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         bool compress;
  61 } Server;
  62
  63 static void fix_perms(JournalFile *f, uid_t uid) {
  64         acl_t acl;
  65         acl_entry_t entry;
  66         acl_permset_t permset;
  67         int r;
  68
  69         assert(f);
  70
  71         r = fchmod_and_fchown(f->fd, 0640, 0, 0);
  72         if (r < 0)
  73                 log_warning("Failed to fix access mode/rights on %s, ignoring: %s", f->path, strerror(-r));
  74
  75         if (uid <= 0)
  76                 return;
  77
  78         acl = acl_get_fd(f->fd);
  79         if (!acl) {
  80                 log_warning("Failed to read ACL on %s, ignoring: %m", f->path);
  81                 return;
  82         }
  83
  84         r = acl_find_uid(acl, uid, &entry);
  85         if (r <= 0) {
  86
  87                 if (acl_create_entry(&acl, &entry) < 0 ||
  88                     acl_set_tag_type(entry, ACL_USER) < 0 ||
  89                     acl_set_qualifier(entry, &uid) < 0) {
  90                         log_warning("Failed to patch ACL on %s, ignoring: %m", f->path);
  91                         goto finish;
  92                 }
  93         }
  94
  95         if (acl_get_permset(entry, &permset) < 0 ||
  96             acl_add_perm(permset, ACL_READ) < 0 ||
  97             acl_calc_mask(&acl) < 0) {
  98                 log_warning("Failed to patch ACL on %s, ignoring: %m", f->path);
  99                 goto finish;
 100         }
 101
 102         if (acl_set_fd(f->fd, acl) < 0)
 103                 log_warning("Failed to set ACL on %s, ignoring: %m", f->path);
 104
 105 finish:
 106         acl_free(acl);
 107 }
 108
 109 static JournalFile* find_journal(Server *s, uid_t uid) {
 110         char *p;
 111         int r;
 112         JournalFile *f;
 113         char ids[33];
 114         sd_id128_t machine;
 115
 116         assert(s);
 117
 118         /* We split up user logs only on /var, not on /run */
 119         if (!s->system_journal)
 120                 return s->runtime_journal;
 121
 122         if (uid <= 0)
 123                 return s->system_journal;
 124
 125         r = sd_id128_get_machine(&machine);
 126         if (r < 0)
 127                 return s->system_journal;
 128
 129         f = hashmap_get(s->user_journals, UINT32_TO_PTR(uid));
 130         if (f)
 131                 return f;
 132
 133         if (asprintf(&p, "/var/log/journal/%s/user-%lu.journal", sd_id128_to_string(machine, ids), (unsigned long) uid) < 0)
 134                 return s->system_journal;
 135
 136         while (hashmap_size(s->user_journals) >= USER_JOURNALS_MAX) {
 137                 /* Too many open? Then let's close one */
 138                 f = hashmap_steal_first(s->user_journals);
 139                 assert(f);
 140                 journal_file_close(f);
 141         }
 142
 143         r = journal_file_open(p, O_RDWR|O_CREAT, 0640, s->system_journal, &f);
 144         free(p);
 145
 146         if (r < 0)
 147                 return s->system_journal;
 148
 149         fix_perms(f, uid);
 150         f->metrics = s->metrics;
 151         f->compress = s->compress;
 152
 153         r = hashmap_put(s->user_journals, UINT32_TO_PTR(uid), f);
 154         if (r < 0) {
 155                 journal_file_close(f);
 156                 return s->system_journal;
 157         }
 158
 159         return f;
 160 }
 161
 162 static void server_vacuum(Server *s) {
 163         Iterator i;
 164         void *k;
 165         char *p;
 166         char ids[33];
 167         sd_id128_t machine;
 168         int r;
 169         JournalFile *f;
 170
 171         log_info("Rotating...");
 172
 173         if (s->runtime_journal) {
 174                 r = journal_file_rotate(&s->runtime_journal);
 175                 if (r < 0)
 176                         log_error("Failed to rotate %s: %s", s->runtime_journal->path, strerror(-r));
 177         }
 178
 179         if (s->system_journal) {
 180                 r = journal_file_rotate(&s->system_journal);
 181                 if (r < 0)
 182                         log_error("Failed to rotate %s: %s", s->system_journal->path, strerror(-r));
 183         }
 184
 185         HASHMAP_FOREACH_KEY(f, k, s->user_journals, i) {
 186                 r = journal_file_rotate(&f);
