src/journal/journald.c

   1 /*-*- Mode: C; c-basic-offset: 8; indent-tabs-mode: nil -*-*/
   2
   3 /***
   4   This file is part of systemd.
   5
   6   Copyright 2011 Lennart Poettering
   7
   8   systemd is free software; you can redistribute it and/or modify it
   9   under the terms of the GNU General Public License as published by
  10   the Free Software Foundation; either version 2 of the License, or
  11   (at your option) any later version.
  12
  13   systemd is distributed in the hope that it will be useful, but
  14   WITHOUT ANY WARRANTY; without even the implied warranty of
  15   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  16   General Public License for more details.
  17
  18   You should have received a copy of the GNU General Public License
  19   along with systemd; If not, see <http://www.gnu.org/licenses/>.
  20 ***/
  21
  22 #include <sys/epoll.h>
  23 #include <sys/socket.h>
  24 #include <errno.h>
  25 #include <sys/signalfd.h>
  26 #include <unistd.h>
  27 #include <fcntl.h>
  28 #include <sys/acl.h>
  29 #include <acl/libacl.h>
  30 #include <stddef.h>
  31 #include <sys/ioctl.h>
  32 #include <linux/sockios.h>
  33 #include <sys/statvfs.h>
  34
  35 #include <systemd/sd-journal.h>
  36 #include <systemd/sd-login.h>
  37 #include <systemd/sd-messages.h>
  38 #include <systemd/sd-daemon.h>
  39
  40 #include "hashmap.h"
  41 #include "journal-file.h"
  42 #include "socket-util.h"
  43 #include "acl-util.h"
  44 #include "cgroup-util.h"
  45 #include "list.h"
  46 #include "journal-rate-limit.h"
  47 #include "journal-internal.h"
  48 #include "conf-parser.h"
  49 #include "journald.h"
  50
  51 #define USER_JOURNALS_MAX 1024
  52 #define STDOUT_STREAMS_MAX 4096
  53
  54 #define DEFAULT_RATE_LIMIT_INTERVAL (10*USEC_PER_SEC)
  55 #define DEFAULT_RATE_LIMIT_BURST 200
  56
  57 #define RECHECK_AVAILABLE_SPACE_USEC (30*USEC_PER_SEC)
  58
  59 #define RECHECK_VAR_AVAILABLE_USEC (30*USEC_PER_SEC)
  60
  61 #define SYSLOG_TIMEOUT_USEC (5*USEC_PER_SEC)
  62
  63 #define N_IOVEC_META_FIELDS 16
  64
  65 typedef enum StdoutStreamState {
  66         STDOUT_STREAM_TAG,
  67         STDOUT_STREAM_PRIORITY,
  68         STDOUT_STREAM_PRIORITY_PREFIX,
  69         STDOUT_STREAM_FORWARD_TO_SYSLOG,
  70         STDOUT_STREAM_FORWARD_TO_KMSG,
  71         STDOUT_STREAM_FORWARD_TO_CONSOLE,
  72         STDOUT_STREAM_RUNNING
  73 } StdoutStreamState;
  74
  75 struct StdoutStream {
  76         Server *server;
  77         StdoutStreamState state;
  78
  79         int fd;
  80
  81         struct ucred ucred;
  82
  83         char *tag;
  84         int priority;
  85         bool priority_prefix:1;
  86         bool forward_to_syslog:1;
  87         bool forward_to_kmsg:1;
  88         bool forward_to_console:1;
  89
  90         char buffer[LINE_MAX+1];
  91         size_t length;
  92
  93         LIST_FIELDS(StdoutStream, stdout_stream);
  94 };
  95
  96 static int server_flush_to_var(Server *s);
  97
  98 static uint64_t available_space(Server *s) {
  99         char ids[33], *p;
 100         const char *f;
 101         sd_id128_t machine;
 102         struct statvfs ss;
 103         uint64_t sum = 0, avail = 0, ss_avail = 0;
 104         int r;
 105         DIR *d;
 106         usec_t ts;
 107         JournalMetrics *m;
 108
 109         ts = now(CLOCK_MONOTONIC);
 110
 111         if (s->cached_available_space_timestamp + RECHECK_AVAILABLE_SPACE_USEC > ts)
 112                 return s->cached_available_space;
 113
 114         r = sd_id128_get_machine(&machine);
 115         if (r < 0)
 116                 return 0;
 117
 118         if (s->system_journal) {
 119                 f = "/var/log/journal/";
 120                 m = &s->system_metrics;
 121         } else {
 122                 f = "/run/log/journal/";
 123                 m = &s->runtime_metrics;
 124         }
 125
 126         assert(m);
 127
 128         p = strappend(f, sd_id128_to_string(machine, ids));
 129         if (!p)
 130                 return 0;
 131
 132         d = opendir(p);
 133         free(p);
 134
 135         if (!d)
 136                 return 0;
 137
 138         if (fstatvfs(dirfd(d), &ss) < 0)
 139                 goto finish;
 140
 141         for (;;) {
 142                 struct stat st;
 143                 struct dirent buf, *de;
 144                 int k;
 145
 146                 k = readdir_r(d, &buf, &de);
 147                 if (k != 0) {
 148                         r = -k;
 149                         goto finish;
 150                 }
 151
 152                 if (!de)
 153                         break;
 154
 155                 if (!dirent_is_file_with_suffix(de, ".journal"))
 156                         continue;
 157
 158                 if (fstatat(dirfd(d), de->d_name, &st, AT_SYMLINK_NOFOLLOW) < 0)
 159                         continue;
 160
 161                 sum += (uint64_t) st.st_blocks * (uint64_t) st.st_blksize;
 162         }
 163
 164         avail = sum >= m->max_use ? 0 : m->max_use - sum;
 165
 166         ss_avail = ss.f_bsize * ss.f_bavail;
 167
 168         ss_avail = ss_avail < m->keep_free ? 0 : ss_avail - m->keep_free;
 169
 170         if (ss_avail < avail)
 171                 avail = ss_avail;
 172
 173         s->cached_available_space = avail;
 174         s->cached_available_space_timestamp = ts;
 175
 176 finish:
 177         closedir(d);
 178
 179         return avail;
 180 }
 181
 182 static void fix_perms(JournalFile *f, uid_t uid) {
 183         acl_t acl;
 184         acl_entry_t entry;
 185         acl_permset_t permset;
 186         int r;
 187
 188         assert(f);
 189
 190         r = fchmod_and_fchown(f->fd, 0640, 0, 0);
 191         if (r < 0)
 192                 log_warning("Failed to fix access mode/rights on %s, ignoring: %s", f->path, strerror(-r));
 193
 194         if (uid <= 0)
 195                 return;
 196
 197         acl = acl_get_fd(f->fd);
 198         if (!acl) {
 199                 log_warning("Failed to read ACL on %s, ignoring: %m", f->path);
 200                 return;
 201         }
 202
 203         r = acl_find_uid(acl, uid, &entry);
 204         if (r <= 0) {
 205
 206                 if (acl_create_entry(&acl, &entry) < 0 ||
 207                     acl_set_tag_type(entry, ACL_USER) < 0 ||
 208                     acl_set_qualifier(entry, &uid) < 0) {
 209                         log_warning("Failed to patch ACL on %s, ignoring: %m", f->path);
 210                         goto finish;
 211                 }
 212         }
 213
 214         if (acl_get_permset(entry, &permset) < 0 ||
 215             acl_add_perm(permset, ACL_READ) < 0 ||
 216             acl_calc_mask(&acl) < 0) {
 217                 log_warning("Failed to patch ACL on %s, ignoring: %m", f->path);
 218                 goto finish;
 219         }
 220
 221         if (acl_set_fd(f->fd, acl) < 0)
 222                 log_warning("Failed to set ACL on %s, ignoring: %m", f->path);
 223
 224 finish:
 225         acl_free(acl);
 226 }
 227
 228 static JournalFile* find_journal(Server *s, uid_t uid) {
 229         char *p;
 230         int r;
 231         JournalFile *f;
 232         char ids[33];
 233         sd_id128_t machine;
 234
 235         assert(s);
 236
 237         /* We split up user logs only on /var, not on /run. If the
 238          * runtime file is open, we write to it exclusively, in order
 239          * to guarantee proper order as soon as we flush /run to
 240          * /var and close the runtime file. */
 241
 242         if (s->runtime_journal)
 243                 return s->runtime_journal;
 244
 245         if (uid <= 0)
 246                 return s->system_journal;
 247
 248         r = sd_id128_get_machine(&machine);
 249         if (r < 0)
 250                 return s->system_journal;
 251
 252         f = hashmap_get(s->user_journals, UINT32_TO_PTR(uid));
 253         if (f)
 254                 return f;
 255
 256         if (asprintf(&p, "/var/log/journal/%s/user-%lu.journal", sd_id128_to_string(machine, ids), (unsigned long) uid) < 0)
 257                 return s->system_journal;
 258
 259         while (hashmap_size(s->user_journals) >= USER_JOURNALS_MAX) {
 260                 /* Too many open? Then let's close one */
 261                 f = hashmap_steal_first(s->user_journals);
 262                 assert(f);
 263                 journal_file_close(f);
 264         }
 265
 266         r = journal_file_open(p, O_RDWR|O_CREAT, 0640, s->system_journal, &f);
 267         free(p);
 268
 269         if (r < 0)
 270                 return s->system_journal;
 271
 272         fix_perms(f, uid);
 273         f->metrics = s->system_metrics;
 274         f->compress = s->compress;
 275
 276         r = hashmap_put(s->user_journals, UINT32_TO_PTR(uid), f);
 277         if (r < 0) {
 278                 journal_file_close(f);
 279                 return s->system_journal;
 280         }
 281
 282         return f;
 283 }
 284
 285 static void server_rotate(Server *s) {
 286         JournalFile *f;
 287         void *k;
 288         Iterator i;
 289         int r;
 290
 291         log_info("Rotating...");
 292
 293         if (s->runtime_journal) {
 294                 r = journal_file_rotate(&s->runtime_journal);
 295                 if (r < 0)
 296                         log_error("Failed to rotate %s: %s", s->runtime_journal->path, strerror(-r));
 297         }
 298
 299         if (s->system_journal) {
 300                 r = journal_file_rotate(&s->system_journal);
 301                 if (r < 0)
 302                         log_error("Failed to rotate %s: %s", s->system_journal->path, strerror(-r));
 303         }
 304
 305         HASHMAP_FOREACH_KEY(f, k, s->user_journals, i) {
 306                 r = journal_file_rotate(&f);
 307                 if (r < 0)
 308                         log_error("Failed to rotate %s: %s", f->path, strerror(-r));
 309                 else
 310                         hashmap_replace(s->user_journals, k, f);
 311         }
 312 }
 313
 314 static void server_vacuum(Server *s) {
 315         char *p;
 316         char ids[33];
 317         sd_id128_t machine;
 318         int r;
