src/journal/journal-file.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 Lesser General Public License as published by
  10   the Free Software Foundation; either version 2.1 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   Lesser General Public License for more details.
  17
  18   You should have received a copy of the GNU Lesser General Public License
  19   along with systemd; If not, see <http://www.gnu.org/licenses/>.
  20 ***/
  21
  22 #include <sys/mman.h>
  23 #include <errno.h>
  24 #include <sys/uio.h>
  25 #include <unistd.h>
  26 #include <sys/statvfs.h>
  27 #include <fcntl.h>
  28 #include <stddef.h>
  29 #include <sys/xattr.h>
  30
  31 #include "journal-def.h"
  32 #include "journal-file.h"
  33 #include "journal-authenticate.h"
  34 #include "lookup3.h"
  35 #include "compress.h"
  36 #include "fsprg.h"
  37
  38 #define DEFAULT_DATA_HASH_TABLE_SIZE (2047ULL*sizeof(HashItem))
  39 #define DEFAULT_FIELD_HASH_TABLE_SIZE (333ULL*sizeof(HashItem))
  40
  41 #define COMPRESSION_SIZE_THRESHOLD (512ULL)
  42
  43 /* This is the minimum journal file size */
  44 #define JOURNAL_FILE_SIZE_MIN (4ULL*1024ULL*1024ULL)           /* 4 MiB */
  45
  46 /* These are the lower and upper bounds if we deduce the max_use value
  47  * from the file system size */
  48 #define DEFAULT_MAX_USE_LOWER (1ULL*1024ULL*1024ULL)           /* 1 MiB */
  49 #define DEFAULT_MAX_USE_UPPER (4ULL*1024ULL*1024ULL*1024ULL)   /* 4 GiB */
  50
  51 /* This is the upper bound if we deduce max_size from max_use */
  52 #define DEFAULT_MAX_SIZE_UPPER (128ULL*1024ULL*1024ULL)        /* 128 MiB */
  53
  54 /* This is the upper bound if we deduce the keep_free value from the
  55  * file system size */
  56 #define DEFAULT_KEEP_FREE_UPPER (4ULL*1024ULL*1024ULL*1024ULL) /* 4 GiB */
  57
  58 /* This is the keep_free value when we can't determine the system
  59  * size */
  60 #define DEFAULT_KEEP_FREE (1024ULL*1024ULL)                    /* 1 MB */
  61
  62 /* n_data was the first entry we added after the initial file format design */
  63 #define HEADER_SIZE_MIN ALIGN64(offsetof(Header, n_data))
  64
  65 /* How many entries to keep in the entry array chain cache at max */
  66 #define CHAIN_CACHE_MAX 20
  67
  68 /* How much to increase the journal file size at once each time we allocate something new. */
  69 #define FILE_SIZE_INCREASE (8ULL*1024ULL*1024ULL)              /* 8MB */
  70
  71 static int journal_file_set_online(JournalFile *f) {
  72         assert(f);
  73
  74         if (!f->writable)
  75                 return -EPERM;
  76
  77         if (!(f->fd >= 0 && f->header))
  78                 return -EINVAL;
  79
  80         switch(f->header->state) {
  81                 case STATE_ONLINE:
  82                         return 0;
  83
  84                 case STATE_OFFLINE:
  85                         f->header->state = STATE_ONLINE;
  86                         fsync(f->fd);
  87                         return 0;
  88
  89                 default:
  90                         return -EINVAL;
  91         }
  92 }
  93
  94 int journal_file_set_offline(JournalFile *f) {
  95         assert(f);
  96
  97         if (!f->writable)
  98                 return -EPERM;
  99
 100         if (!(f->fd >= 0 && f->header))
 101                 return -EINVAL;
 102
 103         if (f->header->state != STATE_ONLINE)
 104                 return 0;
 105
 106         fsync(f->fd);
 107
 108         f->header->state = STATE_OFFLINE;
 109
 110         fsync(f->fd);
 111
 112         return 0;
 113 }
 114
 115 void journal_file_close(JournalFile *f) {
 116         assert(f);
 117
 118 #ifdef HAVE_GCRYPT
 119         /* Write the final tag */
 120         if (f->seal && f->writable)
 121                 journal_file_append_tag(f);
 122 #endif
 123
 124         /* Sync everything to disk, before we mark the file offline */
 125         if (f->mmap && f->fd >= 0)
 126                 mmap_cache_close_fd(f->mmap, f->fd);
 127
 128         journal_file_set_offline(f);
 129
 130         if (f->header)
 131                 munmap(f->header, PAGE_ALIGN(sizeof(Header)));
 132
 133         safe_close(f->fd);
 134         free(f->path);
 135
 136         if (f->mmap)
 137                 mmap_cache_unref(f->mmap);
 138
 139         ordered_hashmap_free_free(f->chain_cache);
 140
 141 #if defined(HAVE_XZ) || defined(HAVE_LZ4)
 142         free(f->compress_buffer);
 143 #endif
 144
 145 #ifdef HAVE_GCRYPT
 146         if (f->fss_file)
 147                 munmap(f->fss_file, PAGE_ALIGN(f->fss_file_size));
 148         else if (f->fsprg_state)
 149                 free(f->fsprg_state);
 150
 151         free(f->fsprg_seed);
 152
 153         if (f->hmac)
 154                 gcry_md_close(f->hmac);
 155 #endif
 156
 157         free(f);
 158 }
 159
 160 static int journal_file_init_header(JournalFile *f, JournalFile *template) {
 161         Header h = {};
 162         ssize_t k;
 163         int r;
 164
 165         assert(f);
 166
 167         memcpy(h.signature, HEADER_SIGNATURE, 8);
 168         h.header_size = htole64(ALIGN64(sizeof(h)));
 169
 170         h.incompatible_flags |= htole32(
 171                 f->compress_xz * HEADER_INCOMPATIBLE_COMPRESSED_XZ |
 172                 f->compress_lz4 * HEADER_INCOMPATIBLE_COMPRESSED_LZ4);
 173
 174         h.compatible_flags = htole32(
 175                 f->seal * HEADER_COMPATIBLE_SEALED);
 176
 177         r = sd_id128_randomize(&h.file_id);
 178         if (r < 0)
 179                 return r;
 180
 181         if (template) {
 182                 h.seqnum_id = template->header->seqnum_id;
 183                 h.tail_entry_seqnum = template->header->tail_entry_seqnum;
 184         } else
 185                 h.seqnum_id = h.file_id;
 186
 187         k = pwrite(f->fd, &h, sizeof(h), 0);
 188         if (k < 0)
 189                 return -errno;
 190
 191         if (k != sizeof(h))
 192                 return -EIO;
 193
 194         return 0;
 195 }
 196
 197 static int journal_file_refresh_header(JournalFile *f) {
 198         int r;
 199         sd_id128_t boot_id;
 200
 201         assert(f);
 202
 203         r = sd_id128_get_machine(&f->header->machine_id);
 204         if (r < 0)
 205                 return r;
 206
 207         r = sd_id128_get_boot(&boot_id);
 208         if (r < 0)
 209                 return r;
 210
 211         if (sd_id128_equal(boot_id, f->header->boot_id))
 212                 f->tail_entry_monotonic_valid = true;
 213
 214         f->header->boot_id = boot_id;
 215
 216         journal_file_set_online(f);
 217
 218         /* Sync the online state to disk */
 219         fsync(f->fd);
 220
 221         return 0;
 222 }
 223
 224 static int journal_file_verify_header(JournalFile *f) {
 225         uint32_t flags;
 226
 227         assert(f);
 228
 229         if (memcmp(f->header->signature, HEADER_SIGNATURE, 8))
 230                 return -EBADMSG;
 231
 232         /* In both read and write mode we refuse to open files with
 233          * incompatible flags we don't know */
 234         flags = le32toh(f->header->incompatible_flags);
 235         if (flags & ~HEADER_INCOMPATIBLE_SUPPORTED) {
 236                 if (flags & ~HEADER_INCOMPATIBLE_ANY)
 237                         log_debug("Journal file %s has unknown incompatible flags %"PRIx32,
 238                                   f->path, flags & ~HEADER_INCOMPATIBLE_ANY);
 239                 flags = (flags & HEADER_INCOMPATIBLE_ANY) & ~HEADER_INCOMPATIBLE_SUPPORTED;
 240                 if (flags)
 241                         log_debug("Journal file %s uses incompatible flags %"PRIx32
 242                                   " disabled at compilation time.", f->path, flags);
 243                 return -EPROTONOSUPPORT;
 244         }
 245
 246         /* When open for writing we refuse to open files with
 247          * compatible flags, too */
 248         flags = le32toh(f->header->compatible_flags);
 249         if (f->writable && (flags & ~HEADER_COMPATIBLE_SUPPORTED)) {
 250                 if (flags & ~HEADER_COMPATIBLE_ANY)
 251                         log_debug("Journal file %s has unknown compatible flags %"PRIx32,
 252                                   f->path, flags & ~HEADER_COMPATIBLE_ANY);
 253                 flags = (flags & HEADER_COMPATIBLE_ANY) & ~HEADER_COMPATIBLE_SUPPORTED;
 254                 if (flags)
 255                         log_debug("Journal file %s uses compatible flags %"PRIx32
 256                                   " disabled at compilation time.", f->path, flags);
 257                 return -EPROTONOSUPPORT;
 258         }
 259
 260         if (f->header->state >= _STATE_MAX)
 261                 return -EBADMSG;
 262
 263         /* The first addition was n_data, so check that we are at least this large */
 264         if (le64toh(f->header->header_size) < HEADER_SIZE_MIN)
 265                 return -EBADMSG;
 266
 267         if (JOURNAL_HEADER_SEALED(f->header) && !JOURNAL_HEADER_CONTAINS(f->header, n_entry_arrays))
 268                 return -EBADMSG;
 269
 270         if ((le64toh(f->header->header_size) + le64toh(f->header->arena_size)) > (uint64_t) f->last_stat.st_size)
 271                 return -ENODATA;
 272
 273         if (le64toh(f->header->tail_object_offset) > (le64toh(f->header->header_size) + le64toh(f->header->arena_size)))
 274                 return -ENODATA;
 275
 276         if (!VALID64(le64toh(f->header->data_hash_table_offset)) ||
 277             !VALID64(le64toh(f->header->field_hash_table_offset)) ||
 278             !VALID64(le64toh(f->header->tail_object_offset)) ||
 279             !VALID64(le64toh(f->header->entry_array_offset)))
 280                 return -ENODATA;
 281
 282         if (f->writable) {
 283                 uint8_t state;
 284                 sd_id128_t machine_id;
 285                 int r;
 286
 287                 r = sd_id128_get_machine(&machine_id);
 288                 if (r < 0)
 289                         return r;
 290
 291                 if (!sd_id128_equal(machine_id, f->header->machine_id))
 292                         return -EHOSTDOWN;
 293
 294                 state = f->header->state;
 295
 296                 if (state == STATE_ONLINE) {
 297                         log_debug("Journal file %s is already online. Assuming unclean closing.", f->path);
 298                         return -EBUSY;
 299                 } else if (state == STATE_ARCHIVED)
 300                         return -ESHUTDOWN;
 301                 else if (state != STATE_OFFLINE) {
 302                         log_debug("Journal file %s has unknown state %u.", f->path, state);
 303                         return -EBUSY;
 304                 }
 305         }
 306
 307         f->compress_xz = JOURNAL_HEADER_COMPRESSED_XZ(f->header);
 308         f->compress_lz4 = JOURNAL_HEADER_COMPRESSED_LZ4(f->header);
 309
 310         f->seal = JOURNAL_HEADER_SEALED(f->header);
 311
 312         return 0;
 313 }
 314
 315 static int journal_file_allocate(JournalFile *f, uint64_t offset, uint64_t size) {
 316         uint64_t old_size, new_size;
 317         int r;
 318
 319         assert(f);
 320
 321         /* We assume that this file is not sparse, and we know that
 322          * for sure, since we always call posix_fallocate()
 323          * ourselves */
 324
 325         old_size =
 326                 le64toh(f->header->header_size) +
 327                 le64toh(f->header->arena_size);