 187                 if (r < 0)
 188                         log_error("Failed to rotate %s: %s", f->path, strerror(-r));
 189                 else
 190                         hashmap_replace(s->user_journals, k, f);
 191         }
 192
 193         log_info("Vacuuming...");
 194
 195         r = sd_id128_get_machine(&machine);
 196         if (r < 0) {
 197                 log_error("Failed to get machine ID: %s", strerror(-r));
 198                 return;
 199         }
 200
 201         if (asprintf(&p, "/var/log/journal/%s", sd_id128_to_string(machine, ids)) < 0) {
 202                 log_error("Out of memory.");
 203                 return;
 204         }
 205
 206         r = journal_directory_vacuum(p, s->max_use, s->metrics.keep_free);
 207         if (r < 0 && r != -ENOENT)
 208                 log_error("Failed to vacuum %s: %s", p, strerror(-r));
 209         free(p);
 210
 211         if (asprintf(&p, "/run/log/journal/%s", ids) < 0) {
 212                 log_error("Out of memory.");
 213                 return;
 214         }
 215
 216         r = journal_directory_vacuum(p, s->max_use, s->metrics.keep_free);
 217         if (r < 0 && r != -ENOENT)
 218                 log_error("Failed to vacuum %s: %s", p, strerror(-r));
 219         free(p);
 220 }
 221
 222 static void dispatch_message(Server *s, struct iovec *iovec, unsigned n, unsigned m, struct ucred *ucred, struct timeval *tv) {
 223         char *pid = NULL, *uid = NULL, *gid = NULL,
 224                 *source_time = NULL, *boot_id = NULL, *machine_id = NULL,
 225                 *comm = NULL, *cmdline = NULL, *hostname = NULL,
 226                 *audit_session = NULL, *audit_loginuid = NULL,
 227                 *exe = NULL, *cgroup = NULL;
 228
 229         char idbuf[33];
 230         sd_id128_t id;
 231         int r;
 232         char *t;
 233         uid_t loginuid = 0, realuid = 0;
 234         JournalFile *f;
 235         bool vacuumed = false;
 236
 237         assert(s);
 238         assert(iovec || n == 0);
 239
 240         if (n == 0)
 241                 return;
 242
 243         assert(n + 13 <= m);
 244
 245         if (ucred) {
 246                 uint32_t session;
 247                 char *path;
 248
 249                 realuid = ucred->uid;
 250
 251                 if (asprintf(&pid, "_PID=%lu", (unsigned long) ucred->pid) >= 0)
 252                         IOVEC_SET_STRING(iovec[n++], pid);
 253
 254                 if (asprintf(&uid, "_UID=%lu", (unsigned long) ucred->uid) >= 0)
 255                         IOVEC_SET_STRING(iovec[n++], uid);
 256
 257                 if (asprintf(&gid, "_GID=%lu", (unsigned long) ucred->gid) >= 0)
 258                         IOVEC_SET_STRING(iovec[n++], gid);
 259
 260                 r = get_process_comm(ucred->pid, &t);
 261                 if (r >= 0) {
 262                         comm = strappend("_COMM=", t);
 263                         if (comm)
 264                                 IOVEC_SET_STRING(iovec[n++], comm);
 265                         free(t);
 266                 }
 267
 268                 r = get_process_exe(ucred->pid, &t);
 269                 if (r >= 0) {
 270                         exe = strappend("_EXE=", t);
 271                         if (comm)
 272                                 IOVEC_SET_STRING(iovec[n++], exe);
 273                         free(t);
 274                 }
 275
 276                 r = get_process_cmdline(ucred->pid, LINE_MAX, false, &t);
 277                 if (r >= 0) {
 278                         cmdline = strappend("_CMDLINE=", t);
 279                         if (cmdline)
 280                                 IOVEC_SET_STRING(iovec[n++], cmdline);
 281                         free(t);
 282                 }
 283
 284                 r = audit_session_from_pid(ucred->pid, &session);
 285                 if (r >= 0)
 286                         if (asprintf(&audit_session, "_AUDIT_SESSION=%lu", (unsigned long) session) >= 0)
 287                                 IOVEC_SET_STRING(iovec[n++], audit_session);
 288
 289                 r = audit_loginuid_from_pid(ucred->pid, &loginuid);
 290                 if (r >= 0)
 291                         if (asprintf(&audit_loginuid, "_AUDIT_LOGINUID=%lu", (unsigned long) loginuid) >= 0)
