src/journal/sd-journal.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 <errno.h>
  23 #include <fcntl.h>
  24 #include <stddef.h>
  25 #include <unistd.h>
  26 #include <sys/inotify.h>
  27
  28 #include "sd-journal.h"
  29 #include "journal-def.h"
  30 #include "journal-file.h"
  31 #include "hashmap.h"
  32 #include "list.h"
  33 #include "lookup3.h"
  34
  35 #define JOURNAL_FILES_MAX 1024
  36
  37 typedef struct Match Match;
  38
  39 struct Match {
  40         char *data;
  41         size_t size;
  42         uint64_t le_hash;
  43
  44         LIST_FIELDS(Match, matches);
  45 };
  46
  47 typedef enum location_type {
  48         LOCATION_HEAD,
  49         LOCATION_TAIL,
  50         LOCATION_DISCRETE
  51 } location_type_t;
  52
  53 typedef struct Location {
  54         location_type_t type;
  55
  56         uint64_t seqnum;
  57         sd_id128_t seqnum_id;
  58         bool seqnum_set;
  59
  60         uint64_t realtime;
  61         bool realtime_set;
  62
  63         uint64_t monotonic;
  64         sd_id128_t boot_id;
  65         bool monotonic_set;
  66
  67         uint64_t xor_hash;
  68         bool xor_hash_set;
  69 } Location;
  70
  71 struct sd_journal {
  72         Hashmap *files;
  73
  74         Location current_location;
  75         JournalFile *current_file;
  76         uint64_t current_field;
  77
  78         int inotify_fd;
  79         Hashmap *inotify_wd_dirs;
  80         Hashmap *inotify_wd_roots;
  81
  82         LIST_HEAD(Match, matches);
  83         unsigned n_matches;
  84 };
  85
  86 static void detach_location(sd_journal *j) {
  87         Iterator i;
  88         JournalFile *f;
  89
  90         assert(j);
  91
  92         j->current_file = NULL;
  93         j->current_field = 0;
  94
  95         HASHMAP_FOREACH(f, j->files, i)
  96                 f->current_offset = 0;
  97 }
  98
  99 static void reset_location(sd_journal *j) {
 100         assert(j);
 101
 102         detach_location(j);
 103         zero(j->current_location);
 104 }
 105
 106 static void init_location(Location *l, JournalFile *f, Object *o) {
 107         assert(l);
 108         assert(f);
 109         assert(o->object.type == OBJECT_ENTRY);
 110
 111         l->type = LOCATION_DISCRETE;
 112         l->seqnum = le64toh(o->entry.seqnum);
 113         l->seqnum_id = f->header->seqnum_id;
 114         l->realtime = le64toh(o->entry.realtime);
 115         l->monotonic = le64toh(o->entry.monotonic);
 116         l->boot_id = le64toh(o->entry.boot_id);
 117         l->xor_hash = le64toh(o->entry.xor_hash);
 118
 119         l->seqnum_set = l->realtime_set = l->monotonic_set = l->xor_hash_set = true;
 120 }
 121
 122 static void set_location(sd_journal *j, JournalFile *f, Object *o, uint64_t offset) {
 123         assert(j);
 124         assert(f);
 125         assert(o);
 126
 127         init_location(&j->current_location, f, o);
 128
 129         j->current_file = f;
 130         j->current_field = 0;
 131
 132         f->current_offset = offset;
 133 }
 134
 135 static int same_field(const void *_a, size_t s, const void *_b, size_t t) {
 136         const uint8_t *a = _a, *b = _b;
 137         size_t j;
 138         bool a_good = false, b_good = false, different = false;
 139
 140         for (j = 0; j < s && j < t; j++) {
 141
 142                 if (a[j] == '=')
 143                         a_good = true;
 144                 if (b[j] == '=')
 145                         b_good = true;
 146                 if (a[j] != b[j])
 147                         different = true;
 148
 149                 if (a_good && b_good)
 150                         return different ? 0 : 1;
 151         }
 152
 153         return -EINVAL;
 154 }
 155
 156 int sd_journal_add_match(sd_journal *j, const void *data, size_t size) {
 157         Match *m, *after = NULL;
 158         uint64_t le_hash;
 159
 160         assert(j);
 161
 162         if (size <= 0)
 163                 return -EINVAL;
 164
 165         assert(data);
 166
 167         le_hash = htole64(hash64(data, size));
 168
 169         LIST_FOREACH(matches, m, j->matches) {
 170                 int r;
 171
 172                 if (m->le_hash == le_hash &&
 173                     m->size == size &&
 174                     memcmp(m->data, data, size) == 0)
 175                         return 0;
 176
 177                 r = same_field(data, size, m->data, m->size);
 178                 if (r < 0)
 179                         return r;
 180                 else if (r > 0)
 181                         after = m;
 182         }
 183
 184         m = new0(Match, 1);
 185         if (!m)
 186                 return -ENOMEM;
 187
 188         m->size = size;
 189
 190         m->data = malloc(m->size);
 191         if (!m->data) {
 192                 free(m);
 193                 return -ENOMEM;
 194         }
 195
 196         memcpy(m->data, data, size);
 197         m->le_hash = le_hash;
 198
 199         /* Matches for the same fields we order adjacent to each
 200          * other */
 201         LIST_INSERT_AFTER(Match, matches, j->matches, after, m);
 202         j->n_matches ++;
 203
 204         detach_location(j);
 205
 206         return 0;
 207 }
 208
 209 void sd_journal_flush_matches(sd_journal *j) {
 210         assert(j);
 211
 212         while (j->matches) {
 213                 Match *m = j->matches;
 214
 215                 LIST_REMOVE(Match, matches, j->matches, m);
 216                 free(m->data);
 217                 free(m);
 218         }
 219
 220         j->n_matches = 0;
 221
 222         detach_location(j);
 223 }
 224
 225 static int compare_order(JournalFile *af, Object *ao,
 226                          JournalFile *bf, Object *bo) {
 227
 228         uint64_t a, b;
 229
 230         assert(af);
 231         assert(ao);
 232         assert(bf);
 233         assert(bo);
 234
 235         /* We operate on two different files here, hence we can access
 236          * two objects at the same time, which we normally can't.