 328
 329         new_size = PAGE_ALIGN(offset + size);
 330         if (new_size < le64toh(f->header->header_size))
 331                 new_size = le64toh(f->header->header_size);
 332
 333         if (new_size <= old_size)
 334                 return 0;
 335
 336         if (f->metrics.max_size > 0 && new_size > f->metrics.max_size)
 337                 return -E2BIG;
 338
 339         if (new_size > f->metrics.min_size && f->metrics.keep_free > 0) {
 340                 struct statvfs svfs;
 341
 342                 if (fstatvfs(f->fd, &svfs) >= 0) {
 343                         uint64_t available;
 344
 345                         available = svfs.f_bfree * svfs.f_bsize;
 346
 347                         if (available >= f->metrics.keep_free)
 348                                 available -= f->metrics.keep_free;
 349                         else
 350                                 available = 0;
 351
 352                         if (new_size - old_size > available)
 353                                 return -E2BIG;
 354                 }
 355         }
 356
 357         /* Increase by larger blocks at once */
 358         new_size = ((new_size+FILE_SIZE_INCREASE-1) / FILE_SIZE_INCREASE) * FILE_SIZE_INCREASE;
 359         if (f->metrics.max_size > 0 && new_size > f->metrics.max_size)
 360                 new_size = f->metrics.max_size;
 361
 362         /* Note that the glibc fallocate() fallback is very
 363            inefficient, hence we try to minimize the allocation area
 364            as we can. */
 365         r = posix_fallocate(f->fd, old_size, new_size - old_size);
 366         if (r != 0)
 367                 return -r;
 368
 369         if (fstat(f->fd, &f->last_stat) < 0)
 370                 return -errno;
 371
 372         f->header->arena_size = htole64(new_size - le64toh(f->header->header_size));
 373
 374         return 0;
 375 }
 376
 377 static unsigned type_to_context(ObjectType type) {
 378         /* One context for each type, plus one catch-all for the rest */
 379         assert_cc(_OBJECT_TYPE_MAX <= MMAP_CACHE_MAX_CONTEXTS);
 380         return type > OBJECT_UNUSED && type < _OBJECT_TYPE_MAX ? type : 0;
 381 }
 382
 383 static int journal_file_move_to(JournalFile *f, ObjectType type, bool keep_always, uint64_t offset, uint64_t size, void **ret) {
 384         assert(f);
 385         assert(ret);
 386
 387         if (size <= 0)
 388                 return -EINVAL;
 389
 390         /* Avoid SIGBUS on invalid accesses */
 391         if (offset + size > (uint64_t) f->last_stat.st_size) {
 392                 /* Hmm, out of range? Let's refresh the fstat() data
 393                  * first, before we trust that check. */
 394
 395                 if (fstat(f->fd, &f->last_stat) < 0 ||
 396                     offset + size > (uint64_t) f->last_stat.st_size)
 397                         return -EADDRNOTAVAIL;
 398         }
 399
 400         return mmap_cache_get(f->mmap, f->fd, f->prot, type_to_context(type), keep_always, offset, size, &f->last_stat, ret);
 401 }
 402
 403 static uint64_t minimum_header_size(Object *o) {
 404
 405         static const uint64_t table[] = {
 406                 [OBJECT_DATA] = sizeof(DataObject),
 407                 [OBJECT_FIELD] = sizeof(FieldObject),
 408                 [OBJECT_ENTRY] = sizeof(EntryObject),
 409                 [OBJECT_DATA_HASH_TABLE] = sizeof(HashTableObject),
 410                 [OBJECT_FIELD_HASH_TABLE] = sizeof(HashTableObject),
 411                 [OBJECT_ENTRY_ARRAY] = sizeof(EntryArrayObject),
 412                 [OBJECT_TAG] = sizeof(TagObject),
 413         };
 414
 415         if (o->object.type >= ELEMENTSOF(table) || table[o->object.type] <= 0)
 416                 return sizeof(ObjectHeader);
 417
 418         return table[o->object.type];
 419 }
 420
 421 int journal_file_move_to_object(JournalFile *f, ObjectType type, uint64_t offset, Object **ret) {
 422         int r;
 423         void *t;
 424         Object *o;
 425         uint64_t s;
 426
 427         assert(f);
 428         assert(ret);
 429
 430         /* Objects may only be located at multiple of 64 bit */
 431         if (!VALID64(offset))
 432                 return -EFAULT;
 433
 434         r = journal_file_move_to(f, type, false, offset, sizeof(ObjectHeader), &t);
 435         if (r < 0)
 436                 return r;
 437
 438         o = (Object*) t;
 439         s = le64toh(o->object.size);
 440
 441         if (s < sizeof(ObjectHeader))
 442                 return -EBADMSG;
 443
 444         if (o->object.type <= OBJECT_UNUSED)
 445                 return -EBADMSG;
 446
 447         if (s < minimum_header_size(o))
 448                 return -EBADMSG;
 449
 450         if (type > OBJECT_UNUSED && o->object.type != type)
 451                 return -EBADMSG;
 452
 453         if (s > sizeof(ObjectHeader)) {
 454                 r = journal_file_move_to(f, type, false, offset, s, &t);
 455                 if (r < 0)
 456                         return r;
 457
 458                 o = (Object*) t;
 459         }
 460
 461         *ret = o;
 462         return 0;
 463 }
 464
 465 static uint64_t journal_file_entry_seqnum(JournalFile *f, uint64_t *seqnum) {
 466         uint64_t r;
 467
 468         assert(f);
 469
 470         r = le64toh(f->header->tail_entry_seqnum) + 1;
 471
 472         if (seqnum) {
 473                 /* If an external seqnum counter was passed, we update
 474                  * both the local and the external one, and set it to
 475                  * the maximum of both */
 476
 477                 if (*seqnum + 1 > r)
 478                         r = *seqnum + 1;
 479
 480                 *seqnum = r;
 481         }
 482
 483         f->header->tail_entry_seqnum = htole64(r);
 484
 485         if (f->header->head_entry_seqnum == 0)
 486                 f->header->head_entry_seqnum = htole64(r);
 487
 488         return r;
 489 }
 490
 491 int journal_file_append_object(JournalFile *f, ObjectType type, uint64_t size, Object **ret, uint64_t *offset) {
 492         int r;
 493         uint64_t p;
 494         Object *tail, *o;
 495         void *t;
 496
 497         assert(f);
 498         assert(type > OBJECT_UNUSED && type < _OBJECT_TYPE_MAX);
 499         assert(size >= sizeof(ObjectHeader));
 500         assert(offset);
 501         assert(ret);
 502
 503         r = journal_file_set_online(f);
 504         if (r < 0)
 505                 return r;
 506
 507         p = le64toh(f->header->tail_object_offset);
 508         if (p == 0)
 509                 p = le64toh(f->header->header_size);
 510         else {
 511                 r = journal_file_move_to_object(f, OBJECT_UNUSED, p, &tail);
 512                 if (r < 0)
 513                         return r;
 514
 515                 p += ALIGN64(le64toh(tail->object.size));
 516         }
 517
 518         r = journal_file_allocate(f, p, size);
 519         if (r < 0)
 520                 return r;
 521
 522         r = journal_file_move_to(f, type, false, p, size, &t);
 523         if (r < 0)
 524                 return r;
 525
 526         o = (Object*) t;
 527
 528         zero(o->object);
 529         o->object.type = type;
 530         o->object.size = htole64(size);
 531
 532         f->header->tail_object_offset = htole64(p);
 533         f->header->n_objects = htole64(le64toh(f->header->n_objects) + 1);
 534
 535         *ret = o;
 536         *offset = p;
 537
 538         return 0;
 539 }
 540
 541 static int journal_file_setup_data_hash_table(JournalFile *f) {
 542         uint64_t s, p;
 543         Object *o;
 544         int r;
 545
 546         assert(f);
 547
 548         /* We estimate that we need 1 hash table entry per 768 of
 549            journal file and we want to make sure we never get beyond
 550            75% fill level. Calculate the hash table size for the
 551            maximum file size based on these metrics. */
 552
 553         s = (f->metrics.max_size * 4 / 768 / 3) * sizeof(HashItem);
 554         if (s < DEFAULT_DATA_HASH_TABLE_SIZE)
 555                 s = DEFAULT_DATA_HASH_TABLE_SIZE;
 556
 557         log_debug("Reserving %"PRIu64" entries in hash table.", s / sizeof(HashItem));
 558
 559         r = journal_file_append_object(f,
 560                                        OBJECT_DATA_HASH_TABLE,
 561                                        offsetof(Object, hash_table.items) + s,
 562                                        &o, &p);
 563         if (r < 0)
 564                 return r;
 565
 566         memzero(o->hash_table.items, s);
 567
 568         f->header->data_hash_table_offset = htole64(p + offsetof(Object, hash_table.items));
 569         f->header->data_hash_table_size = htole64(s);
 570
 571         return 0;
 572 }
 573
 574 static int journal_file_setup_field_hash_table(JournalFile *f) {
 575         uint64_t s, p;
 576         Object *o;
 577         int r;
 578
 579         assert(f);
 580
 581         /* We use a fixed size hash table for the fields as this
 582          * number should grow very slowly only */
 583
 584         s = DEFAULT_FIELD_HASH_TABLE_SIZE;
 585         r = journal_file_append_object(f,
 586                                        OBJECT_FIELD_HASH_TABLE,
 587                                        offsetof(Object, hash_table.items) + s,
 588                                        &o, &p);
 589         if (r < 0)
 590                 return r;
 591
 592         memzero(o->hash_table.items, s);
 593
 594         f->header->field_hash_table_offset = htole64(p + offsetof(Object, hash_table.items));
 595         f->header->field_hash_table_size = htole64(s);
 596
 597         return 0;
 598 }
 599
 600 static int journal_file_map_data_hash_table(JournalFile *f) {
 601         uint64_t s, p;
 602         void *t;
 603         int r;
 604
 605         assert(f);
 606
 607         p = le64toh(f->header->data_hash_table_offset);
 608         s = le64toh(f->header->data_hash_table_size);
 609
 610         r = journal_file_move_to(f,
 611                                  OBJECT_DATA_HASH_TABLE,
 612                                  true,
 613                                  p, s,
 614                                  &t);
 615         if (r < 0)
 616                 return r;
 617
 618         f->data_hash_table = t;
 619         return 0;
 620 }
 621
 622 static int journal_file_map_field_hash_table(JournalFile *f) {
 623         uint64_t s, p;
 624         void *t;
 625         int r;
 626
 627         assert(f);
 628
 629         p = le64toh(f->header->field_hash_table_offset);
 630         s = le64toh(f->header->field_hash_table_size);
 631
 632         r = journal_file_move_to(f,
 633                                  OBJECT_FIELD_HASH_TABLE,
 634                                  true,
 635                                  p, s,
 636                                  &t);
 637         if (r < 0)
 638                 return r;
 639
 640         f->field_hash_table = t;
 641         return 0;
 642 }
 643
 644 static int journal_file_link_field(
 645                 JournalFile *f,
 646                 Object *o,
 647                 uint64_t offset,
 648                 uint64_t hash) {
 649
 650         uint64_t p, h;
 651         int r;
 652
 653         assert(f);
 654         assert(o);
 655         assert(offset > 0);
 656
 657         if (o->object.type != OBJECT_FIELD)
 658                 return -EINVAL;
