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 <stddef.h>
  29 #include <sys/ioctl.h>
  30 #include <linux/sockios.h>
  31 #include <sys/statvfs.h>
  32
  33 #include <systemd/sd-journal.h>
  34 #include <systemd/sd-login.h>
  35 #include <systemd/sd-messages.h>
  36 #include <systemd/sd-daemon.h>
  37
  38 #include "mkdir.h"
  39 #include "hashmap.h"
  40 #include "journal-file.h"
  41 #include "socket-util.h"
  42 #include "cgroup-util.h"
  43 #include "list.h"
  44 #include "journal-rate-limit.h"
  45 #include "journal-internal.h"
  46 #include "conf-parser.h"
  47 #include "journald.h"
  48 #include "virt.h"
  49 #include "missing.h"
  50
  51 #ifdef HAVE_ACL
  52 #include <sys/acl.h>
  53 #include <acl/libacl.h>
  54 #include "acl-util.h"
  55 #endif
  56
  57 #ifdef HAVE_SELINUX
  58 #include <selinux/selinux.h>
  59 #endif
  60
  61 #define USER_JOURNALS_MAX 1024
  62 #define STDOUT_STREAMS_MAX 4096
  63
  64 #define DEFAULT_RATE_LIMIT_INTERVAL (10*USEC_PER_SEC)
  65 #define DEFAULT_RATE_LIMIT_BURST 200
  66
  67 #define RECHECK_AVAILABLE_SPACE_USEC (30*USEC_PER_SEC)
  68
  69 #define RECHECK_VAR_AVAILABLE_USEC (30*USEC_PER_SEC)
  70
  71 #define N_IOVEC_META_FIELDS 17
  72
  73 #define ENTRY_SIZE_MAX (1024*1024*32)
  74
  75 typedef enum StdoutStreamState {
  76         STDOUT_STREAM_IDENTIFIER,
  77         STDOUT_STREAM_PRIORITY,
  78         STDOUT_STREAM_LEVEL_PREFIX,
  79         STDOUT_STREAM_FORWARD_TO_SYSLOG,
  80         STDOUT_STREAM_FORWARD_TO_KMSG,
  81         STDOUT_STREAM_FORWARD_TO_CONSOLE,
  82         STDOUT_STREAM_RUNNING
  83 } StdoutStreamState;
  84
  85 struct StdoutStream {
  86         Server *server;
  87         StdoutStreamState state;
  88
  89         int fd;
  90
  91         struct ucred ucred;
  92 #ifdef HAVE_SELINUX
  93         security_context_t security_context;
  94 #endif
  95
  96         char *identifier;
  97         int priority;
  98         bool level_prefix:1;
  99         bool forward_to_syslog:1;
 100         bool forward_to_kmsg:1;
 101         bool forward_to_console:1;
 102
 103         char buffer[LINE_MAX+1];
 104         size_t length;
 105
 106         LIST_FIELDS(StdoutStream, stdout_stream);
 107 };
 108
 109 static int server_flush_to_var(Server *s);
 110
 111 static uint64_t available_space(Server *s) {
 112         char ids[33], *p;
 113         const char *f;
 114         sd_id128_t machine;
 115         struct statvfs ss;
 116         uint64_t sum = 0, avail = 0, ss_avail = 0;
 117         int r;
 118         DIR *d;
 119         usec_t ts;
 120         JournalMetrics *m;
 121
 122         ts = now(CLOCK_MONOTONIC);
 123
 124         if (s->cached_available_space_timestamp + RECHECK_AVAILABLE_SPACE_USEC > ts)
 125                 return s->cached_available_space;
 126
 127         r = sd_id128_get_machine(&machine);
 128         if (r < 0)
 129                 return 0;
 130
 131         if (s->system_journal) {
 132                 f = "/var/log/journal/";
 133                 m = &s->system_metrics;
 134         } else {
 135                 f = "/run/log/journal/";
 136                 m = &s->runtime_metrics;
 137         }
 138
 139         assert(m);
 140
 141         p = strappend(f, sd_id128_to_string(machine, ids));
 142         if (!p)
 143                 return 0;
 144
 145         d = opendir(p);
 146         free(p);
 147
 148         if (!d)
 149                 return 0;
 150
 151         if (fstatvfs(dirfd(d), &ss) < 0)
 152                 goto finish;
 153
 154         for (;;) {
 155                 struct stat st;
 156                 struct dirent buf, *de;
 157
 158                 r = readdir_r(d, &buf, &de);
 159                 if (r != 0)
 160                         break;
 161
 162                 if (!de)
 163                         break;
 164
 165                 if (!endswith(de->d_name, ".journal") &&
 166                     !endswith(de->d_name, ".journal~"))
 167                         continue;
 168
 169                 if (fstatat(dirfd(d), de->d_name, &st, AT_SYMLINK_NOFOLLOW) < 0)
 170                         continue;
 171
 172                 if (!S_ISREG(st.st_mode))
 173                         continue;
 174
 175                 sum += (uint64_t) st.st_blocks * 512UL;
 176         }
 177
 178         avail = sum >= m->max_use ? 0 : m->max_use - sum;
 179
 180         ss_avail = ss.f_bsize * ss.f_bavail;
 181
 182         ss_avail = ss_avail < m->keep_free ? 0 : ss_avail - m->keep_free;
 183
 184         if (ss_avail < avail)
 185                 avail = ss_avail;
 186
 187         s->cached_available_space = avail;
 188         s->cached_available_space_timestamp = ts;
 189
 190 finish:
 191         closedir(d);
 192
 193         return avail;
 194 }
 195
 196 static void server_read_file_gid(Server *s) {
 197         const char *adm = "adm";
 198         int r;
 199
 200         assert(s);
 201
 202         if (s->file_gid_valid)
 203                 return;
 204
 205         r = get_group_creds(&adm, &s->file_gid);
 206         if (r < 0)
 207                 log_warning("Failed to resolve 'adm' group: %s", strerror(-r));
 208
 209         /* if we couldn't read the gid, then it will be 0, but that's
 210          * fine and we shouldn't try to resolve the group again, so
 211          * let's just pretend it worked right-away. */
 212         s->file_gid_valid = true;
 213 }
 214
 215 static void server_fix_perms(Server *s, JournalFile *f, uid_t uid) {
 216         int r;
 217 #ifdef HAVE_ACL
 218         acl_t acl;
 219         acl_entry_t entry;
 220         acl_permset_t permset;
 221 #endif
 222
 223         assert(f);
 224
 225         server_read_file_gid(s);
 226
 227         r = fchmod_and_fchown(f->fd, 0640, 0, s->file_gid);
 228         if (r < 0)
 229                 log_warning("Failed to fix access mode/rights on %s, ignoring: %s", f->path, strerror(-r));
 230
 231 #ifdef HAVE_ACL
 232         if (uid <= 0)
 233                 return;
 234
 235         acl = acl_get_fd(f->fd);
 236         if (!acl) {
 237                 log_warning("Failed to read ACL on %s, ignoring: %m", f->path);
 238                 return;
 239         }
 240
 241         r = acl_find_uid(acl, uid, &entry);
 242         if (r <= 0) {
 243
 244                 if (acl_create_entry(&acl, &entry) < 0 ||
 245                     acl_set_tag_type(entry, ACL_USER) < 0 ||
 246                     acl_set_qualifier(entry, &uid) < 0) {
 247                         log_warning("Failed to patch ACL on %s, ignoring: %m", f->path);
 248                         goto finish;
 249                 }
 250         }
 251
 252         if (acl_get_permset(entry, &permset) < 0 ||
 253             acl_add_perm(permset, ACL_READ) < 0 ||
 254             acl_calc_mask(&acl) < 0) {
 255                 log_warning("Failed to patch ACL on %s, ignoring: %m", f->path);
 256                 goto finish;
 257         }
 258
 259         if (acl_set_fd(f->fd, acl) < 0)
 260                 log_warning("Failed to set ACL on %s, ignoring: %m", f->path);
 261
 262 finish:
 263         acl_free(acl);
 264 #endif
 265 }
 266
 267 static JournalFile* find_journal(Server *s, uid_t uid) {
 268         char *p;
 269         int r;
 270         JournalFile *f;
 271         char ids[33];
 272         sd_id128_t machine;
 273
 274         assert(s);
 275
 276         /* We split up user logs only on /var, not on /run. If the
 277          * runtime file is open, we write to it exclusively, in order
 278          * to guarantee proper order as soon as we flush /run to
 279          * /var and close the runtime file. */
 280
 281         if (s->runtime_journal)
 282                 return s->runtime_journal;
 283
 284         if (uid <= 0)
 285                 return s->system_journal;
 286
 287         r = sd_id128_get_machine(&machine);
 288         if (r < 0)
 289                 return s->system_journal;
 290
 291         f = hashmap_get(s->user_journals, UINT32_TO_PTR(uid));
 292         if (f)
 293                 return f;
 294
 295         if (asprintf(&p, "/var/log/journal/%s/user-%lu.journal", sd_id128_to_string(machine, ids), (unsigned long) uid) < 0)
 296                 return s->system_journal;
 297
 298         while (hashmap_size(s->user_journals) >= USER_JOURNALS_MAX) {
 299                 /* Too many open? Then let's close one */
 300                 f = hashmap_steal_first(s->user_journals);
 301                 assert(f);
 302                 journal_file_close(f);
 303         }
 304
 305         r = journal_file_open_reliably(p, O_RDWR|O_CREAT, 0640, s->system_journal, &f);
 306         free(p);
 307
 308         if (r < 0)
 309                 return s->system_journal;
 310
 311         server_fix_perms(s, f, uid);
 312
 313         r = hashmap_put(s->user_journals, UINT32_TO_PTR(uid), f);
 314         if (r < 0) {
 315                 journal_file_close(f);
 316                 return s->system_journal;
 317         }
 318
 319         return f;
 320 }
 321
 322 static void server_rotate(Server *s) {
 323         JournalFile *f;
 324         void *k;
 325         Iterator i;
 326         int r;
 327
 328         log_info("Rotating...");
 329
 330         if (s->runtime_journal) {
 331                 r = journal_file_rotate(&s->runtime_journal);
 332                 if (r < 0)
 333                         log_error("Failed to rotate %s: %s", s->runtime_journal->path, strerror(-r));
 334                 else
 335                         server_fix_perms(s, s->runtime_journal, 0);
 336         }
 337
 338         if (s->system_journal) {
 339                 r = journal_file_rotate(&s->system_journal);
 340                 if (r < 0)
 341                         log_error("Failed to rotate %s: %s", s->system_journal->path, strerror(-r));
 342                 else
 343                         server_fix_perms(s, s->system_journal, 0);
 344         }
 345
 346         HASHMAP_FOREACH_KEY(f, k, s->user_journals, i) {
 347                 r = journal_file_rotate(&f);
 348                 if (r < 0)
 349                         log_error("Failed to rotate %s: %s", f->path, strerror(-r));
 350                 else {
 351                         hashmap_replace(s->user_journals, k, f);
 352                         server_fix_perms(s, s->system_journal, PTR_TO_UINT32(k));
 353                 }
 354         }
 355 }
 356
 357 static void server_vacuum(Server *s) {
 358         char *p;
 359         char ids[33];
 360         sd_id128_t machine;
 361         int r;
 362
 363         log_info("Vacuuming...");
 364
 365         r = sd_id128_get_machine(&machine);
 366         if (r < 0) {
 367                 log_error("Failed to get machine ID: %s", strerror(-r));
 368                 return;
 369         }
 370
 371         sd_id128_to_string(machine, ids);
 372
 373         if (s->system_journal) {
 374                 if (asprintf(&p, "/var/log/journal/%s", ids) < 0) {
 375                         log_error("Out of memory.");
 376                         return;
 377                 }
 378
 379                 r = journal_directory_vacuum(p, s->system_metrics.max_use, s->system_metrics.keep_free);
 380                 if (r < 0 && r != -ENOENT)
 381                         log_error("Failed to vacuum %s: %s", p, strerror(-r));
 382                 free(p);
 383         }
 384
 385
 386         if (s->runtime_journal) {
 387                 if (asprintf(&p, "/run/log/journal/%s", ids) < 0) {
 388                         log_error("Out of memory.");
 389                         return;
 390                 }
 391
 392                 r = journal_directory_vacuum(p, s->runtime_metrics.max_use, s->runtime_metrics.keep_free);
 393                 if (r < 0 && r != -ENOENT)
 394                         log_error("Failed to vacuum %s: %s", p, strerror(-r));
 395                 free(p);
 396         }
 397
 398         s->cached_available_space_timestamp = 0;
 399 }
 400
 401 static char *shortened_cgroup_path(pid_t pid) {
 402         int r;
 403         char *process_path, *init_path, *path;
 404
 405         assert(pid > 0);
 406
 407         r = cg_get_by_pid(SYSTEMD_CGROUP_CONTROLLER, pid, &process_path);
 408         if (r < 0)
 409                 return NULL;
 410
 411         r = cg_get_by_pid(SYSTEMD_CGROUP_CONTROLLER, 1, &init_path);
 412         if (r < 0) {
 413                 free(process_path);
 414                 return NULL;
 415         }
 416
 417         if (endswith(init_path, "/system"))
 418                 init_path[strlen(init_path) - 7] = 0;
 419         else if (streq(init_path, "/"))
 420                 init_path[0] = 0;
 421
 422         if (startswith(process_path, init_path)) {
 423                 char *p;
 424
 425                 p = strdup(process_path + strlen(init_path));
 426                 if (!p) {
 427                         free(process_path);