 237          *
 238          * If contents and timestamps match, these entries are
 239          * identical, even if the seqnum does not match */
 240
 241         if (sd_id128_equal(ao->entry.boot_id, bo->entry.boot_id) &&
 242             ao->entry.monotonic == bo->entry.monotonic &&
 243             ao->entry.realtime == bo->entry.realtime &&
 244             ao->entry.xor_hash == bo->entry.xor_hash)
 245                 return 0;
 246
 247         if (sd_id128_equal(af->header->seqnum_id, bf->header->seqnum_id)) {
 248
 249                 /* If this is from the same seqnum source, compare
 250                  * seqnums */
 251                 a = le64toh(ao->entry.seqnum);
 252                 b = le64toh(bo->entry.seqnum);
 253
 254                 if (a < b)
 255                         return -1;
 256                 if (a > b)
 257                         return 1;
 258
 259                 /* Wow! This is weird, different data but the same
 260                  * seqnums? Something is borked, but let's make the
 261                  * best of it and compare by time. */
 262         }
 263
 264         if (sd_id128_equal(ao->entry.boot_id, bo->entry.boot_id)) {
 265
 266                 /* If the boot id matches compare monotonic time */
 267                 a = le64toh(ao->entry.monotonic);
 268                 b = le64toh(bo->entry.monotonic);
 269
 270                 if (a < b)
 271                         return -1;
 272                 if (a > b)
 273                         return 1;
 274         }
 275
 276         /* Otherwise compare UTC time */
 277         a = le64toh(ao->entry.realtime);
 278         b = le64toh(ao->entry.realtime);
 279
 280         if (a < b)
 281                 return -1;
 282         if (a > b)
 283                 return 1;
 284
 285         /* Finally, compare by contents */
 286         a = le64toh(ao->entry.xor_hash);
 287         b = le64toh(ao->entry.xor_hash);
 288
 289         if (a < b)
 290                 return -1;
 291         if (a > b)
 292                 return 1;
 293
 294         return 0;
 295 }
 296
 297 static int compare_with_location(JournalFile *af, Object *ao, Location *l) {
 298         uint64_t a;
 299
 300         assert(af);
 301         assert(ao);
 302         assert(l);
 303         assert(l->type == LOCATION_DISCRETE);
 304
 305         if (l->monotonic_set &&
 306             sd_id128_equal(ao->entry.boot_id, l->boot_id) &&
 307             l->realtime_set &&
 308             le64toh(ao->entry.realtime) == l->realtime &&
 309             l->xor_hash_set &&
 310             le64toh(ao->entry.xor_hash) == l->xor_hash)
 311                 return 0;
 312
 313         if (l->seqnum_set &&
 314             sd_id128_equal(af->header->seqnum_id, l->seqnum_id)) {
 315
 316                 a = le64toh(ao->entry.seqnum);
 317
 318                 if (a < l->seqnum)
 319                         return -1;
 320                 if (a > l->seqnum)
 321                         return 1;
 322         }
 323
 324         if (l->monotonic_set &&
 325             sd_id128_equal(ao->entry.boot_id, l->boot_id)) {
 326
 327                 a = le64toh(ao->entry.monotonic);
 328
 329                 if (a < l->monotonic)
 330                         return -1;
 331                 if (a > l->monotonic)
 332                         return 1;
 333         }
 334
 335         if (l->realtime_set) {
 336
 337                 a = le64toh(ao->entry.realtime);
 338
 339                 if (a < l->realtime)
 340                         return -1;
 341                 if (a > l->realtime)
 342                         return 1;
 343         }
 344
 345         if (l->xor_hash_set) {
 346                 a = le64toh(ao->entry.xor_hash);
 347
 348                 if (a < l->xor_hash)
 349                         return -1;
 350                 if (a > l->xor_hash)
 351                         return 1;
 352         }
 353
 354         return 0;
 355 }
 356
 357 static int find_location(sd_journal *j, JournalFile *f, direction_t direction, Object **ret, uint64_t *offset) {
 358         Object *o = NULL;
 359         uint64_t p = 0;
 360         int r;
 361
 362         assert(j);
 363
 364         if (!j->matches) {
 365                 /* No matches is simple */
 366
 367                 if (j->current_location.type == LOCATION_HEAD)
 368                         r = journal_file_next_entry(f, NULL, 0, DIRECTION_DOWN, &o, &p);
 369                 else if (j->current_location.type == LOCATION_TAIL)
 370                         r = journal_file_next_entry(f, NULL, 0, DIRECTION_UP, &o, &p);
 371                 else if (j->current_location.seqnum_set &&
 372                          sd_id128_equal(j->current_location.seqnum_id, f->header->seqnum_id))
 373                         r = journal_file_move_to_entry_by_seqnum(f, j->current_location.seqnum, direction, &o, &p);
 374                 else if (j->current_location.monotonic_set)
 375                         r = journal_file_move_to_entry_by_monotonic(f, j->current_location.boot_id, j->current_location.monotonic, direction, &o, &p);
 376                 else if (j->current_location.realtime_set)
 377                         r = journal_file_move_to_entry_by_realtime(f, j->current_location.realtime, direction, &o, &p);
 378                 else
 379                         r = journal_file_next_entry(f, NULL, 0, direction, &o, &p);
 380
 381                 if (r <= 0)
 382                         return r;
 383
 384         } else  {
 385                 Match *m, *term_match = NULL;
 386                 Object *to = NULL;
 387                 uint64_t tp = 0;
 388
 389                 /* We have matches, first, let's jump to the monotonic
 390                  * position if we have any, since it implies a
 391                  * match. */
 392
 393                 if (j->current_location.type == LOCATION_DISCRETE &&
 394                     j->current_location.monotonic_set) {
 395
 396                         r = journal_file_move_to_entry_by_monotonic(f, j->current_location.boot_id, j->current_location.monotonic, direction, &o, &p);
 397                         if (r <= 0)
 398                                 return r;
 399                 }
 400
 401                 LIST_FOREACH(matches, m, j->matches) {
 402                         Object *c, *d;
 403                         uint64_t cp, dp;
 404
 405                         r = journal_file_find_data_object_with_hash(f, m->data, m->size, m->le_hash, &d, &dp);
