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
  34 #include "hashmap.h"
  35 #include "journal-file.h"
  36 #include "sd-daemon.h"
  37 #include "socket-util.h"
  38 #include "acl-util.h"
  39 #include "cgroup-util.h"
  40
  41 #define USER_JOURNALS_MAX 1024
  42
  43 typedef struct Server {
  44         int epoll_fd;
  45         int signal_fd;
  46         int syslog_fd;
  47         int native_fd;
  48
  49         JournalFile *runtime_journal;
  50         JournalFile *system_journal;
  51         Hashmap *user_journals;
  52
  53         uint64_t seqnum;
  54
  55         char *buffer;
  56         size_t buffer_size;
  57
  58         JournalMetrics metrics;
  59         uint64_t max_use;
  60 } Server;
  61
  62 static void fix_perms(JournalFile *f, uid_t uid) {
  63         acl_t acl;
  64         acl_entry_t entry;
  65         acl_permset_t permset;
  66         int r;
  67
  68         assert(f);
  69
  70         r = fchmod_and_fchown(f->fd, 0640, 0, 0);
  71         if (r < 0)
  72                 log_warning("Failed to fix access mode/rights on %s, ignoring: %s", f->path, strerror(-r));
  73
  74         if (uid <= 0)
  75                 return;
  76
  77         acl = acl_get_fd(f->fd);
  78         if (!acl) {
  79                 log_warning("Failed to read ACL on %s, ignoring: %m", f->path);
  80                 return;
  81         }
  82
  83         r = acl_find_uid(acl, uid, &entry);
  84         if (r <= 0) {
  85
  86                 if (acl_create_entry(&acl, &entry) < 0 ||
  87                     acl_set_tag_type(entry, ACL_USER) < 0 ||
  88                     acl_set_qualifier(entry, &uid) < 0) {
  89                         log_warning("Failed to patch ACL on %s, ignoring: %m", f->path);
  90                         goto finish;
  91                 }
  92         }
  93
  94         if (acl_get_permset(entry, &permset) < 0 ||
  95             acl_add_perm(permset, ACL_READ) < 0 ||
  96             acl_calc_mask(&acl) < 0) {
  97                 log_warning("Failed to patch ACL on %s, ignoring: %m", f->path);
  98                 goto finish;
  99         }
 100
 101         if (acl_set_fd(f->fd, acl) < 0)
 102                 log_warning("Failed to set ACL on %s, ignoring: %m", f->path);
 103
 104 finish:
 105         acl_free(acl);
 106 }
 107
 108 static JournalFile* find_journal(Server *s, uid_t uid) {
 109         char *p;
 110         int r;
 111         JournalFile *f;
 112         char ids[33];
 113         sd_id128_t machine;
 114
 115         assert(s);
 116
 117         /* We split up user logs only on /var, not on /run */
 118         if (!s->system_journal)
 119                 return s->runtime_journal;
 120
 121         if (uid <= 0)
 122                 return s->system_journal;
 123
 124         r = sd_id128_get_machine(&machine);
 125         if (r < 0)
 126                 return s->system_journal;
 127
 128         f = hashmap_get(s->user_journals, UINT32_TO_PTR(uid));
 129         if (f)
 130                 return f;
 131
 132         if (asprintf(&p, "/var/log/journal/%s/user-%lu.journal", sd_id128_to_string(machine, ids), (unsigned long) uid) < 0)
 133                 return s->system_journal;
 134
 135         while (hashmap_size(s->user_journals) >= USER_JOURNALS_MAX) {
 136                 /* Too many open? Then let's close one */
 137                 f = hashmap_steal_first(s->user_journals);
 138                 assert(f);
 139                 journal_file_close(f);
 140         }
 141
 142         r = journal_file_open(p, O_RDWR|O_CREAT, 0640, s->system_journal, &f);
 143         free(p);
 144
 145         if (r < 0)
 146                 return s->system_journal;
 147
 148         fix_perms(f, uid);
 149
 150         r = hashmap_put(s->user_journals, UINT32_TO_PTR(uid), f);
 151         if (r < 0) {
 152                 journal_file_close(f);
 153                 return s->system_journal;
 154         }
 155
 156         return f;
 157 }
 158
 159 static void server_vacuum(Server *s) {
 160         Iterator i;
 161         void *k;
 162         char *p;
 163         char ids[33];
 164         sd_id128_t machine;
 165         int r;
 166         JournalFile *f;
 167
 168         log_info("Rotating...");
 169
 170         if (s->runtime_journal) {
 171                 r = journal_file_rotate(&s->runtime_journal);