 428                         free(init_path);
 429                         return NULL;
 430                 }
 431                 path = p;
 432         } else {
 433                 path = process_path;
 434                 process_path = NULL;
 435         }
 436
 437         free(process_path);
 438         free(init_path);
 439
 440         return path;
 441 }
 442
 443 static void dispatch_message_real(
 444                 Server *s,
 445                 struct iovec *iovec, unsigned n, unsigned m,
 446                 struct ucred *ucred,
 447                 struct timeval *tv,
 448                 const char *label, size_t label_len) {
 449
 450         char *pid = NULL, *uid = NULL, *gid = NULL,
 451                 *source_time = NULL, *boot_id = NULL, *machine_id = NULL,
 452                 *comm = NULL, *cmdline = NULL, *hostname = NULL,
 453                 *audit_session = NULL, *audit_loginuid = NULL,
 454                 *exe = NULL, *cgroup = NULL, *session = NULL,
 455                 *owner_uid = NULL, *unit = NULL, *selinux_context = NULL;
 456
 457         char idbuf[33];
 458         sd_id128_t id;
 459         int r;
 460         char *t;
 461         uid_t loginuid = 0, realuid = 0;
 462         JournalFile *f;
 463         bool vacuumed = false;
 464
 465         assert(s);
 466         assert(iovec);
 467         assert(n > 0);
 468         assert(n + N_IOVEC_META_FIELDS <= m);
 469
 470         if (ucred) {
 471                 uint32_t audit;
 472                 uid_t owner;
 473
 474                 realuid = ucred->uid;
 475
 476                 if (asprintf(&pid, "_PID=%lu", (unsigned long) ucred->pid) >= 0)
 477                         IOVEC_SET_STRING(iovec[n++], pid);
 478
 479                 if (asprintf(&uid, "_UID=%lu", (unsigned long) ucred->uid) >= 0)
 480                         IOVEC_SET_STRING(iovec[n++], uid);
 481
 482                 if (asprintf(&gid, "_GID=%lu", (unsigned long) ucred->gid) >= 0)
 483                         IOVEC_SET_STRING(iovec[n++], gid);
 484
 485                 r = get_process_comm(ucred->pid, &t);
 486                 if (r >= 0) {
 487                         comm = strappend("_COMM=", t);
 488                         free(t);
 489
 490                         if (comm)
 491                                 IOVEC_SET_STRING(iovec[n++], comm);
 492                 }
 493
 494                 r = get_process_exe(ucred->pid, &t);
 495                 if (r >= 0) {
 496                         exe = strappend("_EXE=", t);
 497                         free(t);
 498
 499                         if (exe)
 500                                 IOVEC_SET_STRING(iovec[n++], exe);
 501                 }
 502
 503                 r = get_process_cmdline(ucred->pid, LINE_MAX, false, &t);
 504                 if (r >= 0) {
 505                         cmdline = strappend("_CMDLINE=", t);
 506                         free(t);
 507
 508                         if (cmdline)
 509                                 IOVEC_SET_STRING(iovec[n++], cmdline);
 510                 }
 511
 512                 r = audit_session_from_pid(ucred->pid, &audit);
 513                 if (r >= 0)
 514                         if (asprintf(&audit_session, "_AUDIT_SESSION=%lu", (unsigned long) audit) >= 0)
 515                                 IOVEC_SET_STRING(iovec[n++], audit_session);
 516
 517                 r = audit_loginuid_from_pid(ucred->pid, &loginuid);
 518                 if (r >= 0)
 519                         if (asprintf(&audit_loginuid, "_AUDIT_LOGINUID=%lu", (unsigned long) loginuid) >= 0)
 520                                 IOVEC_SET_STRING(iovec[n++], audit_loginuid);
 521
 522                 t = shortened_cgroup_path(ucred->pid);
 523                 if (t) {
 524                         cgroup = strappend("_SYSTEMD_CGROUP=", t);
 525                         free(t);
 526
 527                         if (cgroup)
 528                                 IOVEC_SET_STRING(iovec[n++], cgroup);
 529                 }
 530
 531                 if (sd_pid_get_session(ucred->pid, &t) >= 0) {
 532                         session = strappend("_SYSTEMD_SESSION=", t);
 533                         free(t);
 534
 535                         if (session)
 536                                 IOVEC_SET_STRING(iovec[n++], session);
 537                 }
 538
 539                 if (sd_pid_get_unit(ucred->pid, &t) >= 0) {
 540                         unit = strappend("_SYSTEMD_UNIT=", t);
 541                         free(t);
 542
 543                         if (unit)
 544                                 IOVEC_SET_STRING(iovec[n++], unit);
 545                 }
 546
 547                 if (sd_pid_get_owner_uid(ucred->uid, &owner) >= 0)
 548                         if (asprintf(&owner_uid, "_SYSTEMD_OWNER_UID=%lu", (unsigned long) owner) >= 0)
 549                                 IOVEC_SET_STRING(iovec[n++], owner_uid);
 550
 551 #ifdef HAVE_SELINUX
 552                 if (label) {
 553                         selinux_context = malloc(sizeof("_SELINUX_CONTEXT=") + label_len);
 554                         if (selinux_context) {
 555                                 memcpy(selinux_context, "_SELINUX_CONTEXT=", sizeof("_SELINUX_CONTEXT=")-1);
 556                                 memcpy(selinux_context+sizeof("_SELINUX_CONTEXT=")-1, label, label_len);
 557                                 selinux_context[sizeof("_SELINUX_CONTEXT=")-1+label_len] = 0;
 558                                 IOVEC_SET_STRING(iovec[n++], selinux_context);
 559                         }
 560                 } else {
 561                         security_context_t con;
 562
 563                         if (getpidcon(ucred->pid, &con) >= 0) {
 564                                 selinux_context = strappend("_SELINUX_CONTEXT=", con);
 565                                 if (selinux_context)
 566                                         IOVEC_SET_STRING(iovec[n++], selinux_context);
 567
 568                                 freecon(con);
 569                         }
 570                 }
 571 #endif
 572         }
 573
 574         if (tv) {
 575                 if (asprintf(&source_time, "_SOURCE_REALTIME_TIMESTAMP=%llu",
 576                              (unsigned long long) timeval_load(tv)) >= 0)
 577                         IOVEC_SET_STRING(iovec[n++], source_time);
 578         }
 579
 580         /* Note that strictly speaking storing the boot id here is
 581          * redundant since the entry includes this in-line
 582          * anyway. However, we need this indexed, too. */
 583         r = sd_id128_get_boot(&id);
 584         if (r >= 0)
 585                 if (asprintf(&boot_id, "_BOOT_ID=%s", sd_id128_to_string(id, idbuf)) >= 0)
 586                         IOVEC_SET_STRING(iovec[n++], boot_id);
 587
 588         r = sd_id128_get_machine(&id);
 589         if (r >= 0)
 590                 if (asprintf(&machine_id, "_MACHINE_ID=%s", sd_id128_to_string(id, idbuf)) >= 0)
 591                         IOVEC_SET_STRING(iovec[n++], machine_id);
 592
 593         t = gethostname_malloc();
 594         if (t) {
 595                 hostname = strappend("_HOSTNAME=", t);
 596                 free(t);
 597                 if (hostname)
 598                         IOVEC_SET_STRING(iovec[n++], hostname);
 599         }
 600
 601         assert(n <= m);
 602
 603         server_flush_to_var(s);
 604
 605 retry:
 606         f = find_journal(s, realuid == 0 ? 0 : loginuid);
 607         if (!f)
 608                 log_warning("Dropping message, as we can't find a place to store the data.");
 609         else {
 610                 r = journal_file_append_entry(f, NULL, iovec, n, &s->seqnum, NULL, NULL);
 611
 612                 if ((r == -E2BIG || /* hit limit */
 613                      r == -EFBIG || /* hit fs limit */
 614                      r == -EDQUOT || /* quota hit */
 615                      r == -ENOSPC || /* disk full */
 616                      r == -EBADMSG || /* corrupted */
 617                      r == -ENODATA || /* truncated */
 618                      r == -EHOSTDOWN || /* other machine */
 619                      r == -EPROTONOSUPPORT) && /* unsupported feature */
 620                     !vacuumed) {
 621
 622                         if (r == -E2BIG)
 623                                 log_info("Allocation limit reached, rotating.");
 624                         else
 625                                 log_warning("Journal file corrupted, rotating.");
 626
 627                         server_rotate(s);
 628                         server_vacuum(s);
 629                         vacuumed = true;
 630
 631                         log_info("Retrying write.");
 632                         goto retry;
 633                 }
 634
 635                 if (r < 0)
 636                         log_error("Failed to write entry, ignoring: %s", strerror(-r));
 637         }
 638
 639         free(pid);
 640         free(uid);
 641         free(gid);
 642         free(comm);
 643         free(exe);
 644         free(cmdline);
 645         free(source_time);
 646         free(boot_id);
 647         free(machine_id);
 648         free(hostname);
 649         free(audit_session);
 650         free(audit_loginuid);
 651         free(cgroup);
 652         free(session);
 653         free(owner_uid);
 654         free(unit);
 655         free(selinux_context);
 656 }
 657
 658 static void driver_message(Server *s, sd_id128_t message_id, const char *format, ...) {
 659         char mid[11 + 32 + 1];
 660         char buffer[16 + LINE_MAX + 1];
 661         struct iovec iovec[N_IOVEC_META_FIELDS + 4];
 662         int n = 0;
 663         va_list ap;
 664         struct ucred ucred;
 665
 666         assert(s);
 667         assert(format);
 668
 669         IOVEC_SET_STRING(iovec[n++], "PRIORITY=5");
 670         IOVEC_SET_STRING(iovec[n++], "_TRANSPORT=driver");
 671
 672         memcpy(buffer, "MESSAGE=", 8);
 673         va_start(ap, format);
 674         vsnprintf(buffer + 8, sizeof(buffer) - 8, format, ap);
 675         va_end(ap);
 676         char_array_0(buffer);
 677         IOVEC_SET_STRING(iovec[n++], buffer);
 678
 679         snprintf(mid, sizeof(mid), "MESSAGE_ID=" SD_ID128_FORMAT_STR, SD_ID128_FORMAT_VAL(message_id));
 680         char_array_0(mid);
 681         IOVEC_SET_STRING(iovec[n++], mid);
 682
 683         zero(ucred);
 684         ucred.pid = getpid();
 685         ucred.uid = getuid();
 686         ucred.gid = getgid();
 687
 688         dispatch_message_real(s, iovec, n, ELEMENTSOF(iovec), &ucred, NULL, NULL, 0);
 689 }
 690
 691 static void dispatch_message(Server *s,
 692                              struct iovec *iovec, unsigned n, unsigned m,
 693                              struct ucred *ucred,
 694                              struct timeval *tv,
 695                              const char *label, size_t label_len,
 696                              int priority) {
 697         int rl;
 698         char *path = NULL, *c;
 699
 700         assert(s);
 701         assert(iovec || n == 0);
 702
 703         if (n == 0)
 704                 return;
 705
 706         if (!ucred)
 707                 goto finish;
 708
 709         path = shortened_cgroup_path(ucred->pid);
 710         if (!path)
 711                 goto finish;
 712
 713         /* example: /user/lennart/3/foobar
 714          *          /system/dbus.service/foobar
 715          *
 716          * So let's cut of everything past the third /, since that is
 717          * wher user directories start */
 718
 719         c = strchr(path, '/');
 720         if (c) {
 721                 c = strchr(c+1, '/');
 722                 if (c) {
 723                         c = strchr(c+1, '/');
 724                         if (c)
 725                                 *c = 0;
 726                 }
 727         }
 728
 729         rl = journal_rate_limit_test(s->rate_limit, path, priority & LOG_PRIMASK, available_space(s));
 730
 731         if (rl == 0) {
 732                 free(path);
 733                 return;
 734         }
 735
 736         /* Write a suppression message if we suppressed something */
 737         if (rl > 1)
 738                 driver_message(s, SD_MESSAGE_JOURNAL_DROPPED, "Suppressed %u messages from %s", rl - 1, path);
 739
 740         free(path);
 741
 742 finish:
 743         dispatch_message_real(s, iovec, n, m, ucred, tv, label, label_len);
 744 }
 745
 746 static void forward_syslog_iovec(Server *s, const struct iovec *iovec, unsigned n_iovec, struct ucred *ucred, struct timeval *tv) {
 747         struct msghdr msghdr;
 748         struct cmsghdr *cmsg;
 749         union {
 750                 struct cmsghdr cmsghdr;
 751                 uint8_t buf[CMSG_SPACE(sizeof(struct ucred))];
 752         } control;
 753         union sockaddr_union sa;
