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