RUN(m))
3651 return;
3652
3653 if (MANAGER_IS_SYSTEM(m) && detect_container() <= 0) {
3654 char buf[FORMAT_TIMESPAN_MAX + STRLEN(" (firmware) + ") + FORMAT_TIMESPAN_MAX + STRLEN(" (loader) + ")]
3655 = {};
3656 char *p = buf;
3657 size_t size = sizeof buf;
3658
3659 /* Note that MANAGER_TIMESTAMP_KERNEL's monotonic value is always at 0, and
3660 * MANAGER_TIMESTAMP_FIRMWARE's and MANAGER_TIMESTAMP_LOADER's monotonic value should be considered
3661 * negative values. */
3662
3663 firmware_usec = m->timestamps[MANAGER_TIMESTAMP_FIRMWARE].monotonic - m->timestamps[MANAGER_TIMESTAMP_LOADER].monotonic;
3664 loader_usec = m->timestamps[MANAGER_TIMESTAMP_LOADER].monotonic - m->timestamps[MANAGER_TIMESTAMP_KERNEL].monotonic;
3665 userspace_usec = m->timestamps[MANAGER_TIMESTAMP_FINISH].monotonic - m->timestamps[MANAGER_TIMESTAMP_USERSPACE].monotonic;
3666 total_usec = m->timestamps[MANAGER_TIMESTAMP_FIRMWARE].monotonic + m->timestamps[MANAGER_TIMESTAMP_FINISH].monotonic;
3667
3668 if (firmware_usec > 0)
3669 size = strpcpyf(&p, size, "%s (firmware) + ", FORMAT_TIMESPAN(firmware_usec, USEC_PER_MSEC));
3670 if (loader_usec > 0)
3671 size = strpcpyf(&p, size, "%s (loader) + ", FORMAT_TIMESPAN(loader_usec, USEC_PER_MSEC));
3672
3673 if (dual_timestamp_is_set(&m->timestamps[MANAGER_TIMESTAMP_INITRD])) {
3674
3675 /* The initrd case on bare-metal */
3676 kernel_usec = m->timestamps[MANAGER_TIMESTAMP_INITRD].monotonic - m->timestamps[MANAGER_TIMESTAMP_KERNEL].monotonic;
3677 initrd_usec = m->timestamps[MANAGER_TIMESTAMP_USERSPACE].monotonic - m->timestamps[MANAGER_TIMESTAMP_INITRD].monotonic;
3678
3679 log_struct(LOG_INFO,
3680 "MESSAGE_ID=" SD_MESSAGE_STARTUP_FINISHED_STR,
3681 "KERNEL_USEC="USEC_FMT, kernel_usec,
3682 "INITRD_USEC="USEC_FMT, initrd_usec,
3683 "USERSPACE_USEC="USEC_FMT, userspace_usec,
3684 LOG_MESSAGE("Startup finished in %s%s (kernel) + %s (initrd) + %s (userspace) = %s.",
3685 buf,
3686 FORMAT_TIMESPAN(kernel_usec, USEC_PER_MSEC),
3687 FORMAT_TIMESPAN(initrd_usec, USEC_PER_MSEC),
3688 FORMAT_TIMESPAN(userspace_usec, USEC_PER_MSEC),
3689 FORMAT_TIMESPAN(total_usec, USEC_PER_MSEC)));
3690 } else {
3691 /* The initrd-less case on bare-metal */
3692
3693 kernel_usec = m->timestamps[MANAGER_TIMESTAMP_USERSPACE].monotonic - m->timestamps[MANAGER_TIMESTAMP_KERNEL].monotonic;
3694 initrd_usec = 0;
3695
3696 log_struct(LOG_INFO,
3697 "MESSAGE_ID=" SD_MESSAGE_STARTUP_FINISHED_STR,
3698 "KERNEL_USEC="USEC_FMT, kernel_usec,
3699 "USERSPACE_USEC="USEC_FMT, userspace_usec,
3700 LOG_MESSAGE("Startup finished in %s%s (kernel) + %s (userspace) = %s.",
3701 buf,
3702 FORMAT_TIMESPAN(kernel_usec, USEC_PER_MSEC),
3703 FORMAT_TIMESPAN(userspace_usec, USEC_PER_MSEC),
3704 FORMAT_TIMESPAN(total_usec, USEC_PER_MSEC)));
3705 }
3706 } else {
3707 /* The container and --user case */
3708 firmware_usec = loader_usec = initrd_usec = kernel_usec = 0;
3709 total_usec = userspace_usec = m->timestamps[MANAGER_TIMESTAMP_FINISH].monotonic - m->timestamps[MANAGER_TIMESTAMP_USERSPACE].monotonic;
3710
3711 log_struct(LOG_INFO,
3712 "MESSAGE_ID=" SD_MESSAGE_USER_STARTUP_FINISHED_STR,
3713 "USERSPACE_USEC="USEC_FMT, userspace_usec,
3714 LOG_MESSAGE("Startup finished in %s.",
3715 FORMAT_TIMESPAN(total_usec, USEC_PER_MSEC)));
3716 }
3717
3718 bus_manager_send_finished(m, firmware_usec, loader_usec, kernel_usec, initrd_usec, userspace_usec, total_usec);
3719
3720 log_taint_string(m);
3721 }
3722
3723 static void user_manager_send_ready(Manager *m) {
3724 int r;
3725
3726 assert(m);
3727
3728 /* We send READY=1 on reaching basic.target only when running in --user mode. */
3729 if (!MANAGER_IS_USER(m) || m->ready_sent)
3730 return;
3731
3732 r = sd_notify(false,
3733 "READY=1\n"
3734 "STATUS=Reached " SPECIAL_BASIC_TARGET ".");
3735 if (r < 0)
3736 log_warning_errno(r, "Failed to send readiness notification, ignoring: %m");
3737
3738 m->ready_sent = true;
3739 m->status_ready = false;
3740 }
3741
3742 static void manager_send_ready(Manager *m) {
3743 int r;
3744
3745 if (m->ready_sent && m->status_ready)
3746 /* Skip the notification if nothing changed. */
3747 return;
3748
3749 r = sd_notify(false,
3750 "READY=1\n"
3751 "STATUS=Ready.");
3752 if (r < 0)
3753 log_full_errno(m->ready_sent ? LOG_DEBUG : LOG_WARNING, r,
3754 "Failed to send readiness notification, ignoring: %m");
3755
3756 m->ready_sent = m->status_ready = true;
3757 }
3758
3759 static void manager_check_basic_target(Manager *m) {
3760 Unit *u;
3761
3762 assert(m);
3763
3764 /* Small shortcut */
3765 if (m->ready_sent && m->taint_logged)
3766 return;
3767
3768 u = manager_get_unit(m, SPECIAL_BASIC_TARGET);
3769 if (!u || !UNIT_IS_ACTIVE_OR_RELOADING(unit_active_state(u)))
3770 return;
3771
3772 /* For user managers, send out READY=1 as soon as we reach basic.target */
3773 user_manager_send_ready(m);
3774
3775 /* Log the taint string as soon as we reach basic.target */
3776 log_taint_string(m);
3777 }
3778
3779 void manager_check_finished(Manager *m) {
3780 assert(m);
3781
3782 if (MANAGER_IS_RELOADING(m))
3783 return;
3784
3785 /* Verify that we have entered the event loop already, and not left it again. */
3786 if (!MANAGER_IS_RUNNING(m))
3787 return;
3788
3789 manager_check_basic_target(m);
3790
3791 if (hashmap_size(m->jobs) > 0) {
3792 if (m->jobs_in_progress_event_source)
3793 /* Ignore any failure, this is only for feedback */
3794 (void) sd_event_source_set_time(m->jobs_in_progress_event_source,
3795 manager_watch_jobs_next_time(m));
3796 return;
3797 }
3798