 754
 755         assert(s);
 756         assert(iovec);
 757         assert(n_iovec > 0);
 758
 759         zero(msghdr);
 760         msghdr.msg_iov = (struct iovec*) iovec;
 761         msghdr.msg_iovlen = n_iovec;
 762
 763         zero(sa);
 764         sa.un.sun_family = AF_UNIX;
 765         strncpy(sa.un.sun_path, "/run/systemd/journal/syslog", sizeof(sa.un.sun_path));
 766         msghdr.msg_name = &sa;
 767         msghdr.msg_namelen = offsetof(union sockaddr_union, un.sun_path) + strlen(sa.un.sun_path);
 768
 769         if (ucred) {
 770                 zero(control);
 771                 msghdr.msg_control = &control;
 772                 msghdr.msg_controllen = sizeof(control);
 773
 774                 cmsg = CMSG_FIRSTHDR(&msghdr);
 775                 cmsg->cmsg_level = SOL_SOCKET;
 776                 cmsg->cmsg_type = SCM_CREDENTIALS;
 777                 cmsg->cmsg_len = CMSG_LEN(sizeof(struct ucred));
 778                 memcpy(CMSG_DATA(cmsg), ucred, sizeof(struct ucred));
 779                 msghdr.msg_controllen = cmsg->cmsg_len;
 780         }
 781
 782         /* Forward the syslog message we received via /dev/log to
 783          * /run/systemd/syslog. Unfortunately we currently can't set
 784          * the SO_TIMESTAMP auxiliary data, and hence we don't. */
 785
 786         if (sendmsg(s->syslog_fd, &msghdr, MSG_NOSIGNAL) >= 0)
 787                 return;
 788
 789         /* The socket is full? I guess the syslog implementation is
 790          * too slow, and we shouldn't wait for that... */
 791         if (errno == EAGAIN)
 792                 return;
 793
 794         if (ucred && errno == ESRCH) {
 795                 struct ucred u;
 796
 797                 /* Hmm, presumably the sender process vanished
 798                  * by now, so let's fix it as good as we
 799                  * can, and retry */
 800
 801                 u = *ucred;
 802                 u.pid = getpid();
 803                 memcpy(CMSG_DATA(cmsg), &u, sizeof(struct ucred));
 804
 805                 if (sendmsg(s->syslog_fd, &msghdr, MSG_NOSIGNAL) >= 0)
 806                         return;
 807
 808                 if (errno == EAGAIN)
 809                         return;
 810         }
 811
 812         log_debug("Failed to forward syslog message: %m");
 813 }
 814
 815 static void forward_syslog_raw(Server *s, const char *buffer, struct ucred *ucred, struct timeval *tv) {
 816         struct iovec iovec;
 817
 818         assert(s);
 819         assert(buffer);
 820
 821         IOVEC_SET_STRING(iovec, buffer);
 822         forward_syslog_iovec(s, &iovec, 1, ucred, tv);
 823 }
 824
 825 static void forward_syslog(Server *s, int priority, const char *identifier, const char *message, struct ucred *ucred, struct timeval *tv) {
 826         struct iovec iovec[5];
 827         char header_priority[6], header_time[64], header_pid[16];
 828         int n = 0;
 829         time_t t;
 830         struct tm *tm;
 831         char *ident_buf = NULL;
 832
 833         assert(s);
 834         assert(priority >= 0);
 835         assert(priority <= 999);
 836         assert(message);
 837
 838         /* First: priority field */
 839         snprintf(header_priority, sizeof(header_priority), "<%i>", priority);
 840         char_array_0(header_priority);
 841         IOVEC_SET_STRING(iovec[n++], header_priority);
 842
 843         /* Second: timestamp */
 844         t = tv ? tv->tv_sec : ((time_t) (now(CLOCK_REALTIME) / USEC_PER_SEC));
 845         tm = localtime(&t);
 846         if (!tm)
 847                 return;
 848         if (strftime(header_time, sizeof(header_time), "%h %e %T ", tm) <= 0)
 849                 return;
 850         IOVEC_SET_STRING(iovec[n++], header_time);
 851
 852         /* Third: identifier and PID */
 853         if (ucred) {
 854                 if (!identifier) {
 855                         get_process_comm(ucred->pid, &ident_buf);
 856                         identifier = ident_buf;
 857                 }
 858
 859                 snprintf(header_pid, sizeof(header_pid), "[%lu]: ", (unsigned long) ucred->pid);
 860                 char_array_0(header_pid);
 861
 862                 if (identifier)
 863                         IOVEC_SET_STRING(iovec[n++], identifier);
 864
 865                 IOVEC_SET_STRING(iovec[n++], header_pid);
 866         } else if (identifier) {
 867                 IOVEC_SET_STRING(iovec[n++], identifier);
 868                 IOVEC_SET_STRING(iovec[n++], ": ");
 869         }
 870
 871         /* Fourth: message */
 872         IOVEC_SET_STRING(iovec[n++], message);
 873
 874         forward_syslog_iovec(s, iovec, n, ucred, tv);
 875
 876         free(ident_buf);
 877 }
 878
 879 static int fixup_priority(int priority) {
 880
 881         if ((priority & LOG_FACMASK) == 0)
 882                 return (priority & LOG_PRIMASK) | LOG_USER;
 883
 884         return priority;
 885 }
 886
 887 static void forward_kmsg(Server *s, int priority, const char *identifier, const char *message, struct ucred *ucred) {
 888         struct iovec iovec[5];
 889         char header_priority[6], header_pid[16];
 890         int n = 0;
 891         char *ident_buf = NULL;
 892         int fd;
 893
 894         assert(s);
 895         assert(priority >= 0);
 896         assert(priority <= 999);
 897         assert(message);
 898
 899         /* Never allow messages with kernel facility to be written to
 900          * kmsg, regardless where the data comes from. */
 901         priority = fixup_priority(priority);
 902
 903         /* First: priority field */
 904         snprintf(header_priority, sizeof(header_priority), "<%i>", priority);
 905         char_array_0(header_priority);
 906         IOVEC_SET_STRING(iovec[n++], header_priority);
 907
 908         /* Second: identifier and PID */
 909         if (ucred) {
 910                 if (!identifier) {
 911                         get_process_comm(ucred->pid, &ident_buf);
 912                         identifier = ident_buf;
 913                 }
 914
 915                 snprintf(header_pid, sizeof(header_pid), "[%lu]: ", (unsigned long) ucred->pid);
 916                 char_array_0(header_pid);
 917
 918                 if (identifier)
 919                         IOVEC_SET_STRING(iovec[n++], identifier);
 920
 921                 IOVEC_SET_STRING(iovec[n++], header_pid);
 922         } else if (identifier) {
 923                 IOVEC_SET_STRING(iovec[n++], identifier);
 924                 IOVEC_SET_STRING(iovec[n++], ": ");
 925         }
 926
 927         /* Fourth: message */
 928         IOVEC_SET_STRING(iovec[n++], message);
 929         IOVEC_SET_STRING(iovec[n++], "\n");
 930
 931         fd = open("/dev/kmsg", O_WRONLY|O_NOCTTY|O_CLOEXEC);
 932         if (fd < 0) {
 933                 log_debug("Failed to open /dev/kmsg for logging: %s", strerror(errno));
 934                 goto finish;
 935         }
 936
 937         if (writev(fd, iovec, n) < 0)
 938                 log_debug("Failed to write to /dev/kmsg for logging: %s", strerror(errno));
 939
 940         close_nointr_nofail(fd);
 941
 942 finish:
 943         free(ident_buf);
 944 }
 945
 946 static void forward_console(Server *s, const char *identifier, const char *message, struct ucred *ucred) {
 947         struct iovec iovec[4];
 948         char header_pid[16];
 949         int n = 0, fd;
 950         char *ident_buf = NULL;
 951
 952         assert(s);
 953         assert(message);
 954
 955         /* First: identifier and PID */
 956         if (ucred) {
 957                 if (!identifier) {
 958                         get_process_comm(ucred->pid, &ident_buf);
 959                         identifier = ident_buf;
 960                 }
 961
 962                 snprintf(header_pid, sizeof(header_pid), "[%lu]: ", (unsigned long) ucred->pid);
 963                 char_array_0(header_pid);
 964
 965                 if (identifier)
 966                         IOVEC_SET_STRING(iovec[n++], identifier);
 967
 968                 IOVEC_SET_STRING(iovec[n++], header_pid);
 969         } else if (identifier) {
 970                 IOVEC_SET_STRING(iovec[n++], identifier);
 971                 IOVEC_SET_STRING(iovec[n++], ": ");
 972         }
 973
 974         /* Third: message */
 975         IOVEC_SET_STRING(iovec[n++], message);
 976         IOVEC_SET_STRING(iovec[n++], "\n");
 977
 978         fd = open_terminal("/dev/console", O_WRONLY|O_NOCTTY|O_CLOEXEC);
 979         if (fd < 0) {
 980                 log_debug("Failed to open /dev/console for logging: %s", strerror(errno));
 981                 goto finish;
 982         }
 983
 984         if (writev(fd, iovec, n) < 0)
 985                 log_debug("Failed to write to /dev/console for logging: %s", strerror(errno));
 986
 987         close_nointr_nofail(fd);
 988
 989 finish:
 990         free(ident_buf);
 991 }
 992
 993 static void read_identifier(const char **buf, char **identifier, char **pid) {
 994         const char *p;
 995         char *t;
 996         size_t l, e;
 997
 998         assert(buf);
 999         assert(identifier);
1000         assert(pid);
1001
1002         p = *buf;
1003
1004         p += strspn(p, WHITESPACE);
1005         l = strcspn(p, WHITESPACE);
1006
1007         if (l <= 0 ||
1008             p[l-1] != ':')
1009                 return;
1010
1011         e = l;
1012         l--;
1013
1014         if (p[l-1] == ']') {
1015                 size_t k = l-1;
1016
1017                 for (;;) {
1018
1019                         if (p[k] == '[') {
1020                                 t = strndup(p+k+1, l-k-2);
1021                                 if (t)
1022                                         *pid = t;
1023
1024                                 l = k;
1025                                 break;
1026                         }
1027
1028                         if (k == 0)
1029                                 break;
1030
1031                         k--;
1032                 }
1033         }
1034
1035         t = strndup(p, l);
1036         if (t)
1037                 *identifier = t;
1038
1039         *buf = p + e;
1040         *buf += strspn(*buf, WHITESPACE);
1041 }
1042
1043 static void process_syslog_message(Server *s, const char *buf, struct ucred *ucred, struct timeval *tv, const char *label, size_t label_len) {
1044         char *message = NULL, *syslog_priority = NULL, *syslog_facility = NULL, *syslog_identifier = NULL, *syslog_pid = NULL;
1045         struct iovec iovec[N_IOVEC_META_FIELDS + 6];
1046         unsigned n = 0;
1047         int priority = LOG_USER | LOG_INFO;
1048         char *identifier = NULL, *pid = NULL;
1049
1050         assert(s);
1051         assert(buf);
1052
1053         if (s->forward_to_syslog)
1054                 forward_syslog_raw(s, buf, ucred, tv);
1055
1056         parse_syslog_priority((char**) &buf, &priority);
1057         skip_syslog_date((char**) &buf);
1058         read_identifier(&buf, &identifier, &pid);
1059
1060         if (s->forward_to_kmsg)
1061                 forward_kmsg(s, priority, identifier, buf, ucred);
1062
1063         if (s->forward_to_console)
1064                 forward_console(s, identifier, buf, ucred);
1065
1066         IOVEC_SET_STRING(iovec[n++], "_TRANSPORT=syslog");
1067
1068         if (asprintf(&syslog_priority, "PRIORITY=%i", priority & LOG_PRIMASK) >= 0)
1069                 IOVEC_SET_STRING(iovec[n++], syslog_priority);
1070
1071         if (priority & LOG_FACMASK)
1072                 if (asprintf(&syslog_facility, "SYSLOG_FACILITY=%i", LOG_FAC(priority)) >= 0)
1073                         IOVEC_SET_STRING(iovec[n++], syslog_facility);
1074
1075         if (identifier) {
1076                 syslog_identifier = strappend("SYSLOG_IDENTIFIER=", identifier);
1077                 if (syslog_identifier)
1078                         IOVEC_SET_STRING(iovec[n++], syslog_identifier);
1079         }
1080
1081         if (pid) {
1082                 syslog_pid = strappend("SYSLOG_PID=", pid);
1083                 if (syslog_pid)
1084                         IOVEC_SET_STRING(iovec[n++], syslog_pid);
1085         }
1086
1087         message = strappend("MESSAGE=", buf);
1088         if (message)
1089                 IOVEC_SET_STRING(iovec[n++], message);
1090
1091         dispatch_message(s, iovec, n, ELEMENTSOF(iovec), ucred, tv, label, label_len, priority);
1092
1093         free(message);
1094         free(identifier);
1095         free(pid);
1096         free(syslog_priority);
1097         free(syslog_facility);
1098         free(syslog_identifier);
1099 }
1100
1101 static bool valid_user_field(const char *p, size_t l) {
1102         const char *a;
1103
1104         /* We kinda enforce POSIX syntax recommendations for
1105            environment variables here, but make a couple of additional
1106            requirements.