 172                 if (r < 0)
 173                         log_error("Failed to rotate %s: %s", s->runtime_journal->path, strerror(-r));
 174         }
 175
 176         if (s->system_journal) {
 177                 r = journal_file_rotate(&s->system_journal);
 178                 if (r < 0)
 179                         log_error("Failed to rotate %s: %s", s->system_journal->path, strerror(-r));
 180         }
 181
 182         HASHMAP_FOREACH_KEY(f, k, s->user_journals, i) {
 183                 r = journal_file_rotate(&f);
 184                 if (r < 0)
 185                         log_error("Failed to rotate %s: %s", f->path, strerror(-r));
 186                 else
 187                         hashmap_replace(s->user_journals, k, f);
 188         }
 189
 190         log_info("Vacuuming...");
 191
 192         r = sd_id128_get_machine(&machine);
 193         if (r < 0) {
 194                 log_error("Failed to get machine ID: %s", strerror(-r));
 195                 return;
 196         }
 197
 198         if (asprintf(&p, "/var/log/journal/%s", sd_id128_to_string(machine, ids)) < 0) {
 199                 log_error("Out of memory.");
 200                 return;
 201         }
 202
 203         r = journal_directory_vacuum(p, s->max_use, s->metrics.keep_free);
 204         if (r < 0 && r != -ENOENT)
 205                 log_error("Failed to vacuum %s: %s", p, strerror(-r));
 206         free(p);
 207
 208         if (asprintf(&p, "/run/log/journal/%s", ids) < 0) {
 209                 log_error("Out of memory.");
 210                 return;
 211         }
 212
 213         r = journal_directory_vacuum(p, s->max_use, s->metrics.keep_free);
 214         if (r < 0 && r != -ENOENT)
 215                 log_error("Failed to vacuum %s: %s", p, strerror(-r));
 216         free(p);
 217 }
 218
 219 static void dispatch_message(Server *s, struct iovec *iovec, unsigned n, unsigned m, struct ucred *ucred, struct timeval *tv) {
 220         char *pid = NULL, *uid = NULL, *gid = NULL,
 221                 *source_time = NULL, *boot_id = NULL, *machine_id = NULL,
 222                 *comm = NULL, *cmdline = NULL, *hostname = NULL,
 223                 *audit_session = NULL, *audit_loginuid = NULL,
 224                 *exe = NULL, *cgroup = NULL;
 225
 226         char idbuf[33];
 227         sd_id128_t id;
 228         int r;
 229         char *t;
 230         uid_t loginuid = 0, realuid = 0;
 231         JournalFile *f;
 232         bool vacuumed = false;
 233
 234         assert(s);
 235         assert(iovec || n == 0);
 236
 237         if (n == 0)
 238                 return;
 239
 240         assert(n + 13 <= m);
 241
 242         if (ucred) {
 243                 uint32_t session;
 244                 char *path;
 245
 246                 realuid = ucred->uid;
 247
 248                 if (asprintf(&pid, "_PID=%lu", (unsigned long) ucred->pid) >= 0)
 249                         IOVEC_SET_STRING(iovec[n++], pid);
 250
 251                 if (asprintf(&uid, "_UID=%lu", (unsigned long) ucred->uid) >= 0)
 252                         IOVEC_SET_STRING(iovec[n++], uid);
 253
 254                 if (asprintf(&gid, "_GID=%lu", (unsigned long) ucred->gid) >= 0)
 255                         IOVEC_SET_STRING(iovec[n++], gid);
 256
 257                 r = get_process_comm(ucred->pid, &t);
 258                 if (r >= 0) {
 259                         comm = strappend("_COMM=", t);
 260                         if (comm)
 261                                 IOVEC_SET_STRING(iovec[n++], comm);
 262                         free(t);
 263                 }
 264
 265                 r = get_process_exe(ucred->pid, &t);
 266                 if (r >= 0) {
 267                         exe = strappend("_EXE=", t);
 268                         if (comm)
 269                                 IOVEC_SET_STRING(iovec[n++], exe);
 270                         free(t);
 271                 }
 272
 273                 r = get_process_cmdline(ucred->pid, LINE_MAX, false, &t);
 274                 if (r >= 0) {
 275                         cmdline = strappend("_CMDLINE=", t);
 276                         if (cmdline)
 277                                 IOVEC_SET_STRING(iovec[n++], cmdline);
 278                         free(t);
 279                 }
 280
 281                 r = audit_session_from_pid(ucred->pid, &session);