 659
 660         /* This might alter the window we are looking at */
 661
 662         o->field.next_hash_offset = o->field.head_data_offset = 0;
 663
 664         h = hash % (le64toh(f->header->field_hash_table_size) / sizeof(HashItem));
 665         p = le64toh(f->field_hash_table[h].tail_hash_offset);
 666         if (p == 0)
 667                 f->field_hash_table[h].head_hash_offset = htole64(offset);
 668         else {
 669                 r = journal_file_move_to_object(f, OBJECT_FIELD, p, &o);
 670                 if (r < 0)
 671                         return r;
 672
 673                 o->field.next_hash_offset = htole64(offset);
 674         }
 675
 676         f->field_hash_table[h].tail_hash_offset = htole64(offset);
 677
 678         if (JOURNAL_HEADER_CONTAINS(f->header, n_fields))
 679                 f->header->n_fields = htole64(le64toh(f->header->n_fields) + 1);
 680
 681         return 0;
 682 }
 683
 684 static int journal_file_link_data(
 685                 JournalFile *f,
 686                 Object *o,
 687                 uint64_t offset,
 688                 uint64_t hash) {
 689
 690         uint64_t p, h;
 691         int r;
 692
 693         assert(f);
 694         assert(o);
 695         assert(offset > 0);
 696
 697         if (o->object.type != OBJECT_DATA)
 698                 return -EINVAL;
 699
 700         /* This might alter the window we are looking at */
 701
 702         o->data.next_hash_offset = o->data.next_field_offset = 0;
 703         o->data.entry_offset = o->data.entry_array_offset = 0;
 704         o->data.n_entries = 0;
 705
 706         h = hash % (le64toh(f->header->data_hash_table_size) / sizeof(HashItem));
 707         p = le64toh(f->data_hash_table[h].tail_hash_offset);
 708         if (p == 0)
 709                 /* Only entry in the hash table is easy */
 710                 f->data_hash_table[h].head_hash_offset = htole64(offset);
 711         else {
 712                 /* Move back to the previous data object, to patch in
 713                  * pointer */
 714
 715                 r = journal_file_move_to_object(f, OBJECT_DATA, p, &o);
 716                 if (r < 0)
 717                         return r;
 718
 719                 o->data.next_hash_offset = htole64(offset);
 720         }
 721
 722         f->data_hash_table[h].tail_hash_offset = htole64(offset);
 723
 724         if (JOURNAL_HEADER_CONTAINS(f->header, n_data))
 725                 f->header->n_data = htole64(le64toh(f->header->n_data) + 1);
 726
 727         return 0;
 728 }
 729
 730 int journal_file_find_field_object_with_hash(
 731                 JournalFile *f,
 732                 const void *field, uint64_t size, uint64_t hash,
 733                 Object **ret, uint64_t *offset) {
 734
 735         uint64_t p, osize, h;
 736         int r;
 737
 738         assert(f);
 739         assert(field && size > 0);
 740
 741         osize = offsetof(Object, field.payload) + size;
 742
 743         if (f->header->field_hash_table_size == 0)
 744                 return -EBADMSG;
 745
 746         h = hash % (le64toh(f->header->field_hash_table_size) / sizeof(HashItem));
 747         p = le64toh(f->field_hash_table[h].head_hash_offset);
 748
 749         while (p > 0) {
 750                 Object *o;
 751
 752                 r = journal_file_move_to_object(f, OBJECT_FIELD, p, &o);
 753                 if (r < 0)
 754                         return r;
 755
 756                 if (le64toh(o->field.hash) == hash &&
 757                     le64toh(o->object.size) == osize &&
 758                     memcmp(o->field.payload, field, size) == 0) {
 759
 760                         if (ret)
 761                                 *ret = o;
 762                         if (offset)
 763                                 *offset = p;
 764
 765                         return 1;
 766                 }
 767
 768                 p = le64toh(o->field.next_hash_offset);
 769         }
 770
 771         return 0;
 772 }
 773
 774 int journal_file_find_field_object(
 775                 JournalFile *f,
 776                 const void *field, uint64_t size,
 777                 Object **ret, uint64_t *offset) {
 778
 779         uint64_t hash;
 780
 781         assert(f);
 782         assert(field && size > 0);
 783
 784         hash = hash64(field, size);
 785
 786         return journal_file_find_field_object_with_hash(f,
 787                                                         field, size, hash,
 788                                                         ret, offset);
 789 }
 790
 791 int journal_file_find_data_object_with_hash(
 792                 JournalFile *f,
 793                 const void *data, uint64_t size, uint64_t hash,
 794                 Object **ret, uint64_t *offset) {
 795
 796         uint64_t p, osize, h;
 797         int r;
 798
 799         assert(f);
 800         assert(data || size == 0);
 801
 802         osize = offsetof(Object, data.payload) + size;
 803
 804         if (f->header->data_hash_table_size == 0)
 805                 return -EBADMSG;
 806
 807         h = hash % (le64toh(f->header->data_hash_table_size) / sizeof(HashItem));
 808         p = le64toh(f->data_hash_table[h].head_hash_offset);
 809
 810         while (p > 0) {
 811                 Object *o;
 812
 813                 r = journal_file_move_to_object(f, OBJECT_DATA, p, &o);
 814                 if (r < 0)
 815                         return r;
 816
 817                 if (le64toh(o->data.hash) != hash)
 818                         goto next;
 819
 820                 if (o->object.flags & OBJECT_COMPRESSION_MASK) {
 821 #if defined(HAVE_XZ) || defined(HAVE_LZ4)
 822                         uint64_t l;
 823                         size_t rsize;
 824
 825                         l = le64toh(o->object.size);
 826                         if (l <= offsetof(Object, data.payload))
 827                                 return -EBADMSG;
 828
 829                         l -= offsetof(Object, data.payload);
 830
 831                         r = decompress_blob(o->object.flags & OBJECT_COMPRESSION_MASK,
 832                                             o->data.payload, l, &f->compress_buffer, &f->compress_buffer_size, &rsize, 0);
 833                         if (r < 0)
 834                                 return r;
 835
 836                         if (rsize == size &&
 837                             memcmp(f->compress_buffer, data, size) == 0) {
 838
 839                                 if (ret)
 840                                         *ret = o;
 841
 842                                 if (offset)
 843                                         *offset = p;
 844
 845                                 return 1;
 846                         }
 847 #else
 848                         return -EPROTONOSUPPORT;
 849 #endif
 850                 } else if (le64toh(o->object.size) == osize &&
 851                            memcmp(o->data.payload, data, size) == 0) {
 852
 853                         if (ret)
 854                                 *ret = o;
 855
 856                         if (offset)
 857                                 *offset = p;
 858
 859                         return 1;
 860                 }
 861
 862         next:
 863                 p = le64toh(o->data.next_hash_offset);
 864         }
 865
 866         return 0;
 867 }
 868
 869 int journal_file_find_data_object(
 870                 JournalFile *f,
 871                 const void *data, uint64_t size,
 872                 Object **ret, uint64_t *offset) {
 873
 874         uint64_t hash;
 875
 876         assert(f);
 877         assert(data || size == 0);
 878
 879         hash = hash64(data, size);
 880
 881         return journal_file_find_data_object_with_hash(f,
 882                                                        data, size, hash,
 883                                                        ret, offset);
 884 }
 885
 886 static int journal_file_append_field(
 887                 JournalFile *f,
 888                 const void *field, uint64_t size,
 889                 Object **ret, uint64_t *offset) {
 890
 891         uint64_t hash, p;
 892         uint64_t osize;
 893         Object *o;
 894         int r;
 895
 896         assert(f);
 897         assert(field && size > 0);
 898
 899         hash = hash64(field, size);
 900
 901         r = journal_file_find_field_object_with_hash(f, field, size, hash, &o, &p);
 902         if (r < 0)
 903                 return r;
 904         else if (r > 0) {
 905
 906                 if (ret)
 907                         *ret = o;
 908
 909                 if (offset)
 910                         *offset = p;
 911
 912                 return 0;
 913         }
 914
 915         osize = offsetof(Object, field.payload) + size;
 916         r = journal_file_append_object(f, OBJECT_FIELD, osize, &o, &p);
 917         if (r < 0)
 918                 return r;
 919
 920         o->field.hash = htole64(hash);
 921         memcpy(o->field.payload, field, size);
 922
 923         r = journal_file_link_field(f, o, p, hash);
 924         if (r < 0)
 925                 return r;
 926
 927         /* The linking might have altered the window, so let's
 928          * refresh our pointer */
 929         r = journal_file_move_to_object(f, OBJECT_FIELD, p, &o);
 930         if (r < 0)
 931                 return r;
 932
 933 #ifdef HAVE_GCRYPT
 934         r = journal_file_hmac_put_object(f, OBJECT_FIELD, o, p);
 935         if (r < 0)
 936                 return r;
 937 #endif
 938
 939         if (ret)
 940                 *ret = o;
 941
 942         if (offset)
 943                 *offset = p;
 944
 945         return 0;
 946 }
 947
 948 static int journal_file_append_data(
 949                 JournalFile *f,
 950                 const void *data, uint64_t size,
 951                 Object **ret, uint64_t *offset) {
 952
 953         uint64_t hash, p;
 954         uint64_t osize;
 955         Object *o;
 956         int r, compression = 0;
 957         const void *eq;
 958
 959         assert(f);
 960         assert(data || size == 0);
 961
 962         hash = hash64(data, size);
 963
 964         r = journal_file_find_data_object_with_hash(f, data, size, hash, &o, &p);
 965         if (r < 0)
 966                 return r;
 967         else if (r > 0) {
 968
 969                 if (ret)
 970                         *ret = o;
 971
 972                 if (offset)
 973                         *offset = p;
 974
 975                 return 0;
 976         }
 977
 978         osize = offsetof(Object, data.payload) + size;
 979         r = journal_file_append_object(f, OBJECT_DATA, osize, &o, &p);
 980         if (r < 0)
 981                 return r;
 982
 983         o->data.hash = htole64(hash);
 984
 985 #if defined(HAVE_XZ) || defined(HAVE_LZ4)
 986         if (f->compress_xz &&
 987             size >= COMPRESSION_SIZE_THRESHOLD) {
 988                 size_t rsize;
 989
 990                 compression = compress_blob(data, size, o->data.payload, &rsize);
 991
 992                 if (compression) {
 993                         o->object.size = htole64(offsetof(Object, data.payload) + rsize);
 994                         o->object.flags |= compression;
 995
 996                         log_debug("Compressed data object %"PRIu64" -> %zu using %s",
 997                                   size, rsize, object_compressed_to_string(compression));