3799 /* The jobs hashmap tends to grow a lot during boot, and then it's not reused until shutdown. Let's
3800 kill the hashmap if it is relatively large. */
3801 if (hashmap_buckets(m->jobs) > hashmap_size(m->units) / 10)
3802 m->jobs = hashmap_free(m->jobs);
3803
3804 manager_send_ready(m);
3805
3806 /* Notify Type=idle units that we are done now */
3807 manager_close_idle_pipe(m);
3808
3809 if (MANAGER_IS_FINISHED(m))
3810 return;
3811
3812 manager_flip_auto_status(m, false, "boot finished");
3813
3814 /* Turn off confirm spawn now */
3815 m->confirm_spawn = NULL;
3816
3817 /* No need to update ask password status when we're going non-interactive */
3818 manager_close_ask_password(m);
3819
3820 /* This is no longer the first boot */
3821 manager_set_first_boot(m, false);
3822
3823 dual_timestamp_get(m->timestamps + MANAGER_TIMESTAMP_FINISH);
3824
3825 manager_notify_finished(m);
3826
3827 manager_invalidate_startup_units(m);
3828 }
3829
3830 void manager_send_reloading(Manager *m) {
3831 assert(m);
3832
3833 /* Let whoever invoked us know that we are now reloading */
3834 (void) sd_notifyf(/* unset= */ false,
3835 "RELOADING=1\n"
3836 "MONOTONIC_USEC=" USEC_FMT "\n", now(CLOCK_MONOTONIC));
3837
3838 /* And ensure that we'll send READY=1 again as soon as we are ready again */
3839 m->ready_sent = false;
3840 }
3841
3842 static bool generator_path_any(const char* const* paths) {
3843 bool found = false;
3844
3845 /* Optimize by skipping the whole process by not creating output directories
3846 * if no generators are found. */
3847 STRV_FOREACH(path, paths)
3848 if (access(*path, F_OK) == 0)
3849 found = true;
3850 else if (errno != ENOENT)
3851 log_warning_errno(errno, "Failed to open generator directory %s: %m", *path);
3852
3853 return found;
3854 }
3855
3856 static int manager_run_environment_generators(Manager *m) {
3857 char **tmp = NULL; /* this is only used in the forked process, no cleanup here */
3858 _cleanup_strv_free_ char **paths = NULL;
3859 void* args[] = {
3860 [STDOUT_GENERATE] = &tmp,
3861 [STDOUT_COLLECT] = &tmp,
3862 [STDOUT_CONSUME] = &m->transient_environment,
3863 };
3864 int r;
3865
3866 if (MANAGER_IS_TEST_RUN(m) && !(m->test_run_flags & MANAGER_TEST_RUN_ENV_GENERATORS))
3867 return 0;
3868
3869 paths = env_generator_binary_paths(m->runtime_scope);
3870 if (!paths)
3871 return log_oom();
3872
3873 if (!generator_path_any((const char* const*) paths))
3874 return 0;
3875
3876 WITH_UMASK(0022)
3877 r = execute_directories((const char* const*) paths, DEFAULT_TIMEOUT_USEC, gather_environment,
3878 args, NULL, m->transient_environment,
3879 EXEC_DIR_PARALLEL | EXEC_DIR_IGNORE_ERRORS | EXEC_DIR_SET_SYSTEMD_EXEC_PID);
3880 return r;
3881 }
3882
3883 static int build_generator_environment(Manager *m, char ***ret) {
3884 _cleanup_strv_free_ char **nl = NULL;
3885 Virtualization v;
3886 int r;
3887
3888 assert(m);
3889 assert(ret);
3890
3891 /* Generators oftentimes want to know some basic facts about the environment they run in, in order to
3892 * adjust generated units to that. Let's pass down some bits of information that are easy for us to
3893 * determine (but a bit harder for generator scripts to determine), as environment variables. */
3894
3895 nl = strv_copy(m->transient_environment);
3896 if (!nl)
3897 return -ENOMEM;
3898
3899 r = strv_env_assign(&nl, "SYSTEMD_SCOPE", runtime_scope_to_string(m->runtime_scope));
3900 if (r < 0)
3901 return r;
3902
3903 if (MANAGER_IS_SYSTEM(m)) {
3904 /* Note that $SYSTEMD_IN_INITRD may be used to override the initrd detection in much of our
3905 * codebase. This is hence more than purely informational. It will shortcut detection of the
3906 * initrd state if generators invoke our own tools. But that's OK, as it would come to the
3907 * same results (hopefully). */
3908 r = strv_env_assign(&nl, "SYSTEMD_IN_INITRD", one_zero(in_initrd()));
3909 if (r < 0)
3910 return r;
3911
3912 if (m->first_boot >= 0) {
3913 r = strv_env_assign(&nl, "SYSTEMD_FIRST_BOOT", one_zero(m->first_boot));
3914 if (r < 0)
3915 return r;
3916 }
3917 }
3918
3919 v = detect_virtualization();
3920 if (v < 0)
3921 log_debug_errno(v, "Failed to detect virtualization, ignoring: %m");
3922 else if (v > 0) {
3923 const char *s;
3924
3925 s = strjoina(VIRTUALIZATION_IS_VM(v) ? "vm:" :
3926 VIRTUALIZATION_IS_CONTAINER(v) ? "container:" : ":",
3927 virtualization_to_string(v));
3928
3929 r = strv_env_assign(&nl, "SYSTEMD_VIRTUALIZATION", s);
3930 if (r < 0)
3931 return r;
3932 }
3933
3934 r = strv_env_assign(&nl, "SYSTEMD_ARCHITECTURE", architecture_to_string(uname_architecture()));
3935 if (r < 0)
3936 return r;
3937
3938 *ret = TAKE_PTR(nl);
3939 return 0;
3940 }
3941
3942 static int manager_execute_generators(Manager *m, char **paths, bool remount_ro) {
3943 _cleanup_strv_free_ char **ge = NULL;
3944 const char *argv[] = {
3945 NULL, /* Leave this empty, execute_directory() will fill something in */
3946 m->lookup_paths.generator,
3947 m->lookup_paths.generator_early,
3948 m->lookup_paths.generator_late,
3949 NULL,
3950 };
3951 int r;
3952
3953 r = build_generator_environment(m, &ge);
3954 if (r < 0)
3955 return log_error_errno(r, "Failed to build generator environment: %m");
3956
3957 if (remount_ro) {
3958 /* Remount most of the filesystem tree read-only. We leave /sys/ as-is, because our code
3959 * checks whether it is read-only to detect containerized execution environments. We leave
3960 * /run/ as-is too, because that's where our output goes. We also leave /proc/ and /dev/shm/
3961 * because they're API, and /tmp/ that safe_fork() mounted for us.