 406                         if (r <= 0)
 407                                 return r;
 408
 409                         if (j->current_location.type == LOCATION_HEAD)
 410                                 r = journal_file_next_entry_for_data(f, NULL, 0, dp, DIRECTION_DOWN, &c, &cp);
 411                         else if (j->current_location.type == LOCATION_TAIL)
 412                                 r = journal_file_next_entry_for_data(f, NULL, 0, dp, DIRECTION_UP, &c, &cp);
 413                         else if (j->current_location.seqnum_set &&
 414                                  sd_id128_equal(j->current_location.seqnum_id, f->header->seqnum_id))
 415                                 r = journal_file_move_to_entry_by_seqnum_for_data(f, dp, j->current_location.seqnum, direction, &c, &cp);
 416                         else if (j->current_location.realtime_set)
 417                                 r = journal_file_move_to_entry_by_realtime_for_data(f, dp, j->current_location.realtime, direction, &c, &cp);
 418                         else
 419                                 r = journal_file_next_entry_for_data(f, NULL, 0, dp, direction, &c, &cp);
 420
 421                         if (!term_match) {
 422                                 term_match = m;
 423
 424                                 if (r > 0) {
 425                                         to = c;
 426                                         tp = cp;
 427                                 }
 428                         } else if (same_field(term_match->data, term_match->size, m->data, m->size)) {
 429
 430                                 /* Same field as previous match... */
 431                                 if (r > 0) {
 432
 433                                         /* Find the earliest of the OR matches */
 434
 435                                         if (!to ||
 436                                             (direction == DIRECTION_DOWN && cp < tp) ||
 437                                             (direction == DIRECTION_UP && cp > tp)) {
 438                                                 to = c;
 439                                                 tp = tp;
 440                                         }
 441
 442                                 }
 443
 444                         } else {
 445
 446                                 /* Previous term is finished, did anything match? */
 447                                 if (!to)
 448                                         return 0;
 449
 450                                 /* Find the last of the AND matches */
 451                                 if (!o ||
 452                                     (direction == DIRECTION_DOWN && tp > p) ||
 453                                     (direction == DIRECTION_UP && tp < p)) {
 454                                         o = to;
 455                                         p = tp;
 456                                 }
 457
 458                                 term_match = m;
 459
 460                                 if (r > 0) {
 461                                         to = c;
 462                                         tp = cp;
 463                                 } else {
 464                                         to = NULL;
 465                                         tp = 0;
 466                                 }
 467                         }
 468                 }
 469
 470                 /* Last term is finished, did anything match? */
 471                 if (!to)
 472                         return 0;
 473
 474                 if (!o ||
 475                     (direction == DIRECTION_DOWN && tp > p) ||
 476                     (direction == DIRECTION_UP && tp < p)) {
 477                         o = to;
 478                         p = tp;
 479                 }
 480
 481                 if (!o)
 482                         return 0;
 483         }
 484
 485         if (ret)
 486                 *ret = o;
 487
 488         if (offset)
 489                 *offset = p;
 490
 491         return 1;
 492 }
 493
 494 static int next_with_matches(sd_journal *j, JournalFile *f, direction_t direction, Object **ret, uint64_t *offset) {
 495         int r;
 496         uint64_t cp;
 497         Object *c;
 498
 499         assert(j);
 500         assert(f);
 501         assert(ret);
 502         assert(offset);
 503
 504         c = *ret;
 505         cp = *offset;
 506
 507         if (!j->matches) {
 508                 /* No matches is easy */
 509
 510                 r = journal_file_next_entry(f, c, cp, direction, &c, &cp);
 511                 if (r <= 0)
 512                         return r;
 513
 514                 if (ret)
 515                         *ret = c;
 516                 if (offset)
 517                         *offset = cp;
 518                 return 1;
 519         }
 520
 521         /* So there are matches we have to adhere to, let's find the
 522          * first entry that matches all of them */
 523
 524         for (;;) {
 525                 uint64_t np, n;
 526                 bool found, term_result = false;
 527                 Match *m, *term_match = NULL;
 528
 529                 n = journal_file_entry_n_items(c);
 530
 531                 /* Make sure we don't match the entry we are starting
 532                  * from. */
 533                 found = cp > *offset;
 534
 535                 np = 0;
 536                 LIST_FOREACH(matches, m, j->matches) {
 537                         uint64_t q, k;
 538
 539                         /* Let's check if this is the beginning of a
 540                          * new term, i.e. has a different field prefix
 541                          * as the preceeding match. */
 542                         if (!term_match) {
 543                                 term_match = m;
 544                                 term_result = false;
 545                         } else if (!same_field(term_match->data, term_match->size, m->data, m->size)) {
 546                                 if (!term_result)
 547                                         found = false;
 548
 549                                 term_match = m;
 550                                 term_result = false;
 551                         }
 552
 553                         for (k = 0; k < n; k++)
 554                                 if (c->entry.items[k].hash == m->le_hash)
 555                                         break;
 556