 292                                 IOVEC_SET_STRING(iovec[n++], audit_loginuid);
 293
 294                 r = cg_get_by_pid(SYSTEMD_CGROUP_CONTROLLER, ucred->pid, &path);
 295                 if (r >= 0) {
 296                         cgroup = strappend("_SYSTEMD_CGROUP=", path);
 297                         if (cgroup)
 298                                 IOVEC_SET_STRING(iovec[n++], cgroup);
 299                         free(path);
 300                 }
 301         }
 302
 303         if (tv) {
 304                 if (asprintf(&source_time, "_SOURCE_REALTIME_TIMESTAMP=%llu",
 305                              (unsigned long long) timeval_load(tv)) >= 0)
 306                         IOVEC_SET_STRING(iovec[n++], source_time);
 307         }
 308
 309         /* Note that strictly speaking storing the boot id here is
 310          * redundant since the entry includes this in-line
 311          * anyway. However, we need this indexed, too. */
 312         r = sd_id128_get_boot(&id);
 313         if (r >= 0)
 314                 if (asprintf(&boot_id, "_BOOT_ID=%s", sd_id128_to_string(id, idbuf)) >= 0)
 315                         IOVEC_SET_STRING(iovec[n++], boot_id);
 316
 317         r = sd_id128_get_machine(&id);
 318         if (r >= 0)
 319                 if (asprintf(&machine_id, "_MACHINE_ID=%s", sd_id128_to_string(id, idbuf)) >= 0)
 320                         IOVEC_SET_STRING(iovec[n++], machine_id);
 321
 322         t = gethostname_malloc();
 323         if (t) {
 324                 hostname = strappend("_HOSTNAME=", t);
 325                 if (hostname)
 326                         IOVEC_SET_STRING(iovec[n++], hostname);
 327                 free(t);
 328         }
 329
 330         assert(n <= m);
 331
 332 retry:
 333         f = find_journal(s, realuid == 0 ? 0 : loginuid);
 334         if (!f)
 335                 log_warning("Dropping message, as we can't find a place to store the data.");
 336         else {
 337                 r = journal_file_append_entry(f, NULL, iovec, n, &s->seqnum, NULL, NULL);
 338
 339                 if (r == -E2BIG && !vacuumed) {
 340                         log_info("Allocation limit reached.");
 341
 342                         server_vacuum(s);
 343                         vacuumed = true;
 344
 345                         log_info("Retrying write.");
 346                         goto retry;
 347                 }
 348
 349                 if (r < 0)
 350                         log_error("Failed to write entry, ignoring: %s", strerror(-r));
 351         }
 352
 353         free(pid);
 354         free(uid);
 355         free(gid);
 356         free(comm);
 357         free(exe);
 358         free(cmdline);
 359         free(source_time);
 360         free(boot_id);
 361         free(machine_id);
 362         free(hostname);
 363         free(audit_session);
 364         free(audit_loginuid);
 365         free(cgroup);
 366
 367 }
 368
 369 static void process_syslog_message(Server *s, const char *buf, struct ucred *ucred, struct timeval *tv) {
 370         char *message = NULL, *syslog_priority = NULL, *syslog_facility = NULL;
 371         struct iovec iovec[16];
 372         unsigned n = 0;
 373         int priority = LOG_USER | LOG_INFO;
 374
 375         assert(s);
 376         assert(buf);
 377
 378         parse_syslog_priority((char**) &buf, &priority);
 379         skip_syslog_date((char**) &buf);
 380
 381         if (asprintf(&syslog_priority, "PRIORITY=%i", priority & LOG_PRIMASK) >= 0)
 382                 IOVEC_SET_STRING(iovec[n++], syslog_priority);
 383
 384         if (asprintf(&syslog_facility, "SYSLOG_FACILITY=%i", LOG_FAC(priority)) >= 0)
 385                 IOVEC_SET_STRING(iovec[n++], syslog_facility);
 386
 387         message = strappend("MESSAGE=", buf);
 388         if (message)
 389                 IOVEC_SET_STRING(iovec[n++], message);
 390
 391         dispatch_message(s, iovec, n, ELEMENTSOF(iovec), ucred, tv);
 392
 393         free(message);
 394         free(syslog_facility);
 395         free(syslog_priority);
 396 }
 397
 398 static bool valid_user_field(const char *p, size_t l) {
 399         const char *a;
 400
 401         /* We kinda enforce POSIX syntax recommendations for
 402            environment variables here, but make a couple of additional
 403            requirements.