3962 */
3963 r = bind_remount_recursive("/", MS_RDONLY, MS_RDONLY,
3964 STRV_MAKE("/sys", "/run", "/proc", "/dev/shm", "/tmp"));
3965 if (r < 0)
3966 log_warning_errno(r, "Read-only bind remount failed, ignoring: %m");
3967 }
3968
3969 BLOCK_WITH_UMASK(0022);
3970 return execute_directories(
3971 (const char* const*) paths,
3972 DEFAULT_TIMEOUT_USEC,
3973 /* callbacks= */ NULL, /* callback_args= */ NULL,
3974 (char**) argv,
3975 ge,
3976 EXEC_DIR_PARALLEL | EXEC_DIR_IGNORE_ERRORS | EXEC_DIR_SET_SYSTEMD_EXEC_PID);
3977 }
3978
3979 static int manager_run_generators(Manager *m) {
3980 ForkFlags flags = FORK_RESET_SIGNALS | FORK_WAIT | FORK_NEW_MOUNTNS | FORK_MOUNTNS_SLAVE;
3981 _cleanup_strv_free_ char **paths = NULL;
3982 int r;
3983
3984 assert(m);
3985
3986 if (MANAGER_IS_TEST_RUN(m) && !(m->test_run_flags & MANAGER_TEST_RUN_GENERATORS))
3987 return 0;
3988
3989 paths = generator_binary_paths(m->runtime_scope);
3990 if (!paths)
3991 return log_oom();
3992
3993 if (!generator_path_any((const char* const*) paths))
3994 return 0;
3995
3996 r = lookup_paths_mkdir_generator(&m->lookup_paths);
3997 if (r < 0) {
3998 log_error_errno(r, "Failed to create generator directories: %m");
3999 goto finish;
4000 }
4001
4002 /* If we are the system manager, we fork and invoke the generators in a sanitized mount namespace. If
4003 * we are the user manager, let's just execute the generators directly. We might not have the
4004 * necessary privileges, and the system manager has already mounted /tmp/ and everything else for us.
4005 */
4006 if (MANAGER_IS_USER(m)) {
4007 r = manager_execute_generators(m, paths, /* remount_ro= */ false);
4008 goto finish;
4009 }
4010
4011 /* On some systems /tmp/ doesn't exist, and on some other systems we cannot create it at all. Avoid
4012 * trying to mount a private tmpfs on it as there's no one size fits all. */
4013 if (is_dir("/tmp", /* follow= */ false) > 0)
4014 flags |= FORK_PRIVATE_TMP;
4015
4016 r = safe_fork("(sd-gens)", flags, NULL);
4017 if (r == 0) {
4018 r = manager_execute_generators(m, paths, /* remount_ro= */ true);
4019 _exit(r >= 0 ? EXIT_SUCCESS : EXIT_FAILURE);
4020 }
4021 if (r < 0) {
4022 if (!ERRNO_IS_PRIVILEGE(r)) {
4023 log_error_errno(r, "Failed to fork off sandboxing environment for executing generators: %m");
4024 goto finish;
4025 }
4026
4027 /* Failed to fork with new mount namespace? Maybe, running in a container environment with
4028 * seccomp or without capability. */
4029 log_debug_errno(r,
4030 "Failed to fork off sandboxing environment for executing generators. "
4031 "Falling back to execute generators without sandboxing: %m");
4032 r = manager_execute_generators(m, paths, /* remount_ro= */ false);
4033 }
4034
4035 finish:
4036 lookup_paths_trim_generator(&m->lookup_paths);
4037 return r;
4038 }
4039
4040 int manager_transient_environment_add(Manager *m, char **plus) {
4041 char **a;
4042
4043 assert(m);
4044
4045 if (strv_isempty(plus))
4046 return 0;
4047
4048 a = strv_env_merge(m->transient_environment, plus);
4049 if (!a)
4050 return log_oom();
4051
4052 sanitize_environment(a);
4053
4054 return strv_free_and_replace(m->transient_environment, a);
4055 }
4056
4057 int manager_client_environment_modify(
4058 Manager *m,
4059 char **minus,
4060 char **plus) {
4061
4062 char **a = NULL, **b = NULL, **l;
4063
4064 assert(m);
4065
4066 if (strv_isempty(minus) && strv_isempty(plus))
4067 return 0;
4068
4069 l = m->client_environment;
4070
4071 if (!strv_isempty(minus)) {
4072 a = strv_env_delete(l, 1, minus);
4073 if (!a)
4074 return -ENOMEM;
4075
4076 l = a;
4077 }
4078
4079 if (!strv_isempty(plus)) {
4080 b = strv_env_merge(l, plus);
4081 if (!b) {
4082 strv_free(a);
4083 return -ENOMEM;
4084 }
4085
4086 l = b;
4087 }
4088
4089 if (m->client_environment != l)
4090 strv_free(m->client_environment);
4091
4092 if (a != l)
4093 strv_free(a);
4094 if (b != l)
4095 strv_free(b);
4096
4097 m->client_environment = sanitize_environment(l);
4098 return 0;
4099 }
4100
4101 int manager_get_effective_environment(Manager *m, char ***ret) {
4102 char **l;
4103
4104 assert(m);
4105 assert(ret);
4106
4107 l = strv_env_merge(m->transient_environment, m->client_environment);
4108 if (!l)
4109 return -ENOMEM;
4110
4111 *ret = l;
4112 return 0;
4113 }
4114
4115 int manager_set_default_smack_process_label(Manager *m, const char *label) {
4116 assert(m);
4117
4118 #ifdef SMACK_DEFAULT_PROCESS_LABEL
4119 if (!label)
4120 return free_and_strdup(&m->default_smack_process_label, SMACK_DEFAULT_PROCESS_LABEL);
4121 #endif
4122 if (streq_ptr(label, "/"))
4123 return free_and_strdup(&m->default_smack_process_label, NULL);
4124
4125 return free_and_strdup(&m->default_smack_process_label, label);
4126 }
4127
4128 int manager_set_default_rlimits(Manager *m, struct rlimit **default_rlimit) {
4129 assert(m);
4130
4131 for (unsigned i = 0; i < _RLIMIT_MAX; i++) {
4132 m->rlimit[i] = mfree(m->rlimit[i]);
4133
4134 if (!default_rlimit[i])
4135 continue;
4136
4137 m->rlimit[i] = newdup(struct rlimit, default_rlimit[i], 1);
4138 if (!m->rlimit[i])
4139 return log_oom();
4140 }
4141
4142 return 0;
4143 }
4144
4145 void manager_recheck_dbus(Manager *m) {
4146 assert(m);
4147