 319
 320         log_info("Vacuuming...");
 321
 322         r = sd_id128_get_machine(&machine);
 323         if (r < 0) {
 324                 log_error("Failed to get machine ID: %s", strerror(-r));
 325                 return;
 326         }
 327
 328         sd_id128_to_string(machine, ids);
 329
 330         if (s->system_journal) {
 331                 if (asprintf(&p, "/var/log/journal/%s", ids) < 0) {
 332                         log_error("Out of memory.");
 333                         return;
 334                 }
 335
 336                 r = journal_directory_vacuum(p, s->system_metrics.max_use, s->system_metrics.keep_free);
 337                 if (r < 0 && r != -ENOENT)
 338                         log_error("Failed to vacuum %s: %s", p, strerror(-r));
 339                 free(p);
 340         }
 341
 342
 343         if (s->runtime_journal) {
 344                 if (asprintf(&p, "/run/log/journal/%s", ids) < 0) {
 345                         log_error("Out of memory.");
 346                         return;
 347                 }
 348
 349                 r = journal_directory_vacuum(p, s->runtime_metrics.max_use, s->runtime_metrics.keep_free);
 350                 if (r < 0 && r != -ENOENT)
 351                         log_error("Failed to vacuum %s: %s", p, strerror(-r));
 352                 free(p);
 353         }
 354
 355         s->cached_available_space_timestamp = 0;
 356 }
 357
 358 static char *shortened_cgroup_path(pid_t pid) {
 359         int r;
 360         char *process_path, *init_path, *path;
 361
 362         assert(pid > 0);
 363
 364         r = cg_get_by_pid(SYSTEMD_CGROUP_CONTROLLER, pid, &process_path);
 365         if (r < 0)
 366                 return NULL;
 367
 368         r = cg_get_by_pid(SYSTEMD_CGROUP_CONTROLLER, 1, &init_path);
 369         if (r < 0) {
 370                 free(process_path);
 371                 return NULL;
 372         }
 373
 374         if (endswith(init_path, "/system"))
 375                 init_path[strlen(init_path) - 7] = 0;
 376         else if (streq(init_path, "/"))
 377                 init_path[0] = 0;
 378
 379         if (startswith(process_path, init_path)) {
 380                 char *p;
 381
 382                 p = strdup(process_path + strlen(init_path));
 383                 if (!p) {
 384                         free(process_path);
 385                         free(init_path);
 386                         return NULL;
 387                 }
 388                 path = p;
 389         } else {
 390                 path = process_path;
 391                 process_path = NULL;
 392         }
 393
 394         free(process_path);
 395         free(init_path);
 396
 397         return path;
 398 }
 399
 400 static void dispatch_message_real(Server *s,
 401                              struct iovec *iovec, unsigned n, unsigned m,
 402                              struct ucred *ucred,
 403                              struct timeval *tv) {
 404
 405         char *pid = NULL, *uid = NULL, *gid = NULL,
 406                 *source_time = NULL, *boot_id = NULL, *machine_id = NULL,
 407                 *comm = NULL, *cmdline = NULL, *hostname = NULL,
 408                 *audit_session = NULL, *audit_loginuid = NULL,
 409                 *exe = NULL, *cgroup = NULL, *session = NULL,
 410                 *owner_uid = NULL, *unit = NULL;
 411
 412         char idbuf[33];
 413         sd_id128_t id;
 414         int r;
 415         char *t;
 416         uid_t loginuid = 0, realuid = 0;
 417         JournalFile *f;
 418         bool vacuumed = false;
 419
 420         assert(s);
 421         assert(iovec);
 422         assert(n > 0);
 423         assert(n + N_IOVEC_META_FIELDS <= m);
 424
 425         if (ucred) {
 426                 uint32_t audit;
 427                 uid_t owner;
 428
 429                 realuid = ucred->uid;
 430
 431                 if (asprintf(&pid, "_PID=%lu", (unsigned long) ucred->pid) >= 0)
 432                         IOVEC_SET_STRING(iovec[n++], pid);
 433
 434                 if (asprintf(&uid, "_UID=%lu", (unsigned long) ucred->uid) >= 0)
 435                         IOVEC_SET_STRING(iovec[n++], uid);
 436
 437                 if (asprintf(&gid, "_GID=%lu", (unsigned long) ucred->gid) >= 0)
 438                         IOVEC_SET_STRING(iovec[n++], gid);
 439
 440                 r = get_process_comm(ucred->pid, &t);
 441                 if (r >= 0) {
 442                         comm = strappend("_COMM=", t);
 443                         free(t);
 444
 445                         if (comm)
 446                                 IOVEC_SET_STRING(iovec[n++], comm);
 447                 }
 448
 449                 r = get_process_exe(ucred->pid, &t);
 450                 if (r >= 0) {
 451                         exe = strappend("_EXE=", t);
 452                         free(t);
 453
 454                         if (comm)
 455                                 IOVEC_SET_STRING(iovec[n++], exe);
 456                 }
 457
 458                 r = get_process_cmdline(ucred->pid, LINE_MAX, false, &t);
 459                 if (r >= 0) {
 460                         cmdline = strappend("_CMDLINE=", t);
 461                         free(t);
 462
 463                         if (cmdline)
 464                                 IOVEC_SET_STRING(iovec[n++], cmdline);
 465                 }
 466
 467                 r = audit_session_from_pid(ucred->pid, &audit);
 468                 if (r >= 0)
 469                         if (asprintf(&audit_session, "_AUDIT_SESSION=%lu", (unsigned long) audit) >= 0)
 470                                 IOVEC_SET_STRING(iovec[n++], audit_session);
 471
 472                 r = audit_loginuid_from_pid(ucred->pid, &loginuid);
 473                 if (r >= 0)
 474                         if (asprintf(&audit_loginuid, "_AUDIT_LOGINUID=%lu", (unsigned long) loginuid) >= 0)
 475                                 IOVEC_SET_STRING(iovec[n++], audit_loginuid);
 476
 477                 t = shortened_cgroup_path(ucred->pid);
 478                 if (t) {
 479                         cgroup = strappend("_SYSTEMD_CGROUP=", t);
 480                         free(t);
 481
 482                         if (cgroup)
 483                                 IOVEC_SET_STRING(iovec[n++], cgroup);
 484                 }
 485
 486                 if (sd_pid_get_session(ucred->pid, &t) >= 0) {
 487                         session = strappend("_SYSTEMD_SESSION=", t);
 488                         free(t);
 489
 490                         if (session)
 491                                 IOVEC_SET_STRING(iovec[n++], session);
 492                 }
 493
 494                 if (sd_pid_get_unit(ucred->pid, &t) >= 0) {
 495                         unit = strappend("_SYSTEMD_UNIT=", t);
 496                         free(t);
 497
 498                         if (unit)
 499                                 IOVEC_SET_STRING(iovec[n++], unit);
 500                 }
 501
 502                 if (sd_pid_get_owner_uid(ucred->uid, &owner) >= 0)
 503                         if (asprintf(&owner_uid, "_SYSTEMD_OWNER_UID=%lu", (unsigned long) owner) >= 0)
 504                                 IOVEC_SET_STRING(iovec[n++], owner_uid);
 505         }
 506
 507         if (tv) {
 508                 if (asprintf(&source_time, "_SOURCE_REALTIME_TIMESTAMP=%llu",
 509                              (unsigned long long) timeval_load(tv)) >= 0)
 510                         IOVEC_SET_STRING(iovec[n++], source_time);
 511         }
 512
 513         /* Note that strictly speaking storing the boot id here is
 514          * redundant since the entry includes this in-line
 515          * anyway. However, we need this indexed, too. */
 516         r = sd_id128_get_boot(&id);
 517         if (r >= 0)
 518                 if (asprintf(&boot_id, "_BOOT_ID=%s", sd_id128_to_string(id, idbuf)) >= 0)
 519                         IOVEC_SET_STRING(iovec[n++], boot_id);
 520
 521         r = sd_id128_get_machine(&id);
 522         if (r >= 0)
 523                 if (asprintf(&machine_id, "_MACHINE_ID=%s", sd_id128_to_string(id, idbuf)) >= 0)
 524                         IOVEC_SET_STRING(iovec[n++], machine_id);
 525
 526         t = gethostname_malloc();
 527         if (t) {
 528                 hostname = strappend("_HOSTNAME=", t);
 529                 free(t);
 530                 if (hostname)
 531                         IOVEC_SET_STRING(iovec[n++], hostname);
 532         }
 533
 534         assert(n <= m);
 535
 536         server_flush_to_var(s);
 537
 538 retry:
 539         f = find_journal(s, realuid == 0 ? 0 : loginuid);
 540         if (!f)
 541                 log_warning("Dropping message, as we can't find a place to store the data.");
 542         else {
 543                 r = journal_file_append_entry(f, NULL, iovec, n, &s->seqnum, NULL, NULL);
 544
 545                 if (r == -E2BIG && !vacuumed) {
 546                         log_info("Allocation limit reached.");
 547
 548                         server_rotate(s);
 549                         server_vacuum(s);
 550                         vacuumed = true;
 551
 552                         log_info("Retrying write.");
 553                         goto retry;
 554                 }
 555
 556                 if (r < 0)
 557                         log_error("Failed to write entry, ignoring: %s", strerror(-r));
 558         }
 559
 560         free(pid);
 561         free(uid);
 562         free(gid);
 563         free(comm);
 564         free(exe);
 565         free(cmdline);
 566         free(source_time);
 567         free(boot_id);
 568         free(machine_id);
 569         free(hostname);
 570         free(audit_session);
 571         free(audit_loginuid);
 572         free(cgroup);
 573         free(session);
 574         free(owner_uid);
 575         free(unit);
 576 }
 577
 578 static void driver_message(Server *s, sd_id128_t message_id, const char *format, ...) {
 579         char mid[11 + 32 + 1];
 580         char buffer[16 + LINE_MAX + 1];
 581         struct iovec iovec[N_IOVEC_META_FIELDS + 3];
 582         int n = 0;
 583         va_list ap;
 584         struct ucred ucred;
 585