1107
1108            http://pubs.opengroup.org/onlinepubs/000095399/basedefs/xbd_chap08.html */
1109
1110         /* No empty field names */
1111         if (l <= 0)
1112                 return false;
1113
1114         /* Don't allow names longer than 64 chars */
1115         if (l > 64)
1116                 return false;
1117
1118         /* Variables starting with an underscore are protected */
1119         if (p[0] == '_')
1120                 return false;
1121
1122         /* Don't allow digits as first character */
1123         if (p[0] >= '0' && p[0] <= '9')
1124                 return false;
1125
1126         /* Only allow A-Z0-9 and '_' */
1127         for (a = p; a < p + l; a++)
1128                 if (!((*a >= 'A' && *a <= 'Z') ||
1129                       (*a >= '0' && *a <= '9') ||
1130                       *a == '_'))
1131                         return false;
1132
1133         return true;
1134 }
1135
1136 static void process_native_message(
1137                 Server *s,
1138                 const void *buffer, size_t buffer_size,
1139                 struct ucred *ucred,
1140                 struct timeval *tv,
1141                 const char *label, size_t label_len) {
1142
1143         struct iovec *iovec = NULL;
1144         unsigned n = 0, m = 0, j, tn = (unsigned) -1;
1145         const char *p;
1146         size_t remaining;
1147         int priority = LOG_INFO;
1148         char *identifier = NULL, *message = NULL;
1149
1150         assert(s);
1151         assert(buffer || n == 0);
1152
1153         p = buffer;
1154         remaining = buffer_size;
1155
1156         while (remaining > 0) {
1157                 const char *e, *q;
1158
1159                 e = memchr(p, '\n', remaining);
1160
1161                 if (!e) {
1162                         /* Trailing noise, let's ignore it, and flush what we collected */
1163                         log_debug("Received message with trailing noise, ignoring.");
1164                         break;
1165                 }
1166
1167                 if (e == p) {
1168                         /* Entry separator */
1169                         dispatch_message(s, iovec, n, m, ucred, tv, label, label_len, priority);
1170                         n = 0;
1171                         priority = LOG_INFO;
1172
1173                         p++;
1174                         remaining--;
1175                         continue;
1176                 }
1177
1178                 if (*p == '.' || *p == '#') {
1179                         /* Ignore control commands for now, and
1180                          * comments too. */
1181                         remaining -= (e - p) + 1;
1182                         p = e + 1;
1183                         continue;
1184                 }
1185
1186                 /* A property follows */
1187
1188                 if (n+N_IOVEC_META_FIELDS >= m) {
1189                         struct iovec *c;
1190                         unsigned u;
1191
1192                         u = MAX((n+N_IOVEC_META_FIELDS+1) * 2U, 4U);
1193                         c = realloc(iovec, u * sizeof(struct iovec));
1194                         if (!c) {
1195                                 log_error("Out of memory");
1196                                 break;
1197                         }
1198
1199                         iovec = c;
1200                         m = u;
1201                 }
1202
1203                 q = memchr(p, '=', e - p);
1204                 if (q) {
1205                         if (valid_user_field(p, q - p)) {
1206                                 size_t l;
1207
1208                                 l = e - p;
1209
1210                                 /* If the field name starts with an
1211                                  * underscore, skip the variable,
1212                                  * since that indidates a trusted
1213                                  * field */
1214                                 iovec[n].iov_base = (char*) p;
1215                                 iovec[n].iov_len = l;
1216                                 n++;
1217
1218                                 /* We need to determine the priority
1219                                  * of this entry for the rate limiting
1220                                  * logic */
1221                                 if (l == 10 &&
1222                                     memcmp(p, "PRIORITY=", 9) == 0 &&
1223                                     p[9] >= '0' && p[9] <= '9')
1224                                         priority = (priority & LOG_FACMASK) | (p[9] - '0');
1225
1226                                 else if (l == 17 &&
1227                                          memcmp(p, "SYSLOG_FACILITY=", 16) == 0 &&
1228                                          p[16] >= '0' && p[16] <= '9')
1229                                         priority = (priority & LOG_PRIMASK) | ((p[16] - '0') << 3);
1230
1231                                 else if (l == 18 &&
1232                                          memcmp(p, "SYSLOG_FACILITY=", 16) == 0 &&
1233                                          p[16] >= '0' && p[16] <= '9' &&
1234                                          p[17] >= '0' && p[17] <= '9')
1235                                         priority = (priority & LOG_PRIMASK) | (((p[16] - '0')*10 + (p[17] - '0')) << 3);
1236
1237                                 else if (l >= 12 &&
1238                                          memcmp(p, "SYSLOG_IDENTIFIER=", 11) == 0) {
1239                                         char *t;
1240
1241                                         t = strndup(p + 11, l - 11);
1242                                         if (t) {
1243                                                 free(identifier);
1244                                                 identifier = t;
1245                                         }
1246                                 } else if (l >= 8 &&
1247                                            memcmp(p, "MESSAGE=", 8) == 0) {
1248                                         char *t;
1249
1250                                         t = strndup(p + 8, l - 8);
1251                                         if (t) {
1252                                                 free(message);
1253                                                 message = t;
1254                                         }
1255                                 }
1256                         }
1257
1258                         remaining -= (e - p) + 1;
1259                         p = e + 1;
1260                         continue;
1261                 } else {
1262                         le64_t l_le;
1263                         uint64_t l;
1264                         char *k;
1265
1266                         if (remaining < e - p + 1 + sizeof(uint64_t) + 1) {
1267                                 log_debug("Failed to parse message, ignoring.");
1268                                 break;
1269                         }
1270
1271                         memcpy(&l_le, e + 1, sizeof(uint64_t));
1272                         l = le64toh(l_le);
1273
1274                         if (remaining < e - p + 1 + sizeof(uint64_t) + l + 1 ||
1275                             e[1+sizeof(uint64_t)+l] != '\n') {
1276                                 log_debug("Failed to parse message, ignoring.");
1277                                 break;
1278                         }
1279
1280                         k = malloc((e - p) + 1 + l);
1281                         if (!k) {
1282                                 log_error("Out of memory");
1283                                 break;
1284                         }
1285
1286                         memcpy(k, p, e - p);
1287                         k[e - p] = '=';
1288                         memcpy(k + (e - p) + 1, e + 1 + sizeof(uint64_t), l);
1289
1290                         if (valid_user_field(p, e - p)) {
1291                                 iovec[n].iov_base = k;
1292                                 iovec[n].iov_len = (e - p) + 1 + l;
1293                                 n++;
1294                         } else
1295                                 free(k);
1296
1297                         remaining -= (e - p) + 1 + sizeof(uint64_t) + l + 1;
1298                         p = e + 1 + sizeof(uint64_t) + l + 1;
1299                 }
1300         }
1301
1302         if (n <= 0)
1303                 goto finish;
1304
1305         tn = n++;
1306         IOVEC_SET_STRING(iovec[tn], "_TRANSPORT=journal");
1307
1308         if (message) {
1309                 if (s->forward_to_syslog)
1310                         forward_syslog(s, priority, identifier, message, ucred, tv);
1311
1312                 if (s->forward_to_kmsg)
1313                         forward_kmsg(s, priority, identifier, message, ucred);
1314
1315                 if (s->forward_to_console)
1316                         forward_console(s, identifier, message, ucred);
1317         }
1318
1319         dispatch_message(s, iovec, n, m, ucred, tv, label, label_len, priority);
1320
1321 finish:
1322         for (j = 0; j < n; j++)  {
1323                 if (j == tn)
1324                         continue;
1325
1326                 if (iovec[j].iov_base < buffer ||
1327                     (const uint8_t*) iovec[j].iov_base >= (const uint8_t*) buffer + buffer_size)
1328                         free(iovec[j].iov_base);
1329         }
1330
1331         free(iovec);
1332         free(identifier);
1333         free(message);
1334 }
1335
1336 static void process_native_file(
1337                 Server *s,
1338                 int fd,
1339                 struct ucred *ucred,
1340                 struct timeval *tv,
1341                 const char *label, size_t label_len) {
1342
1343         struct stat st;
1344         void *p;
1345         ssize_t n;
1346
1347         assert(s);
1348         assert(fd >= 0);
1349
1350         /* Data is in the passed file, since it didn't fit in a
1351          * datagram. We can't map the file here, since clients might
1352          * then truncate it and trigger a SIGBUS for us. So let's
1353          * stupidly read it */
1354
1355         if (fstat(fd, &st) < 0) {
1356                 log_error("Failed to stat passed file, ignoring: %m");
1357                 return;
1358         }
1359
1360         if (!S_ISREG(st.st_mode)) {
1361                 log_error("File passed is not regular. Ignoring.");
1362                 return;
1363         }
1364
1365         if (st.st_size <= 0)
1366                 return;
1367
1368         if (st.st_size > ENTRY_SIZE_MAX) {
1369                 log_error("File passed too large. Ignoring.");
1370                 return;
1371         }
1372
1373         p = malloc(st.st_size);
1374         if (!p) {
1375                 log_error("Out of memory");
1376                 return;
1377         }
1378
1379         n = pread(fd, p, st.st_size, 0);
1380         if (n < 0)
1381                 log_error("Failed to read file, ignoring: %s", strerror(-n));
1382         else if (n > 0)
1383                 process_native_message(s, p, n, ucred, tv, label, label_len);
1384
1385         free(p);
1386 }
1387
1388 static int stdout_stream_log(StdoutStream *s, const char *p) {