4148 /* Connects to the bus if the dbus service and socket are running. If we are running in user mode
4149 * this is all it does. In system mode we'll also connect to the system bus (which will most likely
4150 * just reuse the connection of the API bus). That's because the system bus after all runs as service
4151 * of the system instance, while in the user instance we can assume it's already there. */
4152
4153 if (MANAGER_IS_RELOADING(m))
4154 return; /* don't check while we are reloading… */
4155
4156 if (manager_dbus_is_running(m, false)) {
4157 (void) bus_init_api(m);
4158
4159 if (MANAGER_IS_SYSTEM(m))
4160 (void) bus_init_system(m);
4161 } else {
4162 (void) bus_done_api(m);
4163
4164 if (MANAGER_IS_SYSTEM(m))
4165 (void) bus_done_system(m);
4166 }
4167 }
4168
4169 static bool manager_journal_is_running(Manager *m) {
4170 Unit *u;
4171
4172 assert(m);
4173
4174 if (MANAGER_IS_TEST_RUN(m))
4175 return false;
4176
4177 /* If we are the user manager we can safely assume that the journal is up */
4178 if (!MANAGER_IS_SYSTEM(m))
4179 return true;
4180
4181 /* Check that the socket is not only up, but in RUNNING state */
4182 u = manager_get_unit(m, SPECIAL_JOURNALD_SOCKET);
4183 if (!u)
4184 return false;
4185 if (SOCKET(u)->state != SOCKET_RUNNING)
4186 return false;
4187
4188 /* Similar, check if the daemon itself is fully up, too */
4189 u = manager_get_unit(m, SPECIAL_JOURNALD_SERVICE);
4190 if (!u)
4191 return false;
4192 if (!IN_SET(SERVICE(u)->state, SERVICE_RELOAD, SERVICE_RUNNING))
4193 return false;
4194
4195 return true;
4196 }
4197
4198 void disable_printk_ratelimit(void) {
4199 /* Disable kernel's printk ratelimit.
4200 *
4201 * Logging to /dev/kmsg is most useful during early boot and shutdown, where normal logging
4202 * mechanisms are not available. The semantics of this sysctl are such that any kernel command-line
4203 * setting takes precedence. */
4204 int r;
4205
4206 r = sysctl_write("kernel/printk_devkmsg", "on");
4207 if (r < 0)
4208 log_debug_errno(r, "Failed to set sysctl kernel.printk_devkmsg=on: %m");
4209 }
4210
4211 void manager_recheck_journal(Manager *m) {
4212
4213 assert(m);
4214
4215 /* Don't bother with this unless we are in the special situation of being PID 1 */
4216 if (getpid_cached() != 1)
4217 return;
4218
4219 /* Don't check this while we are reloading, things might still change */
4220 if (MANAGER_IS_RELOADING(m))
4221 return;
4222
4223 /* The journal is fully and entirely up? If so, let's permit logging to it, if that's configured. If
4224 * the journal is down, don't ever log to it, otherwise we might end up deadlocking ourselves as we
4225 * might trigger an activation ourselves we can't fulfill. */
4226 log_set_prohibit_ipc(!manager_journal_is_running(m));
4227 log_open();
4228 }
4229
4230 static ShowStatus manager_get_show_status(Manager *m) {
4231 assert(m);
4232
4233 if (MANAGER_IS_USER(m))
4234 return _SHOW_STATUS_INVALID;
4235
4236 if (m->show_status_overridden != _SHOW_STATUS_INVALID)
4237 return m->show_status_overridden;
4238
4239 return m->show_status;
4240 }
4241
4242 bool manager_get_show_status_on(Manager *m) {
4243 assert(m);
4244
4245 return show_status_on(manager_get_show_status(m));
4246 }
4247
4248 static void set_show_status_marker(bool b) {
4249 if (b)
4250 (void) touch("/run/systemd/show-status");
4251 else
4252 (void) unlink("/run/systemd/show-status");
4253 }
4254
4255 void manager_set_show_status(Manager *m, ShowStatus mode, const char *reason) {
4256 assert(m);
4257 assert(reason);
4258 assert(mode >= 0 && mode < _SHOW_STATUS_MAX);
4259
4260 if (MANAGER_IS_USER(m))
4261 return;
4262
4263 if (mode == m->show_status)
4264 return;
4265
4266 if (m->show_status_overridden == _SHOW_STATUS_INVALID) {
4267 bool enabled;
4268
4269 enabled = show_status_on(mode);
4270 log_debug("%s (%s) showing of status (%s).",
4271 enabled ? "Enabling" : "Disabling",
4272 strna(show_status_to_string(mode)),
4273 reason);
4274
4275 set_show_status_marker(enabled);
4276 }
4277
4278 m->show_status = mode;
4279 }
4280
4281 void manager_override_show_status(Manager *m, ShowStatus mode, const char *reason) {
4282 assert(m);
4283 assert(mode < _SHOW_STATUS_MAX);
4284
4285 if (MANAGER_IS_USER(m))
4286 return;
4287
4288 if (mode == m->show_status_overridden)
4289 return;
4290
4291 m->show_status_overridden = mode;
4292
4293 if (mode == _SHOW_STATUS_INVALID)
4294 mode = m->show_status;
4295
4296 log_debug("%s (%s) showing of status (%s).",
4297 m->show_status_overridden != _SHOW_STATUS_INVALID ? "Overriding" : "Restoring",
4298 strna(show_status_to_string(mode)),
4299 reason);
4300
4301 set_show_status_marker(show_status_on(mode));
4302 }
4303
4304 const char *manager_get_confirm_spawn(Manager *m) {
4305 static int last_errno = 0;
4306 struct stat st;
4307 int r;
4308
4309 assert(m);
4310
4311 /* Here's the deal: we want to test the validity of the console but don't want
4312 * PID1 to go through the whole console process which might block. But we also
4313 * want to warn the user only once if something is wrong with the console so we
4314 * cannot do the sanity checks after spawning our children. So here we simply do
4315 * really basic tests to hopefully trap common errors.