 282                 if (r >= 0)
 283                         if (asprintf(&audit_session, "_AUDIT_SESSION=%lu", (unsigned long) session) >= 0)
 284                                 IOVEC_SET_STRING(iovec[n++], audit_session);
 285
 286                 r = audit_loginuid_from_pid(ucred->pid, &loginuid);
 287                 if (r >= 0)
 288                         if (asprintf(&audit_loginuid, "_AUDIT_LOGINUID=%lu", (unsigned long) loginuid) >= 0)
 289                                 IOVEC_SET_STRING(iovec[n++], audit_loginuid);
 290
 291                 r = cg_get_by_pid(SYSTEMD_CGROUP_CONTROLLER, ucred->pid, &path);
 292                 if (r >= 0) {
 293                         cgroup = strappend("_SYSTEMD_CGROUP=", path);
 294                         if (cgroup)
 295                                 IOVEC_SET_STRING(iovec[n++], cgroup);
 296                         free(path);
 297                 }
 298         }
 299
 300         if (tv) {
 301                 if (asprintf(&source_time, "_SOURCE_REALTIME_TIMESTAMP=%llu",
 302                              (unsigned long long) timeval_load(tv)) >= 0)
 303                         IOVEC_SET_STRING(iovec[n++], source_time);
 304         }
 305
 306         /* Note that strictly speaking storing the boot id here is
 307          * redundant since the entry includes this in-line
 308          * anyway. However, we need this indexed, too. */
 309         r = sd_id128_get_boot(&id);
 310         if (r >= 0)
 311                 if (asprintf(&boot_id, "_BOOT_ID=%s", sd_id128_to_string(id, idbuf)) >= 0)
 312                         IOVEC_SET_STRING(iovec[n++], boot_id);
 313
 314         r = sd_id128_get_machine(&id);
 315         if (r >= 0)
 316                 if (asprintf(&machine_id, "_MACHINE_ID=%s", sd_id128_to_string(id, idbuf)) >= 0)
 317                         IOVEC_SET_STRING(iovec[n++], machine_id);
 318
 319         t = gethostname_malloc();
 320         if (t) {
 321                 hostname = strappend("_HOSTNAME=", t);
 322                 if (hostname)
 323                         IOVEC_SET_STRING(iovec[n++], hostname);
 324                 free(t);
 325         }
 326
 327         assert(n <= m);
 328
 329 retry:
 330         f = find_journal(s, realuid == 0 ? 0 : loginuid);
 331         if (!f)
 332                 log_warning("Dropping message, as we can't find a place to store the data.");
 333         else {
 334                 r = journal_file_append_entry(f, NULL, iovec, n, &s->seqnum, NULL, NULL);
 335
 336                 if (r == -E2BIG && !vacuumed) {
 337                         log_info("Allocation limit reached.");
 338
 339                         server_vacuum(s);
 340                         vacuumed = true;
 341
 342                         log_info("Retrying write.");
 343                         goto retry;
 344                 }
 345
 346                 if (r < 0)
 347                         log_error("Failed to write entry, ignoring: %s", strerror(-r));
 348         }
 349
 350         free(pid);
 351         free(uid);
 352         free(gid);
 353         free(comm);
 354         free(exe);
 355         free(cmdline);
 356         free(source_time);
 357         free(boot_id);
 358         free(machine_id);
 359         free(hostname);
 360         free(audit_session);
 361         free(audit_loginuid);
 362         free(cgroup);
 363
 364 }
 365
 366 static void process_syslog_message(Server *s, const char *buf, struct ucred *ucred, struct timeval *tv) {
 367         char *message = NULL, *syslog_priority = NULL, *syslog_facility = NULL;
 368         struct iovec iovec[16];
 369         unsigned n = 0;
 370         int priority = LOG_USER | LOG_INFO;
 371
 372         assert(s);
 373         assert(buf);
 374
 375         parse_syslog_priority((char**) &buf, &priority);
 376         skip_syslog_date((char**) &buf);
 377
 378         if (asprintf(&syslog_priority, "PRIORITY=%i", priority & LOG_PRIMASK) >= 0)
 379                 IOVEC_SET_STRING(iovec[n++], syslog_priority);
 380
 381         if (asprintf(&syslog_facility, "SYSLOG_FACILITY=%i", LOG_FAC(priority)) >= 0)
 382                 IOVEC_SET_STRING(iovec[n++], syslog_facility);
 383
 384         message = strappend("MESSAGE=", buf);
 385         if (message)
 386                 IOVEC_SET_STRING(iovec[n++], message);