 404
 405            http://pubs.opengroup.org/onlinepubs/000095399/basedefs/xbd_chap08.html */
 406
 407         /* No empty field names */
 408         if (l <= 0)
 409                 return false;
 410
 411         /* Don't allow names longer than 64 chars */
 412         if (l > 64)
 413                 return false;
 414
 415         /* Variables starting with an underscore are protected */
 416         if (p[0] == '_')
 417                 return false;
 418
 419         /* Don't allow digits as first character */
 420         if (p[0] >= '0' && p[0] <= '9')
 421                 return false;
 422
 423         /* Only allow A-Z0-9 and '_' */
 424         for (a = p; a < p + l; a++)
 425                 if (!((*a >= 'A' && *a <= 'Z') ||
 426                       (*a >= '0' && *a <= '9') ||
 427                       *a == '_'))
 428                         return false;
 429
 430         return true;
 431 }
 432
 433 static void process_native_message(Server *s, const void *buffer, size_t buffer_size, struct ucred *ucred, struct timeval *tv) {
 434         struct iovec *iovec = NULL;
 435         unsigned n = 0, m = 0, j;
 436         const char *p;
 437         size_t remaining;
 438
 439         assert(s);
 440         assert(buffer || n == 0);
 441
 442         p = buffer;
 443         remaining = buffer_size;
 444
 445         while (remaining > 0) {
 446                 const char *e, *q;
 447
 448                 e = memchr(p, '\n', remaining);
 449
 450                 if (!e) {
 451                         /* Trailing noise, let's ignore it, and flush what we collected */
 452                         log_debug("Received message with trailing noise, ignoring.");
 453                         break;
 454                 }
 455
 456                 if (e == p) {
 457                         /* Entry separator */
 458                         dispatch_message(s, iovec, n, m, ucred, tv);
 459                         n = 0;
 460
 461                         p++;
 462                         remaining--;
 463                         continue;
 464                 }
 465
 466                 if (*p == '.' || *p == '#') {
 467                         /* Ignore control commands for now, and
 468                          * comments too. */
 469                         remaining -= (e - p) + 1;
 470                         p = e + 1;
 471                         continue;
 472                 }
 473
 474                 /* A property follows */
 475
 476                 if (n+13 >= m) {
 477                         struct iovec *c;
 478                         unsigned u;
 479
 480                         u = MAX((n+13U) * 2U, 4U);
 481                         c = realloc(iovec, u * sizeof(struct iovec));
 482                         if (!c) {
 483                                 log_error("Out of memory");
 484                                 break;
 485                         }
 486
 487                         iovec = c;
 488                         m = u;
 489                 }
 490
 491                 q = memchr(p, '=', e - p);
 492                 if (q) {
 493                         if (valid_user_field(p, q - p)) {
 494                                 /* If the field name starts with an
 495                                  * underscore, skip the variable,
 496                                  * since that indidates a trusted
 497                                  * field */
 498                                 iovec[n].iov_base = (char*) p;
 499                                 iovec[n].iov_len = e - p;
 500                                 n++;
 501                         }
 502
 503                         remaining -= (e - p) + 1;
 504                         p = e + 1;
 505                         continue;
 506                 } else {
 507                         uint64_t l;
 508                         char *k;
 509
 510                         if (remaining < e - p + 1 + sizeof(uint64_t) + 1) {
 511                                 log_debug("Failed to parse message, ignoring.");
 512                                 break;
 513                         }
 514
 515                         memcpy(&l, e + 1, sizeof(uint64_t));
 516                         l = le64toh(l);
 517
 518                         if (remaining < e - p + 1 + sizeof(uint64_t) + l + 1 ||
 519                             e[1+sizeof(uint64_t)+l] != '\n') {
 520                                 log_debug("Failed to parse message, ignoring.");
 521                                 break;
 522                         }