4316 *
4317 * If the console suddenly disappear at the time our children will really it
4318 * then they will simply fail to acquire it and a positive answer will be
4319 * assumed. New children will fall back to /dev/console though.
4320 *
4321 * Note: TTYs are devices that can come and go any time, and frequently aren't
4322 * available yet during early boot (consider a USB rs232 dongle...). If for any
4323 * reason the configured console is not ready, we fall back to the default
4324 * console. */
4325
4326 if (!m->confirm_spawn || path_equal(m->confirm_spawn, "/dev/console"))
4327 return m->confirm_spawn;
4328
4329 if (stat(m->confirm_spawn, &st) < 0) {
4330 r = -errno;
4331 goto fail;
4332 }
4333
4334 if (!S_ISCHR(st.st_mode)) {
4335 r = -ENOTTY;
4336 goto fail;
4337 }
4338
4339 last_errno = 0;
4340 return m->confirm_spawn;
4341
4342 fail:
4343 if (last_errno != r)
4344 last_errno = log_warning_errno(r, "Failed to open %s, using default console: %m", m->confirm_spawn);
4345
4346 return "/dev/console";
4347 }
4348
4349 void manager_set_first_boot(Manager *m, bool b) {
4350 assert(m);
4351
4352 if (!MANAGER_IS_SYSTEM(m))
4353 return;
4354
4355 if (m->first_boot != (int) b) {
4356 if (b)
4357 (void) touch("/run/systemd/first-boot");
4358 else
4359 (void) unlink("/run/systemd/first-boot");
4360 }
4361
4362 m->first_boot = b;
4363 }
4364
4365 void manager_disable_confirm_spawn(void) {
4366 (void) touch("/run/systemd/confirm_spawn_disabled");
4367 }
4368
4369 bool manager_is_confirm_spawn_disabled(Manager *m) {
4370 if (!m->confirm_spawn)
4371 return true;
4372
4373 return access("/run/systemd/confirm_spawn_disabled", F_OK) >= 0;
4374 }
4375
4376 static bool manager_should_show_status(Manager *m, StatusType type) {
4377 assert(m);
4378
4379 if (!MANAGER_IS_SYSTEM(m))
4380 return false;
4381
4382 if (m->no_console_output)
4383 return false;
4384
4385 if (!IN_SET(manager_state(m), MANAGER_INITIALIZING, MANAGER_STARTING, MANAGER_STOPPING))
4386 return false;
4387
4388 /* If we cannot find out the status properly, just proceed. */
4389 if (type != STATUS_TYPE_EMERGENCY && manager_check_ask_password(m) > 0)
4390 return false;
4391
4392 if (type == STATUS_TYPE_NOTICE && m->show_status != SHOW_STATUS_NO)
4393 return true;
4394
4395 return manager_get_show_status_on(m);
4396 }
4397
4398 void manager_status_printf(Manager *m, StatusType type, const char *status, const char *format, ...) {
4399 va_list ap;
4400
4401 /* If m is NULL, assume we're after shutdown and let the messages through. */
4402
4403 if (m && !manager_should_show_status(m, type))
4404 return;
4405
4406 /* XXX We should totally drop the check for ephemeral here
4407 * and thus effectively make 'Type=idle' pointless. */
4408 if (type == STATUS_TYPE_EPHEMERAL && m && m->n_on_console > 0)
4409 return;
4410
4411 va_start(ap, format);
4412 status_vprintf(status, SHOW_STATUS_ELLIPSIZE|(type == STATUS_TYPE_EPHEMERAL ? SHOW_STATUS_EPHEMERAL : 0), format, ap);
4413 va_end(ap);
4414 }
4415
4416 Set* manager_get_units_requiring_mounts_for(Manager *m, const char *path) {
4417 assert(m);
4418 assert(path);
4419
4420 if (path_equal(path, "/"))
4421 path = "";
4422
4423 return hashmap_get(m->units_requiring_mounts_for, path);
4424 }
4425
4426 int manager_update_failed_units(Manager *m, Unit *u, bool failed) {
4427 unsigned size;
4428 int r;
4429
4430 assert(m);
4431 assert(u->manager == m);
4432
4433 size = set_size(m->failed_units);
4434
4435 if (failed) {
4436 r = set_ensure_put(&m->failed_units, NULL, u);
4437 if (r < 0)
4438 return log_oom();
4439 } else
4440 (void) set_remove(m->failed_units, u);
4441
4442 if (set_size(m->failed_units) != size)
4443 bus_manager_send_change_signal(m);
4444
4445 return 0;
4446 }
4447
4448 ManagerState manager_state(Manager *m) {
4449 Unit *u;
4450
4451 assert(m);
4452
4453 /* Is the special shutdown target active or queued? If so, we are in shutdown state */
4454 u = manager_get_unit(m, SPECIAL_SHUTDOWN_TARGET);
4455 if (u && unit_active_or_pending(u))
4456 return MANAGER_STOPPING;
4457
4458 /* Did we ever finish booting? If not then we are still starting up */
4459 if (!MANAGER_IS_FINISHED(m)) {
4460
4461 u = manager_get_unit(m, SPECIAL_BASIC_TARGET);
4462 if (!u || !UNIT_IS_ACTIVE_OR_RELOADING(unit_active_state(u)))
4463 return MANAGER_INITIALIZING;
4464
4465 return MANAGER_STARTING;
4466 }
4467
4468 if (MANAGER_IS_SYSTEM(m)) {
4469 /* Are the rescue or emergency targets active or queued? If so we are in maintenance state */
4470 u = manager_get_unit(m, SPECIAL_RESCUE_TARGET);
4471 if (u && unit_active_or_pending(u))
4472 return MANAGER_MAINTENANCE;
4473
4474 u = manager_get_unit(m, SPECIAL_EMERGENCY_TARGET);
4475 if (u && unit_active_or_pending(u))
4476 return MANAGER_MAINTENANCE;
4477 }
4478
4479 /* Are there any failed units? If so, we are in degraded mode */
4480 if (set_size(m->failed_units) > 0)
4481 return MANAGER_DEGRADED;
4482
4483 return MANAGER_RUNNING;
4484 }
4485
4486 static void manager_unref_uid_internal(
4487 Hashmap *uid_refs,
4488 uid_t uid,
4489 bool destroy_now,
4490 int (*_clean_ipc)(uid_t uid)) {
4491
4492 uint32_t c, n;
4493
4494 assert(uid_is_valid(uid));
4495 assert(_clean_ipc);
4496
4497 /* A generic implementation, covering both manager_unref_uid() and manager_unref_gid(), under the
4498 * assumption that uid_t and gid_t are actually defined the same way, with the same validity rules.