 586         assert(s);
 587         assert(format);
 588
 589         IOVEC_SET_STRING(iovec[n++], "PRIORITY=5");
 590
 591         memcpy(buffer, "MESSAGE=", 8);
 592         va_start(ap, format);
 593         vsnprintf(buffer + 8, sizeof(buffer) - 8, format, ap);
 594         va_end(ap);
 595         char_array_0(buffer);
 596         IOVEC_SET_STRING(iovec[n++], buffer);
 597
 598         snprintf(mid, sizeof(mid), "MESSAGE_ID=" SD_ID128_FORMAT_STR, SD_ID128_FORMAT_VAL(message_id));
 599         char_array_0(mid);
 600         IOVEC_SET_STRING(iovec[n++], mid);
 601
 602         zero(ucred);
 603         ucred.pid = getpid();
 604         ucred.uid = getuid();
 605         ucred.gid = getgid();
 606
 607         dispatch_message_real(s, iovec, n, ELEMENTSOF(iovec), &ucred, NULL);
 608 }
 609
 610 static void dispatch_message(Server *s,
 611                              struct iovec *iovec, unsigned n, unsigned m,
 612                              struct ucred *ucred,
 613                              struct timeval *tv,
 614                              int priority) {
 615         int rl;
 616         char *path = NULL, *c;
 617
 618         assert(s);
 619         assert(iovec || n == 0);
 620
 621         if (n == 0)
 622                 return;
 623
 624         if (!ucred)
 625                 goto finish;
 626
 627         path = shortened_cgroup_path(ucred->pid);
 628         if (!path)
 629                 goto finish;
 630
 631         /* example: /user/lennart/3/foobar
 632          *          /system/dbus.service/foobar
 633          *
 634          * So let's cut of everything past the third /, since that is
 635          * wher user directories start */
 636
 637         c = strchr(path, '/');
 638         if (c) {
 639                 c = strchr(c+1, '/');
 640                 if (c) {
 641                         c = strchr(c+1, '/');
 642                         if (c)
 643                                 *c = 0;
 644                 }
 645         }
 646
 647         rl = journal_rate_limit_test(s->rate_limit, path, priority & LOG_PRIMASK, available_space(s));
 648
 649         if (rl == 0) {
 650                 free(path);
 651                 return;
 652         }
 653
 654         /* Write a suppression message if we suppressed something */
 655         if (rl > 1)
 656                 driver_message(s, SD_MESSAGE_JOURNAL_DROPPED, "Suppressed %u messages from %s", rl - 1, path);
 657
 658         free(path);
 659
 660 finish:
 661         dispatch_message_real(s, iovec, n, m, ucred, tv);
 662 }
 663
 664 static void forward_syslog_iovec(Server *s, const struct iovec *iovec, unsigned n_iovec, struct ucred *ucred, struct timeval *tv) {
 665         struct msghdr msghdr;
 666         struct cmsghdr *cmsg;
 667         union {
 668                 struct cmsghdr cmsghdr;
 669                 uint8_t buf[CMSG_SPACE(sizeof(struct ucred))];
 670         } control;
 671         union sockaddr_union sa;
 672
 673         assert(s);
 674         assert(iovec);
 675         assert(n_iovec > 0);
 676
 677         zero(msghdr);
 678         msghdr.msg_iov = (struct iovec*) iovec;
 679         msghdr.msg_iovlen = n_iovec;
 680
 681         zero(sa);
 682         sa.un.sun_family = AF_UNIX;
 683         strncpy(sa.un.sun_path, "/run/systemd/journal/syslog", sizeof(sa.un.sun_path));
 684         msghdr.msg_name = &sa;
 685         msghdr.msg_namelen = offsetof(union sockaddr_union, un.sun_path) + strlen(sa.un.sun_path);
 686
 687         if (ucred) {
 688                 zero(control);
 689                 msghdr.msg_control = &control;
 690                 msghdr.msg_controllen = sizeof(control);
 691
 692                 cmsg = CMSG_FIRSTHDR(&msghdr);
 693                 cmsg->cmsg_level = SOL_SOCKET;
 694                 cmsg->cmsg_type = SCM_CREDENTIALS;
 695                 cmsg->cmsg_len = CMSG_LEN(sizeof(struct ucred));
 696                 memcpy(CMSG_DATA(cmsg), ucred, sizeof(struct ucred));
 697                 msghdr.msg_controllen = cmsg->cmsg_len;
 698         }
 699
 700         /* Forward the syslog message we received via /dev/log to
 701          * /run/systemd/syslog. Unfortunately we currently can't set
 702          * the SO_TIMESTAMP auxiliary data, and hence we don't. */
 703
 704         if (sendmsg(s->syslog_fd, &msghdr, MSG_NOSIGNAL) >= 0)
 705                 return;
 706
 707         if (ucred && errno == ESRCH) {
 708                 struct ucred u;
 709
 710                 /* Hmm, presumably the sender process vanished
 711                  * by now, so let's fix it as good as we
 712                  * can, and retry */
 713
 714                 u = *ucred;
 715                 u.pid = getpid();
 716                 memcpy(CMSG_DATA(cmsg), &u, sizeof(struct ucred));
 717
 718                 if (sendmsg(s->syslog_fd, &msghdr, MSG_NOSIGNAL) >= 0)
 719                         return;
 720         }
 721
 722         log_debug("Failed to forward syslog message: %m");
 723 }
 724
 725 static void forward_syslog_raw(Server *s, const char *buffer, struct ucred *ucred, struct timeval *tv) {
 726         struct iovec iovec;
 727
 728         assert(s);
 729         assert(buffer);
 730
 731         IOVEC_SET_STRING(iovec, buffer);
 732         forward_syslog_iovec(s, &iovec, 1, ucred, tv);
 733 }
 734
 735 static void forward_syslog(Server *s, int priority, const char *tag, const char *message, struct ucred *ucred, struct timeval *tv) {
 736         struct iovec iovec[5];
 737         char header_priority[6], header_time[64], header_pid[16];
 738         int n = 0;
 739         time_t t;
 740         struct tm *tm;
 741         char *tag_buf = NULL;
 742
 743         assert(s);
 744         assert(priority >= 0);
 745         assert(priority <= 999);
 746         assert(message);
 747
 748         /* First: priority field */
 749         snprintf(header_priority, sizeof(header_priority), "<%i>", priority);
 750         char_array_0(header_priority);
 751         IOVEC_SET_STRING(iovec[n++], header_priority);
 752
 753         /* Second: timestamp */
 754         t = tv ? tv->tv_sec : ((time_t) (now(CLOCK_REALTIME) / USEC_PER_SEC));
 755         tm = localtime(&t);
 756         if (!tm)
 757                 return;
 758         if (strftime(header_time, sizeof(header_time), "%h %e %T ", tm) <= 0)
 759                 return;
 760         IOVEC_SET_STRING(iovec[n++], header_time);
 761
 762         /* Third: tag and PID */
 763         if (ucred) {
 764                 if (!tag) {
 765                         get_process_comm(ucred->pid, &tag_buf);
 766                         tag = tag_buf;
 767                 }
 768
 769                 snprintf(header_pid, sizeof(header_pid), "[%lu]: ", (unsigned long) ucred->pid);
 770                 char_array_0(header_pid);
 771
 772                 if (tag)
 773                         IOVEC_SET_STRING(iovec[n++], tag);
 774
 775                 IOVEC_SET_STRING(iovec[n++], header_pid);
 776         } else if (tag) {
 777                 IOVEC_SET_STRING(iovec[n++], tag);
 778                 IOVEC_SET_STRING(iovec[n++], ": ");
 779         }
 780
 781         /* Fourth: message */
 782         IOVEC_SET_STRING(iovec[n++], message);
 783
 784         forward_syslog_iovec(s, iovec, n, ucred, tv);
 785
 786         free(tag_buf);
 787 }
 788
 789 static int fixup_priority(int priority) {
 790
 791         if ((priority & LOG_FACMASK) == 0)
 792                 return (priority & LOG_PRIMASK) | LOG_USER;
 793
 794         return priority;
 795 }
 796
 797 static void forward_kmsg(Server *s, int priority, const char *tag, const char *message, struct ucred *ucred) {
 798         struct iovec iovec[5];
 799         char header_priority[6], header_pid[16];
 800         int n = 0;
 801         char *tag_buf = NULL;
 802         int fd;
 803
 804         assert(s);
 805         assert(priority >= 0);
 806         assert(priority <= 999);
 807         assert(message);
 808
 809         /* Never allow messages with kernel facility to be written to
 810          * kmsg, regardless where the data comes from. */
 811         priority = fixup_priority(priority);
 812
 813         /* First: priority field */
 814         snprintf(header_priority, sizeof(header_priority), "<%i>", priority);
 815         char_array_0(header_priority);
 816         IOVEC_SET_STRING(iovec[n++], header_priority);
 817
 818         /* Second: tag and PID */
 819         if (ucred) {
 820                 if (!tag) {
 821                         get_process_comm(ucred->pid, &tag_buf);
 822                         tag = tag_buf;
 823                 }
 824
 825                 snprintf(header_pid, sizeof(header_pid), "[%lu]: ", (unsigned long) ucred->pid);
 826                 char_array_0(header_pid);
 827
 828                 if (tag)
 829                         IOVEC_SET_STRING(iovec[n++], tag);
 830
 831                 IOVEC_SET_STRING(iovec[n++], header_pid);
 832         } else if (tag) {
 833                 IOVEC_SET_STRING(iovec[n++], tag);
 834                 IOVEC_SET_STRING(iovec[n++], ": ");
 835         }
 836
 837         /* Fourth: message */
 838         IOVEC_SET_STRING(iovec[n++], message);
 839         IOVEC_SET_STRING(iovec[n++], "\n");
 840
 841         fd = open("/dev/kmsg", O_WRONLY|O_NOCTTY|O_CLOEXEC);
 842         if (fd < 0) {
 843                 log_debug("Failed to open /dev/kmsg for logging: %s", strerror(errno));
 844                 goto finish;
 845         }
 846
 847         if (writev(fd, iovec, n) < 0)
 848                 log_debug("Failed to write to /dev/kmsg for logging: %s", strerror(errno));
 849
 850         close_nointr_nofail(fd);
 851
 852 finish:
 853         free(tag_buf);
 854 }
 855
 856 static void forward_console(Server *s, const char *tag, const char *message, struct ucred *ucred) {
 857         struct iovec iovec[4];
 858         char header_pid[16];
 859         int n = 0, fd;
 860         char *tag_buf = NULL;
 861
 862         assert(s);
 863         assert(message);
 864
 865         /* First: tag and PID */
 866         if (ucred) {
 867                 if (!tag) {
 868                         get_process_comm(ucred->pid, &tag_buf);