 998                 }
 999         }
1000 #endif
1001
1002         if (!compression && size > 0)
1003                 memcpy(o->data.payload, data, size);
1004
1005         r = journal_file_link_data(f, o, p, hash);
1006         if (r < 0)
1007                 return r;
1008
1009         /* The linking might have altered the window, so let's
1010          * refresh our pointer */
1011         r = journal_file_move_to_object(f, OBJECT_DATA, p, &o);
1012         if (r < 0)
1013                 return r;
1014
1015         if (!data)
1016                 eq = NULL;
1017         else
1018                 eq = memchr(data, '=', size);
1019         if (eq && eq > data) {
1020                 Object *fo = NULL;
1021                 uint64_t fp;
1022
1023                 /* Create field object ... */
1024                 r = journal_file_append_field(f, data, (uint8_t*) eq - (uint8_t*) data, &fo, &fp);
1025                 if (r < 0)
1026                         return r;
1027
1028                 /* ... and link it in. */
1029                 o->data.next_field_offset = fo->field.head_data_offset;
1030                 fo->field.head_data_offset = le64toh(p);
1031         }
1032
1033 #ifdef HAVE_GCRYPT
1034         r = journal_file_hmac_put_object(f, OBJECT_DATA, o, p);
1035         if (r < 0)
1036                 return r;
1037 #endif
1038
1039         if (ret)
1040                 *ret = o;
1041
1042         if (offset)
1043                 *offset = p;
1044
1045         return 0;
1046 }
1047
1048 uint64_t journal_file_entry_n_items(Object *o) {
1049         assert(o);
1050
1051         if (o->object.type != OBJECT_ENTRY)
1052                 return 0;
1053
1054         return (le64toh(o->object.size) - offsetof(Object, entry.items)) / sizeof(EntryItem);
1055 }
1056
1057 uint64_t journal_file_entry_array_n_items(Object *o) {
1058         assert(o);
1059
1060         if (o->object.type != OBJECT_ENTRY_ARRAY)
1061                 return 0;
1062
1063         return (le64toh(o->object.size) - offsetof(Object, entry_array.items)) / sizeof(uint64_t);
1064 }
1065
1066 uint64_t journal_file_hash_table_n_items(Object *o) {
1067         assert(o);
1068
1069         if (o->object.type != OBJECT_DATA_HASH_TABLE &&
1070             o->object.type != OBJECT_FIELD_HASH_TABLE)
1071                 return 0;
1072
1073         return (le64toh(o->object.size) - offsetof(Object, hash_table.items)) / sizeof(HashItem);
1074 }
1075
1076 static int link_entry_into_array(JournalFile *f,
1077                                  le64_t *first,
1078                                  le64_t *idx,
1079                                  uint64_t p) {
1080         int r;
1081         uint64_t n = 0, ap = 0, q, i, a, hidx;
1082         Object *o;
1083
1084         assert(f);
1085         assert(first);
1086         assert(idx);
1087         assert(p > 0);
1088
1089         a = le64toh(*first);
1090         i = hidx = le64toh(*idx);
1091         while (a > 0) {
1092
1093                 r = journal_file_move_to_object(f, OBJECT_ENTRY_ARRAY, a, &o);
1094                 if (r < 0)
1095                         return r;
1096
1097                 n = journal_file_entry_array_n_items(o);
1098                 if (i < n) {
1099                         o->entry_array.items[i] = htole64(p);
1100                         *idx = htole64(hidx + 1);
1101                         return 0;
1102                 }
1103
1104                 i -= n;
1105                 ap = a;
1106                 a = le64toh(o->entry_array.next_entry_array_offset);
1107         }
1108
1109         if (hidx > n)
1110                 n = (hidx+1) * 2;
1111         else
1112                 n = n * 2;
1113
1114         if (n < 4)
1115                 n = 4;
1116
1117         r = journal_file_append_object(f, OBJECT_ENTRY_ARRAY,
1118                                        offsetof(Object, entry_array.items) + n * sizeof(uint64_t),
1119                                        &o, &q);
1120         if (r < 0)
1121                 return r;
1122
1123 #ifdef HAVE_GCRYPT
1124         r = journal_file_hmac_put_object(f, OBJECT_ENTRY_ARRAY, o, q);
1125         if (r < 0)
1126                 return r;
1127 #endif
1128
1129         o->entry_array.items[i] = htole64(p);
1130
1131         if (ap == 0)
1132                 *first = htole64(q);
1133         else {
1134                 r = journal_file_move_to_object(f, OBJECT_ENTRY_ARRAY, ap, &o);
1135                 if (r < 0)
1136                         return r;
1137
1138                 o->entry_array.next_entry_array_offset = htole64(q);
1139         }
1140
1141         if (JOURNAL_HEADER_CONTAINS(f->header, n_entry_arrays))
1142                 f->header->n_entry_arrays = htole64(le64toh(f->header->n_entry_arrays) + 1);
1143
1144         *idx = htole64(hidx + 1);
1145
1146         return 0;
1147 }
1148
1149 static int link_entry_into_array_plus_one(JournalFile *f,
1150                                           le64_t *extra,
1151                                           le64_t *first,
1152                                           le64_t *idx,
1153                                           uint64_t p) {
1154
1155         int r;
1156
1157         assert(f);
1158         assert(extra);
1159         assert(first);
1160         assert(idx);
1161         assert(p > 0);
1162
1163         if (*idx == 0)
1164                 *extra = htole64(p);
1165         else {
1166                 le64_t i;
1167
1168                 i = htole64(le64toh(*idx) - 1);
1169                 r = link_entry_into_array(f, first, &i, p);
1170                 if (r < 0)
1171                         return r;
1172         }
1173
1174         *idx = htole64(le64toh(*idx) + 1);
1175         return 0;
1176 }
1177
1178 static int journal_file_link_entry_item(JournalFile *f, Object *o, uint64_t offset, uint64_t i) {
1179         uint64_t p;
1180         int r;
1181         assert(f);
1182         assert(o);
1183         assert(offset > 0);
1184
1185         p = le64toh(o->entry.items[i].object_offset);
1186         if (p == 0)
1187                 return -EINVAL;
1188
1189         r = journal_file_move_to_object(f, OBJECT_DATA, p, &o);
1190         if (r < 0)
1191                 return r;
1192
1193         return link_entry_into_array_plus_one(f,
1194                                               &o->data.entry_offset,
1195                                               &o->data.entry_array_offset,
1196                                               &o->data.n_entries,
1197                                               offset);
1198 }
1199
1200 static int journal_file_link_entry(JournalFile *f, Object *o, uint64_t offset) {
1201         uint64_t n, i;
1202         int r;
1203
1204         assert(f);
1205         assert(o);
1206         assert(offset > 0);
1207
1208         if (o->object.type != OBJECT_ENTRY)
1209                 return -EINVAL;
1210
1211         __sync_synchronize();
1212
1213         /* Link up the entry itself */
1214         r = link_entry_into_array(f,
1215                                   &f->header->entry_array_offset,
1216                                   &f->header->n_entries,
1217                                   offset);
1218         if (r < 0)
1219                 return r;
1220
1221         /* log_debug("=> %s seqnr=%"PRIu64" n_entries=%"PRIu64, f->path, o->entry.seqnum, f->header->n_entries); */
1222
1223         if (f->header->head_entry_realtime == 0)
1224                 f->header->head_entry_realtime = o->entry.realtime;
1225
1226         f->header->tail_entry_realtime = o->entry.realtime;
1227         f->header->tail_entry_monotonic = o->entry.monotonic;
1228
1229         f->tail_entry_monotonic_valid = true;
1230
1231         /* Link up the items */
1232         n = journal_file_entry_n_items(o);
1233         for (i = 0; i < n; i++) {
1234                 r = journal_file_link_entry_item(f, o, offset, i);
1235                 if (r < 0)
1236                         return r;
1237         }
1238
1239         return 0;
1240 }
1241
1242 static int journal_file_append_entry_internal(
1243                 JournalFile *f,
1244                 const dual_timestamp *ts,
1245                 uint64_t xor_hash,
1246                 const EntryItem items[], unsigned n_items,
1247                 uint64_t *seqnum,
1248                 Object **ret, uint64_t *offset) {
1249         uint64_t np;
1250         uint64_t osize;
1251         Object *o;
1252         int r;
1253
1254         assert(f);
1255         assert(items || n_items == 0);
1256         assert(ts);
1257
1258         osize = offsetof(Object, entry.items) + (n_items * sizeof(EntryItem));
1259
1260         r = journal_file_append_object(f, OBJECT_ENTRY, osize, &o, &np);
1261         if (r < 0)
1262                 return r;
1263
1264         o->entry.seqnum = htole64(journal_file_entry_seqnum(f, seqnum));
1265         memcpy(o->entry.items, items, n_items * sizeof(EntryItem));
1266         o->entry.realtime = htole64(ts->realtime);
1267         o->entry.monotonic = htole64(ts->monotonic);
1268         o->entry.xor_hash = htole64(xor_hash);
1269         o->entry.boot_id = f->header->boot_id;
1270
1271 #ifdef HAVE_GCRYPT
1272         r = journal_file_hmac_put_object(f, OBJECT_ENTRY, o, np);
1273         if (r < 0)
1274                 return r;
1275 #endif
1276
1277         r = journal_file_link_entry(f, o, np);
1278         if (r < 0)
1279                 return r;
1280
1281         if (ret)
1282                 *ret = o;
1283
1284         if (offset)
1285                 *offset = np;
1286
1287         return 0;
1288 }
1289
1290 void journal_file_post_change(JournalFile *f) {
1291         assert(f);
1292
1293         /* inotify() does not receive IN_MODIFY events from file
1294          * accesses done via mmap(). After each access we hence
1295          * trigger IN_MODIFY by truncating the journal file to its
1296          * current size which triggers IN_MODIFY. */
1297
1298         __sync_synchronize();
1299
1300         if (ftruncate(f->fd, f->last_stat.st_size) < 0)
1301                 log_error_errno(errno, "Failed to truncate file to its own size: %m");
1302 }
1303
1304 static int entry_item_cmp(const void *_a, const void *_b) {
1305         const EntryItem *a = _a, *b = _b;
1306
1307         if (le64toh(a->object_offset) < le64toh(b->object_offset))
1308                 return -1;
1309         if (le64toh(a->object_offset) > le64toh(b->object_offset))
1310                 return 1;
1311         return 0;
1312 }
1313
1314 int journal_file_append_entry(JournalFile *f, const dual_timestamp *ts, const struct iovec iovec[], unsigned n_iovec, uint64_t *seqnum, Object **ret, uint64_t *offset) {
1315         unsigned i;
1316         EntryItem *items;
1317         int r;
1318         uint64_t xor_hash = 0;
1319         struct dual_timestamp _ts;
1320
1321         assert(f);
1322         assert(iovec || n_iovec == 0);
1323
1324         if (!ts) {
1325                 dual_timestamp_get(&_ts);
1326                 ts = &_ts;
1327         }
1328
1329         if (f->tail_entry_monotonic_valid &&
1330             ts->monotonic < le64toh(f->header->tail_entry_monotonic))
1331                 return -EINVAL;
1332
1333 #ifdef HAVE_GCRYPT
1334         r = journal_file_maybe_append_tag(f, ts->realtime);
1335         if (r < 0)
1336                 return r;
1337 #endif
1338
1339         /* alloca() can't take 0, hence let's allocate at least one */