4499 *
4500 * We store a hashmap where the key is the UID/GID and the value is a 32-bit reference counter, whose
4501 * highest bit is used as flag for marking UIDs/GIDs whose IPC objects to remove when the last
4502 * reference to the UID/GID is dropped. The flag is set to on, once at least one reference from a
4503 * unit where RemoveIPC= is set is added on a UID/GID. It is reset when the UID's/GID's reference
4504 * counter drops to 0 again. */
4505
4506 assert_cc(sizeof(uid_t) == sizeof(gid_t));
4507 assert_cc(UID_INVALID == (uid_t) GID_INVALID);
4508
4509 if (uid == 0) /* We don't keep track of root, and will never destroy it */
4510 return;
4511
4512 c = PTR_TO_UINT32(hashmap_get(uid_refs, UID_TO_PTR(uid)));
4513
4514 n = c & ~DESTROY_IPC_FLAG;
4515 assert(n > 0);
4516 n--;
4517
4518 if (destroy_now && n == 0) {
4519 hashmap_remove(uid_refs, UID_TO_PTR(uid));
4520
4521 if (c & DESTROY_IPC_FLAG) {
4522 log_debug("%s " UID_FMT " is no longer referenced, cleaning up its IPC.",
4523 _clean_ipc == clean_ipc_by_uid ? "UID" : "GID",
4524 uid);
4525 (void) _clean_ipc(uid);
4526 }
4527 } else {
4528 c = n | (c & DESTROY_IPC_FLAG);
4529 assert_se(hashmap_update(uid_refs, UID_TO_PTR(uid), UINT32_TO_PTR(c)) >= 0);
4530 }
4531 }
4532
4533 void manager_unref_uid(Manager *m, uid_t uid, bool destroy_now) {
4534 manager_unref_uid_internal(m->uid_refs, uid, destroy_now, clean_ipc_by_uid);
4535 }
4536
4537 void manager_unref_gid(Manager *m, gid_t gid, bool destroy_now) {
4538 manager_unref_uid_internal(m->gid_refs, (uid_t) gid, destroy_now, clean_ipc_by_gid);
4539 }
4540
4541 static int manager_ref_uid_internal(
4542 Hashmap **uid_refs,
4543 uid_t uid,
4544 bool clean_ipc) {
4545
4546 uint32_t c, n;
4547 int r;
4548
4549 assert(uid_refs);
4550 assert(uid_is_valid(uid));
4551
4552 /* A generic implementation, covering both manager_ref_uid() and manager_ref_gid(), under the
4553 * assumption that uid_t and gid_t are actually defined the same way, with the same validity
4554 * rules. */
4555
4556 assert_cc(sizeof(uid_t) == sizeof(gid_t));
4557 assert_cc(UID_INVALID == (uid_t) GID_INVALID);
4558
4559 if (uid == 0) /* We don't keep track of root, and will never destroy it */
4560 return 0;
4561
4562 r = hashmap_ensure_allocated(uid_refs, &trivial_hash_ops);
4563 if (r < 0)
4564 return r;
4565
4566 c = PTR_TO_UINT32(hashmap_get(*uid_refs, UID_TO_PTR(uid)));
4567
4568 n = c & ~DESTROY_IPC_FLAG;
4569 n++;
4570
4571 if (n & DESTROY_IPC_FLAG) /* check for overflow */
4572 return -EOVERFLOW;
4573
4574 c = n | (c & DESTROY_IPC_FLAG) | (clean_ipc ? DESTROY_IPC_FLAG : 0);
4575
4576 return hashmap_replace(*uid_refs, UID_TO_PTR(uid), UINT32_TO_PTR(c));
4577 }
4578
4579 int manager_ref_uid(Manager *m, uid_t uid, bool clean_ipc) {
4580 return manager_ref_uid_internal(&m->uid_refs, uid, clean_ipc);
4581 }
4582
4583 int manager_ref_gid(Manager *m, gid_t gid, bool clean_ipc) {
4584 return manager_ref_uid_internal(&m->gid_refs, (uid_t) gid, clean_ipc);
4585 }
4586
4587 static void manager_vacuum_uid_refs_internal(
4588 Hashmap *uid_refs,
4589 int (*_clean_ipc)(uid_t uid)) {
4590
4591 void *p, *k;
4592
4593 assert(_clean_ipc);
4594
4595 HASHMAP_FOREACH_KEY(p, k, uid_refs) {
4596 uint32_t c, n;
4597 uid_t uid;
4598
4599 uid = PTR_TO_UID(k);
4600 c = PTR_TO_UINT32(p);
4601
4602 n = c & ~DESTROY_IPC_FLAG;
4603 if (n > 0)
4604 continue;
4605
4606 if (c & DESTROY_IPC_FLAG) {
4607 log_debug("Found unreferenced %s " UID_FMT " after reload/reexec. Cleaning up.",
4608 _clean_ipc == clean_ipc_by_uid ? "UID" : "GID",
4609 uid);
4610 (void) _clean_ipc(uid);
4611 }
4612
4613 assert_se(hashmap_remove(uid_refs, k) == p);
4614 }
4615 }
4616
4617 static void manager_vacuum_uid_refs(Manager *m) {
4618 manager_vacuum_uid_refs_internal(m->uid_refs, clean_ipc_by_uid);
4619 }
4620
4621 static void manager_vacuum_gid_refs(Manager *m) {
4622 manager_vacuum_uid_refs_internal(m->gid_refs, clean_ipc_by_gid);
4623 }
4624
4625 static void manager_vacuum(Manager *m) {
4626 assert(m);
4627
4628 /* Release any dynamic users no longer referenced */
4629 dynamic_user_vacuum(m, true);
4630
4631 /* Release any references to UIDs/GIDs no longer referenced, and destroy any IPC owned by them */
4632 manager_vacuum_uid_refs(m);
4633 manager_vacuum_gid_refs(m);
4634
4635 /* Release any runtimes no longer referenced */
4636 exec_shared_runtime_vacuum(m);
4637 }
4638