 523
 524                         k = malloc((e - p) + 1 + l);
 525                         if (!k) {
 526                                 log_error("Out of memory");
 527                                 break;
 528                         }
 529
 530                         memcpy(k, p, e - p);
 531                         k[e - p] = '=';
 532                         memcpy(k + (e - p) + 1, e + 1 + sizeof(uint64_t), l);
 533
 534                         if (valid_user_field(p, e - p)) {
 535                                 iovec[n].iov_base = k;
 536                                 iovec[n].iov_len = (e - p) + 1 + l;
 537                                 n++;
 538                         } else
 539                                 free(k);
 540
 541                         remaining -= (e - p) + 1 + sizeof(uint64_t) + l + 1;
 542                         p = e + 1 + sizeof(uint64_t) + l + 1;
 543                 }
 544         }
 545
 546         dispatch_message(s, iovec, n, m, ucred, tv);
 547
 548         for (j = 0; j < n; j++)
 549                 if (iovec[j].iov_base < buffer ||
 550                     (const uint8_t*) iovec[j].iov_base >= (const uint8_t*) buffer + buffer_size)
 551                         free(iovec[j].iov_base);
 552 }
 553
 554 static int process_event(Server *s, struct epoll_event *ev) {
 555         assert(s);
 556
 557         if (ev->events != EPOLLIN) {
 558                 log_info("Got invalid event from epoll.");
 559                 return -EIO;
 560         }
 561
 562         if (ev->data.fd == s->signal_fd) {
 563                 struct signalfd_siginfo sfsi;
 564                 ssize_t n;
 565
 566                 n = read(s->signal_fd, &sfsi, sizeof(sfsi));
 567                 if (n != sizeof(sfsi)) {
 568
 569                         if (n >= 0)
 570                                 return -EIO;
 571
 572                         if (errno == EINTR || errno == EAGAIN)
 573                                 return 0;
 574
 575                         return -errno;
 576                 }
 577
 578                 log_debug("Received SIG%s", signal_to_string(sfsi.ssi_signo));
 579                 return 0;
 580
 581         }
 582
 583         if (ev->data.fd == s->native_fd ||
 584             ev->data.fd == s->syslog_fd) {
 585                 for (;;) {
 586                         struct msghdr msghdr;
 587                         struct iovec iovec;
 588                         struct ucred *ucred = NULL;
 589                         struct timeval *tv = NULL;
 590                         struct cmsghdr *cmsg;
 591                         union {
 592                                 struct cmsghdr cmsghdr;
 593                                 uint8_t buf[CMSG_SPACE(sizeof(struct ucred)) +
 594                                             CMSG_SPACE(sizeof(struct timeval))];
 595                         } control;
 596                         ssize_t n;
 597                         int v;
 598
 599                         if (ioctl(ev->data.fd, SIOCINQ, &v) < 0) {
 600                                 log_error("SIOCINQ failed: %m");
 601                                 return -errno;
 602                         }
 603
 604                         if (v <= 0)
 605                                 return 1;
 606
 607                         if (s->buffer_size < (size_t) v) {
 608                                 void *b;
 609                                 size_t l;
 610
 611                                 l = MAX(LINE_MAX + (size_t) v, s->buffer_size * 2);
 612                                 b = realloc(s->buffer, l+1);
 613
 614                                 if (!b) {
 615                                         log_error("Couldn't increase buffer.");
 616                                         return -ENOMEM;
 617                                 }
 618
 619                                 s->buffer_size = l;
 620                                 s->buffer = b;
 621                         }
 622
 623                         zero(iovec);
 624                         iovec.iov_base = s->buffer;
 625                         iovec.iov_len = s->buffer_size;
 626
 627                         zero(control);
 628                         zero(msghdr);
 629                         msghdr.msg_iov = &iovec;
 630                         msghdr.msg_iovlen = 1;
 631                         msghdr.msg_control = &control;
 632                         msghdr.msg_controllen = sizeof(control);
 633
 634                         n = recvmsg(ev->data.fd, &msghdr, MSG_DONTWAIT);