1340         items = alloca(sizeof(EntryItem) * MAX(1u, n_iovec));
1341
1342         for (i = 0; i < n_iovec; i++) {
1343                 uint64_t p;
1344                 Object *o;
1345
1346                 r = journal_file_append_data(f, iovec[i].iov_base, iovec[i].iov_len, &o, &p);
1347                 if (r < 0)
1348                         return r;
1349
1350                 xor_hash ^= le64toh(o->data.hash);
1351                 items[i].object_offset = htole64(p);
1352                 items[i].hash = o->data.hash;
1353         }
1354
1355         /* Order by the position on disk, in order to improve seek
1356          * times for rotating media. */
1357         qsort_safe(items, n_iovec, sizeof(EntryItem), entry_item_cmp);
1358
1359         r = journal_file_append_entry_internal(f, ts, xor_hash, items, n_iovec, seqnum, ret, offset);
1360
1361         journal_file_post_change(f);
1362
1363         return r;
1364 }
1365
1366 typedef struct ChainCacheItem {
1367         uint64_t first; /* the array at the beginning of the chain */
1368         uint64_t array; /* the cached array */
1369         uint64_t begin; /* the first item in the cached array */
1370         uint64_t total; /* the total number of items in all arrays before this one in the chain */
1371         uint64_t last_index; /* the last index we looked at, to optimize locality when bisecting */
1372 } ChainCacheItem;
1373
1374 static void chain_cache_put(
1375                 OrderedHashmap *h,
1376                 ChainCacheItem *ci,
1377                 uint64_t first,
1378                 uint64_t array,
1379                 uint64_t begin,
1380                 uint64_t total,
1381                 uint64_t last_index) {
1382
1383         if (!ci) {
1384                 /* If the chain item to cache for this chain is the
1385                  * first one it's not worth caching anything */
1386                 if (array == first)
1387                         return;
1388
1389                 if (ordered_hashmap_size(h) >= CHAIN_CACHE_MAX) {
1390                         ci = ordered_hashmap_steal_first(h);
1391                         assert(ci);
1392                 } else {
1393                         ci = new(ChainCacheItem, 1);
1394                         if (!ci)
1395                                 return;
1396                 }
1397
1398                 ci->first = first;
1399
1400                 if (ordered_hashmap_put(h, &ci->first, ci) < 0) {
1401                         free(ci);
1402                         return;
1403                 }
1404         } else
1405                 assert(ci->first == first);
1406
1407         ci->array = array;
1408         ci->begin = begin;
1409         ci->total = total;
1410         ci->last_index = last_index;
1411 }
1412
1413 static int generic_array_get(
1414                 JournalFile *f,
1415                 uint64_t first,
1416                 uint64_t i,
1417                 Object **ret, uint64_t *offset) {
1418
1419         Object *o;
1420         uint64_t p = 0, a, t = 0;
1421         int r;
1422         ChainCacheItem *ci;
1423
1424         assert(f);
1425
1426         a = first;
1427
1428         /* Try the chain cache first */
1429         ci = ordered_hashmap_get(f->chain_cache, &first);
1430         if (ci && i > ci->total) {
1431                 a = ci->array;
1432                 i -= ci->total;
1433                 t = ci->total;
1434         }
1435
1436         while (a > 0) {
1437                 uint64_t k;
1438
1439                 r = journal_file_move_to_object(f, OBJECT_ENTRY_ARRAY, a, &o);
1440                 if (r < 0)
1441                         return r;
1442
1443                 k = journal_file_entry_array_n_items(o);
1444                 if (i < k) {
1445                         p = le64toh(o->entry_array.items[i]);
1446                         goto found;
1447                 }
1448
1449                 i -= k;
1450                 t += k;
1451                 a = le64toh(o->entry_array.next_entry_array_offset);
1452         }
1453
1454         return 0;
1455
1456 found:
1457         /* Let's cache this item for the next invocation */
1458         chain_cache_put(f->chain_cache, ci, first, a, le64toh(o->entry_array.items[0]), t, i);
1459
1460         r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o);
1461         if (r < 0)
1462                 return r;
1463
1464         if (ret)
1465                 *ret = o;
1466
1467         if (offset)
1468                 *offset = p;
1469
1470         return 1;
1471 }
1472
1473 static int generic_array_get_plus_one(
1474                 JournalFile *f,
1475                 uint64_t extra,
1476                 uint64_t first,
1477                 uint64_t i,
1478                 Object **ret, uint64_t *offset) {
1479
1480         Object *o;
1481
1482         assert(f);
1483
1484         if (i == 0) {
1485                 int r;
1486
1487                 r = journal_file_move_to_object(f, OBJECT_ENTRY, extra, &o);
1488                 if (r < 0)
1489                         return r;
1490
1491                 if (ret)
1492                         *ret = o;
1493
1494                 if (offset)
1495                         *offset = extra;
1496
1497                 return 1;
1498         }
1499
1500         return generic_array_get(f, first, i-1, ret, offset);
1501 }
1502
1503 enum {
1504         TEST_FOUND,
1505         TEST_LEFT,
1506         TEST_RIGHT
1507 };
1508
1509 static int generic_array_bisect(
1510                 JournalFile *f,
1511                 uint64_t first,
1512                 uint64_t n,
1513                 uint64_t needle,
1514                 int (*test_object)(JournalFile *f, uint64_t p, uint64_t needle),
1515                 direction_t direction,
1516                 Object **ret,
1517                 uint64_t *offset,
1518                 uint64_t *idx) {
1519
1520         uint64_t a, p, t = 0, i = 0, last_p = 0, last_index = (uint64_t) -1;
1521         bool subtract_one = false;
1522         Object *o, *array = NULL;
1523         int r;
1524         ChainCacheItem *ci;
1525
1526         assert(f);
1527         assert(test_object);
1528
1529         /* Start with the first array in the chain */
1530         a = first;
1531
1532         ci = ordered_hashmap_get(f->chain_cache, &first);
1533         if (ci && n > ci->total) {
1534                 /* Ah, we have iterated this bisection array chain
1535                  * previously! Let's see if we can skip ahead in the
1536                  * chain, as far as the last time. But we can't jump
1537                  * backwards in the chain, so let's check that
1538                  * first. */
1539
1540                 r = test_object(f, ci->begin, needle);
1541                 if (r < 0)
1542                         return r;
1543
1544                 if (r == TEST_LEFT) {
1545                         /* OK, what we are looking for is right of the
1546                          * begin of this EntryArray, so let's jump
1547                          * straight to previously cached array in the
1548                          * chain */
1549
1550                         a = ci->array;
1551                         n -= ci->total;
1552                         t = ci->total;
1553                         last_index = ci->last_index;
1554                 }
1555         }
1556
1557         while (a > 0) {
1558                 uint64_t left, right, k, lp;
1559
1560                 r = journal_file_move_to_object(f, OBJECT_ENTRY_ARRAY, a, &array);
1561                 if (r < 0)
1562                         return r;
1563
1564                 k = journal_file_entry_array_n_items(array);
1565                 right = MIN(k, n);
1566                 if (right <= 0)
1567                         return 0;
1568
1569                 i = right - 1;
1570                 lp = p = le64toh(array->entry_array.items[i]);
1571                 if (p <= 0)
1572                         return -EBADMSG;
1573
1574                 r = test_object(f, p, needle);
1575                 if (r < 0)
1576                         return r;
1577
1578                 if (r == TEST_FOUND)
1579                         r = direction == DIRECTION_DOWN ? TEST_RIGHT : TEST_LEFT;
1580
1581                 if (r == TEST_RIGHT) {
1582                         left = 0;
1583                         right -= 1;
1584
1585                         if (last_index != (uint64_t) -1) {
1586                                 assert(last_index <= right);
1587
1588                                 /* If we cached the last index we
1589                                  * looked at, let's try to not to jump
1590                                  * too wildly around and see if we can
1591                                  * limit the range to look at early to
1592                                  * the immediate neighbors of the last
1593                                  * index we looked at. */
1594
1595                                 if (last_index > 0) {
1596                                         uint64_t x = last_index - 1;
1597
1598                                         p = le64toh(array->entry_array.items[x]);
1599                                         if (p <= 0)
1600                                                 return -EBADMSG;
1601
1602                                         r = test_object(f, p, needle);
1603                                         if (r < 0)
1604                                                 return r;
1605
1606                                         if (r == TEST_FOUND)
1607                                                 r = direction == DIRECTION_DOWN ? TEST_RIGHT : TEST_LEFT;
1608
1609                                         if (r == TEST_RIGHT)
1610                                                 right = x;
1611                                         else
1612                                                 left = x + 1;
1613                                 }
1614
1615                                 if (last_index < right) {
1616                                         uint64_t y = last_index + 1;
1617
1618                                         p = le64toh(array->entry_array.items[y]);
1619                                         if (p <= 0)
1620                                                 return -EBADMSG;
1621
1622                                         r = test_object(f, p, needle);
1623                                         if (r < 0)
1624                                                 return r;
1625
1626                                         if (r == TEST_FOUND)
1627                                                 r = direction == DIRECTION_DOWN ? TEST_RIGHT : TEST_LEFT;
1628
1629                                         if (r == TEST_RIGHT)
1630                                                 right = y;
1631                                         else
1632                                                 left = y + 1;
1633                                 }
1634                         }
1635
1636                         for (;;) {
1637                                 if (left == right) {
1638                                         if (direction == DIRECTION_UP)
1639                                                 subtract_one = true;
1640
1641                                         i = left;
1642                                         goto found;
1643                                 }
1644