 557                         if (k >= n) {
 558                                 /* Hmm, didn't find any field that
 559                                  * matched this rule, so ignore this
 560                                  * match. Go on with next match */
 561                                 continue;
 562                         }
 563
 564                         term_result = true;
 565
 566                         /* Hmm, so, this field matched, let's remember
 567                          * where we'd have to try next, in case the other
 568                          * matches are not OK */
 569
 570                         r = journal_file_next_entry_for_data(f, c, cp, le64toh(c->entry.items[k].object_offset), direction, NULL, &q);
 571                         if (r > 0) {
 572
 573                                 if (direction == DIRECTION_DOWN) {
 574                                         if (q > np)
 575                                                 np = q;
 576                                 } else {
 577                                         if (np == 0 || q < np)
 578                                                 np = q;
 579                                 }
 580                         }
 581                 }
 582
 583                 /* Check the last term */
 584                 if (term_match && term_result)
 585                         found = true;
 586
 587                 /* Did this entry match against all matches? */
 588                 if (found) {
 589                         if (ret)
 590                                 *ret = c;
 591                         if (offset)
 592                                 *offset = cp;
 593                         return 1;
 594                 }
 595
 596                 /* Did we find a subsequent entry? */
 597                 if (np == 0)
 598                         return 0;
 599
 600                 /* Hmm, ok, this entry only matched partially, so
 601                  * let's try another one */
 602                 cp = np;
 603         }
 604 }
 605
 606 static int next_beyond_location(sd_journal *j, JournalFile *f, direction_t direction, Object **ret, uint64_t *offset) {
 607         Object *c;
 608         uint64_t cp;
 609         int compare_value, r;
 610
 611         assert(j);
 612         assert(f);
 613
 614         if (f->current_offset > 0) {
 615                 r = journal_file_move_to_object(f, OBJECT_ENTRY, f->current_offset, &c);
 616                 if (r < 0)
 617                         return r;
 618
 619                 cp = f->current_offset;
 620
 621                 r = next_with_matches(j, f, direction, &c, &cp);
 622                 if (r <= 0)
 623                         return r;
 624
 625                 compare_value = 1;
 626         } else {
 627                 r = find_location(j, f, direction, &c, &cp);
 628                 if (r <= 0)
 629                         return r;
 630
 631                 compare_value = 0;
 632         }
 633
 634         for (;;) {
 635                 bool found;
 636
 637                 if (j->current_location.type == LOCATION_DISCRETE) {
 638                         int k;
 639
 640                         k = compare_with_location(f, c, &j->current_location);
 641                         if (direction == DIRECTION_DOWN)
 642                                 found = k >= compare_value;
 643                         else
 644                                 found = k <= -compare_value;
 645                 } else
 646                         found = true;
 647
 648                 if (found) {
 649                         if (ret)
 650                                 *ret = c;
 651                         if (offset)
 652                                 *offset = cp;
 653                         return 1;
 654                 }
 655
 656                 r = next_with_matches(j, f, direction, &c, &cp);
 657                 if (r <= 0)
 658                         return r;
 659         }
 660 }
 661
 662 static int real_journal_next(sd_journal *j, direction_t direction) {
 663         JournalFile *f, *new_current = NULL;
 664         Iterator i;
 665         int r;
 666         uint64_t new_offset = 0;
 667         Object *new_entry = NULL;
 668
 669         assert(j);
 670
 671         HASHMAP_FOREACH(f, j->files, i) {
 672                 Object *o;
 673                 uint64_t p;
 674                 bool found;
 675
 676                 r = next_beyond_location(j, f, direction, &o, &p);
 677                 if (r < 0)
 678                         return r;
 679                 else if (r == 0)
 680                         continue;
 681
 682                 if (!new_current)
 683                         found = true;
 684                 else {
 685                         int k;
 686
 687                         k = compare_order(f, o, new_current, new_entry);
 688
 689                         if (direction == DIRECTION_DOWN)
 690                                 found = k < 0;
 691                         else
 692                                 found = k > 0;
 693                 }
 694
 695                 if (found) {
 696                         new_current = f;
 697                         new_entry = o;
 698                         new_offset = p;
 699                 }
 700         }
 701
 702         if (!new_current)
 703                 return 0;
 704
 705         set_location(j, new_current, new_entry, new_offset);
 706
 707         return 1;
 708 }
 709
 710 int sd_journal_next(sd_journal *j) {
 711         return real_journal_next(j, DIRECTION_DOWN);
 712 }
 713
 714 int sd_journal_previous(sd_journal *j) {
 715         return real_journal_next(j, DIRECTION_UP);
 716 }
 717
 718 int sd_journal_next_skip(sd_journal *j, uint64_t skip) {
 719         int c = 0, r;
 720
 721         assert(j);
 722
 723         while (skip > 0) {
 724                 r = sd_journal_next(j);
 725                 if (r < 0)
 726                         return r;
 727
 728                 if (r == 0)
 729                         return c;
 730
 731                 skip--;
 732                 c++;
 733         }
 734
 735         return c;
 736 }
 737
 738 int sd_journal_previous_skip(sd_journal *j, uint64_t skip) {
 739         int c = 0, r;
 740
 741         assert(j);
 742
 743         while (skip > 0) {
 744                 r = sd_journal_previous(j);
 745                 if (r < 0)
 746                         return r;
 747