 635                         if (n < 0) {
 636
 637                                 if (errno == EINTR || errno == EAGAIN)
 638                                         return 1;
 639
 640                                 log_error("recvmsg() failed: %m");
 641                                 return -errno;
 642                         }
 643
 644                         for (cmsg = CMSG_FIRSTHDR(&msghdr); cmsg; cmsg = CMSG_NXTHDR(&msghdr, cmsg)) {
 645
 646                                 if (cmsg->cmsg_level == SOL_SOCKET &&
 647                                     cmsg->cmsg_type == SCM_CREDENTIALS &&
 648                                     cmsg->cmsg_len == CMSG_LEN(sizeof(struct ucred)))
 649                                         ucred = (struct ucred*) CMSG_DATA(cmsg);
 650                                 else if (cmsg->cmsg_level == SOL_SOCKET &&
 651                                          cmsg->cmsg_type == SO_TIMESTAMP &&
 652                                          cmsg->cmsg_len == CMSG_LEN(sizeof(struct timeval)))
 653                                         tv = (struct timeval*) CMSG_DATA(cmsg);
 654                         }
 655
 656                         if (ev->data.fd == s->syslog_fd) {
 657                                 char *e;
 658
 659                                 e = memchr(s->buffer, '\n', n);
 660                                 if (e)
 661                                         *e = 0;
 662                                 else
 663                                         s->buffer[n] = 0;
 664
 665                                 process_syslog_message(s, strstrip(s->buffer), ucred, tv);
 666                         } else
 667                                 process_native_message(s, s->buffer, n, ucred, tv);
 668                 }
 669
 670                 return 1;
 671         }
 672
 673         log_error("Unknown event.");
 674         return 0;
 675 }
 676
 677 static int system_journal_open(Server *s) {
 678         int r;
 679         char *fn;
 680         sd_id128_t machine;
 681         char ids[33];
 682
 683         r = sd_id128_get_machine(&machine);
 684         if (r < 0)
 685                 return r;
 686
 687         /* First try to create the machine path, but not the prefix */
 688         fn = strappend("/var/log/journal/", sd_id128_to_string(machine, ids));
 689         if (!fn)
 690                 return -ENOMEM;
 691         (void) mkdir(fn, 0755);
 692         free(fn);
 693
 694         /* The create the system journal file */
 695         fn = join("/var/log/journal/", ids, "/system.journal", NULL);
 696         if (!fn)
 697                 return -ENOMEM;
 698
 699         r = journal_file_open(fn, O_RDWR|O_CREAT, 0640, NULL, &s->system_journal);
 700         free(fn);
 701
 702         if (r >= 0) {
 703                 s->system_journal->metrics = s->metrics;
 704                 s->system_journal->compress = s->compress;
 705
 706                 fix_perms(s->system_journal, 0);
 707                 return r;
 708         }
 709
 710         if (r < 0 && r != -ENOENT) {
 711                 log_error("Failed to open system journal: %s", strerror(-r));
 712                 return r;
 713         }
 714
 715         /* /var didn't work, so try /run, but this time we
 716          * create the prefix too */
 717         fn = join("/run/log/journal/", ids, "/system.journal", NULL);
 718         if (!fn)
 719                 return -ENOMEM;
 720
 721         (void) mkdir_parents(fn, 0755);
 722         r = journal_file_open(fn, O_RDWR|O_CREAT, 0640, NULL, &s->runtime_journal);
 723         free(fn);
 724
 725         if (r < 0) {
 726                 log_error("Failed to open runtime journal: %s", strerror(-r));
 727                 return r;
 728         }
 729
 730         s->runtime_journal->metrics = s->metrics;
 731         s->runtime_journal->compress = s->compress;
 732
 733         fix_perms(s->runtime_journal, 0);
 734         return r;
 735 }
 736
 737 static int open_syslog_socket(Server *s) {
 738         union sockaddr_union sa;
 739         int one, r;
 740
 741         assert(s);
 742
 743         if (s->syslog_fd < 0) {
 744
 745                 s->syslog_fd = socket(AF_UNIX, SOCK_DGRAM|SOCK_CLOEXEC, 0);
 746                 if (s->syslog_fd < 0) {
 747                         log_error("socket() failed: %m");
 748                         return -errno;
 749                 }
 750
 751                 zero(sa);
 752                 sa.un.sun_family = AF_UNIX;
 753                 strncpy(sa.un.sun_path, "/run/systemd/syslog", sizeof(sa.un.sun_path));