1645                                 assert(left < right);
1646                                 i = (left + right) / 2;
1647
1648                                 p = le64toh(array->entry_array.items[i]);
1649                                 if (p <= 0)
1650                                         return -EBADMSG;
1651
1652                                 r = test_object(f, p, needle);
1653                                 if (r < 0)
1654                                         return r;
1655
1656                                 if (r == TEST_FOUND)
1657                                         r = direction == DIRECTION_DOWN ? TEST_RIGHT : TEST_LEFT;
1658
1659                                 if (r == TEST_RIGHT)
1660                                         right = i;
1661                                 else
1662                                         left = i + 1;
1663                         }
1664                 }
1665
1666                 if (k >= n) {
1667                         if (direction == DIRECTION_UP) {
1668                                 i = n;
1669                                 subtract_one = true;
1670                                 goto found;
1671                         }
1672
1673                         return 0;
1674                 }
1675
1676                 last_p = lp;
1677
1678                 n -= k;
1679                 t += k;
1680                 last_index = (uint64_t) -1;
1681                 a = le64toh(array->entry_array.next_entry_array_offset);
1682         }
1683
1684         return 0;
1685
1686 found:
1687         if (subtract_one && t == 0 && i == 0)
1688                 return 0;
1689
1690         /* Let's cache this item for the next invocation */
1691         chain_cache_put(f->chain_cache, ci, first, a, le64toh(array->entry_array.items[0]), t, subtract_one ? (i > 0 ? i-1 : (uint64_t) -1) : i);
1692
1693         if (subtract_one && i == 0)
1694                 p = last_p;
1695         else if (subtract_one)
1696                 p = le64toh(array->entry_array.items[i-1]);
1697         else
1698                 p = le64toh(array->entry_array.items[i]);
1699
1700         r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o);
1701         if (r < 0)
1702                 return r;
1703
1704         if (ret)
1705                 *ret = o;
1706
1707         if (offset)
1708                 *offset = p;
1709
1710         if (idx)
1711                 *idx = t + i + (subtract_one ? -1 : 0);
1712
1713         return 1;
1714 }
1715
1716
1717 static int generic_array_bisect_plus_one(
1718                 JournalFile *f,
1719                 uint64_t extra,
1720                 uint64_t first,
1721                 uint64_t n,
1722                 uint64_t needle,
1723                 int (*test_object)(JournalFile *f, uint64_t p, uint64_t needle),
1724                 direction_t direction,
1725                 Object **ret,
1726                 uint64_t *offset,
1727                 uint64_t *idx) {
1728
1729         int r;
1730         bool step_back = false;
1731         Object *o;
1732
1733         assert(f);
1734         assert(test_object);
1735
1736         if (n <= 0)
1737                 return 0;
1738
1739         /* This bisects the array in object 'first', but first checks
1740          * an extra  */
1741         r = test_object(f, extra, needle);
1742         if (r < 0)
1743                 return r;
1744
1745         if (r == TEST_FOUND)
1746                 r = direction == DIRECTION_DOWN ? TEST_RIGHT : TEST_LEFT;
1747
1748         /* if we are looking with DIRECTION_UP then we need to first
1749            see if in the actual array there is a matching entry, and
1750            return the last one of that. But if there isn't any we need
1751            to return this one. Hence remember this, and return it
1752            below. */
1753         if (r == TEST_LEFT)
1754                 step_back = direction == DIRECTION_UP;
1755
1756         if (r == TEST_RIGHT) {
1757                 if (direction == DIRECTION_DOWN)
1758                         goto found;
1759                 else
1760                         return 0;
1761         }
1762
1763         r = generic_array_bisect(f, first, n-1, needle, test_object, direction, ret, offset, idx);
1764
1765         if (r == 0 && step_back)
1766                 goto found;
1767
1768         if (r > 0 && idx)
1769                 (*idx) ++;
1770
1771         return r;
1772
1773 found:
1774         r = journal_file_move_to_object(f, OBJECT_ENTRY, extra, &o);
1775         if (r < 0)
1776                 return r;
1777
1778         if (ret)
1779                 *ret = o;
1780
1781         if (offset)
1782                 *offset = extra;
1783
1784         if (idx)
1785                 *idx = 0;
1786
1787         return 1;
1788 }
1789
1790 _pure_ static int test_object_offset(JournalFile *f, uint64_t p, uint64_t needle) {
1791         assert(f);
1792         assert(p > 0);
1793
1794         if (p == needle)
1795                 return TEST_FOUND;
1796         else if (p < needle)
1797                 return TEST_LEFT;
1798         else
1799                 return TEST_RIGHT;
1800 }
1801
1802 static int test_object_seqnum(JournalFile *f, uint64_t p, uint64_t needle) {
1803         Object *o;
1804         int r;
1805
1806         assert(f);
1807         assert(p > 0);
1808
1809         r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o);
1810         if (r < 0)
1811                 return r;
1812
1813         if (le64toh(o->entry.seqnum) == needle)
1814                 return TEST_FOUND;
1815         else if (le64toh(o->entry.seqnum) < needle)
1816                 return TEST_LEFT;
1817         else
1818                 return TEST_RIGHT;
1819 }
1820
1821 int journal_file_move_to_entry_by_seqnum(
1822                 JournalFile *f,
1823                 uint64_t seqnum,
1824                 direction_t direction,
1825                 Object **ret,
1826                 uint64_t *offset) {
1827
1828         return generic_array_bisect(f,
1829                                     le64toh(f->header->entry_array_offset),
1830                                     le64toh(f->header->n_entries),
1831                                     seqnum,
1832                                     test_object_seqnum,
1833                                     direction,
1834                                     ret, offset, NULL);
1835 }
1836
1837 static int test_object_realtime(JournalFile *f, uint64_t p, uint64_t needle) {
1838         Object *o;
1839         int r;
1840
1841         assert(f);
1842         assert(p > 0);
1843
1844         r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o);
1845         if (r < 0)
1846                 return r;
1847
1848         if (le64toh(o->entry.realtime) == needle)
1849                 return TEST_FOUND;
1850         else if (le64toh(o->entry.realtime) < needle)
1851                 return TEST_LEFT;
1852         else
1853                 return TEST_RIGHT;
1854 }
1855
1856 int journal_file_move_to_entry_by_realtime(
1857                 JournalFile *f,
1858                 uint64_t realtime,
1859                 direction_t direction,
1860                 Object **ret,
1861                 uint64_t *offset) {
1862
1863         return generic_array_bisect(f,
1864                                     le64toh(f->header->entry_array_offset),
1865                                     le64toh(f->header->n_entries),
1866                                     realtime,
1867                                     test_object_realtime,
1868                                     direction,
1869                                     ret, offset, NULL);
1870 }
1871
1872 static int test_object_monotonic(JournalFile *f, uint64_t p, uint64_t needle) {
1873         Object *o;
1874         int r;
1875
1876         assert(f);
1877         assert(p > 0);
1878
1879         r = journal_file_move_to_object(f, OBJECT_ENTRY, p, &o);
1880         if (r < 0)
1881                 return r;
1882
1883         if (le64toh(o->entry.monotonic) == needle)
1884                 return TEST_FOUND;
1885         else if (le64toh(o->entry.monotonic) < needle)
1886                 return TEST_LEFT;
1887         else
1888                 return TEST_RIGHT;
1889 }
1890
1891 static inline int find_data_object_by_boot_id(
1892                 JournalFile *f,
1893                 sd_id128_t boot_id,
1894                 Object **o,
1895                 uint64_t *b) {
1896         char t[sizeof("_BOOT_ID=")-1 + 32 + 1] = "_BOOT_ID=";
1897
1898         sd_id128_to_string(boot_id, t + 9);
1899         return journal_file_find_data_object(f, t, sizeof(t) - 1, o, b);
1900 }
1901
1902 int journal_file_move_to_entry_by_monotonic(
1903                 JournalFile *f,
1904                 sd_id128_t boot_id,
1905                 uint64_t monotonic,
1906                 direction_t direction,
1907                 Object **ret,
1908                 uint64_t *offset) {
1909
1910         Object *o;
1911         int r;
1912
1913         assert(f);
1914
1915         r = find_data_object_by_boot_id(f, boot_id, &o, NULL);
1916         if (r < 0)
1917                 return r;
1918         if (r == 0)
1919                 return -ENOENT;
1920
1921         return generic_array_bisect_plus_one(f,
1922                                              le64toh(o->data.entry_offset),
1923                                              le64toh(o->data.entry_array_offset),
1924                                              le64toh(o->data.n_entries),
1925                                              monotonic,
1926                                              test_object_monotonic,
1927                                              direction,
1928                                              ret, offset, NULL);
1929 }
1930
1931 void journal_file_reset_location(JournalFile *f) {
1932         f->location_type = LOCATION_HEAD;
1933         f->current_offset = 0;
1934         f->current_seqnum = 0;
1935         f->current_realtime = 0;
1936         f->current_monotonic = 0;
1937         zero(f->current_boot_id);
1938         f->current_xor_hash = 0;
1939 }
1940
1941 void journal_file_save_location(JournalFile *f, direction_t direction, Object *o, uint64_t offset) {
1942         f->last_direction = direction;
1943         f->location_type = LOCATION_SEEK;
1944         f->current_offset = offset;
1945         f->current_seqnum = le64toh(o->entry.seqnum);
1946         f->current_realtime = le64toh(o->entry.realtime);
1947         f->current_monotonic = le64toh(o->entry.monotonic);
1948         f->current_boot_id = o->entry.boot_id;
1949         f->current_xor_hash = le64toh(o->entry.xor_hash);
1950 }
1951
1952 int journal_file_compare_locations(JournalFile *af, JournalFile *bf) {
1953         assert(af);
1954         assert(bf);
1955         assert(af->location_type == LOCATION_SEEK);
1956         assert(bf->location_type == LOCATION_SEEK);
1957
1958         /* If contents and timestamps match, these entries are
1959          * identical, even if the seqnum does not match */
1960         if (sd_id128_equal(af->current_boot_id, bf->current_boot_id) &&
1961             af->current_monotonic == bf->current_monotonic &&
1962             af->current_realtime == bf->current_realtime &&
1963             af->current_xor_hash == bf->current_xor_hash)
1964                 return 0;
1965
1966         if (sd_id128_equal(af->header->seqnum_id, bf->header->seqnum_id)) {