 869                         tag = tag_buf;
 870                 }
 871
 872                 snprintf(header_pid, sizeof(header_pid), "[%lu]: ", (unsigned long) ucred->pid);
 873                 char_array_0(header_pid);
 874
 875                 if (tag)
 876                         IOVEC_SET_STRING(iovec[n++], tag);
 877
 878                 IOVEC_SET_STRING(iovec[n++], header_pid);
 879         } else if (tag) {
 880                 IOVEC_SET_STRING(iovec[n++], tag);
 881                 IOVEC_SET_STRING(iovec[n++], ": ");
 882         }
 883
 884         /* Third: message */
 885         IOVEC_SET_STRING(iovec[n++], message);
 886         IOVEC_SET_STRING(iovec[n++], "\n");
 887
 888         fd = open_terminal("/dev/console", O_WRONLY|O_NOCTTY|O_CLOEXEC);
 889         if (fd < 0) {
 890                 log_debug("Failed to open /dev/console for logging: %s", strerror(errno));
 891                 goto finish;
 892         }
 893
 894         if (writev(fd, iovec, n) < 0)
 895                 log_debug("Failed to write to /dev/console for logging: %s", strerror(errno));
 896
 897         close_nointr_nofail(fd);
 898
 899 finish:
 900         free(tag_buf);
 901 }
 902
 903 static void read_tag(const char **buf, char **tag) {
 904         const char *p;
 905         char *t;
 906         size_t l, e;
 907
 908         assert(buf);
 909         assert(tag);
 910
 911         p = *buf;
 912
 913         p += strspn(p, WHITESPACE);
 914         l = strcspn(p, WHITESPACE);
 915
 916         if (l <= 0 ||
 917             p[l-1] != ':')
 918                 return;
 919
 920         e = l;
 921         l--;
 922
 923         if (p[l-1] == ']') {
 924                 size_t k = l-1;
 925
 926                 for (;;) {
 927
 928                         if (p[k] == '[') {
 929                                 l = k;
 930                                 break;
 931                         }
 932
 933                         if (k == 0)
 934                                 break;
 935
 936                         k--;
 937                 }
 938         }
 939
 940         t = strndup(p, l);
 941         if (t)
 942                 *tag = t;
 943
 944         *buf = p + e;
 945         *buf += strspn(*buf, WHITESPACE);
 946 }
 947
 948 static void process_syslog_message(Server *s, const char *buf, struct ucred *ucred, struct timeval *tv) {
 949         char *message = NULL, *syslog_priority = NULL, *syslog_facility = NULL, *syslog_tag = NULL;
 950         struct iovec iovec[N_IOVEC_META_FIELDS + 4];
 951         unsigned n = 0;
 952         int priority = LOG_USER | LOG_INFO;
 953         char *tag = NULL;
 954
 955         assert(s);
 956         assert(buf);
 957
 958         if (s->forward_to_syslog)
 959                 forward_syslog_raw(s, buf, ucred, tv);
 960
 961         parse_syslog_priority((char**) &buf, &priority);
 962         skip_syslog_date((char**) &buf);
 963         read_tag(&buf, &tag);
 964
 965         if (s->forward_to_kmsg)
 966                 forward_kmsg(s, priority, tag, buf, ucred);
 967
 968         if (s->forward_to_console)
 969                 forward_console(s, tag, buf, ucred);
 970
 971         if (asprintf(&syslog_priority, "PRIORITY=%i", priority & LOG_PRIMASK) >= 0)
 972                 IOVEC_SET_STRING(iovec[n++], syslog_priority);
 973
 974         if (priority & LOG_FACMASK)
 975                 if (asprintf(&syslog_facility, "SYSLOG_FACILITY=%i", LOG_FAC(priority)) >= 0)
 976                         IOVEC_SET_STRING(iovec[n++], syslog_facility);
 977
 978         if (tag) {
 979                 syslog_tag = strappend("SYSLOG_TAG=", tag);
 980                 if (syslog_tag)
 981                         IOVEC_SET_STRING(iovec[n++], syslog_tag);
 982         }
 983
 984         message = strappend("MESSAGE=", buf);
 985         if (message)
 986                 IOVEC_SET_STRING(iovec[n++], message);
 987
 988         dispatch_message(s, iovec, n, ELEMENTSOF(iovec), ucred, tv, priority);
 989
 990         free(message);
 991         free(tag);
 992         free(syslog_priority);
 993         free(syslog_facility);
 994         free(syslog_tag);
 995 }
 996
 997 static bool valid_user_field(const char *p, size_t l) {
 998         const char *a;
 999
1000         /* We kinda enforce POSIX syntax recommendations for
1001            environment variables here, but make a couple of additional
1002            requirements.
1003
1004            http://pubs.opengroup.org/onlinepubs/000095399/basedefs/xbd_chap08.html */
1005
1006         /* No empty field names */
1007         if (l <= 0)
1008                 return false;
1009
1010         /* Don't allow names longer than 64 chars */
1011         if (l > 64)
1012                 return false;
1013
1014         /* Variables starting with an underscore are protected */
1015         if (p[0] == '_')
1016                 return false;
1017
1018         /* Don't allow digits as first character */
1019         if (p[0] >= '0' && p[0] <= '9')
1020                 return false;
1021
1022         /* Only allow A-Z0-9 and '_' */
1023         for (a = p; a < p + l; a++)
1024                 if (!((*a >= 'A' && *a <= 'Z') ||
1025                       (*a >= '0' && *a <= '9') ||
1026                       *a == '_'))
1027                         return false;
1028
1029         return true;
1030 }
1031
1032 static void process_native_message(Server *s, const void *buffer, size_t buffer_size, struct ucred *ucred, struct timeval *tv) {
1033         struct iovec *iovec = NULL;
1034         unsigned n = 0, m = 0, j;
1035         const char *p;
1036         size_t remaining;
1037         int priority = LOG_INFO;
1038         char *tag = NULL, *message = NULL;
1039
1040         assert(s);
1041         assert(buffer || n == 0);
1042
1043         p = buffer;
1044         remaining = buffer_size;
1045
1046         while (remaining > 0) {
1047                 const char *e, *q;
1048
1049                 e = memchr(p, '\n', remaining);
1050
1051                 if (!e) {
1052                         /* Trailing noise, let's ignore it, and flush what we collected */
1053                         log_debug("Received message with trailing noise, ignoring.");
1054                         break;
1055                 }
1056
1057                 if (e == p) {
1058                         /* Entry separator */
1059                         dispatch_message(s, iovec, n, m, ucred, tv, priority);
1060                         n = 0;
1061                         priority = LOG_INFO;
1062
1063                         p++;
1064                         remaining--;
1065                         continue;
1066                 }
1067
1068                 if (*p == '.' || *p == '#') {
1069                         /* Ignore control commands for now, and
1070                          * comments too. */
1071                         remaining -= (e - p) + 1;
1072                         p = e + 1;
1073                         continue;
1074                 }
1075
1076                 /* A property follows */
1077
1078                 if (n+N_IOVEC_META_FIELDS >= m) {
1079                         struct iovec *c;
1080                         unsigned u;
1081
1082                         u = MAX((n+N_IOVEC_META_FIELDS) * 2U, 4U);
1083                         c = realloc(iovec, u * sizeof(struct iovec));
1084                         if (!c) {
1085                                 log_error("Out of memory");
1086                                 break;
1087                         }
1088
1089                         iovec = c;
1090                         m = u;
1091                 }
1092
1093                 q = memchr(p, '=', e - p);
1094                 if (q) {
1095                         if (valid_user_field(p, q - p)) {
1096                                 size_t l;
1097
1098                                 l = e - p;
1099
1100                                 /* If the field name starts with an
1101                                  * underscore, skip the variable,
1102                                  * since that indidates a trusted
1103                                  * field */
1104                                 iovec[n].iov_base = (char*) p;
1105                                 iovec[n].iov_len = l;
1106                                 n++;
1107
1108                                 /* We need to determine the priority
1109                                  * of this entry for the rate limiting
1110                                  * logic */
1111                                 if (l == 10 &&
1112                                     memcmp(p, "PRIORITY=", 9) == 0 &&
1113                                     p[9] >= '0' && p[9] <= '9')
1114                                         priority = (priority & LOG_FACMASK) | (p[9] - '0');
1115
1116                                 else if (l == 17 &&
1117                                          memcmp(p, "SYSLOG_FACILITY=", 16) == 0 &&
1118                                          p[16] >= '0' && p[16] <= '9')
1119                                         priority = (priority & LOG_PRIMASK) | ((p[16] - '0') << 3);
1120
1121                                 else if (l == 18 &&
1122                                          memcmp(p, "SYSLOG_FACILITY=", 16) == 0 &&
1123                                          p[16] >= '0' && p[16] <= '9' &&
1124                                          p[17] >= '0' && p[17] <= '9')
1125                                         priority = (priority & LOG_PRIMASK) | (((p[16] - '0')*10 + (p[17] - '0')) << 3);
1126
1127                                 else if (l >= 12 &&
1128                                          memcmp(p, "SYSLOG_TAG=", 11) == 0) {
1129                                         char *t;
1130
1131                                         t = strndup(p + 11, l - 11);
1132                                         if (t) {
1133                                                 free(tag);
1134                                                 tag = t;
1135                                         }
1136                                 } else if (l >= 8 &&
1137                                            memcmp(p, "MESSAGE=", 8) == 0) {