 754
 755                 unlink(sa.un.sun_path);
 756
 757                 r = bind(s->syslog_fd, &sa.sa, offsetof(union sockaddr_union, un.sun_path) + strlen(sa.un.sun_path));
 758                 if (r < 0) {
 759                         log_error("bind() failed: %m");
 760                         return -errno;
 761                 }
 762
 763                 chmod(sa.un.sun_path, 0666);
 764         }
 765
 766         one = 1;
 767         r = setsockopt(s->syslog_fd, SOL_SOCKET, SO_PASSCRED, &one, sizeof(one));
 768         if (r < 0) {
 769                 log_error("SO_PASSCRED failed: %m");
 770                 return -errno;
 771         }
 772
 773         one = 1;
 774         r = setsockopt(s->syslog_fd, SOL_SOCKET, SO_TIMESTAMP, &one, sizeof(one));
 775         if (r < 0) {
 776                 log_error("SO_TIMESTAMP failed: %m");
 777                 return -errno;
 778         }
 779
 780         return 0;
 781 }
 782
 783 static int open_native_socket(Server*s) {
 784         union sockaddr_union sa;
 785         int one, r;
 786
 787         assert(s);
 788
 789         if (s->native_fd < 0) {
 790
 791                 s->native_fd = socket(AF_UNIX, SOCK_DGRAM|SOCK_CLOEXEC, 0);
 792                 if (s->native_fd < 0) {
 793                         log_error("socket() failed: %m");
 794                         return -errno;
 795                 }
 796
 797                 zero(sa);
 798                 sa.un.sun_family = AF_UNIX;
 799                 strncpy(sa.un.sun_path, "/run/systemd/journal", sizeof(sa.un.sun_path));
 800
 801                 unlink(sa.un.sun_path);
 802
 803                 r = bind(s->native_fd, &sa.sa, offsetof(union sockaddr_union, un.sun_path) + strlen(sa.un.sun_path));
 804                 if (r < 0) {
 805                         log_error("bind() failed: %m");
 806                         return -errno;
 807                 }
 808
 809                 chmod(sa.un.sun_path, 0666);
 810         }
 811
 812         one = 1;
 813         r = setsockopt(s->native_fd, SOL_SOCKET, SO_PASSCRED, &one, sizeof(one));
 814         if (r < 0) {
 815                 log_error("SO_PASSCRED failed: %m");
 816                 return -errno;
 817         }
 818
 819         one = 1;
 820         r = setsockopt(s->native_fd, SOL_SOCKET, SO_TIMESTAMP, &one, sizeof(one));
 821         if (r < 0) {
 822                 log_error("SO_TIMESTAMP failed: %m");
 823                 return -errno;
 824         }
 825
 826         return 0;
 827 }
 828
 829 static int server_init(Server *s) {
 830         int n, r, fd;
 831         struct epoll_event ev;
 832         sigset_t mask;
 833
 834         assert(s);
 835
 836         zero(*s);
 837         s->syslog_fd = s->native_fd = s->signal_fd = -1;
 838         s->metrics.max_size = DEFAULT_MAX_SIZE;
 839         s->metrics.min_size = DEFAULT_MIN_SIZE;
 840         s->metrics.keep_free = DEFAULT_KEEP_FREE;
 841         s->max_use = DEFAULT_MAX_USE;
 842         s->compress = true;
 843
 844         s->epoll_fd = epoll_create1(EPOLL_CLOEXEC);
 845         if (s->epoll_fd < 0) {
 846                 log_error("Failed to create epoll object: %m");
 847                 return -errno;
 848         }
 849
 850         n = sd_listen_fds(true);
 851         if (n < 0) {
 852                 log_error("Failed to read listening file descriptors from environment: %s", strerror(-n));
 853                 return n;
 854         }
 855
 856         for (fd = SD_LISTEN_FDS_START; fd < SD_LISTEN_FDS_START + n; fd++) {
 857
 858                 if (sd_is_socket_unix(fd, SOCK_DGRAM, -1, "/dev/log", 0) > 0) {
 859
 860                         if (s->syslog_fd >= 0) {
 861                                 log_error("Too many /dev/log sockets passed.");
 862                                 return -EINVAL;
 863                         }
 864
 865                         s->syslog_fd = fd;
 866
 867                 } else if (sd_is_socket(fd, AF_UNIX, SOCK_DGRAM, -1) > 0) {
 868
 869                         if (s->native_fd >= 0) {
 870                                 log_error("Too many native sockets passed.");
 871                                 return -EINVAL;
 872                         }
 873
 874                         s->native_fd = fd;
 875                 } else {
 876                         log_error("Unknown socket passed.");
 877                         return -EINVAL;
 878                 }
 879         }
 880
 881         r = open_syslog_socket(s);
 882         if (r < 0)
 883                 return r;