1967
1968                 /* If this is from the same seqnum source, compare
1969                  * seqnums */
1970                 if (af->current_seqnum < bf->current_seqnum)
1971                         return -1;
1972                 if (af->current_seqnum > bf->current_seqnum)
1973                         return 1;
1974
1975                 /* Wow! This is weird, different data but the same
1976                  * seqnums? Something is borked, but let's make the
1977                  * best of it and compare by time. */
1978         }
1979
1980         if (sd_id128_equal(af->current_boot_id, bf->current_boot_id)) {
1981
1982                 /* If the boot id matches, compare monotonic time */
1983                 if (af->current_monotonic < bf->current_monotonic)
1984                         return -1;
1985                 if (af->current_monotonic > bf->current_monotonic)
1986                         return 1;
1987         }
1988
1989         /* Otherwise, compare UTC time */
1990         if (af->current_realtime < bf->current_realtime)
1991                 return -1;
1992         if (af->current_realtime > bf->current_realtime)
1993                 return 1;
1994
1995         /* Finally, compare by contents */
1996         if (af->current_xor_hash < bf->current_xor_hash)
1997                 return -1;
1998         if (af->current_xor_hash > bf->current_xor_hash)
1999                 return 1;
2000
2001         return 0;
2002 }
2003
2004 int journal_file_next_entry(
2005                 JournalFile *f,
2006                 uint64_t p,
2007                 direction_t direction,
2008                 Object **ret, uint64_t *offset) {
2009
2010         uint64_t i, n, ofs;
2011         int r;
2012
2013         assert(f);
2014
2015         n = le64toh(f->header->n_entries);
2016         if (n <= 0)
2017                 return 0;
2018
2019         if (p == 0)
2020                 i = direction == DIRECTION_DOWN ? 0 : n - 1;
2021         else {
2022                 r = generic_array_bisect(f,
2023                                          le64toh(f->header->entry_array_offset),
2024                                          le64toh(f->header->n_entries),
2025                                          p,
2026                                          test_object_offset,
2027                                          DIRECTION_DOWN,
2028                                          NULL, NULL,
2029                                          &i);
2030                 if (r <= 0)
2031                         return r;
2032
2033                 if (direction == DIRECTION_DOWN) {
2034                         if (i >= n - 1)
2035                                 return 0;
2036
2037                         i++;
2038                 } else {
2039                         if (i <= 0)
2040                                 return 0;
2041
2042                         i--;
2043                 }
2044         }
2045
2046         /* And jump to it */
2047         r = generic_array_get(f,
2048                               le64toh(f->header->entry_array_offset),
2049                               i,
2050                               ret, &ofs);
2051         if (r <= 0)
2052                 return r;
2053
2054         if (p > 0 &&
2055             (direction == DIRECTION_DOWN ? ofs <= p : ofs >= p)) {
2056                 log_debug("%s: entry array corrupted at entry %"PRIu64,
2057                           f->path, i);
2058                 return -EBADMSG;
2059         }
2060
2061         if (offset)
2062                 *offset = ofs;
2063
2064         return 1;
2065 }
2066
2067 int journal_file_next_entry_for_data(
2068                 JournalFile *f,
2069                 Object *o, uint64_t p,
2070                 uint64_t data_offset,
2071                 direction_t direction,
2072                 Object **ret, uint64_t *offset) {
2073
2074         uint64_t n, i;
2075         int r;
2076         Object *d;
2077
2078         assert(f);
2079         assert(p > 0 || !o);
2080
2081         r = journal_file_move_to_object(f, OBJECT_DATA, data_offset, &d);
2082         if (r < 0)
2083                 return r;
2084
2085         n = le64toh(d->data.n_entries);
2086         if (n <= 0)
2087                 return n;
2088
2089         if (!o)
2090                 i = direction == DIRECTION_DOWN ? 0 : n - 1;
2091         else {
2092                 if (o->object.type != OBJECT_ENTRY)
2093                         return -EINVAL;
2094
2095                 r = generic_array_bisect_plus_one(f,
2096                                                   le64toh(d->data.entry_offset),
2097                                                   le64toh(d->data.entry_array_offset),
2098                                                   le64toh(d->data.n_entries),
2099                                                   p,
2100                                                   test_object_offset,
2101                                                   DIRECTION_DOWN,
2102                                                   NULL, NULL,
2103                                                   &i);
2104
2105                 if (r <= 0)
2106                         return r;
2107
2108                 if (direction == DIRECTION_DOWN) {
2109                         if (i >= n - 1)
2110                                 return 0;
2111
2112                         i++;
2113                 } else {
2114                         if (i <= 0)
2115                                 return 0;
2116
2117                         i--;
2118                 }
2119
2120         }
2121
2122         return generic_array_get_plus_one(f,
2123                                           le64toh(d->data.entry_offset),
2124                                           le64toh(d->data.entry_array_offset),
2125                                           i,
2126                                           ret, offset);
2127 }
2128
2129 int journal_file_move_to_entry_by_offset_for_data(
2130                 JournalFile *f,
2131                 uint64_t data_offset,
2132                 uint64_t p,
2133                 direction_t direction,
2134                 Object **ret, uint64_t *offset) {
2135
2136         int r;
2137         Object *d;
2138
2139         assert(f);
2140
2141         r = journal_file_move_to_object(f, OBJECT_DATA, data_offset, &d);
2142         if (r < 0)
2143                 return r;
2144
2145         return generic_array_bisect_plus_one(f,
2146                                              le64toh(d->data.entry_offset),
2147                                              le64toh(d->data.entry_array_offset),
2148                                              le64toh(d->data.n_entries),
2149                                              p,
2150                                              test_object_offset,
2151                                              direction,
2152                                              ret, offset, NULL);
2153 }
2154
2155 int journal_file_move_to_entry_by_monotonic_for_data(
2156                 JournalFile *f,
2157                 uint64_t data_offset,
2158                 sd_id128_t boot_id,
2159                 uint64_t monotonic,
2160                 direction_t direction,
2161                 Object **ret, uint64_t *offset) {
2162
2163         Object *o, *d;
2164         int r;
2165         uint64_t b, z;
2166
2167         assert(f);
2168
2169         /* First, seek by time */
2170         r = find_data_object_by_boot_id(f, boot_id, &o, &b);
2171         if (r < 0)
2172                 return r;
2173         if (r == 0)
2174                 return -ENOENT;
2175
2176         r = generic_array_bisect_plus_one(f,
2177                                           le64toh(o->data.entry_offset),
2178                                           le64toh(o->data.entry_array_offset),
2179                                           le64toh(o->data.n_entries),
2180                                           monotonic,
2181                                           test_object_monotonic,
2182                                           direction,
2183                                           NULL, &z, NULL);
2184         if (r <= 0)
2185                 return r;
2186
2187         /* And now, continue seeking until we find an entry that
2188          * exists in both bisection arrays */
2189
2190         for (;;) {
2191                 Object *qo;
2192                 uint64_t p, q;
2193
2194                 r = journal_file_move_to_object(f, OBJECT_DATA, data_offset, &d);
2195                 if (r < 0)
2196                         return r;
2197
2198                 r = generic_array_bisect_plus_one(f,
2199                                                   le64toh(d->data.entry_offset),
2200                                                   le64toh(d->data.entry_array_offset),
2201                                                   le64toh(d->data.n_entries),
2202                                                   z,
2203                                                   test_object_offset,
2204                                                   direction,
2205                                                   NULL, &p, NULL);
2206                 if (r <= 0)
2207                         return r;
2208
2209                 r = journal_file_move_to_object(f, OBJECT_DATA, b, &o);
2210                 if (r < 0)
2211                         return r;
2212
2213                 r = generic_array_bisect_plus_one(f,
2214                                                   le64toh(o->data.entry_offset),
2215                                                   le64toh(o->data.entry_array_offset),
2216                                                   le64toh(o->data.n_entries),
2217                                                   p,
2218                                                   test_object_offset,
2219                                                   direction,
2220                                                   &qo, &q, NULL);
2221
2222                 if (r <= 0)
2223                         return r;
2224
2225                 if (p == q) {
2226                         if (ret)
2227                                 *ret = qo;
2228                         if (offset)
2229                                 *offset = q;
2230
2231                         return 1;
2232                 }
2233
2234                 z = q;
2235         }
2236 }
2237
2238 int journal_file_move_to_entry_by_seqnum_for_data(
2239                 JournalFile *f,
2240                 uint64_t data_offset,
2241                 uint64_t seqnum,
2242                 direction_t direction,
2243                 Object **ret, uint64_t *offset) {
2244
2245         Object *d;
2246         int r;
2247
2248         assert(f);
2249
2250         r = journal_file_move_to_object(f, OBJECT_DATA, data_offset, &d);
2251         if (r < 0)
2252                 return r;
2253
2254         return generic_array_bisect_plus_one(f,
2255                                              le64toh(d->data.entry_offset),
2256                                              le64toh(d->data.entry_array_offset),
2257                                              le64toh(d->data.n_entries),
2258                                              seqnum,
2259                                              test_object_seqnum,
2260                                              direction,