 387
 388         dispatch_message(s, iovec, n, ELEMENTSOF(iovec), ucred, tv);
 389
 390         free(message);
 391         free(syslog_facility);
 392         free(syslog_priority);
 393 }
 394
 395 static void process_native_message(Server *s, const void *buffer, size_t buffer_size, struct ucred *ucred, struct timeval *tv) {
 396         struct iovec *iovec = NULL;
 397         unsigned n = 0, m = 0, j;
 398         const char *p;
 399         size_t remaining;
 400
 401         assert(s);
 402         assert(buffer || n == 0);
 403
 404         p = buffer;
 405         remaining = buffer_size;
 406
 407         while (remaining > 0) {
 408                 const char *e, *q;
 409
 410                 e = memchr(p, '\n', remaining);
 411
 412                 if (!e) {
 413                         /* Trailing noise, let's ignore it, and flush what we collected */
 414                         log_debug("Received message with trailing noise, ignoring.");
 415                         break;
 416                 }
 417
 418                 if (e == p) {
 419                         /* Entry separator */
 420                         dispatch_message(s, iovec, n, m, ucred, tv);
 421                         n = 0;
 422
 423                         p++;
 424                         remaining--;
 425                         continue;
 426                 }
 427
 428                 if (*p == '.') {
 429                         /* Control command, ignore for now */
 430                         remaining -= (e - p) + 1;
 431                         p = e + 1;
 432                         continue;
 433                 }
 434
 435                 /* A property follows */
 436
 437                 if (n+13 >= m) {
 438                         struct iovec *c;
 439                         unsigned u;
 440
 441                         u = MAX((n+13U) * 2U, 4U);
 442                         c = realloc(iovec, u * sizeof(struct iovec));
 443                         if (!c) {
 444                                 log_error("Out of memory");
 445                                 break;
 446                         }
 447
 448                         iovec = c;
 449                         m = u;
 450                 }
 451
 452                 q = memchr(p, '=', e - p);
 453                 if (q) {
 454                         if (p[0] != '_') {
 455                                 /* If the field name starts with an
 456                                  * underscore, skip the variable,
 457                                  * since that indidates a trusted
 458                                  * field */
 459                                 iovec[n].iov_base = (char*) p;
 460                                 iovec[n].iov_len = e - p;
 461                                 n++;
 462                         }
 463
 464                         remaining -= (e - p) + 1;
 465                         p = e + 1;
 466                         continue;
 467                 } else {
 468                         uint64_t l;
 469                         char *k;
 470
 471                         if (remaining < e - p + 1 + sizeof(uint64_t) + 1) {
 472                                 log_debug("Failed to parse message, ignoring.");
 473                                 break;
 474                         }
 475
 476                         memcpy(&l, e + 1, sizeof(uint64_t));
 477                         l = le64toh(l);
 478
 479                         if (remaining < e - p + 1 + sizeof(uint64_t) + l + 1 ||
 480                             e[1+sizeof(uint64_t)+l] != '\n') {
 481                                 log_debug("Failed to parse message, ignoring.");
 482                                 break;
 483                         }
 484
 485                         k = malloc((e - p) + 1 + l);
 486                         if (!k) {
 487                                 log_error("Out of memory");
 488                                 break;
 489                         }
 490
 491                         memcpy(k, p, e - p);
 492                         k[e - p] = '=';
 493                         memcpy(k + (e - p) + 1, e + 1 + sizeof(uint64_t), l);
 494
 495                         if (k[0] != '_') {
 496                                 iovec[n].iov_base = k;
 497                                 iovec[n].iov_len = (e - p) + 1 + l;