1138                                         char *t;
1139
1140                                         t = strndup(p + 8, l - 8);
1141                                         if (t) {
1142                                                 free(message);
1143                                                 message = t;
1144                                         }
1145                                 }
1146                         }
1147
1148                         remaining -= (e - p) + 1;
1149                         p = e + 1;
1150                         continue;
1151                 } else {
1152                         uint64_t l;
1153                         char *k;
1154
1155                         if (remaining < e - p + 1 + sizeof(uint64_t) + 1) {
1156                                 log_debug("Failed to parse message, ignoring.");
1157                                 break;
1158                         }
1159
1160                         memcpy(&l, e + 1, sizeof(uint64_t));
1161                         l = le64toh(l);
1162
1163                         if (remaining < e - p + 1 + sizeof(uint64_t) + l + 1 ||
1164                             e[1+sizeof(uint64_t)+l] != '\n') {
1165                                 log_debug("Failed to parse message, ignoring.");
1166                                 break;
1167                         }
1168
1169                         k = malloc((e - p) + 1 + l);
1170                         if (!k) {
1171                                 log_error("Out of memory");
1172                                 break;
1173                         }
1174
1175                         memcpy(k, p, e - p);
1176                         k[e - p] = '=';
1177                         memcpy(k + (e - p) + 1, e + 1 + sizeof(uint64_t), l);
1178
1179                         if (valid_user_field(p, e - p)) {
1180                                 iovec[n].iov_base = k;
1181                                 iovec[n].iov_len = (e - p) + 1 + l;
1182                                 n++;
1183                         } else
1184                                 free(k);
1185
1186                         remaining -= (e - p) + 1 + sizeof(uint64_t) + l + 1;
1187                         p = e + 1 + sizeof(uint64_t) + l + 1;
1188                 }
1189         }
1190
1191         if (message) {
1192                 if (s->forward_to_syslog)
1193                         forward_syslog(s, priority, tag, message, ucred, tv);
1194
1195                 if (s->forward_to_kmsg)
1196                         forward_kmsg(s, priority, tag, message, ucred);
1197
1198                 if (s->forward_to_console)
1199                         forward_console(s, tag, message, ucred);
1200         }
1201
1202         dispatch_message(s, iovec, n, m, ucred, tv, priority);
1203
1204         for (j = 0; j < n; j++)
1205                 if (iovec[j].iov_base < buffer ||
1206                     (const uint8_t*) iovec[j].iov_base >= (const uint8_t*) buffer + buffer_size)
1207                         free(iovec[j].iov_base);
1208
1209         free(tag);
1210         free(message);
1211 }
1212
1213 static int stdout_stream_log(StdoutStream *s, const char *p) {
1214         struct iovec iovec[N_IOVEC_META_FIELDS + 4];
1215         char *message = NULL, *syslog_priority = NULL, *syslog_facility = NULL, *syslog_tag = NULL;
1216         unsigned n = 0;
1217         int priority;
1218
1219         assert(s);
1220         assert(p);
1221
1222         priority = s->priority;
1223
1224         if (s->priority_prefix)
1225                 parse_syslog_priority((char**) &p, &priority);
1226
1227         if (s->forward_to_syslog || s->server->forward_to_syslog)
1228                 forward_syslog(s->server, fixup_priority(priority), s->tag, p, &s->ucred, NULL);
1229
1230         if (s->forward_to_kmsg || s->server->forward_to_kmsg)
1231                 forward_kmsg(s->server, priority, s->tag, p, &s->ucred);
1232
1233         if (s->forward_to_console || s->server->forward_to_console)
1234                 forward_console(s->server, s->tag, p, &s->ucred);
1235
1236         if (asprintf(&syslog_priority, "PRIORITY=%i", priority & LOG_PRIMASK) >= 0)
1237                 IOVEC_SET_STRING(iovec[n++], syslog_priority);
1238
1239         if (priority & LOG_FACMASK)
1240                 if (asprintf(&syslog_facility, "SYSLOG_FACILITY=%i", LOG_FAC(priority)) >= 0)
1241                         IOVEC_SET_STRING(iovec[n++], syslog_facility);
1242
1243         if (s->tag) {
1244                 syslog_tag = strappend("SYSLOG_TAG=", s->tag);
1245                 if (syslog_tag)
1246                         IOVEC_SET_STRING(iovec[n++], syslog_tag);
1247         }
1248
1249         message = strappend("MESSAGE=", p);
1250         if (message)
1251                 IOVEC_SET_STRING(iovec[n++], message);
1252
1253         dispatch_message(s->server, iovec, n, ELEMENTSOF(iovec), &s->ucred, NULL, priority);
1254
1255         free(message);
1256         free(syslog_priority);
1257         free(syslog_facility);
1258         free(syslog_tag);
1259
1260         return 0;
1261 }
1262
1263 static int stdout_stream_line(StdoutStream *s, char *p) {
1264         int r;
1265
1266         assert(s);
1267         assert(p);
1268
1269         p = strstrip(p);
1270
1271         switch (s->state) {
1272
1273         case STDOUT_STREAM_TAG:
1274                 s->tag = strdup(p);
1275                 if (!s->tag) {
1276                         log_error("Out of memory");
1277                         return -ENOMEM;
1278                 }
1279
1280                 s->state = STDOUT_STREAM_PRIORITY;
1281                 return 0;
1282
1283         case STDOUT_STREAM_PRIORITY:
1284                 r = safe_atoi(p, &s->priority);
1285                 if (r < 0 || s->priority <= 0 || s->priority >= 999) {
1286                         log_warning("Failed to parse log priority line.");
1287                         return -EINVAL;
1288                 }
1289
1290                 s->state = STDOUT_STREAM_PRIORITY_PREFIX;
1291                 return 0;
1292
1293         case STDOUT_STREAM_PRIORITY_PREFIX:
1294                 r = parse_boolean(p);
1295                 if (r < 0) {
1296                         log_warning("Failed to parse priority prefix line.");
1297                         return -EINVAL;
1298                 }
1299
1300                 s->priority_prefix = !!r;
1301                 s->state = STDOUT_STREAM_FORWARD_TO_SYSLOG;
1302                 return 0;
1303
1304         case STDOUT_STREAM_FORWARD_TO_SYSLOG:
1305                 r = parse_boolean(p);
1306                 if (r < 0) {
1307                         log_warning("Failed to parse forward to syslog line.");
1308                         return -EINVAL;
1309                 }
1310
1311                 s->forward_to_syslog = !!r;
1312                 s->state = STDOUT_STREAM_FORWARD_TO_KMSG;
1313                 return 0;
1314
1315         case STDOUT_STREAM_FORWARD_TO_KMSG:
1316                 r = parse_boolean(p);
1317                 if (r < 0) {
1318                         log_warning("Failed to parse copy to kmsg line.");
1319                         return -EINVAL;
1320                 }
1321
1322                 s->forward_to_kmsg = !!r;
1323                 s->state = STDOUT_STREAM_FORWARD_TO_CONSOLE;
1324                 return 0;
1325
1326         case STDOUT_STREAM_FORWARD_TO_CONSOLE:
1327                 r = parse_boolean(p);
1328                 if (r < 0) {
1329                         log_warning("Failed to parse copy to console line.");
1330                         return -EINVAL;
1331                 }
1332
1333                 s->forward_to_console = !!r;
1334                 s->state = STDOUT_STREAM_RUNNING;
1335                 return 0;
1336
1337         case STDOUT_STREAM_RUNNING:
1338                 return stdout_stream_log(s, p);
1339         }
1340
1341         assert_not_reached("Unknown stream state");
1342 }
1343
1344 static int stdout_stream_scan(StdoutStream *s, bool force_flush) {
1345         char *p;
1346         size_t remaining;
1347         int r;
1348
1349         assert(s);
1350
1351         p = s->buffer;
1352         remaining = s->length;
1353         for (;;) {
1354                 char *end;
1355                 size_t skip;
1356
1357                 end = memchr(p, '\n', remaining);
1358                 if (end)
1359                         skip = end - p + 1;
1360                 else if (remaining >= sizeof(s->buffer) - 1) {
1361                         end = p + sizeof(s->buffer) - 1;
1362                         skip = sizeof(s->buffer) - 1;
1363                 } else
1364                         break;
1365
1366                 *end = 0;
1367
1368                 r = stdout_stream_line(s, p);
1369                 if (r < 0)
1370                         return r;
1371
1372                 remaining -= skip;
1373                 p += skip;
1374         }
1375
1376         if (force_flush && remaining > 0) {
1377                 p[remaining] = 0;
1378                 r = stdout_stream_line(s, p);
1379                 if (r < 0)
1380                         return r;
1381
1382                 p += remaining;
1383                 remaining = 0;
1384         }
1385
1386         if (p > s->buffer) {
1387                 memmove(s->buffer, p, remaining);
1388                 s->length = remaining;
1389         }
1390
1391         return 0;
1392 }
1393
1394 static int stdout_stream_process(StdoutStream *s) {
1395         ssize_t l;
1396         int r;
1397
1398         assert(s);
1399
1400         l = read(s->fd, s->buffer+s->length, sizeof(s->buffer)-1-s->length);
1401         if (l < 0) {
1402
1403                 if (errno == EAGAIN)
1404                         return 0;
1405
1406                 log_warning("Failed to read from stream: %m");
1407                 return -errno;
1408         }
1409
1410         if (l == 0) {
1411                 r = stdout_stream_scan(s, true);
1412                 if (r < 0)
1413                         return r;
1414
1415                 return 0;
1416         }
1417
1418         s->length += l;