 748                 if (r == 0)
 749                         return c;
 750
 751                 skip--;
 752                 c++;
 753         }
 754
 755         return 1;
 756 }
 757
 758 int sd_journal_get_cursor(sd_journal *j, char **cursor) {
 759         Object *o;
 760         int r;
 761         char bid[33], sid[33];
 762
 763         assert(j);
 764         assert(cursor);
 765
 766         if (!j->current_file || j->current_file->current_offset <= 0)
 767                 return -EADDRNOTAVAIL;
 768
 769         r = journal_file_move_to_object(j->current_file, OBJECT_ENTRY, j->current_file->current_offset, &o);
 770         if (r < 0)
 771                 return r;
 772
 773         sd_id128_to_string(j->current_file->header->seqnum_id, sid);
 774         sd_id128_to_string(o->entry.boot_id, bid);
 775
 776         if (asprintf(cursor,
 777                      "s=%s;i=%llx;b=%s;m=%llx;t=%llx;x=%llx;p=%s",
 778                      sid, (unsigned long long) le64toh(o->entry.seqnum),
 779                      bid, (unsigned long long) le64toh(o->entry.monotonic),
 780                      (unsigned long long) le64toh(o->entry.realtime),
 781                      (unsigned long long) le64toh(o->entry.xor_hash),
 782                      file_name_from_path(j->current_file->path)) < 0)
 783                 return -ENOMEM;
 784
 785         return 1;
 786 }
 787
 788 int sd_journal_seek_cursor(sd_journal *j, const char *cursor) {
 789         char *w;
 790         size_t l;
 791         char *state;
 792         unsigned long long seqnum, monotonic, realtime, xor_hash;
 793         bool
 794                 seqnum_id_set = false,
 795                 seqnum_set = false,
 796                 boot_id_set = false,
 797                 monotonic_set = false,
 798                 realtime_set = false,
 799                 xor_hash_set = false;
 800         sd_id128_t seqnum_id, boot_id;
 801
 802         assert(j);
 803         assert(cursor);
 804
 805         FOREACH_WORD_SEPARATOR(w, l, cursor, ";", state) {
 806                 char *item;
 807                 int k = 0;
 808
 809                 if (l < 2 || w[1] != '=')
 810                         return -EINVAL;
 811
 812                 item = strndup(w, l);
 813                 if (!item)
 814                         return -ENOMEM;
 815
 816                 switch (w[0]) {
 817
 818                 case 's':
 819                         seqnum_id_set = true;
 820                         k = sd_id128_from_string(w+2, &seqnum_id);
 821                         break;
 822
 823                 case 'i':
 824                         seqnum_set = true;
 825                         if (sscanf(w+2, "%llx", &seqnum) != 1)
 826                                 k = -EINVAL;
 827                         break;
 828
 829                 case 'b':
 830                         boot_id_set = true;
 831                         k = sd_id128_from_string(w+2, &boot_id);
 832                         break;
 833
 834                 case 'm':
 835                         monotonic_set = true;
 836                         if (sscanf(w+2, "%llx", &monotonic) != 1)
 837                                 k = -EINVAL;
 838                         break;
 839
 840                 case 't':
 841                         realtime_set = true;
 842                         if (sscanf(w+2, "%llx", &realtime) != 1)
 843                                 k = -EINVAL;
 844                         break;
 845
 846                 case 'x':
 847                         xor_hash_set = true;
 848                         if (sscanf(w+2, "%llx", &xor_hash) != 1)
 849                                 k = -EINVAL;
 850                         break;
 851                 }
 852
 853                 free(item);
 854
 855                 if (k < 0)
 856                         return k;
 857         }
 858
 859         if ((!seqnum_set || !seqnum_id_set) &&
 860             (!monotonic_set || !boot_id_set) &&
 861             !realtime_set)
 862                 return -EINVAL;
 863
 864         reset_location(j);
 865
 866         j->current_location.type = LOCATION_DISCRETE;
 867
 868         if (realtime_set) {
 869                 j->current_location.realtime = (uint64_t) realtime;
 870                 j->current_location.realtime_set = true;
 871         }
 872
 873         if (seqnum_set && seqnum_id_set) {
 874                 j->current_location.seqnum = (uint64_t) seqnum;
 875                 j->current_location.seqnum_id = seqnum_id;
 876                 j->current_location.seqnum_set = true;
 877         }
 878
 879         if (monotonic_set && boot_id_set) {
 880                 j->current_location.monotonic = (uint64_t) monotonic;
 881                 j->current_location.boot_id = boot_id;
 882                 j->current_location.monotonic_set = true;
 883         }
 884
 885         if (xor_hash_set) {
 886                 j->current_location.xor_hash = (uint64_t) xor_hash;
 887                 j->current_location.xor_hash_set = true;
 888         }
 889
 890         return 0;
 891 }
 892
 893 int sd_journal_seek_monotonic_usec(sd_journal *j, sd_id128_t boot_id, uint64_t usec) {
 894         assert(j);
 895
 896         reset_location(j);
 897         j->current_location.type = LOCATION_DISCRETE;
 898         j->current_location.boot_id = boot_id;
 899         j->current_location.monotonic = usec;
 900         j->current_location.monotonic_set = true;
 901
 902         return 0;
 903 }
 904
 905 int sd_journal_seek_realtime_usec(sd_journal *j, uint64_t usec) {
 906         assert(j);
 907
 908         reset_location(j);
 909         j->current_location.type = LOCATION_DISCRETE;
 910         j->current_location.realtime = usec;
 911         j->current_location.realtime_set = true;
 912
 913         return 0;
 914 }
 915
 916 int sd_journal_seek_head(sd_journal *j) {
 917         assert(j);
 918
 919         reset_location(j);
 920         j->current_location.type = LOCATION_HEAD;
 921
 922         return 0;
 923 }
 924
 925 int sd_journal_seek_tail(sd_journal *j) {
 926         assert(j);
 927
 928         reset_location(j);
 929         j->current_location.type = LOCATION_TAIL;
 930
 931         return 0;
 932 }
 933
 934 static int add_file(sd_journal *j, const char *prefix, const char *dir, const char *filename) {
 935         char *fn;
 936         int r;
 937         JournalFile *f;
 938
 939         assert(j);
 940         assert(prefix);