1389         struct iovec iovec[N_IOVEC_META_FIELDS + 5];
1390         char *message = NULL, *syslog_priority = NULL, *syslog_facility = NULL, *syslog_identifier = NULL;
1391         unsigned n = 0;
1392         int priority;
1393         char *label = NULL;
1394         size_t label_len = 0;
1395
1396         assert(s);
1397         assert(p);
1398
1399         if (isempty(p))
1400                 return 0;
1401
1402         priority = s->priority;
1403
1404         if (s->level_prefix)
1405                 parse_syslog_priority((char**) &p, &priority);
1406
1407         if (s->forward_to_syslog || s->server->forward_to_syslog)
1408                 forward_syslog(s->server, fixup_priority(priority), s->identifier, p, &s->ucred, NULL);
1409
1410         if (s->forward_to_kmsg || s->server->forward_to_kmsg)
1411                 forward_kmsg(s->server, priority, s->identifier, p, &s->ucred);
1412
1413         if (s->forward_to_console || s->server->forward_to_console)
1414                 forward_console(s->server, s->identifier, p, &s->ucred);
1415
1416         IOVEC_SET_STRING(iovec[n++], "_TRANSPORT=stdout");
1417
1418         if (asprintf(&syslog_priority, "PRIORITY=%i", priority & LOG_PRIMASK) >= 0)
1419                 IOVEC_SET_STRING(iovec[n++], syslog_priority);
1420
1421         if (priority & LOG_FACMASK)
1422                 if (asprintf(&syslog_facility, "SYSLOG_FACILITY=%i", LOG_FAC(priority)) >= 0)
1423                         IOVEC_SET_STRING(iovec[n++], syslog_facility);
1424
1425         if (s->identifier) {
1426                 syslog_identifier = strappend("SYSLOG_IDENTIFIER=", s->identifier);
1427                 if (syslog_identifier)
1428                         IOVEC_SET_STRING(iovec[n++], syslog_identifier);
1429         }
1430
1431         message = strappend("MESSAGE=", p);
1432         if (message)
1433                 IOVEC_SET_STRING(iovec[n++], message);
1434
1435 #ifdef HAVE_SELINUX
1436         if (s->security_context) {
1437                 label = (char*) s->security_context;
1438                 label_len = strlen((char*) s->security_context);
1439         }
1440 #endif
1441
1442         dispatch_message(s->server, iovec, n, ELEMENTSOF(iovec), &s->ucred, NULL, label, label_len, priority);
1443
1444         free(message);
1445         free(syslog_priority);
1446         free(syslog_facility);
1447         free(syslog_identifier);
1448
1449         return 0;
1450 }
1451
1452 static int stdout_stream_line(StdoutStream *s, char *p) {
1453         int r;
1454
1455         assert(s);
1456         assert(p);
1457
1458         p = strstrip(p);
1459
1460         switch (s->state) {
1461
1462         case STDOUT_STREAM_IDENTIFIER:
1463                 if (isempty(p))
1464                         s->identifier = NULL;
1465                 else  {
1466                         s->identifier = strdup(p);
1467                         if (!s->identifier) {
1468                                 log_error("Out of memory");
1469                                 return -ENOMEM;
1470                         }
1471                 }
1472
1473                 s->state = STDOUT_STREAM_PRIORITY;
1474                 return 0;
1475
1476         case STDOUT_STREAM_PRIORITY:
1477                 r = safe_atoi(p, &s->priority);
1478                 if (r < 0 || s->priority <= 0 || s->priority >= 999) {
1479                         log_warning("Failed to parse log priority line.");
1480                         return -EINVAL;
1481                 }
1482
1483                 s->state = STDOUT_STREAM_LEVEL_PREFIX;
1484                 return 0;
1485
1486         case STDOUT_STREAM_LEVEL_PREFIX:
1487                 r = parse_boolean(p);
1488                 if (r < 0) {
1489                         log_warning("Failed to parse level prefix line.");
1490                         return -EINVAL;
1491                 }
1492
1493                 s->level_prefix = !!r;
1494                 s->state = STDOUT_STREAM_FORWARD_TO_SYSLOG;
1495                 return 0;
1496
1497         case STDOUT_STREAM_FORWARD_TO_SYSLOG:
1498                 r = parse_boolean(p);
1499                 if (r < 0) {
1500                         log_warning("Failed to parse forward to syslog line.");
1501                         return -EINVAL;
1502                 }
1503
1504                 s->forward_to_syslog = !!r;
1505                 s->state = STDOUT_STREAM_FORWARD_TO_KMSG;
1506                 return 0;
1507
1508         case STDOUT_STREAM_FORWARD_TO_KMSG:
1509                 r = parse_boolean(p);
1510                 if (r < 0) {
1511                         log_warning("Failed to parse copy to kmsg line.");
1512                         return -EINVAL;
1513                 }
1514
1515                 s->forward_to_kmsg = !!r;
1516                 s->state = STDOUT_STREAM_FORWARD_TO_CONSOLE;
1517                 return 0;
1518
1519         case STDOUT_STREAM_FORWARD_TO_CONSOLE:
1520                 r = parse_boolean(p);
1521                 if (r < 0) {
1522                         log_warning("Failed to parse copy to console line.");
1523                         return -EINVAL;
1524                 }
1525
1526                 s->forward_to_console = !!r;
1527                 s->state = STDOUT_STREAM_RUNNING;
1528                 return 0;
1529
1530         case STDOUT_STREAM_RUNNING:
1531                 return stdout_stream_log(s, p);
1532         }
1533
1534         assert_not_reached("Unknown stream state");
1535 }
1536
1537 static int stdout_stream_scan(StdoutStream *s, bool force_flush) {
1538         char *p;
1539         size_t remaining;
1540         int r;
1541
1542         assert(s);
1543
1544         p = s->buffer;
1545         remaining = s->length;
1546         for (;;) {
1547                 char *end;
1548                 size_t skip;
1549
1550                 end = memchr(p, '\n', remaining);
1551                 if (end)
1552                         skip = end - p + 1;
1553                 else if (remaining >= sizeof(s->buffer) - 1) {
1554                         end = p + sizeof(s->buffer) - 1;
1555                         skip = remaining;
1556                 } else
1557                         break;
1558
1559                 *end = 0;
1560
1561                 r = stdout_stream_line(s, p);
1562                 if (r < 0)
1563                         return r;
1564
1565                 remaining -= skip;
1566                 p += skip;
1567         }
1568
1569         if (force_flush && remaining > 0) {
1570                 p[remaining] = 0;
1571                 r = stdout_stream_line(s, p);
1572                 if (r < 0)
1573                         return r;
1574
1575                 p += remaining;
1576                 remaining = 0;
1577         }
1578
1579         if (p > s->buffer) {
1580                 memmove(s->buffer, p, remaining);
1581                 s->length = remaining;
1582         }
1583
1584         return 0;
1585 }
1586
1587 static int stdout_stream_process(StdoutStream *s) {
1588         ssize_t l;
1589         int r;
1590
1591         assert(s);
1592
1593         l = read(s->fd, s->buffer+s->length, sizeof(s->buffer)-1-s->length);
1594         if (l < 0) {
1595
1596                 if (errno == EAGAIN)
1597                         return 0;
1598
1599                 log_warning("Failed to read from stream: %m");
1600                 return -errno;
1601         }
1602
1603         if (l == 0) {
1604                 r = stdout_stream_scan(s, true);
1605                 if (r < 0)
1606                         return r;
1607
1608                 return 0;
1609         }
1610
1611         s->length += l;
1612         r = stdout_stream_scan(s, false);
1613         if (r < 0)
1614                 return r;
1615
1616         return 1;
1617
1618 }
1619
1620 static void stdout_stream_free(StdoutStream *s) {
1621         assert(s);
1622
1623         if (s->server) {
1624                 assert(s->server->n_stdout_streams > 0);
1625                 s->server->n_stdout_streams --;
1626                 LIST_REMOVE(StdoutStream, stdout_stream, s->server->stdout_streams, s);
1627         }
1628
1629         if (s->fd >= 0) {
1630                 if (s->server)
1631                         epoll_ctl(s->server->epoll_fd, EPOLL_CTL_DEL, s->fd, NULL);
1632
1633                 close_nointr_nofail(s->fd);
1634         }
1635
1636 #ifdef HAVE_SELINUX
1637         if (s->security_context)
1638                 freecon(s->security_context);
1639 #endif
1640
1641         free(s->identifier);
1642         free(s);
1643 }
1644
1645 static int stdout_stream_new(Server *s) {
1646         StdoutStream *stream;
1647         int fd, r;
1648         socklen_t len;
1649         struct epoll_event ev;
1650
1651         assert(s);
1652
1653         fd = accept4(s->stdout_fd, NULL, NULL, SOCK_NONBLOCK|SOCK_CLOEXEC);
1654         if (fd < 0) {
1655                 if (errno == EAGAIN)
1656                         return 0;
1657
1658                 log_error("Failed to accept stdout connection: %m");
1659                 return -errno;
1660         }
1661
1662         if (s->n_stdout_streams >= STDOUT_STREAMS_MAX) {
1663                 log_warning("Too many stdout streams, refusing connection.");
1664                 close_nointr_nofail(fd);
1665                 return 0;
1666         }
1667
1668         stream = new0(StdoutStream, 1);
1669         if (!stream) {
1670                 log_error("Out of memory.");
1671                 close_nointr_nofail(fd);
1672                 return -ENOMEM;
1673         }
1674
1675         stream->fd = fd;
1676
1677         len = sizeof(stream->ucred);
1678         if (getsockopt(fd, SOL_SOCKET, SO_PEERCRED, &stream->ucred, &len) < 0) {
1679                 log_error("Failed to determine peer credentials: %m");
1680                 r = -errno;
1681                 goto fail;
1682         }
1683
1684 #ifdef HAVE_SELINUX
1685         if (getpeercon(fd, &stream->security_context) < 0)
1686                 log_error("Failed to determine peer security context.");
1687 #endif
1688
1689         if (shutdown(fd, SHUT_WR) < 0) {
1690                 log_error("Failed to shutdown writing side of socket: %m");
1691                 r = -errno;
1692                 goto fail;
1693         }
1694
1695         zero(ev);
1696         ev.data.ptr = stream;
1697         ev.events = EPOLLIN;
1698         if (epoll_ctl(s->epoll_fd, EPOLL_CTL_ADD, fd, &ev) < 0) {
1699                 log_error("Failed to add stream to event loop: %m");
1700                 r = -errno;
1701                 goto fail;
1702         }
1703
1704         stream->server = s;
1705         LIST_PREPEND(StdoutStream, stdout_stream, s->stdout_streams, stream);
1706         s->n_stdout_streams ++;
1707
1708         return 0;
1709
1710 fail:
1711         stdout_stream_free(stream);
1712         return r;
1713 }
1714
1715 static int parse_kernel_timestamp(char **_p, usec_t *t) {
1716         usec_t r;
1717         int k, i;
1718         char *p;
1719
1720         assert(_p);
1721         assert(*_p);
1722         assert(t);
1723
1724         p = *_p;
1725
1726         if (strlen(p) < 14 || p[0] != '[' || p[13] != ']' || p[6] != '.')