 498                                 n++;
 499                         } else
 500                                 free(k);
 501
 502                         remaining -= (e - p) + 1 + sizeof(uint64_t) + l + 1;
 503                         p = e + 1 + sizeof(uint64_t) + l + 1;
 504                 }
 505         }
 506
 507         dispatch_message(s, iovec, n, m, ucred, tv);
 508
 509         for (j = 0; j < n; j++)
 510                 if (iovec[j].iov_base < buffer ||
 511                     (const uint8_t*) iovec[j].iov_base >= (const uint8_t*) buffer + buffer_size)
 512                         free(iovec[j].iov_base);
 513 }
 514
 515 static int process_event(Server *s, struct epoll_event *ev) {
 516         assert(s);
 517
 518         if (ev->events != EPOLLIN) {
 519                 log_info("Got invalid event from epoll.");
 520                 return -EIO;
 521         }
 522
 523         if (ev->data.fd == s->signal_fd) {
 524                 struct signalfd_siginfo sfsi;
 525                 ssize_t n;
 526
 527                 n = read(s->signal_fd, &sfsi, sizeof(sfsi));
 528                 if (n != sizeof(sfsi)) {
 529
 530                         if (n >= 0)
 531                                 return -EIO;
 532
 533                         if (errno == EINTR || errno == EAGAIN)
 534                                 return 0;
 535
 536                         return -errno;
 537                 }
 538
 539                 log_debug("Received SIG%s", signal_to_string(sfsi.ssi_signo));
 540                 return 0;
 541
 542         }
 543
 544         if (ev->data.fd == s->native_fd ||
 545             ev->data.fd == s->syslog_fd) {
 546                 for (;;) {
 547                         struct msghdr msghdr;
 548                         struct iovec iovec;
 549                         struct ucred *ucred = NULL;
 550                         struct timeval *tv = NULL;
 551                         struct cmsghdr *cmsg;
 552                         union {
 553                                 struct cmsghdr cmsghdr;
 554                                 uint8_t buf[CMSG_SPACE(sizeof(struct ucred)) +
 555                                             CMSG_SPACE(sizeof(struct timeval))];
 556                         } control;
 557                         ssize_t n;
 558                         int v;
 559
 560                         if (ioctl(ev->data.fd, SIOCINQ, &v) < 0) {
 561                                 log_error("SIOCINQ failed: %m");
 562                                 return -errno;
 563                         }
 564
 565                         if (v <= 0)
 566                                 return 1;
 567
 568                         if (s->buffer_size < (size_t) v) {
 569                                 void *b;
 570                                 size_t l;
 571
 572                                 l = MAX(LINE_MAX + (size_t) v, s->buffer_size * 2);
 573                                 b = realloc(s->buffer, l+1);
 574
 575                                 if (!b) {
 576                                         log_error("Couldn't increase buffer.");
 577                                         return -ENOMEM;
 578                                 }
 579
 580                                 s->buffer_size = l;
 581                                 s->buffer = b;
 582                         }
 583
 584                         zero(iovec);
 585                         iovec.iov_base = s->buffer;
 586                         iovec.iov_len = s->buffer_size;
 587
 588                         zero(control);
 589                         zero(msghdr);
 590                         msghdr.msg_iov = &iovec;
 591                         msghdr.msg_iovlen = 1;
 592                         msghdr.msg_control = &control;
 593                         msghdr.msg_controllen = sizeof(control);
 594
 595                         n = recvmsg(ev->data.fd, &msghdr, MSG_DONTWAIT);
 596                         if (n < 0) {
 597
 598                                 if (errno == EINTR || errno == EAGAIN)
 599                                         return 1;
 600
 601                                 log_error("recvmsg() failed: %m");
 602                                 return -errno;
 603                         }
 604