 884
 885         zero(ev);
 886         ev.events = EPOLLIN;
 887         ev.data.fd = s->syslog_fd;
 888         if (epoll_ctl(s->epoll_fd, EPOLL_CTL_ADD, s->syslog_fd, &ev) < 0) {
 889                 log_error("Failed to add syslog server fd to epoll object: %m");
 890                 return -errno;
 891         }
 892
 893         r = open_native_socket(s);
 894         if (r < 0)
 895                 return r;
 896
 897         zero(ev);
 898         ev.events = EPOLLIN;
 899         ev.data.fd = s->native_fd;
 900         if (epoll_ctl(s->epoll_fd, EPOLL_CTL_ADD, s->native_fd, &ev) < 0) {
 901                 log_error("Failed to add native server fd to epoll object: %m");
 902                 return -errno;
 903         }
 904
 905         s->user_journals = hashmap_new(trivial_hash_func, trivial_compare_func);
 906         if (!s->user_journals) {
 907                 log_error("Out of memory.");
 908                 return -ENOMEM;
 909         }
 910
 911         r = system_journal_open(s);
 912         if (r < 0)
 913                 return r;
 914
 915         assert_se(sigemptyset(&mask) == 0);
 916         sigset_add_many(&mask, SIGINT, SIGTERM, -1);
 917         assert_se(sigprocmask(SIG_SETMASK, &mask, NULL) == 0);
 918
 919         s->signal_fd = signalfd(-1, &mask, SFD_NONBLOCK|SFD_CLOEXEC);
 920         if (s->signal_fd < 0) {
 921                 log_error("signalfd(): %m");
 922                 return -errno;
 923         }
 924
 925         zero(ev);
 926         ev.events = EPOLLIN;
 927         ev.data.fd = s->signal_fd;
 928
 929         if (epoll_ctl(s->epoll_fd, EPOLL_CTL_ADD, s->signal_fd, &ev) < 0) {
 930                 log_error("epoll_ctl(): %m");
 931                 return -errno;
 932         }
 933
 934         return 0;
 935 }
 936
 937 static void server_done(Server *s) {
 938         JournalFile *f;
 939         assert(s);
 940
 941         if (s->system_journal)
 942                 journal_file_close(s->system_journal);
 943
 944         if (s->runtime_journal)
 945                 journal_file_close(s->runtime_journal);
 946
 947         while ((f = hashmap_steal_first(s->user_journals)))
 948                 journal_file_close(f);
 949
 950         hashmap_free(s->user_journals);
 951
 952         if (s->epoll_fd >= 0)
 953                 close_nointr_nofail(s->epoll_fd);
 954
 955         if (s->signal_fd >= 0)
 956                 close_nointr_nofail(s->signal_fd);
 957
 958         if (s->syslog_fd >= 0)
 959                 close_nointr_nofail(s->syslog_fd);
 960
 961         if (s->native_fd >= 0)
 962                 close_nointr_nofail(s->native_fd);
 963 }
 964
 965 int main(int argc, char *argv[]) {
 966         Server server;
 967         int r;
 968
 969         /* if (getppid() != 1) { */
 970         /*         log_error("This program should be invoked by init only."); */
 971         /*         return EXIT_FAILURE; */
 972         /* } */
 973
 974         if (argc > 1) {
 975                 log_error("This program does not take arguments.");
 976                 return EXIT_FAILURE;
 977         }
 978
 979         log_set_target(LOG_TARGET_CONSOLE);
 980         log_set_max_level(LOG_DEBUG);
 981         log_parse_environment();
 982         log_open();
 983
 984         umask(0022);
 985
 986         r = server_init(&server);
 987         if (r < 0)
 988                 goto finish;
 989
 990         log_debug("systemd-journald running as pid %lu", (unsigned long) getpid());
 991
 992         sd_notify(false,
 993                   "READY=1\n"
 994                   "STATUS=Processing messages...");
 995
 996         for (;;) {
 997                 struct epoll_event event;
 998
 999                 r = epoll_wait(server.epoll_fd, &event, 1, -1);
1000                 if (r < 0) {
1001
1002                         if (errno == EINTR)
1003                                 continue;
1004
1005                         log_error("epoll_wait() failed: %m");
1006                         r = -errno;
1007                         goto finish;
1008                 } else if (r == 0)
1009                         break;
1010
1011                 r = process_event(&server, &event);
1012                 if (r < 0)
1013                         goto finish;
1014                 else if (r == 0)
1015                         break;
1016         }
1017
1018 finish:
1019         sd_notify(false,
1020                   "STATUS=Shutting down...");
1021
1022         server_done(&server);
1023
1024         return r < 0 ? EXIT_FAILURE : EXIT_SUCCESS;
1025 }