1727                 return 0;
1728
1729         r = 0;
1730
1731         for (i = 1; i <= 5; i++) {
1732                 r *= 10;
1733
1734                 if (p[i] == ' ')
1735                         continue;
1736
1737                 k = undecchar(p[i]);
1738                 if (k < 0)
1739                         return 0;
1740
1741                 r += k;
1742         }
1743
1744         for (i = 7; i <= 12; i++) {
1745                 r *= 10;
1746
1747                 k = undecchar(p[i]);
1748                 if (k < 0)
1749                         return 0;
1750
1751                 r += k;
1752         }
1753
1754         *t = r;
1755         *_p += 14;
1756         *_p += strspn(*_p, WHITESPACE);
1757
1758         return 1;
1759 }
1760
1761 static void proc_kmsg_line(Server *s, const char *p) {
1762         struct iovec iovec[N_IOVEC_META_FIELDS + 7];
1763         char *message = NULL, *syslog_priority = NULL, *syslog_pid = NULL, *syslog_facility = NULL, *syslog_identifier = NULL, *source_time = NULL;
1764         int priority = LOG_KERN | LOG_INFO;
1765         unsigned n = 0;
1766         usec_t usec;
1767         char *identifier = NULL, *pid = NULL;
1768
1769         assert(s);
1770         assert(p);
1771
1772         if (isempty(p))
1773                 return;
1774
1775         parse_syslog_priority((char **) &p, &priority);
1776
1777         if (s->forward_to_kmsg && (priority & LOG_FACMASK) != LOG_KERN)
1778                 return;
1779
1780         if (parse_kernel_timestamp((char **) &p, &usec) > 0) {
1781                 if (asprintf(&source_time, "_SOURCE_MONOTONIC_TIMESTAMP=%llu",
1782                              (unsigned long long) usec) >= 0)
1783                         IOVEC_SET_STRING(iovec[n++], source_time);
1784         }
1785
1786         IOVEC_SET_STRING(iovec[n++], "_TRANSPORT=kernel");
1787
1788         if (asprintf(&syslog_priority, "PRIORITY=%i", priority & LOG_PRIMASK) >= 0)
1789                 IOVEC_SET_STRING(iovec[n++], syslog_priority);
1790
1791         if ((priority & LOG_FACMASK) == LOG_KERN) {
1792
1793                 if (s->forward_to_syslog)
1794                         forward_syslog(s, priority, "kernel", p, NULL, NULL);
1795
1796                 IOVEC_SET_STRING(iovec[n++], "SYSLOG_IDENTIFIER=kernel");
1797         } else {
1798                 read_identifier(&p, &identifier, &pid);
1799
1800                 if (s->forward_to_syslog)
1801                         forward_syslog(s, priority, identifier, p, NULL, NULL);
1802
1803                 if (identifier) {
1804                         syslog_identifier = strappend("SYSLOG_IDENTIFIER=", identifier);
1805                         if (syslog_identifier)
1806                                 IOVEC_SET_STRING(iovec[n++], syslog_identifier);
1807                 }
1808
1809                 if (pid) {
1810                         syslog_pid = strappend("SYSLOG_PID=", pid);
1811                         if (syslog_pid)
1812                                 IOVEC_SET_STRING(iovec[n++], syslog_pid);
1813                 }
1814
1815                 if (asprintf(&syslog_facility, "SYSLOG_FACILITY=%i", LOG_FAC(priority)) >= 0)
1816                         IOVEC_SET_STRING(iovec[n++], syslog_facility);
1817         }
1818
1819         message = strappend("MESSAGE=", p);
1820         if (message)
1821                 IOVEC_SET_STRING(iovec[n++], message);
1822
1823         dispatch_message(s, iovec, n, ELEMENTSOF(iovec), NULL, NULL, NULL, 0, priority);
1824
1825         free(message);
1826         free(syslog_priority);
1827         free(syslog_identifier);
1828         free(syslog_pid);
1829         free(syslog_facility);
1830         free(source_time);
1831         free(identifier);
1832         free(pid);
1833 }
1834
1835 static void proc_kmsg_scan(Server *s) {
1836         char *p;
1837         size_t remaining;
1838
1839         assert(s);
1840
1841         p = s->proc_kmsg_buffer;
1842         remaining = s->proc_kmsg_length;
1843         for (;;) {
1844                 char *end;
1845                 size_t skip;
1846
1847                 end = memchr(p, '\n', remaining);
1848                 if (end)
1849                         skip = end - p + 1;
1850                 else if (remaining >= sizeof(s->proc_kmsg_buffer) - 1) {
1851                         end = p + sizeof(s->proc_kmsg_buffer) - 1;
1852                         skip = remaining;
1853                 } else
1854                         break;
1855
1856                 *end = 0;
1857
1858                 proc_kmsg_line(s, p);
1859
1860                 remaining -= skip;
1861                 p += skip;
1862         }
1863
1864         if (p > s->proc_kmsg_buffer) {
1865                 memmove(s->proc_kmsg_buffer, p, remaining);
1866                 s->proc_kmsg_length = remaining;
1867         }
1868 }
1869
1870 static int system_journal_open(Server *s) {
1871         int r;
1872         char *fn;
1873         sd_id128_t machine;
1874         char ids[33];
1875
1876         r = sd_id128_get_machine(&machine);
1877         if (r < 0)
1878                 return r;
1879
1880         sd_id128_to_string(machine, ids);
1881
1882         if (!s->system_journal) {
1883
1884                 /* First try to create the machine path, but not the prefix */
1885                 fn = strappend("/var/log/journal/", ids);
1886                 if (!fn)
1887                         return -ENOMEM;
1888                 (void) mkdir(fn, 0755);
1889                 free(fn);
1890
1891                 /* The create the system journal file */
1892                 fn = join("/var/log/journal/", ids, "/system.journal", NULL);
1893                 if (!fn)
1894                         return -ENOMEM;
1895
1896                 r = journal_file_open_reliably(fn, O_RDWR|O_CREAT, 0640, NULL, &s->system_journal);
1897                 free(fn);
1898
1899                 if (r >= 0) {
1900                         journal_default_metrics(&s->system_metrics, s->system_journal->fd);
1901
1902                         s->system_journal->metrics = s->system_metrics;
1903                         s->system_journal->compress = s->compress;
1904
1905                         server_fix_perms(s, s->system_journal, 0);
1906                 } else if (r < 0) {
1907
1908                         if (r != -ENOENT && r != -EROFS)
1909                                 log_warning("Failed to open system journal: %s", strerror(-r));
1910
1911                         r = 0;
1912                 }
1913         }
1914
1915         if (!s->runtime_journal) {
1916
1917                 fn = join("/run/log/journal/", ids, "/system.journal", NULL);
1918                 if (!fn)
1919                         return -ENOMEM;
1920
1921                 if (s->system_journal) {
1922
1923                         /* Try to open the runtime journal, but only
1924                          * if it already exists, so that we can flush
1925                          * it into the system journal */
1926
1927                         r = journal_file_open(fn, O_RDWR, 0640, NULL, &s->runtime_journal);
1928                         free(fn);
1929
1930                         if (r < 0) {
1931                                 if (r != -ENOENT)
1932                                         log_warning("Failed to open runtime journal: %s", strerror(-r));
1933
1934                                 r = 0;
1935                         }
1936
1937                 } else {
1938
1939                         /* OK, we really need the runtime journal, so create
1940                          * it if necessary. */
1941
1942                         (void) mkdir_parents(fn, 0755);
1943                         r = journal_file_open_reliably(fn, O_RDWR|O_CREAT, 0640, NULL, &s->runtime_journal);
1944                         free(fn);
1945
1946                         if (r < 0) {
1947                                 log_error("Failed to open runtime journal: %s", strerror(-r));
1948                                 return r;
1949                         }
1950                 }
1951
1952                 if (s->runtime_journal) {
1953                         journal_default_metrics(&s->runtime_metrics, s->runtime_journal->fd);
1954
1955                         s->runtime_journal->metrics = s->runtime_metrics;
1956                         s->runtime_journal->compress = s->compress;
1957
1958                         server_fix_perms(s, s->runtime_journal, 0);
1959                 }
1960         }
1961
1962         return r;
1963 }
1964
1965 static int server_flush_to_var(Server *s) {
1966         char path[] = "/run/log/journal/xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx";
1967         Object *o = NULL;
1968         int r;
1969         sd_id128_t machine;
1970         sd_journal *j;
1971         usec_t ts;
1972
1973         assert(s);
1974
1975         if (!s->runtime_journal)
1976                 return 0;
1977
1978         ts = now(CLOCK_MONOTONIC);
1979         if (s->var_available_timestamp + RECHECK_VAR_AVAILABLE_USEC > ts)
1980                 return 0;
1981
1982         s->var_available_timestamp = ts;
1983
1984         system_journal_open(s);
1985
1986         if (!s->system_journal)
1987                 return 0;
1988
1989         log_info("Flushing to /var...");
1990
1991         r = sd_id128_get_machine(&machine);
1992         if (r < 0) {
1993                 log_error("Failed to get machine id: %s", strerror(-r));
1994                 return r;
1995         }
1996
1997         r = sd_journal_open(&j, SD_JOURNAL_RUNTIME_ONLY);
1998         if (r < 0) {
1999                 log_error("Failed to read runtime journal: %s", strerror(-r));
2000                 return r;
2001         }
2002
2003         SD_JOURNAL_FOREACH(j) {
2004                 JournalFile *f;
2005
2006                 f = j->current_file;
2007                 assert(f && f->current_offset > 0);
2008
2009                 r = journal_file_move_to_object(f, OBJECT_ENTRY, f->current_offset, &o);
2010                 if (r < 0) {
2011                         log_error("Can't read entry: %s", strerror(-r));
2012                         goto finish;
2013                 }
2014
2015                 r = journal_file_copy_entry(f, s->system_journal, o, f->current_offset, NULL, NULL, NULL);
2016                 if (r == -E2BIG) {
2017                         log_info("Allocation limit reached.");
2018
2019                         journal_file_post_change(s->system_journal);
2020                         server_rotate(s);
2021                         server_vacuum(s);
2022
2023                         r = journal_file_copy_entry(f, s->system_journal, o, f->current_offset, NULL, NULL, NULL);
2024                 }
2025
2026                 if (r < 0) {
2027                         log_error("Can't write entry: %s", strerror(-r));
2028                         goto finish;
2029                 }
2030         }
2031
2032 finish:
2033         journal_file_post_change(s->system_journal);
2034
2035         journal_file_close(s->runtime_journal);
2036         s->runtime_journal = NULL;
2037
2038         if (r >= 0) {
2039                 sd_id128_to_string(machine, path + 17);
2040                 rm_rf(path, false, true, false);
2041         }
2042
2043         return r;
2044 }
2045
2046 static int server_read_proc_kmsg(Server *s) {
2047         ssize_t l;
2048         assert(s);
2049         assert(s->proc_kmsg_fd >= 0);
2050
2051         l = read(s->proc_kmsg_fd, s->proc_kmsg_buffer + s->proc_kmsg_length, sizeof(s->proc_kmsg_buffer) - 1 - s->proc_kmsg_length);
2052         if (l < 0) {
2053
2054                 if (errno == EAGAIN || errno == EINTR)
2055                         return 0;
2056
2057                 log_error("Failed to read from kernel: %m");
2058                 return -errno;
2059         }
2060
2061         s->proc_kmsg_length += l;
2062
2063         proc_kmsg_scan(s);
2064         return 1;
2065 }
2066
2067 static int server_flush_proc_kmsg(Server *s) {
2068         int r;
2069
2070         assert(s);
2071
2072         if (s->proc_kmsg_fd < 0)
2073                 return 0;
2074
2075         log_info("Flushing /proc/kmsg...");
2076
2077         for (;;) {
2078                 r = server_read_proc_kmsg(s);
2079                 if (r < 0)
2080                         return r;
2081
2082                 if (r == 0)
2083                         break;
2084         }
2085
2086         return 0;
2087 }
2088
2089 static int process_event(Server *s, struct epoll_event *ev) {
2090         assert(s);
2091
2092         if (ev->data.fd == s->signal_fd) {
2093                 struct signalfd_siginfo sfsi;
2094                 ssize_t n;
2095
2096                 if (ev->events != EPOLLIN) {
2097                         log_info("Got invalid event from epoll.");
2098                         return -EIO;
2099                 }
2100
2101                 n = read(s->signal_fd, &sfsi, sizeof(sfsi));
2102                 if (n != sizeof(sfsi)) {
2103
2104                         if (n >= 0)
2105                                 return -EIO;
2106
2107                         if (errno == EINTR || errno == EAGAIN)
2108                                 return 1;
2109
2110                         return -errno;
2111                 }
2112
2113                 if (sfsi.ssi_signo == SIGUSR1) {
2114                         server_flush_to_var(s);
2115                         return 0;