 605                         for (cmsg = CMSG_FIRSTHDR(&msghdr); cmsg; cmsg = CMSG_NXTHDR(&msghdr, cmsg)) {
 606
 607                                 if (cmsg->cmsg_level == SOL_SOCKET &&
 608                                     cmsg->cmsg_type == SCM_CREDENTIALS &&
 609                                     cmsg->cmsg_len == CMSG_LEN(sizeof(struct ucred)))
 610                                         ucred = (struct ucred*) CMSG_DATA(cmsg);
 611                                 else if (cmsg->cmsg_level == SOL_SOCKET &&
 612                                          cmsg->cmsg_type == SO_TIMESTAMP &&
 613                                          cmsg->cmsg_len == CMSG_LEN(sizeof(struct timeval)))
 614                                         tv = (struct timeval*) CMSG_DATA(cmsg);
 615                         }
 616
 617                         if (ev->data.fd == s->syslog_fd) {
 618                                 char *e;
 619
 620                                 e = memchr(s->buffer, '\n', n);
 621                                 if (e)
 622                                         *e = 0;
 623                                 else
 624                                         s->buffer[n] = 0;
 625
 626                                 process_syslog_message(s, strstrip(s->buffer), ucred, tv);
 627                         } else
 628                                 process_native_message(s, s->buffer, n, ucred, tv);
 629                 }
 630
 631                 return 1;
 632         }
 633
 634         log_error("Unknown event.");
 635         return 0;
 636 }
 637
 638 static int system_journal_open(Server *s) {
 639         int r;
 640         char *fn;
 641         sd_id128_t machine;
 642         char ids[33];
 643
 644         r = sd_id128_get_machine(&machine);
 645         if (r < 0)
 646                 return r;
 647
 648         /* First try to create the machine path, but not the prefix */
 649         fn = strappend("/var/log/journal/", sd_id128_to_string(machine, ids));
 650         if (!fn)
 651                 return -ENOMEM;
 652         (void) mkdir(fn, 0755);
 653         free(fn);
 654
 655         /* The create the system journal file */
 656         fn = join("/var/log/journal/", ids, "/system.journal", NULL);
 657         if (!fn)
 658                 return -ENOMEM;
 659
 660         r = journal_file_open(fn, O_RDWR|O_CREAT, 0640, NULL, &s->system_journal);
 661         free(fn);
 662
 663         if (r >= 0) {
 664                 fix_perms(s->system_journal, 0);
 665                 return r;
 666         }
 667
 668         if (r < 0 && r != -ENOENT) {
 669                 log_error("Failed to open system journal: %s", strerror(-r));
 670                 return r;
 671         }
 672
 673         /* /var didn't work, so try /run, but this time we
 674          * create the prefix too */
 675         fn = join("/run/log/journal/", ids, "/system.journal", NULL);
 676         if (!fn)
 677                 return -ENOMEM;
 678
 679         (void) mkdir_parents(fn, 0755);
 680         r = journal_file_open(fn, O_RDWR|O_CREAT, 0640, NULL, &s->runtime_journal);
 681         free(fn);
 682
 683         if (r < 0) {
 684                 log_error("Failed to open runtime journal: %s", strerror(-r));
 685                 return r;
 686         }
 687
 688         fix_perms(s->runtime_journal, 0);
 689         return r;
 690 }
 691
 692 static int open_syslog_socket(Server *s) {
 693         union sockaddr_union sa;
 694         int one, r;
 695
 696         assert(s);
 697
 698         if (s->syslog_fd < 0) {
 699
 700                 s->syslog_fd = socket(AF_UNIX, SOCK_DGRAM|SOCK_CLOEXEC, 0);
 701                 if (s->syslog_fd < 0) {
 702                         log_error("socket() failed: %m");
 703                         return -errno;
 704                 }
 705
 706                 zero(sa);
 707                 sa.un.sun_family = AF_UNIX;
 708                 strncpy(sa.un.sun_path, "/run/systemd/syslog", sizeof(sa.un.sun_path));
 709
 710                 unlink(sa.un.sun_path);
 711
 712                 r = bind(s->syslog_fd, &sa.sa, offsetof(union sockaddr_union, un.sun_path) + strlen(sa.un.sun_path));
 713                 if (r < 0) {
 714                         log_error("bind() failed: %m");
 715                         return -errno;
 716                 }
 717