 941         assert(filename);
 942
 943         if (dir)
 944                 fn = join(prefix, "/", dir, "/", filename, NULL);
 945         else
 946                 fn = join(prefix, "/", filename, NULL);
 947
 948         if (!fn)
 949                 return -ENOMEM;
 950
 951         if (hashmap_get(j->files, fn)) {
 952                 free(fn);
 953                 return 0;
 954         }
 955
 956         if (hashmap_size(j->files) >= JOURNAL_FILES_MAX) {
 957                 log_debug("Too many open journal files, not adding %s, ignoring.", fn);
 958                 free(fn);
 959                 return 0;
 960         }
 961
 962         r = journal_file_open(fn, O_RDONLY, 0, NULL, &f);
 963         free(fn);
 964
 965         if (r < 0) {
 966                 if (errno == ENOENT)
 967                         return 0;
 968
 969                 return r;
 970         }
 971
 972         journal_file_dump(f);
 973
 974         r = hashmap_put(j->files, f->path, f);
 975         if (r < 0) {
 976                 journal_file_close(f);
 977                 return r;
 978         }
 979
 980         log_debug("File %s got added.", f->path);
 981
 982         return 0;
 983 }
 984
 985 static int remove_file(sd_journal *j, const char *prefix, const char *dir, const char *filename) {
 986         char *fn;
 987         JournalFile *f;
 988
 989         assert(j);
 990         assert(prefix);
 991         assert(filename);
 992
 993         if (dir)
 994                 fn = join(prefix, "/", dir, "/", filename, NULL);
 995         else
 996                 fn = join(prefix, "/", filename, NULL);
 997
 998         if (!fn)
 999                 return -ENOMEM;
1000
1001         f = hashmap_get(j->files, fn);
1002         free(fn);
1003
1004         if (!f)
1005                 return 0;
1006
1007         hashmap_remove(j->files, f->path);
1008         journal_file_close(f);
1009
1010         log_debug("File %s got removed.", f->path);
1011         return 0;
1012 }
1013
1014 static int add_directory(sd_journal *j, const char *prefix, const char *dir) {
1015         char *fn;
1016         int r;
1017         DIR *d;
1018         int wd;
1019
1020         assert(j);
1021         assert(prefix);
1022         assert(dir);
1023
1024         fn = join(prefix, "/", dir, NULL);
1025         if (!fn)
1026                 return -ENOMEM;
1027
1028         d = opendir(fn);
1029
1030         if (!d) {
1031                 free(fn);
1032                 if (errno == ENOENT)
1033                         return 0;
1034
1035                 return -errno;
1036         }
1037
1038         wd = inotify_add_watch(j->inotify_fd, fn,
1039                                IN_CREATE|IN_MOVED_TO|IN_MODIFY|IN_ATTRIB|IN_DELETE|
1040                                IN_DELETE_SELF|IN_MOVE_SELF|IN_UNMOUNT|
1041                                IN_DONT_FOLLOW|IN_ONLYDIR);
1042         if (wd > 0) {
1043                 if (hashmap_put(j->inotify_wd_dirs, INT_TO_PTR(wd), fn) < 0)
1044                         inotify_rm_watch(j->inotify_fd, wd);
1045                 else
1046                         fn = NULL;
1047         }
1048
1049         free(fn);
1050
1051         for (;;) {
1052                 struct dirent buf, *de;
1053
1054                 r = readdir_r(d, &buf, &de);
1055                 if (r != 0 || !de)
1056                         break;
1057
1058                 if (!dirent_is_file_with_suffix(de, ".journal"))
1059                         continue;
1060
1061                 r = add_file(j, prefix, dir, de->d_name);
1062                 if (r < 0)
1063                         log_debug("Failed to add file %s/%s/%s: %s", prefix, dir, de->d_name, strerror(-r));
1064         }
1065
1066         closedir(d);
1067
1068         log_debug("Directory %s/%s got added.", prefix, dir);
1069
1070         return 0;
1071 }
1072
1073 static void remove_directory_wd(sd_journal *j, int wd) {
1074         char *p;
1075
1076         assert(j);
1077         assert(wd > 0);
1078
1079         if (j->inotify_fd >= 0)
1080                 inotify_rm_watch(j->inotify_fd, wd);
1081
1082         p = hashmap_remove(j->inotify_wd_dirs, INT_TO_PTR(wd));
1083
1084         if (p) {
1085                 log_debug("Directory %s got removed.", p);
1086                 free(p);
1087         }
1088 }
1089
1090 static void add_root_wd(sd_journal *j, const char *p) {
1091         int wd;
1092         char *k;
1093
1094         assert(j);
1095         assert(p);
1096
1097         wd = inotify_add_watch(j->inotify_fd, p,
1098                                IN_CREATE|IN_MOVED_TO|IN_MODIFY|IN_ATTRIB|IN_DELETE|
1099                                IN_DONT_FOLLOW|IN_ONLYDIR);
1100         if (wd <= 0)
1101                 return;
1102
1103         k = strdup(p);
1104         if (!k || hashmap_put(j->inotify_wd_roots, INT_TO_PTR(wd), k) < 0) {
1105                 inotify_rm_watch(j->inotify_fd, wd);
1106                 free(k);
1107         }
1108 }
1109
1110 static void remove_root_wd(sd_journal *j, int wd) {
1111         char *p;
1112
1113         assert(j);
1114         assert(wd > 0);
1115
1116         if (j->inotify_fd >= 0)
1117                 inotify_rm_watch(j->inotify_fd, wd);
1118
1119         p = hashmap_remove(j->inotify_wd_roots, INT_TO_PTR(wd));
1120
1121         if (p) {
1122                 log_debug("Root %s got removed.", p);
1123                 free(p);
1124         }
1125 }
1126
1127 int sd_journal_open(sd_journal **ret) {
1128         sd_journal *j;
1129         const char *p;
1130         const char search_paths[] =
1131                 "/run/log/journal\0"
1132                 "/var/log/journal\0";
1133         int r;
1134
1135         assert(ret);
1136
1137         j = new0(sd_journal, 1);
1138         if (!j)
1139                 return -ENOMEM;
1140
1141         j->inotify_fd = inotify_init1(IN_NONBLOCK|IN_CLOEXEC);
1142         if (j->inotify_fd < 0) {
1143                 r = -errno;
1144                 goto fail;
1145         }
1146
1147         j->files = hashmap_new(string_hash_func, string_compare_func);
1148         if (!j->files) {
1149                 r = -ENOMEM;
1150                 goto fail;
1151         }
1152
1153         j->inotify_wd_dirs = hashmap_new(trivial_hash_func, trivial_compare_func);
1154         j->inotify_wd_roots = hashmap_new(trivial_hash_func, trivial_compare_func);
1155
1156         if (!j->inotify_wd_dirs || !j->inotify_wd_roots) {
1157                 r = -ENOMEM;
1158                 goto fail;