2116                 }
2117
2118                 log_debug("Received SIG%s", signal_to_string(sfsi.ssi_signo));
2119                 return 0;
2120
2121         } else if (ev->data.fd == s->proc_kmsg_fd) {
2122                 int r;
2123
2124                 if (ev->events != EPOLLIN) {
2125                         log_info("Got invalid event from epoll.");
2126                         return -EIO;
2127                 }
2128
2129                 r = server_read_proc_kmsg(s);
2130                 if (r < 0)
2131                         return r;
2132
2133                 return 1;
2134
2135         } else if (ev->data.fd == s->native_fd ||
2136                    ev->data.fd == s->syslog_fd) {
2137
2138                 if (ev->events != EPOLLIN) {
2139                         log_info("Got invalid event from epoll.");
2140                         return -EIO;
2141                 }
2142
2143                 for (;;) {
2144                         struct msghdr msghdr;
2145                         struct iovec iovec;
2146                         struct ucred *ucred = NULL;
2147                         struct timeval *tv = NULL;
2148                         struct cmsghdr *cmsg;
2149                         char *label = NULL;
2150                         size_t label_len = 0;
2151                         union {
2152                                 struct cmsghdr cmsghdr;
2153
2154                                 /* We use NAME_MAX space for the
2155                                  * SELinux label here. The kernel
2156                                  * currently enforces no limit, but
2157                                  * according to suggestions from the
2158                                  * SELinux people this will change and
2159                                  * it will probably be identical to
2160                                  * NAME_MAX. For now we use that, but
2161                                  * this should be updated one day when
2162                                  * the final limit is known.*/
2163                                 uint8_t buf[CMSG_SPACE(sizeof(struct ucred)) +
2164                                             CMSG_SPACE(sizeof(struct timeval)) +
2165                                             CMSG_SPACE(sizeof(int)) + /* fd */
2166                                             CMSG_SPACE(NAME_MAX)]; /* selinux label */
2167                         } control;
2168                         ssize_t n;
2169                         int v;
2170                         int *fds = NULL;
2171                         unsigned n_fds = 0;
2172
2173                         if (ioctl(ev->data.fd, SIOCINQ, &v) < 0) {
2174                                 log_error("SIOCINQ failed: %m");
2175                                 return -errno;
2176                         }
2177
2178                         if (s->buffer_size < (size_t) v) {
2179                                 void *b;
2180                                 size_t l;
2181
2182                                 l = MAX(LINE_MAX + (size_t) v, s->buffer_size * 2);
2183                                 b = realloc(s->buffer, l+1);
2184
2185                                 if (!b) {
2186                                         log_error("Couldn't increase buffer.");
2187                                         return -ENOMEM;
2188                                 }
2189
2190                                 s->buffer_size = l;
2191                                 s->buffer = b;
2192                         }
2193
2194                         zero(iovec);
2195                         iovec.iov_base = s->buffer;
2196                         iovec.iov_len = s->buffer_size;
2197
2198                         zero(control);
2199                         zero(msghdr);
2200                         msghdr.msg_iov = &iovec;
2201                         msghdr.msg_iovlen = 1;
2202                         msghdr.msg_control = &control;
2203                         msghdr.msg_controllen = sizeof(control);
2204
2205                         n = recvmsg(ev->data.fd, &msghdr, MSG_DONTWAIT|MSG_CMSG_CLOEXEC);
2206                         if (n < 0) {
2207
2208                                 if (errno == EINTR || errno == EAGAIN)
2209                                         return 1;
2210
2211                                 log_error("recvmsg() failed: %m");
2212                                 return -errno;
2213                         }
2214
2215                         for (cmsg = CMSG_FIRSTHDR(&msghdr); cmsg; cmsg = CMSG_NXTHDR(&msghdr, cmsg)) {
2216
2217                                 if (cmsg->cmsg_level == SOL_SOCKET &&
2218                                     cmsg->cmsg_type == SCM_CREDENTIALS &&
2219                                     cmsg->cmsg_len == CMSG_LEN(sizeof(struct ucred)))
2220                                         ucred = (struct ucred*) CMSG_DATA(cmsg);
2221                                 else if (cmsg->cmsg_level == SOL_SOCKET &&
2222                                          cmsg->cmsg_type == SCM_SECURITY) {
2223                                         label = (char*) CMSG_DATA(cmsg);
2224                                         label_len = cmsg->cmsg_len - CMSG_LEN(0);
2225                                 } else if (cmsg->cmsg_level == SOL_SOCKET &&
2226                                          cmsg->cmsg_type == SO_TIMESTAMP &&
2227                                          cmsg->cmsg_len == CMSG_LEN(sizeof(struct timeval)))
2228                                         tv = (struct timeval*) CMSG_DATA(cmsg);
2229                                 else if (cmsg->cmsg_level == SOL_SOCKET &&
2230                                          cmsg->cmsg_type == SCM_RIGHTS) {
2231                                         fds = (int*) CMSG_DATA(cmsg);
2232                                         n_fds = (cmsg->cmsg_len - CMSG_LEN(0)) / sizeof(int);
2233                                 }
2234                         }
2235
2236                         if (ev->data.fd == s->syslog_fd) {
2237                                 char *e;
2238
2239                                 if (n > 0 && n_fds == 0) {
2240                                         e = memchr(s->buffer, '\n', n);
2241                                         if (e)
2242                                                 *e = 0;
2243                                         else
2244                                                 s->buffer[n] = 0;
2245
2246                                         process_syslog_message(s, strstrip(s->buffer), ucred, tv, label, label_len);
2247                                 } else if (n_fds > 0)
2248                                         log_warning("Got file descriptors via syslog socket. Ignoring.");
2249
2250                         } else {
2251                                 if (n > 0 && n_fds == 0)
2252                                         process_native_message(s, s->buffer, n, ucred, tv, label, label_len);
2253                                 else if (n == 0 && n_fds == 1)
2254                                         process_native_file(s, fds[0], ucred, tv, label, label_len);
2255                                 else if (n_fds > 0)
2256                                         log_warning("Got too many file descriptors via native socket. Ignoring.");
2257                         }
2258
2259                         close_many(fds, n_fds);
2260                 }
2261
2262                 return 1;
2263
2264         } else if (ev->data.fd == s->stdout_fd) {
2265
2266                 if (ev->events != EPOLLIN) {
2267                         log_info("Got invalid event from epoll.");
2268                         return -EIO;
2269                 }
2270
2271                 stdout_stream_new(s);
2272                 return 1;
2273
2274         } else {
2275                 StdoutStream *stream;
2276
2277                 if ((ev->events|EPOLLIN|EPOLLHUP) != (EPOLLIN|EPOLLHUP)) {
2278                         log_info("Got invalid event from epoll.");
2279                         return -EIO;
2280                 }
2281
2282                 /* If it is none of the well-known fds, it must be an
2283                  * stdout stream fd. Note that this is a bit ugly here
2284                  * (since we rely that none of the well-known fds
2285                  * could be interpreted as pointer), but nonetheless
2286                  * safe, since the well-known fds would never get an
2287                  * fd > 4096, i.e. beyond the first memory page */
2288
2289                 stream = ev->data.ptr;
2290
2291                 if (stdout_stream_process(stream) <= 0)
2292                         stdout_stream_free(stream);
2293
2294                 return 1;
2295         }
2296
2297         log_error("Unknown event.");
2298         return 0;
2299 }
2300
2301 static int open_syslog_socket(Server *s) {
2302         union sockaddr_union sa;
2303         int one, r;
2304         struct epoll_event ev;
2305
2306         assert(s);
2307
2308         if (s->syslog_fd < 0) {
2309
2310                 s->syslog_fd = socket(AF_UNIX, SOCK_DGRAM|SOCK_CLOEXEC|SOCK_NONBLOCK, 0);
2311                 if (s->syslog_fd < 0) {
2312                         log_error("socket() failed: %m");
2313                         return -errno;
2314                 }
2315
2316                 zero(sa);
2317                 sa.un.sun_family = AF_UNIX;
2318                 strncpy(sa.un.sun_path, "/dev/log", sizeof(sa.un.sun_path));
2319
2320                 unlink(sa.un.sun_path);
2321
2322                 r = bind(s->syslog_fd, &sa.sa, offsetof(union sockaddr_union, un.sun_path) + strlen(sa.un.sun_path));
2323                 if (r < 0) {
2324                         log_error("bind() failed: %m");
2325                         return -errno;
2326                 }
2327
2328                 chmod(sa.un.sun_path, 0666);
2329         } else
2330                 fd_nonblock(s->syslog_fd, 1);
2331
2332         one = 1;
2333         r = setsockopt(s->syslog_fd, SOL_SOCKET, SO_PASSCRED, &one, sizeof(one));
2334         if (r < 0) {
2335                 log_error("SO_PASSCRED failed: %m");
2336                 return -errno;
2337         }
2338
2339 #ifdef HAVE_SELINUX
2340         one = 1;
2341         r = setsockopt(s->syslog_fd, SOL_SOCKET, SO_PASSSEC, &one, sizeof(one));
2342         if (r < 0)
2343                 log_warning("SO_PASSSEC failed: %m");
2344 #endif
2345
2346         one = 1;
2347         r = setsockopt(s->syslog_fd, SOL_SOCKET, SO_TIMESTAMP, &one, sizeof(one));
2348         if (r < 0) {
2349                 log_error("SO_TIMESTAMP failed: %m");
2350                 return -errno;
2351         }
2352
2353         zero(ev);
2354         ev.events = EPOLLIN;
2355         ev.data.fd = s->syslog_fd;
2356         if (epoll_ctl(s->epoll_fd, EPOLL_CTL_ADD, s->syslog_fd, &ev) < 0) {
2357                 log_error("Failed to add syslog server fd to epoll object: %m");
2358                 return -errno;
2359         }
2360
2361         return 0;
2362 }
2363
2364 static int open_native_socket(Server*s) {
2365         union sockaddr_union sa;
2366         int one, r;
2367         struct epoll_event ev;
2368
2369         assert(s);
2370
2371         if (s->native_fd < 0) {
2372
2373                 s->native_fd = socket(AF_UNIX, SOCK_DGRAM|SOCK_CLOEXEC|SOCK_NONBLOCK, 0);
2374                 if (s->native_fd < 0) {
2375                         log_error("socket() failed: %m");
2376                         return -errno;
2377                 }
2378
2379                 zero(sa);
2380                 sa.un.sun_family = AF_UNIX;
2381                 strncpy(sa.un.sun_path, "/run/systemd/journal/socket", sizeof(sa.un.sun_path));
2382
2383                 unlink(sa.un.sun_path);
2384
2385                 r = bind(s->native_fd, &sa.sa, offsetof(union sockaddr_union, un.sun_path) + strlen(sa.un.sun_path));
2386                 if (r < 0) {
2387                         log_error("bind() failed: %m");
2388                         return -errno;
2389                 }
2390
2391                 chmod(sa.un.sun_path, 0666);
2392         } else
2393                 fd_nonblock(s->native_fd, 1);
2394
2395         one = 1;
2396         r = setsockopt(s->native_fd, SOL_SOCKET, SO_PASSCRED, &one, sizeof(one));
2397         if (r < 0) {
2398                 log_error("SO_PASSCRED failed: %m");
2399                 return -errno;
2400         }
2401
2402 #ifdef HAVE_SELINUX
2403         one = 1;
2404         r = setsockopt(s->syslog_fd, SOL_SOCKET, SO_PASSSEC, &one, sizeof(one));
2405         if (r < 0)
2406                 log_warning("SO_PASSSEC failed: %m");
2407 #endif
2408
2409         one = 1;
2410         r = setsockopt(s->native_fd, SOL_SOCKET, SO_TIMESTAMP, &one, sizeof(one));
2411         if (r < 0) {
2412                 log_error("SO_TIMESTAMP failed: %m");
2413                 return -errno;
2414         }
2415
2416         zero(ev);
2417         ev.events = EPOLLIN;
2418         ev.data.fd = s->native_fd;
2419         if (epoll_ctl(s->epoll_fd, EPOLL_CTL_ADD, s->native_fd, &ev) < 0) {
2420                 log_error("Failed to add native server fd to epoll object: %m");
2421                 return -errno;
2422         }
2423
2424         return 0;
2425 }
2426
2427 static int open_stdout_socket(Server *s) {
2428         union sockaddr_union sa;