 718                 chmod(sa.un.sun_path, 0666);
 719         }
 720
 721         one = 1;
 722         r = setsockopt(s->syslog_fd, SOL_SOCKET, SO_PASSCRED, &one, sizeof(one));
 723         if (r < 0) {
 724                 log_error("SO_PASSCRED failed: %m");
 725                 return -errno;
 726         }
 727
 728         one = 1;
 729         r = setsockopt(s->syslog_fd, SOL_SOCKET, SO_TIMESTAMP, &one, sizeof(one));
 730         if (r < 0) {
 731                 log_error("SO_TIMESTAMP failed: %m");
 732                 return -errno;
 733         }
 734
 735         return 0;
 736 }
 737
 738 static int open_native_socket(Server*s) {
 739         union sockaddr_union sa;
 740         int one, r;
 741
 742         assert(s);
 743
 744         if (s->native_fd < 0) {
 745
 746                 s->native_fd = socket(AF_UNIX, SOCK_DGRAM|SOCK_CLOEXEC, 0);
 747                 if (s->native_fd < 0) {
 748                         log_error("socket() failed: %m");
 749                         return -errno;
 750                 }
 751
 752                 zero(sa);
 753                 sa.un.sun_family = AF_UNIX;
 754                 strncpy(sa.un.sun_path, "/run/systemd/journal", sizeof(sa.un.sun_path));
 755
 756                 unlink(sa.un.sun_path);
 757
 758                 r = bind(s->native_fd, &sa.sa, offsetof(union sockaddr_union, un.sun_path) + strlen(sa.un.sun_path));
 759                 if (r < 0) {
 760                         log_error("bind() failed: %m");
 761                         return -errno;
 762                 }
 763
 764                 chmod(sa.un.sun_path, 0666);
 765         }
 766
 767         one = 1;
 768         r = setsockopt(s->native_fd, SOL_SOCKET, SO_PASSCRED, &one, sizeof(one));
 769         if (r < 0) {
 770                 log_error("SO_PASSCRED failed: %m");
 771                 return -errno;
 772         }
 773
 774         one = 1;
 775         r = setsockopt(s->native_fd, SOL_SOCKET, SO_TIMESTAMP, &one, sizeof(one));
 776         if (r < 0) {
 777                 log_error("SO_TIMESTAMP failed: %m");
 778                 return -errno;
 779         }
 780
 781         return 0;
 782 }
 783
 784 static int server_init(Server *s) {
 785         int n, r, fd;
 786         struct epoll_event ev;
 787         sigset_t mask;
 788
 789         assert(s);
 790
 791         zero(*s);
 792         s->syslog_fd = s->native_fd = s->signal_fd = -1;
 793         s->metrics.max_size = DEFAULT_MAX_SIZE;
 794         s->metrics.min_size = DEFAULT_MIN_SIZE;
 795         s->metrics.keep_free = DEFAULT_KEEP_FREE;
 796         s->max_use = DEFAULT_MAX_USE;
 797
 798         s->epoll_fd = epoll_create1(EPOLL_CLOEXEC);
 799         if (s->epoll_fd < 0) {
 800                 log_error("Failed to create epoll object: %m");
 801                 return -errno;
 802         }
 803
 804         n = sd_listen_fds(true);
 805         if (n < 0) {
 806                 log_error("Failed to read listening file descriptors from environment: %s", strerror(-n));
 807                 return n;
 808         }
 809
 810         for (fd = SD_LISTEN_FDS_START; fd < SD_LISTEN_FDS_START + n; fd++) {
 811
 812                 if (sd_is_socket_unix(fd, SOCK_DGRAM, -1, "/dev/log", 0) > 0) {
 813
 814                         if (s->syslog_fd >= 0) {
 815                                 log_error("Too many /dev/log sockets passed.");
 816                                 return -EINVAL;
 817                         }
 818
 819                         s->syslog_fd = fd;
 820
 821                 } else if (sd_is_socket(fd, AF_UNIX, SOCK_DGRAM, -1) > 0) {
 822
 823                         if (s->native_fd >= 0) {
 824                                 log_error("Too many native sockets passed.");
 825                                 return -EINVAL;
 826                         }
 827
 828                         s->native_fd = fd;
 829                 } else {
 830                         log_error("Unknown socket passed.");
 831                         return -EINVAL;
 832                 }
 833         }
 834
 835         r = open_syslog_socket(s);
 836         if (r < 0)
 837                 return r;
 838
 839         zero(ev);
 840         ev.events = EPOLLIN;
 841         ev.data.fd = s->syslog_fd;
 842         if (epoll_ctl(s->epoll_fd, EPOLL_CTL_ADD, s->syslog_fd, &ev) < 0) {