1419         r = stdout_stream_scan(s, false);
1420         if (r < 0)
1421                 return r;
1422
1423         return 1;
1424
1425 }
1426
1427 static void stdout_stream_free(StdoutStream *s) {
1428         assert(s);
1429
1430         if (s->server) {
1431                 assert(s->server->n_stdout_streams > 0);
1432                 s->server->n_stdout_streams --;
1433                 LIST_REMOVE(StdoutStream, stdout_stream, s->server->stdout_streams, s);
1434         }
1435
1436         if (s->fd >= 0) {
1437                 if (s->server)
1438                         epoll_ctl(s->server->epoll_fd, EPOLL_CTL_DEL, s->fd, NULL);
1439
1440                 close_nointr_nofail(s->fd);
1441         }
1442
1443         free(s->tag);
1444         free(s);
1445 }
1446
1447 static int stdout_stream_new(Server *s) {
1448         StdoutStream *stream;
1449         int fd, r;
1450         socklen_t len;
1451         struct epoll_event ev;
1452
1453         assert(s);
1454
1455         fd = accept4(s->stdout_fd, NULL, NULL, SOCK_NONBLOCK|SOCK_CLOEXEC);
1456         if (fd < 0) {
1457                 if (errno == EAGAIN)
1458                         return 0;
1459
1460                 log_error("Failed to accept stdout connection: %m");
1461                 return -errno;
1462         }
1463
1464         if (s->n_stdout_streams >= STDOUT_STREAMS_MAX) {
1465                 log_warning("Too many stdout streams, refusing connection.");
1466                 close_nointr_nofail(fd);
1467                 return 0;
1468         }
1469
1470         stream = new0(StdoutStream, 1);
1471         if (!stream) {
1472                 log_error("Out of memory.");
1473                 close_nointr_nofail(fd);
1474                 return -ENOMEM;
1475         }
1476
1477         stream->fd = fd;
1478
1479         len = sizeof(stream->ucred);
1480         if (getsockopt(fd, SOL_SOCKET, SO_PEERCRED, &stream->ucred, &len) < 0) {
1481                 log_error("Failed to determine peer credentials: %m");
1482                 r = -errno;
1483                 goto fail;
1484         }
1485
1486         if (shutdown(fd, SHUT_WR) < 0) {
1487                 log_error("Failed to shutdown writing side of socket: %m");
1488                 r = -errno;
1489                 goto fail;
1490         }
1491
1492         zero(ev);
1493         ev.data.ptr = stream;
1494         ev.events = EPOLLIN;
1495         if (epoll_ctl(s->epoll_fd, EPOLL_CTL_ADD, fd, &ev) < 0) {
1496                 log_error("Failed to add stream to event loop: %m");
1497                 r = -errno;
1498                 goto fail;
1499         }
1500
1501         stream->server = s;
1502         LIST_PREPEND(StdoutStream, stdout_stream, s->stdout_streams, stream);
1503         s->n_stdout_streams ++;
1504
1505         return 0;
1506
1507 fail:
1508         stdout_stream_free(stream);
1509         return r;
1510 }
1511
1512 static int system_journal_open(Server *s) {
1513         int r;
1514         char *fn;
1515         sd_id128_t machine;
1516         char ids[33];
1517
1518         r = sd_id128_get_machine(&machine);
1519         if (r < 0)
1520                 return r;
1521
1522         sd_id128_to_string(machine, ids);
1523
1524         if (!s->system_journal) {
1525
1526                 /* First try to create the machine path, but not the prefix */
1527                 fn = strappend("/var/log/journal/", ids);
1528                 if (!fn)
1529                         return -ENOMEM;
1530                 (void) mkdir(fn, 0755);
1531                 free(fn);
1532
1533                 /* The create the system journal file */
1534                 fn = join("/var/log/journal/", ids, "/system.journal", NULL);
1535                 if (!fn)
1536                         return -ENOMEM;
1537
1538                 r = journal_file_open(fn, O_RDWR|O_CREAT, 0640, NULL, &s->system_journal);
1539                 free(fn);
1540
1541                 if (r >= 0) {
1542                         journal_default_metrics(&s->system_metrics, s->system_journal->fd);
1543
1544                         s->system_journal->metrics = s->system_metrics;
1545                         s->system_journal->compress = s->compress;
1546
1547                         fix_perms(s->system_journal, 0);
1548                 } else if (r < 0) {
1549
1550                         if (r != -ENOENT && r != -EROFS)
1551                                 log_warning("Failed to open system journal: %s", strerror(-r));
1552
1553                         r = 0;
1554                 }
1555         }
1556
1557         if (!s->runtime_journal) {
1558
1559                 fn = join("/run/log/journal/", ids, "/system.journal", NULL);
1560                 if (!fn)
1561                         return -ENOMEM;
1562
1563                 if (s->system_journal) {
1564
1565                         /* Try to open the runtime journal, but only
1566                          * if it already exists, so that we can flush
1567                          * it into the system journal */
1568
1569                         r = journal_file_open(fn, O_RDWR, 0640, NULL, &s->runtime_journal);
1570                         free(fn);
1571
1572                         if (r < 0) {
1573                                 if (r != -ENOENT)
1574                                         log_warning("Failed to open runtime journal: %s", strerror(-r));
1575
1576                                 r = 0;
1577                         }
1578
1579                 } else {
1580
1581                         /* OK, we really need the runtime journal, so create
1582                          * it if necessary. */
1583
1584                         (void) mkdir_parents(fn, 0755);
1585                         r = journal_file_open(fn, O_RDWR|O_CREAT, 0640, NULL, &s->runtime_journal);
1586                         free(fn);
1587
1588                         if (r < 0) {
1589                                 log_error("Failed to open runtime journal: %s", strerror(-r));
1590                                 return r;
1591                         }
1592                 }
1593
1594                 if (s->runtime_journal) {
1595                         journal_default_metrics(&s->runtime_metrics, s->runtime_journal->fd);
1596
1597                         s->runtime_journal->metrics = s->runtime_metrics;
1598                         s->runtime_journal->compress = s->compress;
1599
1600                         fix_perms(s->runtime_journal, 0);
1601                 }
1602         }
1603
1604         return r;
1605 }
1606
1607 static int server_flush_to_var(Server *s) {
1608         char path[] = "/run/log/journal/xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx";
1609         Object *o = NULL;
1610         int r;
1611         sd_id128_t machine;
1612         sd_journal *j;
1613         usec_t ts;
1614
1615         assert(s);
1616
1617         if (!s->runtime_journal)
1618                 return 0;
1619
1620         ts = now(CLOCK_MONOTONIC);
1621         if (s->var_available_timestamp + RECHECK_VAR_AVAILABLE_USEC > ts)
1622                 return 0;
1623
1624         s->var_available_timestamp = ts;
1625
1626         system_journal_open(s);
1627
1628         if (!s->system_journal)
1629                 return 0;
1630
1631         r = sd_id128_get_machine(&machine);
1632         if (r < 0) {
1633                 log_error("Failed to get machine id: %s", strerror(-r));
1634                 return r;
1635         }
1636
1637         r = sd_journal_open(&j, SD_JOURNAL_RUNTIME_ONLY);
1638         if (r < 0) {
1639                 log_error("Failed to read runtime journal: %s", strerror(-r));
1640                 return r;
1641         }
1642
1643         SD_JOURNAL_FOREACH(j) {
1644                 JournalFile *f;
1645
1646                 f = j->current_file;
1647                 assert(f && f->current_offset > 0);
1648
1649                 r = journal_file_move_to_object(f, OBJECT_ENTRY, f->current_offset, &o);
1650                 if (r < 0) {
1651                         log_error("Can't read entry: %s", strerror(-r));
1652                         goto finish;
1653                 }
1654
1655                 r = journal_file_copy_entry(f, s->system_journal, o, f->current_offset, NULL, NULL, NULL);
1656                 if (r == -E2BIG) {
1657                         log_info("Allocation limit reached.");
1658
1659                         journal_file_post_change(s->system_journal);
1660                         server_rotate(s);
1661                         server_vacuum(s);
1662
1663                         r = journal_file_copy_entry(f, s->system_journal, o, f->current_offset, NULL, NULL, NULL);
1664                 }
1665
1666                 if (r < 0) {
1667                         log_error("Can't write entry: %s", strerror(-r));
1668                         goto finish;
1669                 }
1670         }
1671
1672 finish:
1673         journal_file_post_change(s->system_journal);
1674
1675         journal_file_close(s->runtime_journal);
1676         s->runtime_journal = NULL;
1677
1678         if (r >= 0) {
1679                 sd_id128_to_string(machine, path + 17);
1680                 rm_rf(path, false, true, false);
1681         }
1682
1683         return r;
1684 }
1685
1686 static int process_event(Server *s, struct epoll_event *ev) {
1687         assert(s);
1688
1689         if (ev->data.fd == s->signal_fd) {
1690                 struct signalfd_siginfo sfsi;
1691                 ssize_t n;
1692
1693                 if (ev->events != EPOLLIN) {
1694                         log_info("Got invalid event from epoll.");
1695                         return -EIO;
1696                 }
1697
1698                 n = read(s->signal_fd, &sfsi, sizeof(sfsi));
1699                 if (n != sizeof(sfsi)) {
1700
1701                         if (n >= 0)
1702                                 return -EIO;
1703
1704                         if (errno == EINTR || errno == EAGAIN)
1705                                 return 0;
1706
1707                         return -errno;
1708                 }
1709
1710                 if (sfsi.ssi_signo == SIGUSR1) {
1711                         server_flush_to_var(s);
1712                         return 0;