2429         int r;
2430         struct epoll_event ev;
2431
2432         assert(s);
2433
2434         if (s->stdout_fd < 0) {
2435
2436                 s->stdout_fd = socket(AF_UNIX, SOCK_STREAM|SOCK_CLOEXEC|SOCK_NONBLOCK, 0);
2437                 if (s->stdout_fd < 0) {
2438                         log_error("socket() failed: %m");
2439                         return -errno;
2440                 }
2441
2442                 zero(sa);
2443                 sa.un.sun_family = AF_UNIX;
2444                 strncpy(sa.un.sun_path, "/run/systemd/journal/stdout", sizeof(sa.un.sun_path));
2445
2446                 unlink(sa.un.sun_path);
2447
2448                 r = bind(s->stdout_fd, &sa.sa, offsetof(union sockaddr_union, un.sun_path) + strlen(sa.un.sun_path));
2449                 if (r < 0) {
2450                         log_error("bind() failed: %m");
2451                         return -errno;
2452                 }
2453
2454                 chmod(sa.un.sun_path, 0666);
2455
2456                 if (listen(s->stdout_fd, SOMAXCONN) < 0) {
2457                         log_error("liste() failed: %m");
2458                         return -errno;
2459                 }
2460         } else
2461                 fd_nonblock(s->stdout_fd, 1);
2462
2463         zero(ev);
2464         ev.events = EPOLLIN;
2465         ev.data.fd = s->stdout_fd;
2466         if (epoll_ctl(s->epoll_fd, EPOLL_CTL_ADD, s->stdout_fd, &ev) < 0) {
2467                 log_error("Failed to add stdout server fd to epoll object: %m");
2468                 return -errno;
2469         }
2470
2471         return 0;
2472 }
2473
2474 static int open_proc_kmsg(Server *s) {
2475         struct epoll_event ev;
2476
2477         assert(s);
2478
2479         if (!s->import_proc_kmsg)
2480                 return 0;
2481
2482
2483         s->proc_kmsg_fd = open("/proc/kmsg", O_CLOEXEC|O_NONBLOCK|O_NOCTTY);
2484         if (s->proc_kmsg_fd < 0) {
2485                 log_warning("Failed to open /proc/kmsg, ignoring: %m");
2486                 return 0;
2487         }
2488
2489         zero(ev);
2490         ev.events = EPOLLIN;
2491         ev.data.fd = s->proc_kmsg_fd;
2492         if (epoll_ctl(s->epoll_fd, EPOLL_CTL_ADD, s->proc_kmsg_fd, &ev) < 0) {
2493                 log_error("Failed to add /proc/kmsg fd to epoll object: %m");
2494                 return -errno;
2495         }
2496
2497         return 0;
2498 }
2499
2500 static int open_signalfd(Server *s) {
2501         sigset_t mask;
2502         struct epoll_event ev;
2503
2504         assert(s);
2505
2506         assert_se(sigemptyset(&mask) == 0);
2507         sigset_add_many(&mask, SIGINT, SIGTERM, SIGUSR1, -1);
2508         assert_se(sigprocmask(SIG_SETMASK, &mask, NULL) == 0);
2509
2510         s->signal_fd = signalfd(-1, &mask, SFD_NONBLOCK|SFD_CLOEXEC);
2511         if (s->signal_fd < 0) {
2512                 log_error("signalfd(): %m");
2513                 return -errno;
2514         }
2515
2516         zero(ev);
2517         ev.events = EPOLLIN;
2518         ev.data.fd = s->signal_fd;
2519
2520         if (epoll_ctl(s->epoll_fd, EPOLL_CTL_ADD, s->signal_fd, &ev) < 0) {
2521                 log_error("epoll_ctl(): %m");
2522                 return -errno;
2523         }
2524
2525         return 0;
2526 }
2527
2528 static int server_parse_proc_cmdline(Server *s) {
2529         char *line, *w, *state;
2530         int r;
2531         size_t l;
2532
2533         if (detect_container(NULL) > 0)
2534                 return 0;
2535
2536         r = read_one_line_file("/proc/cmdline", &line);
2537         if (r < 0) {
2538                 log_warning("Failed to read /proc/cmdline, ignoring: %s", strerror(-r));
2539                 return 0;
2540         }
2541
2542         FOREACH_WORD_QUOTED(w, l, line, state) {
2543                 char *word;
2544
2545                 word = strndup(w, l);
2546                 if (!word) {
2547                         r = -ENOMEM;
2548                         goto finish;
2549                 }
2550
2551                 if (startswith(word, "systemd_journald.forward_to_syslog=")) {
2552                         r = parse_boolean(word + 35);
2553                         if (r < 0)
2554                                 log_warning("Failed to parse forward to syslog switch %s. Ignoring.", word + 35);
2555                         else
2556                                 s->forward_to_syslog = r;
2557                 } else if (startswith(word, "systemd_journald.forward_to_kmsg=")) {
2558                         r = parse_boolean(word + 33);
2559                         if (r < 0)
2560                                 log_warning("Failed to parse forward to kmsg switch %s. Ignoring.", word + 33);
2561                         else
2562                                 s->forward_to_kmsg = r;
2563                 } else if (startswith(word, "systemd_journald.forward_to_console=")) {
2564                         r = parse_boolean(word + 36);
2565                         if (r < 0)
2566                                 log_warning("Failed to parse forward to console switch %s. Ignoring.", word + 36);
2567                         else
2568                                 s->forward_to_console = r;
2569                 }
2570
2571                 free(word);
2572         }
2573
2574         r = 0;
2575
2576 finish:
2577         free(line);
2578         return r;
2579 }
2580
2581 static int server_parse_config_file(Server *s) {
2582         FILE *f;
2583         const char *fn;
2584         int r;
2585
2586         assert(s);
2587
2588         fn = "/etc/systemd/journald.conf";
2589         f = fopen(fn, "re");
2590         if (!f) {
2591                 if (errno == ENOENT)
2592                         return 0;
2593
2594                 log_warning("Failed to open configuration file %s: %m", fn);
2595                 return -errno;
2596         }
2597
2598         r = config_parse(fn, f, "Journal\0", config_item_perf_lookup, (void*) journald_gperf_lookup, false, s);
2599         if (r < 0)
2600                 log_warning("Failed to parse configuration file: %s", strerror(-r));
2601
2602         fclose(f);
2603
2604         return r;
2605 }
2606
2607 static int server_init(Server *s) {
2608         int n, r, fd;
2609
2610         assert(s);
2611
2612         zero(*s);
2613         s->syslog_fd = s->native_fd = s->stdout_fd = s->signal_fd = s->epoll_fd = s->proc_kmsg_fd = -1;
2614         s->compress = true;
2615
2616         s->rate_limit_interval = DEFAULT_RATE_LIMIT_INTERVAL;
2617         s->rate_limit_burst = DEFAULT_RATE_LIMIT_BURST;
2618
2619         s->forward_to_syslog = true;
2620         s->import_proc_kmsg = true;
2621
2622         memset(&s->system_metrics, 0xFF, sizeof(s->system_metrics));
2623         memset(&s->runtime_metrics, 0xFF, sizeof(s->runtime_metrics));
2624
2625         server_parse_config_file(s);
2626         server_parse_proc_cmdline(s);
2627
2628         s->user_journals = hashmap_new(trivial_hash_func, trivial_compare_func);
2629         if (!s->user_journals) {
2630                 log_error("Out of memory.");
2631                 return -ENOMEM;
2632         }
2633
2634         s->epoll_fd = epoll_create1(EPOLL_CLOEXEC);
2635         if (s->epoll_fd < 0) {
2636                 log_error("Failed to create epoll object: %m");
2637                 return -errno;
2638         }
2639
2640         n = sd_listen_fds(true);
2641         if (n < 0) {
2642                 log_error("Failed to read listening file descriptors from environment: %s", strerror(-n));
2643                 return n;
2644         }
2645
2646         for (fd = SD_LISTEN_FDS_START; fd < SD_LISTEN_FDS_START + n; fd++) {
2647
2648                 if (sd_is_socket_unix(fd, SOCK_DGRAM, -1, "/run/systemd/journal/socket", 0) > 0) {
2649
2650                         if (s->native_fd >= 0) {
2651                                 log_error("Too many native sockets passed.");
2652                                 return -EINVAL;
2653                         }
2654
2655                         s->native_fd = fd;
2656
2657                 } else if (sd_is_socket_unix(fd, SOCK_STREAM, 1, "/run/systemd/journal/stdout", 0) > 0) {
2658
2659                         if (s->stdout_fd >= 0) {
2660                                 log_error("Too many stdout sockets passed.");
2661                                 return -EINVAL;
2662                         }
2663
2664                         s->stdout_fd = fd;
2665
2666                 } else if (sd_is_socket_unix(fd, SOCK_DGRAM, -1, "/dev/log", 0) > 0) {
2667
2668                         if (s->syslog_fd >= 0) {
2669                                 log_error("Too many /dev/log sockets passed.");
2670                                 return -EINVAL;
2671                         }
2672
2673                         s->syslog_fd = fd;
2674
2675                 } else {
2676                         log_error("Unknown socket passed.");
2677                         return -EINVAL;
2678                 }
2679         }
2680
2681         r = open_syslog_socket(s);
2682         if (r < 0)
2683                 return r;
2684
2685         r = open_native_socket(s);
2686         if (r < 0)
2687                 return r;
2688
2689         r = open_stdout_socket(s);
2690         if (r < 0)
2691                 return r;
2692
2693         r = open_proc_kmsg(s);
2694         if (r < 0)
2695                 return r;
2696
2697         r = open_signalfd(s);
2698         if (r < 0)
2699                 return r;
2700
2701         s->rate_limit = journal_rate_limit_new(s->rate_limit_interval, s->rate_limit_burst);
2702         if (!s->rate_limit)
2703                 return -ENOMEM;
2704
2705         r = system_journal_open(s);
2706         if (r < 0)
2707                 return r;
2708
2709         return 0;
2710 }
2711
2712 static void server_done(Server *s) {
2713         JournalFile *f;
2714         assert(s);
2715
2716         while (s->stdout_streams)
2717                 stdout_stream_free(s->stdout_streams);
2718
2719         if (s->system_journal)
2720                 journal_file_close(s->system_journal);
2721
2722         if (s->runtime_journal)
2723                 journal_file_close(s->runtime_journal);
2724
2725         while ((f = hashmap_steal_first(s->user_journals)))
2726                 journal_file_close(f);
2727
2728         hashmap_free(s->user_journals);
2729
2730         if (s->epoll_fd >= 0)
2731                 close_nointr_nofail(s->epoll_fd);
2732
2733         if (s->signal_fd >= 0)
2734                 close_nointr_nofail(s->signal_fd);
2735
2736         if (s->syslog_fd >= 0)
2737                 close_nointr_nofail(s->syslog_fd);
2738
2739         if (s->native_fd >= 0)
2740                 close_nointr_nofail(s->native_fd);
2741
2742         if (s->stdout_fd >= 0)
2743                 close_nointr_nofail(s->stdout_fd);
2744
2745         if (s->proc_kmsg_fd >= 0)
2746                 close_nointr_nofail(s->proc_kmsg_fd);
2747
2748         if (s->rate_limit)
2749                 journal_rate_limit_free(s->rate_limit);
2750
2751         free(s->buffer);
2752 }
2753
2754 int main(int argc, char *argv[]) {
2755         Server server;
2756         int r;
2757
2758         /* if (getppid() != 1) { */
2759         /*         log_error("This program should be invoked by init only."); */
2760         /*         return EXIT_FAILURE; */
2761         /* } */
2762
2763         if (argc > 1) {
2764                 log_error("This program does not take arguments.");
2765                 return EXIT_FAILURE;
2766         }
2767
2768         log_set_target(LOG_TARGET_CONSOLE);
2769         log_set_facility(LOG_SYSLOG);
2770         log_parse_environment();
2771         log_open();
2772
2773         umask(0022);
2774
2775         r = server_init(&server);
2776         if (r < 0)
2777                 goto finish;
2778
2779         server_vacuum(&server);
2780         server_flush_to_var(&server);
2781         server_flush_proc_kmsg(&server);
2782
2783         log_debug("systemd-journald running as pid %lu", (unsigned long) getpid());
2784         driver_message(&server, SD_MESSAGE_JOURNAL_START, "Journal started");
2785
2786         sd_notify(false,
2787                   "READY=1\n"
2788                   "STATUS=Processing requests...");
2789
2790         for (;;) {
2791                 struct epoll_event event;
2792
2793                 r = epoll_wait(server.epoll_fd, &event, 1, -1);
2794                 if (r < 0) {
2795
2796                         if (errno == EINTR)
2797                                 continue;
2798
2799                         log_error("epoll_wait() failed: %m");
2800                         r = -errno;
2801                         goto finish;
2802                 } else if (r == 0)
2803                         break;
2804
2805                 r = process_event(&server, &event);
2806                 if (r < 0)
2807                         goto finish;
2808                 else if (r == 0)
2809                         break;
2810         }
2811
2812         log_debug("systemd-journald stopped as pid %lu", (unsigned long) getpid());
2813         driver_message(&server, SD_MESSAGE_JOURNAL_STOP, "Journal stopped");
2814
2815 finish:
2816         sd_notify(false,
2817                   "STATUS=Shutting down...");
2818
2819         server_done(&server);
2820
2821         return r < 0 ? EXIT_FAILURE : EXIT_SUCCESS;
2822 }