1159         }
1160
1161         /* We ignore most errors here, since the idea is to only open
1162          * what's actually accessible, and ignore the rest. */
1163
1164         NULSTR_FOREACH(p, search_paths) {
1165                 DIR *d;
1166
1167                 d = opendir(p);
1168                 if (!d) {
1169                         if (errno != ENOENT)
1170                                 log_debug("Failed to open %s: %m", p);
1171                         continue;
1172                 }
1173
1174                 add_root_wd(j, p);
1175
1176                 for (;;) {
1177                         struct dirent buf, *de;
1178                         sd_id128_t id;
1179
1180                         r = readdir_r(d, &buf, &de);
1181                         if (r != 0 || !de)
1182                                 break;
1183
1184                         if (dirent_is_file_with_suffix(de, ".journal")) {
1185                                 r = add_file(j, p, NULL, de->d_name);
1186                                 if (r < 0)
1187                                         log_debug("Failed to add file %s/%s: %s", p, de->d_name, strerror(-r));
1188
1189                         } else if ((de->d_type == DT_DIR || de->d_type == DT_UNKNOWN) &&
1190                                    sd_id128_from_string(de->d_name, &id) >= 0) {
1191
1192                                 r = add_directory(j, p, de->d_name);
1193                                 if (r < 0)
1194                                         log_debug("Failed to add directory %s/%s: %s", p, de->d_name, strerror(-r));
1195                         }
1196                 }
1197
1198                 closedir(d);
1199         }
1200
1201         *ret = j;
1202         return 0;
1203
1204 fail:
1205         sd_journal_close(j);
1206
1207         return r;
1208 };
1209
1210 void sd_journal_close(sd_journal *j) {
1211         assert(j);
1212
1213         if (j->inotify_wd_dirs) {
1214                 void *k;
1215
1216                 while ((k = hashmap_first_key(j->inotify_wd_dirs)))
1217                         remove_directory_wd(j, PTR_TO_INT(k));
1218
1219                 hashmap_free(j->inotify_wd_dirs);
1220         }
1221
1222         if (j->inotify_wd_roots) {
1223                 void *k;
1224
1225                 while ((k = hashmap_first_key(j->inotify_wd_roots)))
1226                         remove_root_wd(j, PTR_TO_INT(k));
1227
1228                 hashmap_free(j->inotify_wd_roots);
1229         }
1230
1231         if (j->files) {
1232                 JournalFile *f;
1233
1234                 while ((f = hashmap_steal_first(j->files)))
1235                         journal_file_close(f);
1236
1237                 hashmap_free(j->files);
1238         }
1239
1240         sd_journal_flush_matches(j);
1241
1242         if (j->inotify_fd >= 0)
1243                 close_nointr_nofail(j->inotify_fd);
1244
1245         free(j);
1246 }
1247
1248 int sd_journal_get_realtime_usec(sd_journal *j, uint64_t *ret) {
1249         Object *o;
1250         JournalFile *f;
1251         int r;
1252
1253         assert(j);
1254         assert(ret);
1255
1256         f = j->current_file;
1257         if (!f)
1258                 return -EADDRNOTAVAIL;
1259
1260         if (f->current_offset <= 0)
1261                 return -EADDRNOTAVAIL;
1262
1263         r = journal_file_move_to_object(f, OBJECT_ENTRY, f->current_offset, &o);
1264         if (r < 0)
1265                 return r;
1266
1267         *ret = le64toh(o->entry.realtime);
1268         return 0;
1269 }
1270
1271 int sd_journal_get_monotonic_usec(sd_journal *j, uint64_t *ret, sd_id128_t *ret_boot_id) {
1272         Object *o;
1273         JournalFile *f;
1274         int r;
1275         sd_id128_t id;
1276
1277         assert(j);
1278         assert(ret);
1279
1280         f = j->current_file;
1281         if (!f)
1282                 return -EADDRNOTAVAIL;
1283
1284         if (f->current_offset <= 0)
1285                 return -EADDRNOTAVAIL;
1286
1287         r = journal_file_move_to_object(f, OBJECT_ENTRY, f->current_offset, &o);
1288         if (r < 0)
1289                 return r;
1290
1291         if (ret_boot_id)
1292                 *ret_boot_id = o->entry.boot_id;
1293         else {
1294                 r = sd_id128_get_boot(&id);
1295                 if (r < 0)
1296                         return r;
1297
1298                 if (!sd_id128_equal(id, o->entry.boot_id))
1299                         return -ENOENT;
1300         }
1301
1302         *ret = le64toh(o->entry.monotonic);
1303         return 0;
1304 }
1305
1306 int sd_journal_get_data(sd_journal *j, const char *field, const void **data, size_t *size) {
1307         JournalFile *f;
1308         uint64_t i, n;
1309         size_t field_length;
1310         int r;
1311         Object *o;
1312
1313         assert(j);
1314         assert(field);
1315         assert(data);
1316         assert(size);
1317
1318         if (isempty(field) || strchr(field, '='))
1319                 return -EINVAL;
1320
1321         f = j->current_file;
1322         if (!f)
1323                 return -EADDRNOTAVAIL;
1324
1325         if (f->current_offset <= 0)
1326                 return -EADDRNOTAVAIL;
1327
1328         r = journal_file_move_to_object(f, OBJECT_ENTRY, f->current_offset, &o);
1329         if (r < 0)
1330                 return r;
1331
1332         field_length = strlen(field);
1333
1334         n = journal_file_entry_n_items(o);
1335         for (i = 0; i < n; i++) {
1336                 uint64_t p, l, le_hash;
1337                 size_t t;
1338
1339                 p = le64toh(o->entry.items[i].object_offset);
1340                 le_hash = o->entry.items[j->current_field].hash;
1341                 r = journal_file_move_to_object(f, OBJECT_DATA, p, &o);
1342                 if (r < 0)
1343                         return r;
1344
1345                 if (le_hash != o->data.hash)
1346                         return -EBADMSG;
1347
1348                 l = le64toh(o->object.size) - offsetof(Object, data.payload);
1349
1350                 if (l >= field_length+1 &&
1351                     memcmp(o->data.payload, field, field_length) == 0 &&
1352                     o->data.payload[field_length] == '=') {
1353
1354                         t = (size_t) l;
1355
1356                         if ((uint64_t) t != l)
1357                                 return -E2BIG;
1358
1359                         *data = o->data.payload;