 843                 log_error("Failed to add syslog server fd to epoll object: %m");
 844                 return -errno;
 845         }
 846
 847         r = open_native_socket(s);
 848         if (r < 0)
 849                 return r;
 850
 851         zero(ev);
 852         ev.events = EPOLLIN;
 853         ev.data.fd = s->native_fd;
 854         if (epoll_ctl(s->epoll_fd, EPOLL_CTL_ADD, s->native_fd, &ev) < 0) {
 855                 log_error("Failed to add native server fd to epoll object: %m");
 856                 return -errno;
 857         }
 858
 859         s->user_journals = hashmap_new(trivial_hash_func, trivial_compare_func);
 860         if (!s->user_journals) {
 861                 log_error("Out of memory.");
 862                 return -ENOMEM;
 863         }
 864
 865         r = system_journal_open(s);
 866         if (r < 0)
 867                 return r;
 868
 869         assert_se(sigemptyset(&mask) == 0);
 870         sigset_add_many(&mask, SIGINT, SIGTERM, -1);
 871         assert_se(sigprocmask(SIG_SETMASK, &mask, NULL) == 0);
 872
 873         s->signal_fd = signalfd(-1, &mask, SFD_NONBLOCK|SFD_CLOEXEC);
 874         if (s->signal_fd < 0) {
 875                 log_error("signalfd(): %m");
 876                 return -errno;
 877         }
 878
 879         zero(ev);
 880         ev.events = EPOLLIN;
 881         ev.data.fd = s->signal_fd;
 882
 883         if (epoll_ctl(s->epoll_fd, EPOLL_CTL_ADD, s->signal_fd, &ev) < 0) {
 884                 log_error("epoll_ctl(): %m");
 885                 return -errno;
 886         }
 887
 888         return 0;
 889 }
 890
 891 static void server_done(Server *s) {
 892         JournalFile *f;
 893         assert(s);
 894
 895         if (s->system_journal)
 896                 journal_file_close(s->system_journal);
 897
 898         if (s->runtime_journal)
 899                 journal_file_close(s->runtime_journal);
 900
 901         while ((f = hashmap_steal_first(s->user_journals)))
 902                 journal_file_close(f);
 903
 904         hashmap_free(s->user_journals);
 905
 906         if (s->epoll_fd >= 0)
 907                 close_nointr_nofail(s->epoll_fd);
 908
 909         if (s->signal_fd >= 0)
 910                 close_nointr_nofail(s->signal_fd);
 911
 912         if (s->syslog_fd >= 0)
 913                 close_nointr_nofail(s->syslog_fd);
 914
 915         if (s->native_fd >= 0)
 916                 close_nointr_nofail(s->native_fd);
 917 }
 918
 919 int main(int argc, char *argv[]) {
 920         Server server;
 921         int r;
 922
 923         /* if (getppid() != 1) { */
 924         /*         log_error("This program should be invoked by init only."); */
 925         /*         return EXIT_FAILURE; */
 926         /* } */
 927
 928         if (argc > 1) {
 929                 log_error("This program does not take arguments.");
 930                 return EXIT_FAILURE;
 931         }
 932
 933         log_set_target(LOG_TARGET_CONSOLE);
 934         log_parse_environment();
 935         log_open();
 936
 937         umask(0022);
 938
 939         r = server_init(&server);
 940         if (r < 0)
 941                 goto finish;
 942
 943         log_debug("systemd-journald running as pid %lu", (unsigned long) getpid());
 944
 945         sd_notify(false,
 946                   "READY=1\n"
 947                   "STATUS=Processing messages...");
 948
 949         for (;;) {
 950                 struct epoll_event event;
 951
 952                 r = epoll_wait(server.epoll_fd, &event, 1, -1);
 953                 if (r < 0) {
 954
 955                         if (errno == EINTR)
 956                                 continue;
 957
 958                         log_error("epoll_wait() failed: %m");
 959                         r = -errno;
 960                         goto finish;
 961                 } else if (r == 0)
 962                         break;
 963
 964                 r = process_event(&server, &event);
 965                 if (r < 0)
 966                         goto finish;
 967                 else if (r == 0)
 968                         break;
 969         }
 970
 971 finish:
 972         sd_notify(false,
 973                   "STATUS=Shutting down...");
 974
 975         server_done(&server);
 976
 977         return r < 0 ? EXIT_FAILURE : EXIT_SUCCESS;
 978 }