1713                 }
1714
1715                 log_debug("Received SIG%s", signal_to_string(sfsi.ssi_signo));
1716                 return 0;
1717
1718         } else if (ev->data.fd == s->native_fd ||
1719                    ev->data.fd == s->syslog_fd) {
1720
1721                 if (ev->events != EPOLLIN) {
1722                         log_info("Got invalid event from epoll.");
1723                         return -EIO;
1724                 }
1725
1726                 for (;;) {
1727                         struct msghdr msghdr;
1728                         struct iovec iovec;
1729                         struct ucred *ucred = NULL;
1730                         struct timeval *tv = NULL;
1731                         struct cmsghdr *cmsg;
1732                         union {
1733                                 struct cmsghdr cmsghdr;
1734                                 uint8_t buf[CMSG_SPACE(sizeof(struct ucred)) +
1735                                             CMSG_SPACE(sizeof(struct timeval))];
1736                         } control;
1737                         ssize_t n;
1738                         int v;
1739
1740                         if (ioctl(ev->data.fd, SIOCINQ, &v) < 0) {
1741                                 log_error("SIOCINQ failed: %m");
1742                                 return -errno;
1743                         }
1744
1745                         if (v <= 0)
1746                                 return 1;
1747
1748                         if (s->buffer_size < (size_t) v) {
1749                                 void *b;
1750                                 size_t l;
1751
1752                                 l = MAX(LINE_MAX + (size_t) v, s->buffer_size * 2);
1753                                 b = realloc(s->buffer, l+1);
1754
1755                                 if (!b) {
1756                                         log_error("Couldn't increase buffer.");
1757                                         return -ENOMEM;
1758                                 }
1759
1760                                 s->buffer_size = l;
1761                                 s->buffer = b;
1762                         }
1763
1764                         zero(iovec);
1765                         iovec.iov_base = s->buffer;
1766                         iovec.iov_len = s->buffer_size;
1767
1768                         zero(control);
1769                         zero(msghdr);
1770                         msghdr.msg_iov = &iovec;
1771                         msghdr.msg_iovlen = 1;
1772                         msghdr.msg_control = &control;
1773                         msghdr.msg_controllen = sizeof(control);
1774
1775                         n = recvmsg(ev->data.fd, &msghdr, MSG_DONTWAIT);
1776                         if (n < 0) {
1777
1778                                 if (errno == EINTR || errno == EAGAIN)
1779                                         return 1;
1780
1781                                 log_error("recvmsg() failed: %m");
1782                                 return -errno;
1783                         }
1784
1785                         for (cmsg = CMSG_FIRSTHDR(&msghdr); cmsg; cmsg = CMSG_NXTHDR(&msghdr, cmsg)) {
1786
1787                                 if (cmsg->cmsg_level == SOL_SOCKET &&
1788                                     cmsg->cmsg_type == SCM_CREDENTIALS &&
1789                                     cmsg->cmsg_len == CMSG_LEN(sizeof(struct ucred)))
1790                                         ucred = (struct ucred*) CMSG_DATA(cmsg);
1791                                 else if (cmsg->cmsg_level == SOL_SOCKET &&
1792                                          cmsg->cmsg_type == SO_TIMESTAMP &&
1793                                          cmsg->cmsg_len == CMSG_LEN(sizeof(struct timeval)))
1794                                         tv = (struct timeval*) CMSG_DATA(cmsg);
1795                         }
1796
1797                         if (ev->data.fd == s->syslog_fd) {
1798                                 char *e;
1799
1800                                 e = memchr(s->buffer, '\n', n);
1801                                 if (e)
1802                                         *e = 0;
1803                                 else
1804                                         s->buffer[n] = 0;
1805
1806                                 process_syslog_message(s, strstrip(s->buffer), ucred, tv);
1807                         } else
1808                                 process_native_message(s, s->buffer, n, ucred, tv);
1809                 }
1810
1811                 return 1;
1812
1813         } else if (ev->data.fd == s->stdout_fd) {
1814
1815                 if (ev->events != EPOLLIN) {
1816                         log_info("Got invalid event from epoll.");
1817                         return -EIO;
1818                 }
1819
1820                 stdout_stream_new(s);
1821                 return 1;
1822
1823         } else {
1824                 StdoutStream *stream;
1825
1826                 if ((ev->events|EPOLLIN|EPOLLHUP) != (EPOLLIN|EPOLLHUP)) {
1827                         log_info("Got invalid event from epoll.");
1828                         return -EIO;
1829                 }
1830
1831                 /* If it is none of the well-known fds, it must be an
1832                  * stdout stream fd. Note that this is a bit ugly here
1833                  * (since we rely that none of the well-known fds
1834                  * could be interpreted as pointer), but nonetheless
1835                  * safe, since the well-known fds would never get an
1836                  * fd > 4096, i.e. beyond the first memory page */
1837
1838                 stream = ev->data.ptr;
1839
1840                 if (stdout_stream_process(stream) <= 0)
1841                         stdout_stream_free(stream);
1842
1843                 return 1;
1844         }
1845
1846         log_error("Unknown event.");
1847         return 0;
1848 }
1849
1850 static int open_syslog_socket(Server *s) {
1851         union sockaddr_union sa;
1852         int one, r;
1853         struct epoll_event ev;
1854         struct timeval tv;
1855
1856         assert(s);
1857
1858         if (s->syslog_fd < 0) {
1859
1860                 s->syslog_fd = socket(AF_UNIX, SOCK_DGRAM|SOCK_CLOEXEC, 0);
1861                 if (s->syslog_fd < 0) {
1862                         log_error("socket() failed: %m");
1863                         return -errno;
1864                 }
1865
1866                 zero(sa);
1867                 sa.un.sun_family = AF_UNIX;
1868                 strncpy(sa.un.sun_path, "/dev/log", sizeof(sa.un.sun_path));
1869
1870                 unlink(sa.un.sun_path);
1871
1872                 r = bind(s->syslog_fd, &sa.sa, offsetof(union sockaddr_union, un.sun_path) + strlen(sa.un.sun_path));
1873                 if (r < 0) {
1874                         log_error("bind() failed: %m");
1875                         return -errno;
1876                 }
1877
1878                 chmod(sa.un.sun_path, 0666);
1879         }
1880
1881         one = 1;
1882         r = setsockopt(s->syslog_fd, SOL_SOCKET, SO_PASSCRED, &one, sizeof(one));
1883         if (r < 0) {
1884                 log_error("SO_PASSCRED failed: %m");
1885                 return -errno;
1886         }
1887
1888         one = 1;
1889         r = setsockopt(s->syslog_fd, SOL_SOCKET, SO_TIMESTAMP, &one, sizeof(one));
1890         if (r < 0) {
1891                 log_error("SO_TIMESTAMP failed: %m");
1892                 return -errno;
1893         }
1894
1895         /* Since we use the same socket for forwarding this to some
1896          * other syslog implementation, make sure we don't hang
1897          * forever */
1898         timeval_store(&tv, SYSLOG_TIMEOUT_USEC);
1899         if (setsockopt(s->syslog_fd, SOL_SOCKET, SO_SNDTIMEO, &tv, sizeof(tv)) < 0) {
1900                 log_error("SO_SNDTIMEO failed: %m");
1901                 return -errno;
1902         }
1903
1904         zero(ev);
1905         ev.events = EPOLLIN;
1906         ev.data.fd = s->syslog_fd;
1907         if (epoll_ctl(s->epoll_fd, EPOLL_CTL_ADD, s->syslog_fd, &ev) < 0) {
1908                 log_error("Failed to add syslog server fd to epoll object: %m");
1909                 return -errno;
1910         }
1911
1912         return 0;
1913 }
1914
1915 static int open_native_socket(Server*s) {
1916         union sockaddr_union sa;
1917         int one, r;
1918         struct epoll_event ev;
1919
1920         assert(s);
1921
1922         if (s->native_fd < 0) {
1923
1924                 s->native_fd = socket(AF_UNIX, SOCK_DGRAM|SOCK_CLOEXEC, 0);
1925                 if (s->native_fd < 0) {
1926                         log_error("socket() failed: %m");
1927                         return -errno;
1928                 }
1929
1930                 zero(sa);
1931                 sa.un.sun_family = AF_UNIX;
1932                 strncpy(sa.un.sun_path, "/run/systemd/journal/socket", sizeof(sa.un.sun_path));
1933
1934                 unlink(sa.un.sun_path);
1935
1936                 r = bind(s->native_fd, &sa.sa, offsetof(union sockaddr_union, un.sun_path) + strlen(sa.un.sun_path));
1937                 if (r < 0) {
1938                         log_error("bind() failed: %m");
1939                         return -errno;
1940                 }
1941
1942                 chmod(sa.un.sun_path, 0666);
1943         }
1944
1945         one = 1;
1946         r = setsockopt(s->native_fd, SOL_SOCKET, SO_PASSCRED, &one, sizeof(one));
1947         if (r < 0) {
1948                 log_error("SO_PASSCRED failed: %m");
1949                 return -errno;
1950         }
1951
1952         one = 1;
1953         r = setsockopt(s->native_fd, SOL_SOCKET, SO_TIMESTAMP, &one, sizeof(one));
1954         if (r < 0) {
1955                 log_error("SO_TIMESTAMP failed: %m");
1956                 return -errno;
1957         }
1958
1959         zero(ev);
1960         ev.events = EPOLLIN;
1961         ev.data.fd = s->native_fd;
1962         if (epoll_ctl(s->epoll_fd, EPOLL_CTL_ADD, s->native_fd, &ev) < 0) {
1963                 log_error("Failed to add native server fd to epoll object: %m");
1964                 return -errno;