1360                         *size = t;
1361
1362                         return 0;
1363                 }
1364
1365                 r = journal_file_move_to_object(f, OBJECT_ENTRY, f->current_offset, &o);
1366                 if (r < 0)
1367                         return r;
1368         }
1369
1370         return -ENOENT;
1371 }
1372
1373 int sd_journal_enumerate_data(sd_journal *j, const void **data, size_t *size) {
1374         JournalFile *f;
1375         uint64_t p, l, n, le_hash;
1376         int r;
1377         Object *o;
1378         size_t t;
1379
1380         assert(j);
1381         assert(data);
1382         assert(size);
1383
1384         f = j->current_file;
1385         if (!f)
1386                 return -EADDRNOTAVAIL;
1387
1388         if (f->current_offset <= 0)
1389                 return -EADDRNOTAVAIL;
1390
1391         r = journal_file_move_to_object(f, OBJECT_ENTRY, f->current_offset, &o);
1392         if (r < 0)
1393                 return r;
1394
1395         n = journal_file_entry_n_items(o);
1396         if (j->current_field >= n)
1397                 return 0;
1398
1399         p = le64toh(o->entry.items[j->current_field].object_offset);
1400         le_hash = o->entry.items[j->current_field].hash;
1401         r = journal_file_move_to_object(f, OBJECT_DATA, p, &o);
1402         if (r < 0)
1403                 return r;
1404
1405         if (le_hash != o->data.hash)
1406                 return -EBADMSG;
1407
1408         l = le64toh(o->object.size) - offsetof(Object, data.payload);
1409         t = (size_t) l;
1410
1411         /* We can't read objects larger than 4G on a 32bit machine */
1412         if ((uint64_t) t != l)
1413                 return -E2BIG;
1414
1415         *data = o->data.payload;
1416         *size = t;
1417
1418         j->current_field ++;
1419
1420         return 1;
1421 }
1422
1423 void sd_journal_restart_data(sd_journal *j) {
1424         assert(j);
1425
1426         j->current_field = 0;
1427 }
1428
1429 int sd_journal_get_fd(sd_journal *j) {
1430         assert(j);
1431
1432         return j->inotify_fd;
1433 }
1434
1435 static void process_inotify_event(sd_journal *j, struct inotify_event *e) {
1436         char *p;
1437         int r;
1438
1439         assert(j);
1440         assert(e);
1441
1442         /* Is this a subdirectory we watch? */
1443         p = hashmap_get(j->inotify_wd_dirs, INT_TO_PTR(e->wd));
1444         if (p) {
1445
1446                 if (!(e->mask & IN_ISDIR) && e->len > 0 && endswith(e->name, ".journal")) {
1447
1448                         /* Event for a journal file */
1449
1450                         if (e->mask & (IN_CREATE|IN_MOVED_TO|IN_MODIFY|IN_ATTRIB)) {
1451                                 r = add_file(j, p, NULL, e->name);
1452                                 if (r < 0)
1453                                         log_debug("Failed to add file %s/%s: %s", p, e->name, strerror(-r));
1454                         } else if (e->mask & (IN_DELETE|IN_UNMOUNT)) {
1455
1456                                 r = remove_file(j, p, NULL, e->name);
1457                                 if (r < 0)
1458                                         log_debug("Failed to remove file %s/%s: %s", p, e->name, strerror(-r));
1459                         }
1460
1461                 } else if (e->len == 0) {
1462
1463                         /* Event for the directory itself */
1464
1465                         if (e->mask & (IN_DELETE_SELF|IN_MOVE_SELF|IN_UNMOUNT))
1466                                 remove_directory_wd(j, e->wd);
1467                 }
1468
1469                 return;
1470         }
1471
1472         /* Must be the root directory then? */
1473         p = hashmap_get(j->inotify_wd_roots, INT_TO_PTR(e->wd));
1474         if (p) {
1475                 sd_id128_t id;
1476
1477                 if (!(e->mask & IN_ISDIR) && e->len > 0 && endswith(e->name, ".journal")) {
1478
1479                         /* Event for a journal file */
1480
1481                         if (e->mask & (IN_CREATE|IN_MOVED_TO|IN_MODIFY|IN_ATTRIB)) {
1482                                 r = add_file(j, p, NULL, e->name);
1483                                 if (r < 0)
1484                                         log_debug("Failed to add file %s/%s: %s", p, e->name, strerror(-r));
1485                         } else if (e->mask & (IN_DELETE|IN_UNMOUNT)) {
1486
1487                                 r = remove_file(j, p, NULL, e->name);
1488                                 if (r < 0)
1489                                         log_debug("Failed to remove file %s/%s: %s", p, e->name, strerror(-r));
1490                         }
1491
1492                 } else if ((e->mask & IN_ISDIR) && e->len > 0 && sd_id128_from_string(e->name, &id) >= 0) {
1493
1494                         /* Event for subdirectory */
1495
1496                         if (e->mask & (IN_CREATE|IN_MOVED_TO|IN_MODIFY|IN_ATTRIB)) {
1497
1498                                 r = add_directory(j, p, e->name);
1499                                 if (r < 0)
1500                                         log_debug("Failed to add directory %s/%s: %s", p, e->name, strerror(-r));
1501                         }
1502                 }
1503
1504                 return;
1505         }
1506
1507         if (e->mask & IN_IGNORED)
1508                 return;
1509
1510         log_warning("Unknown inotify event.");
1511 }
1512
1513 int sd_journal_process(sd_journal *j) {
1514         uint8_t buffer[sizeof(struct inotify_event) + FILENAME_MAX];
1515
1516         assert(j);
1517
1518         for (;;) {
1519                 struct inotify_event *e;
1520                 ssize_t l;
1521
1522                 l = read(j->inotify_fd, buffer, sizeof(buffer));
1523                 if (l < 0) {
1524                         if (errno == EINTR || errno == EAGAIN)
1525                                 return 0;
1526
1527                         return -errno;
1528                 }
1529
1530                 e = (struct inotify_event*) buffer;
1531                 while (l > 0) {
1532                         size_t step;
1533
1534                         process_inotify_event(j, e);
1535
1536                         step = sizeof(struct inotify_event) + e->len;
1537                         assert(step <= (size_t) l);
1538
1539                         e = (struct inotify_event*) ((uint8_t*) e + step);
1540                         l -= step;
1541                 }
1542         }
1543 }