4639 int manager_dispatch_user_lookup_fd(sd_event_source *source, int fd, uint32_t revents, void *userdata) {
4640 struct buffer {
4641 uid_t uid;
4642 gid_t gid;
4643 char unit_name[UNIT_NAME_MAX+1];
4644 } _packed_ buffer;
4645
4646 Manager *m = userdata;
4647 ssize_t l;
4648 size_t n;
4649 Unit *u;
4650
4651 assert_se(source);
4652 assert_se(m);
4653
4654 /* Invoked whenever a child process succeeded resolving its user/group to use and sent us the
4655 * resulting UID/GID in a datagram. We parse the datagram here and pass it off to the unit, so that
4656 * it can add a reference to the UID/GID so that it can destroy the UID/GID's IPC objects when the
4657 * reference counter drops to 0. */
4658
4659 l = recv(fd, &buffer, sizeof(buffer), MSG_DONTWAIT);
4660 if (l < 0) {
4661 if (ERRNO_IS_TRANSIENT(errno))
4662 return 0;
4663
4664 return log_error_errno(errno, "Failed to read from user lookup fd: %m");
4665 }
4666
4667 if ((size_t) l <= offsetof(struct buffer, unit_name)) {
4668 log_warning("Received too short user lookup message, ignoring.");
4669 return 0;
4670 }
4671
4672 if ((size_t) l > offsetof(struct buffer, unit_name) + UNIT_NAME_MAX) {
4673 log_warning("Received too long user lookup message, ignoring.");
4674 return 0;
4675 }
4676
4677 if (!uid_is_valid(buffer.uid) && !gid_is_valid(buffer.gid)) {
4678 log_warning("Got user lookup message with invalid UID/GID pair, ignoring.");
4679 return 0;
4680 }
4681
4682 n = (size_t) l - offsetof(struct buffer, unit_name);
4683 if (memchr(buffer.unit_name, 0, n)) {
4684 log_warning("Received lookup message with embedded NUL character, ignoring.");
4685 return 0;
4686 }
4687
4688 buffer.unit_name[n] = 0;
4689 u = manager_get_unit(m, buffer.unit_name);
4690 if (!u) {
4691 log_debug("Got user lookup message but unit doesn't exist, ignoring.");
4692 return 0;
4693 }
4694
4695 log_unit_debug(u, "User lookup succeeded: uid=" UID_FMT " gid=" GID_FMT, buffer.uid, buffer.gid);
4696
4697 unit_notify_user_lookup(u, buffer.uid, buffer.gid);
4698 return 0;
4699 }
4700
4701 static int short_uid_range(const char *path) {
4702 _cleanup_(uid_range_freep) UidRange *p = NULL;
4703 int r;
4704
4705 assert(path);
4706
4707 /* Taint systemd if we the UID range assigned to this environment doesn't at least cover 0…65534,
4708 * i.e. from root to nobody. */
4709
4710 r = uid_range_load_userns(&p, path);
4711 if (r < 0) {
4712 if (ERRNO_IS_NOT_SUPPORTED(r))
4713 return false;
4714 return log_debug_errno(r, "Failed to load %s: %m", path);
4715 }
4716
4717 return !uid_range_covers(p, 0, 65535);
4718 }
4719
4720 char* manager_taint_string(const Manager *m) {
4721 /* Returns a "taint string", e.g. "local-hwclock:var-run-bad". Only things that are detected at
4722 * runtime should be tagged here. For stuff that is known during compilation, emit a warning in the
4723 * configuration phase. */
4724
4725 assert(m);
4726
4727 const char* stage[13] = {};
4728 size_t n = 0;
4729
4730 if (m->taint_usr)
4731 stage[n++] = "split-usr";
4732
4733 _cleanup_free_ char *usrbin = NULL;
4734 if (readlink_malloc("/bin", &usrbin) < 0 || !PATH_IN_SET(usrbin, "usr/bin", "/usr/bin"))
4735 stage[n++] = "unmerged-usr";
4736
4737 if (access("/proc/cgroups", F_OK) < 0)
4738 stage[n++] = "cgroups-missing";
4739
4740 if (cg_all_unified() == 0)
4741 stage[n++] = "cgroupsv1";
4742
4743 if (clock_is_localtime(NULL) > 0)
4744 stage[n++] = "local-hwclock";
4745
4746 if (os_release_support_ended(NULL, /* quiet= */ true, NULL) > 0)
4747 stage[n++] = "support-ended";
4748
4749 _cleanup_free_ char *destination = NULL;
4750 if (readlink_malloc("/var/run", &destination) < 0 ||
4751 !PATH_IN_SET(destination, "../run", "/run"))
4752 stage[n++] = "var-run-bad";
4753
4754 _cleanup_free_ char *overflowuid = NULL, *overflowgid = NULL;
4755 if (read_one_line_file("/proc/sys/kernel/overflowuid", &overflowuid) >= 0 &&
4756 !streq(overflowuid, "65534"))
4757 stage[n++] = "overflowuid-not-65534";
4758 if (read_one_line_file("/proc/sys/kernel/overflowgid", &overflowgid) >= 0 &&
4759 !streq(overflowgid, "65534"))
4760 stage[n++] = "overflowgid-not-65534";
4761
4762 struct utsname uts;
4763 assert_se(uname(&uts) >= 0);
4764 if (strverscmp_improved(uts.release, KERNEL_BASELINE_VERSION) < 0)
4765 stage[n++] = "old-kernel";
4766
4767 if (short_uid_range("/proc/self/uid_map") > 0)
4768 stage[n++] = "short-uid-range";
4769 if (short_uid_range("/proc/self/gid_map") > 0)