1965         }
1966
1967         return 0;
1968 }
1969
1970 static int open_stdout_socket(Server *s) {
1971         union sockaddr_union sa;
1972         int r;
1973         struct epoll_event ev;
1974
1975         assert(s);
1976
1977         if (s->stdout_fd < 0) {
1978
1979                 s->stdout_fd = socket(AF_UNIX, SOCK_STREAM|SOCK_CLOEXEC, 0);
1980                 if (s->stdout_fd < 0) {
1981                         log_error("socket() failed: %m");
1982                         return -errno;
1983                 }
1984
1985                 zero(sa);
1986                 sa.un.sun_family = AF_UNIX;
1987                 strncpy(sa.un.sun_path, "/run/systemd/journal/stdout", sizeof(sa.un.sun_path));
1988
1989                 unlink(sa.un.sun_path);
1990
1991                 r = bind(s->stdout_fd, &sa.sa, offsetof(union sockaddr_union, un.sun_path) + strlen(sa.un.sun_path));
1992                 if (r < 0) {
1993                         log_error("bind() failed: %m");
1994                         return -errno;
1995                 }
1996
1997                 chmod(sa.un.sun_path, 0666);
1998
1999                 if (listen(s->stdout_fd, SOMAXCONN) < 0) {
2000                         log_error("liste() failed: %m");
2001                         return -errno;
2002                 }
2003         }
2004
2005         zero(ev);
2006         ev.events = EPOLLIN;
2007         ev.data.fd = s->stdout_fd;
2008         if (epoll_ctl(s->epoll_fd, EPOLL_CTL_ADD, s->stdout_fd, &ev) < 0) {
2009                 log_error("Failed to add stdout server fd to epoll object: %m");
2010                 return -errno;
2011         }
2012
2013         return 0;
2014 }
2015
2016 static int open_signalfd(Server *s) {
2017         sigset_t mask;
2018         struct epoll_event ev;
2019
2020         assert(s);
2021
2022         assert_se(sigemptyset(&mask) == 0);
2023         sigset_add_many(&mask, SIGINT, SIGTERM, SIGUSR1, -1);
2024         assert_se(sigprocmask(SIG_SETMASK, &mask, NULL) == 0);
2025
2026         s->signal_fd = signalfd(-1, &mask, SFD_NONBLOCK|SFD_CLOEXEC);
2027         if (s->signal_fd < 0) {
2028                 log_error("signalfd(): %m");
2029                 return -errno;
2030         }
2031
2032         zero(ev);
2033         ev.events = EPOLLIN;
2034         ev.data.fd = s->signal_fd;
2035
2036         if (epoll_ctl(s->epoll_fd, EPOLL_CTL_ADD, s->signal_fd, &ev) < 0) {
2037                 log_error("epoll_ctl(): %m");
2038                 return -errno;
2039         }
2040
2041         return 0;
2042 }
2043
2044 static int server_parse_config_file(Server *s) {
2045         FILE *f;
2046         const char *fn;
2047         int r;
2048
2049         assert(s);
2050
2051         fn = "/etc/systemd/systemd-journald.conf";
2052         f = fopen(fn, "re");
2053         if (!f) {
2054                 if (errno == ENOENT)
2055                         return 0;
2056
2057                 log_warning("Failed to open configuration file %s: %m", fn);
2058                 return -errno;
2059         }
2060
2061         r = config_parse(fn, f, "Journal\0", config_item_perf_lookup, (void*) journald_gperf_lookup, false, s);
2062         if (r < 0)
2063                 log_warning("Failed to parse configuration file: %s", strerror(-r));
2064
2065         fclose(f);
2066
2067         return r;
2068 }
2069
2070 static int server_init(Server *s) {
2071         int n, r, fd;
2072
2073         assert(s);
2074
2075         zero(*s);
2076         s->syslog_fd = s->native_fd = s->stdout_fd = s->signal_fd = s->epoll_fd = -1;
2077         s->compress = true;
2078
2079         s->rate_limit_interval = DEFAULT_RATE_LIMIT_INTERVAL;
2080         s->rate_limit_burst = DEFAULT_RATE_LIMIT_BURST;
2081
2082         s->forward_to_syslog = true;
2083
2084         memset(&s->system_metrics, 0xFF, sizeof(s->system_metrics));
2085         memset(&s->runtime_metrics, 0xFF, sizeof(s->runtime_metrics));
2086
2087         server_parse_config_file(s);
2088
2089         s->user_journals = hashmap_new(trivial_hash_func, trivial_compare_func);
2090         if (!s->user_journals) {
2091                 log_error("Out of memory.");
2092                 return -ENOMEM;
2093         }
2094
2095         s->epoll_fd = epoll_create1(EPOLL_CLOEXEC);
2096         if (s->epoll_fd < 0) {
2097                 log_error("Failed to create epoll object: %m");
2098                 return -errno;
2099         }
2100
2101         n = sd_listen_fds(true);
2102         if (n < 0) {
2103                 log_error("Failed to read listening file descriptors from environment: %s", strerror(-n));
2104                 return n;
2105         }
2106
2107         for (fd = SD_LISTEN_FDS_START; fd < SD_LISTEN_FDS_START + n; fd++) {
2108
2109                 if (sd_is_socket_unix(fd, SOCK_DGRAM, -1, "/run/systemd/journal/socket", 0) > 0) {
2110
2111                         if (s->native_fd >= 0) {
2112                                 log_error("Too many native sockets passed.");
2113                                 return -EINVAL;
2114                         }
2115
2116                         s->native_fd = fd;
2117
2118                 } else if (sd_is_socket_unix(fd, SOCK_STREAM, 1, "/run/systemd/journal/stdout", 0) > 0) {
2119
2120                         if (s->stdout_fd >= 0) {
2121                                 log_error("Too many stdout sockets passed.");
2122                                 return -EINVAL;
2123                         }
2124
2125                         s->stdout_fd = fd;
2126
2127                 } else if (sd_is_socket_unix(fd, SOCK_DGRAM, -1, "/dev/log", 0) > 0) {
2128
2129                         if (s->syslog_fd >= 0) {
2130                                 log_error("Too many /dev/log sockets passed.");
2131                                 return -EINVAL;
2132                         }
2133
2134                         s->syslog_fd = fd;
2135
2136                 } else {
2137                         log_error("Unknown socket passed.");
2138                         return -EINVAL;
2139                 }
2140         }
2141
2142         r = open_syslog_socket(s);
2143         if (r < 0)
2144                 return r;
2145
2146         r = open_native_socket(s);
2147         if (r < 0)
2148                 return r;
2149
2150         r = open_stdout_socket(s);
2151         if (r < 0)
2152                 return r;
2153
2154         r = system_journal_open(s);
2155         if (r < 0)
2156                 return r;
2157
2158         r = open_signalfd(s);
2159         if (r < 0)
2160                 return r;
2161
2162         s->rate_limit = journal_rate_limit_new(s->rate_limit_interval, s->rate_limit_burst);
2163         if (!s->rate_limit)
2164                 return -ENOMEM;
2165
2166         return 0;
2167 }
2168
2169 static void server_done(Server *s) {
2170         JournalFile *f;
2171         assert(s);
2172
2173         while (s->stdout_streams)
2174                 stdout_stream_free(s->stdout_streams);
2175
2176         if (s->system_journal)
2177                 journal_file_close(s->system_journal);
2178
2179         if (s->runtime_journal)
2180                 journal_file_close(s->runtime_journal);
2181
2182         while ((f = hashmap_steal_first(s->user_journals)))
2183                 journal_file_close(f);
2184
2185         hashmap_free(s->user_journals);
2186
2187         if (s->epoll_fd >= 0)
2188                 close_nointr_nofail(s->epoll_fd);
2189
2190         if (s->signal_fd >= 0)
2191                 close_nointr_nofail(s->signal_fd);
2192
2193         if (s->syslog_fd >= 0)
2194                 close_nointr_nofail(s->syslog_fd);
2195
2196         if (s->native_fd >= 0)
2197                 close_nointr_nofail(s->native_fd);
2198
2199         if (s->stdout_fd >= 0)
2200                 close_nointr_nofail(s->stdout_fd);
2201
2202         if (s->rate_limit)
2203                 journal_rate_limit_free(s->rate_limit);
2204
2205         free(s->buffer);
2206 }
2207
2208 int main(int argc, char *argv[]) {
2209         Server server;
2210         int r;
2211
2212         /* if (getppid() != 1) { */
2213         /*         log_error("This program should be invoked by init only."); */
2214         /*         return EXIT_FAILURE; */
2215         /* } */
2216
2217         if (argc > 1) {
2218                 log_error("This program does not take arguments.");
2219                 return EXIT_FAILURE;
2220         }
2221
2222         log_set_target(LOG_TARGET_CONSOLE);
2223         log_parse_environment();
2224         log_open();
2225
2226         umask(0022);
2227
2228         r = server_init(&server);
2229         if (r < 0)
2230                 goto finish;
2231
2232         server_vacuum(&server);
2233         server_flush_to_var(&server);
2234
2235         log_debug("systemd-journald running as pid %lu", (unsigned long) getpid());
2236         driver_message(&server, SD_MESSAGE_JOURNAL_START, "Journal started");
2237
2238         sd_notify(false,
2239                   "READY=1\n"
2240                   "STATUS=Processing requests...");
2241
2242         for (;;) {
2243                 struct epoll_event event;
2244
2245                 r = epoll_wait(server.epoll_fd, &event, 1, -1);
2246                 if (r < 0) {
2247
2248                         if (errno == EINTR)
2249                                 continue;
2250
2251                         log_error("epoll_wait() failed: %m");
2252                         r = -errno;
2253                         goto finish;
2254                 } else if (r == 0)
2255                         break;
2256
2257                 r = process_event(&server, &event);
2258                 if (r < 0)
2259                         goto finish;
2260                 else if (r == 0)
2261                         break;
2262         }
2263
2264         log_debug("systemd-journald stopped as pid %lu", (unsigned long) getpid());
2265         driver_message(&server, SD_MESSAGE_JOURNAL_STOP, "Journal stopped");
2266
2267 finish:
2268         sd_notify(false,
2269                   "STATUS=Shutting down...");
2270
2271         server_done(&server);
2272
2273         return r < 0 ? EXIT_FAILURE : EXIT_SUCCESS;
2274 }