4770 stage[n++] = "short-gid-range";
4771
4772 assert(n < ELEMENTSOF(stage) - 1); /* One extra for NULL terminator */
4773
4774 return strv_join((char**) stage, ":");
4775 }
4776
4777 void manager_ref_console(Manager *m) {
4778 assert(m);
4779
4780 m->n_on_console++;
4781 }
4782
4783 void manager_unref_console(Manager *m) {
4784
4785 assert(m->n_on_console > 0);
4786 m->n_on_console--;
4787
4788 if (m->n_on_console == 0)
4789 m->no_console_output = false; /* unset no_console_output flag, since the console is definitely free now */
4790 }
4791
4792 void manager_override_log_level(Manager *m, int level) {
4793 _cleanup_free_ char *s = NULL;
4794 assert(m);
4795
4796 if (!m->log_level_overridden) {
4797 m->original_log_level = log_get_max_level();
4798 m->log_level_overridden = true;
4799 }
4800
4801 (void) log_level_to_string_alloc(level, &s);
4802 log_info("Setting log level to %s.", strna(s));
4803
4804 log_set_max_level(level);
4805 }
4806
4807 void manager_restore_original_log_level(Manager *m) {
4808 _cleanup_free_ char *s = NULL;
4809 assert(m);
4810
4811 if (!m->log_level_overridden)
4812 return;
4813
4814 (void) log_level_to_string_alloc(m->original_log_level, &s);
4815 log_info("Restoring log level to original (%s).", strna(s));
4816
4817 log_set_max_level(m->original_log_level);
4818 m->log_level_overridden = false;
4819 }
4820
4821 void manager_override_log_target(Manager *m, LogTarget target) {
4822 assert(m);
4823
4824 if (!m->log_target_overridden) {
4825 m->original_log_target = log_get_target();
4826 m->log_target_overridden = true;
4827 }
4828
4829 log_info("Setting log target to %s.", log_target_to_string(target));
4830 log_set_target(target);
4831 }
4832
4833 void manager_restore_original_log_target(Manager *m) {
4834 assert(m);
4835
4836 if (!m->log_target_overridden)
4837 return;
4838
4839 log_info("Restoring log target to original %s.", log_target_to_string(m->original_log_target));
4840
4841 log_set_target(m->original_log_target);
4842 m->log_target_overridden = false;
4843 }
4844
4845 ManagerTimestamp manager_timestamp_initrd_mangle(ManagerTimestamp s) {
4846 if (in_initrd() &&
4847 s >= MANAGER_TIMESTAMP_SECURITY_START &&
4848 s <= MANAGER_TIMESTAMP_UNITS_LOAD_FINISH)
4849 return s - MANAGER_TIMESTAMP_SECURITY_START + MANAGER_TIMESTAMP_INITRD_SECURITY_START;
4850 return s;
4851 }
4852
4853 static const char *const manager_state_table[_MANAGER_STATE_MAX] = {
4854 [MANAGER_INITIALIZING] = "initializing",
4855 [MANAGER_STARTING] = "starting",
4856 [MANAGER_RUNNING] = "running",
4857 [MANAGER_DEGRADED] = "degraded",
4858 [MANAGER_MAINTENANCE] = "maintenance",
4859 [MANAGER_STOPPING] = "stopping",
4860 };
4861
4862 DEFINE_STRING_TABLE_LOOKUP(manager_state, ManagerState);
4863
4864 static const char *const manager_timestamp_table[_MANAGER_TIMESTAMP_MAX] = {
4865 [MANAGER_TIMESTAMP_FIRMWARE] = "firmware",
4866 [MANAGER_TIMESTAMP_LOADER] = "loader",
4867 [MANAGER_TIMESTAMP_KERNEL] = "kernel",
4868 [MANAGER_TIMESTAMP_INITRD] = "initrd",
4869 [MANAGER_TIMESTAMP_USERSPACE] = "userspace",
4870 [MANAGER_TIMESTAMP_FINISH] = "finish",
4871 [MANAGER_TIMESTAMP_SECURITY_START] = "security-start",
4872 [MANAGER_TIMESTAMP_SECURITY_FINISH] = "security-finish",
4873 [MANAGER_TIMESTAMP_GENERATORS_START] = "generators-start",
4874 [MANAGER_TIMESTAMP_GENERATORS_FINISH] = "generators-finish",
4875 [MANAGER_TIMESTAMP_UNITS_LOAD_START] = "units-load-start",
4876 [MANAGER_TIMESTAMP_UNITS_LOAD_FINISH] = "units-load-finish",
4877 [MANAGER_TIMESTAMP_UNITS_LOAD] = "units-load",
4878 [MANAGER_TIMESTAMP_INITRD_SECURITY_START] = "initrd-security-start",
4879 [MANAGER_TIMESTAMP_INITRD_SECURITY_FINISH] = "initrd-security-finish",
4880 [MANAGER_TIMESTAMP_INITRD_GENERATORS_START] = "initrd-generators-start",
4881 [MANAGER_TIMESTAMP_INITRD_GENERATORS_FINISH] = "initrd-generators-finish",
4882 [MANAGER_TIMESTAMP_INITRD_UNITS_LOAD_START] = "initrd-units-load-start",
4883 [MANAGER_TIMESTAMP_INITRD_UNITS_LOAD_FINISH] = "initrd-units-load-finish",
4884 };
4885
4886 DEFINE_STRING_TABLE_LOOKUP(manager_timestamp, ManagerTimestamp);
4887
4888 static const char* const oom_policy_table[_OOM_POLICY_MAX] = {
4889 [OOM_CONTINUE] = "continue",
4890 [OOM_STOP] = "stop",
4891 [OOM_KILL] = "kill",
4892 };
4893
4894 DEFINE_STRING_TABLE_LOOKUP(oom_policy, OOMPolicy);
Unnamed repository; edit this file 'description' to name the repository.