2261                                              ret, offset, NULL);
2262 }
2263
2264 int journal_file_move_to_entry_by_realtime_for_data(
2265                 JournalFile *f,
2266                 uint64_t data_offset,
2267                 uint64_t realtime,
2268                 direction_t direction,
2269                 Object **ret, uint64_t *offset) {
2270
2271         Object *d;
2272         int r;
2273
2274         assert(f);
2275
2276         r = journal_file_move_to_object(f, OBJECT_DATA, data_offset, &d);
2277         if (r < 0)
2278                 return r;
2279
2280         return generic_array_bisect_plus_one(f,
2281                                              le64toh(d->data.entry_offset),
2282                                              le64toh(d->data.entry_array_offset),
2283                                              le64toh(d->data.n_entries),
2284                                              realtime,
2285                                              test_object_realtime,
2286                                              direction,
2287                                              ret, offset, NULL);
2288 }
2289
2290 void journal_file_dump(JournalFile *f) {
2291         Object *o;
2292         int r;
2293         uint64_t p;
2294
2295         assert(f);
2296
2297         journal_file_print_header(f);
2298
2299         p = le64toh(f->header->header_size);
2300         while (p != 0) {
2301                 r = journal_file_move_to_object(f, OBJECT_UNUSED, p, &o);
2302                 if (r < 0)
2303                         goto fail;
2304
2305                 switch (o->object.type) {
2306
2307                 case OBJECT_UNUSED:
2308                         printf("Type: OBJECT_UNUSED\n");
2309                         break;
2310
2311                 case OBJECT_DATA:
2312                         printf("Type: OBJECT_DATA\n");
2313                         break;
2314
2315                 case OBJECT_FIELD:
2316                         printf("Type: OBJECT_FIELD\n");
2317                         break;
2318
2319                 case OBJECT_ENTRY:
2320                         printf("Type: OBJECT_ENTRY seqnum=%"PRIu64" monotonic=%"PRIu64" realtime=%"PRIu64"\n",
2321                                le64toh(o->entry.seqnum),
2322                                le64toh(o->entry.monotonic),
2323                                le64toh(o->entry.realtime));
2324                         break;
2325
2326                 case OBJECT_FIELD_HASH_TABLE:
2327                         printf("Type: OBJECT_FIELD_HASH_TABLE\n");
2328                         break;
2329
2330                 case OBJECT_DATA_HASH_TABLE:
2331                         printf("Type: OBJECT_DATA_HASH_TABLE\n");
2332                         break;
2333
2334                 case OBJECT_ENTRY_ARRAY:
2335                         printf("Type: OBJECT_ENTRY_ARRAY\n");
2336                         break;
2337
2338                 case OBJECT_TAG:
2339                         printf("Type: OBJECT_TAG seqnum=%"PRIu64" epoch=%"PRIu64"\n",
2340                                le64toh(o->tag.seqnum),
2341                                le64toh(o->tag.epoch));
2342                         break;
2343
2344                 default:
2345                         printf("Type: unknown (%u)\n", o->object.type);
2346                         break;
2347                 }
2348
2349                 if (o->object.flags & OBJECT_COMPRESSION_MASK)
2350                         printf("Flags: %s\n",
2351                                object_compressed_to_string(o->object.flags & OBJECT_COMPRESSION_MASK));
2352
2353                 if (p == le64toh(f->header->tail_object_offset))
2354                         p = 0;
2355                 else
2356                         p = p + ALIGN64(le64toh(o->object.size));
2357         }
2358
2359         return;
2360 fail:
2361         log_error("File corrupt");
2362 }
2363
2364 static const char* format_timestamp_safe(char *buf, size_t l, usec_t t) {
2365         const char *x;
2366
2367         x = format_timestamp(buf, l, t);
2368         if (x)
2369                 return x;
2370         return " --- ";
2371 }
2372
2373 void journal_file_print_header(JournalFile *f) {
2374         char a[33], b[33], c[33], d[33];
2375         char x[FORMAT_TIMESTAMP_MAX], y[FORMAT_TIMESTAMP_MAX], z[FORMAT_TIMESTAMP_MAX];
2376         struct stat st;
2377         char bytes[FORMAT_BYTES_MAX];
2378
2379         assert(f);
2380
2381         printf("File Path: %s\n"
2382                "File ID: %s\n"
2383                "Machine ID: %s\n"
2384                "Boot ID: %s\n"
2385                "Sequential Number ID: %s\n"
2386                "State: %s\n"
2387                "Compatible Flags:%s%s\n"
2388                "Incompatible Flags:%s%s%s\n"
2389                "Header size: %"PRIu64"\n"
2390                "Arena size: %"PRIu64"\n"
2391                "Data Hash Table Size: %"PRIu64"\n"
2392                "Field Hash Table Size: %"PRIu64"\n"
2393                "Rotate Suggested: %s\n"
2394                "Head Sequential Number: %"PRIu64"\n"
2395                "Tail Sequential Number: %"PRIu64"\n"
2396                "Head Realtime Timestamp: %s\n"
2397                "Tail Realtime Timestamp: %s\n"
2398                "Tail Monotonic Timestamp: %s\n"
2399                "Objects: %"PRIu64"\n"
2400                "Entry Objects: %"PRIu64"\n",
2401                f->path,
2402                sd_id128_to_string(f->header->file_id, a),
2403                sd_id128_to_string(f->header->machine_id, b),
2404                sd_id128_to_string(f->header->boot_id, c),
2405                sd_id128_to_string(f->header->seqnum_id, d),
2406                f->header->state == STATE_OFFLINE ? "OFFLINE" :
2407                f->header->state == STATE_ONLINE ? "ONLINE" :
2408                f->header->state == STATE_ARCHIVED ? "ARCHIVED" : "UNKNOWN",
2409                JOURNAL_HEADER_SEALED(f->header) ? " SEALED" : "",
2410                (le32toh(f->header->compatible_flags) & ~HEADER_COMPATIBLE_ANY) ? " ???" : "",
2411                JOURNAL_HEADER_COMPRESSED_XZ(f->header) ? " COMPRESSED-XZ" : "",
2412                JOURNAL_HEADER_COMPRESSED_LZ4(f->header) ? " COMPRESSED-LZ4" : "",
2413                (le32toh(f->header->incompatible_flags) & ~HEADER_INCOMPATIBLE_ANY) ? " ???" : "",
2414                le64toh(f->header->header_size),
2415                le64toh(f->header->arena_size),
2416                le64toh(f->header->data_hash_table_size) / sizeof(HashItem),
2417                le64toh(f->header->field_hash_table_size) / sizeof(HashItem),
2418                yes_no(journal_file_rotate_suggested(f, 0)),
2419                le64toh(f->header->head_entry_seqnum),
2420                le64toh(f->header->tail_entry_seqnum),
2421                format_timestamp_safe(x, sizeof(x), le64toh(f->header->head_entry_realtime)),
2422                format_timestamp_safe(y, sizeof(y), le64toh(f->header->tail_entry_realtime)),
2423                format_timespan(z, sizeof(z), le64toh(f->header->tail_entry_monotonic), USEC_PER_MSEC),
2424                le64toh(f->header->n_objects),
2425                le64toh(f->header->n_entries));
2426
2427         if (JOURNAL_HEADER_CONTAINS(f->header, n_data))
2428                 printf("Data Objects: %"PRIu64"\n"
2429                        "Data Hash Table Fill: %.1f%%\n",
2430                        le64toh(f->header->n_data),
2431                        100.0 * (double) le64toh(f->header->n_data) / ((double) (le64toh(f->header->data_hash_table_size) / sizeof(HashItem))));
2432
2433         if (JOURNAL_HEADER_CONTAINS(f->header, n_fields))
2434                 printf("Field Objects: %"PRIu64"\n"
2435                        "Field Hash Table Fill: %.1f%%\n",
2436                        le64toh(f->header->n_fields),
2437                        100.0 * (double) le64toh(f->header->n_fields) / ((double) (le64toh(f->header->field_hash_table_size) / sizeof(HashItem))));
2438
2439         if (JOURNAL_HEADER_CONTAINS(f->header, n_tags))
2440                 printf("Tag Objects: %"PRIu64"\n",
2441                        le64toh(f->header->n_tags));
2442         if (JOURNAL_HEADER_CONTAINS(f->header, n_entry_arrays))
2443                 printf("Entry Array Objects: %"PRIu64"\n",
2444                        le64toh(f->header->n_entry_arrays));
2445
2446         if (fstat(f->fd, &st) >= 0)
2447                 printf("Disk usage: %s\n", format_bytes(bytes, sizeof(bytes), (off_t) st.st_blocks * 512ULL));
2448 }
2449
2450 int journal_file_open(
2451                 const char *fname,
2452                 int flags,
2453                 mode_t mode,
2454                 bool compress,
2455                 bool seal,
2456                 JournalMetrics *metrics,
2457                 MMapCache *mmap_cache,
2458                 JournalFile *template,
2459                 JournalFile **ret) {
2460
2461         JournalFile *f;
2462         int r;
2463         bool newly_created = false;
2464
2465         assert(fname);
2466         assert(ret);
2467
2468         if ((flags & O_ACCMODE) != O_RDONLY &&
2469             (flags & O_ACCMODE) != O_RDWR)
2470                 return -EINVAL;
2471
2472         if (!endswith(fname, ".journal") &&
2473             !endswith(fname, ".journal~"))
2474                 return -EINVAL;
2475
2476         f = new0(JournalFile, 1);
2477         if (!f)
2478                 return -ENOMEM;
2479
2480         f->fd = -1;
2481         f->mode = mode;
2482
2483         f->flags = flags;
2484         f->prot = prot_from_flags(flags);
2485         f->writable = (flags & O_ACCMODE) != O_RDONLY;
2486 #if defined(HAVE_LZ4)
2487         f->compress_lz4 = compress;
2488 #elif defined(HAVE_XZ)
2489         f->compress_xz = compress;
2490 #endif
2491 #ifdef HAVE_GCRYPT
2492         f->seal = seal;
2493 #endif
2494
2495         if (mmap_cache)
2496                 f->mmap = mmap_cache_ref(mmap_cache);
2497         else {
2498                 f->mmap = mmap_cache_new();
2499                 if (!f->mmap) {
2500                         r = -ENOMEM;
2501                         goto fail;
2502                 }
2503         }
2504
2505         f->path = strdup(fname);
2506         if (!f->path) {
2507                 r = -ENOMEM;
2508                 goto fail;
2509         }
2510
2511         f->chain_cache = ordered_hashmap_new(&uint64_hash_ops);
2512         if (!f->chain_cache) {
2513                 r = -ENOMEM;
2514                 goto fail;
2515         }
2516
2517         f->fd = open(f->path, f->flags|O_CLOEXEC, f->mode);
2518         if (f->fd < 0) {
2519                 r = -errno;
2520                 goto fail;
2521         }
2522
2523         if (fstat(f->fd, &f->last_stat) < 0) {
2524                 r = -errno;
2525                 goto fail;
2526         }
2527
2528         if (f->last_stat.st_size == 0 && f->writable) {
2529                 uint64_t crtime;
2530
2531                 /* Let's attach the creation time to the journal file,
2532                  * so that the vacuuming code knows the age of this
2533                  * file even if the file might end up corrupted one
2534                  * day... Ideally we'd just use the creation time many
2535                  * file systems maintain for each file, but there is
2536                  * currently no usable API to query this, hence let's
2537                  * emulate this via extended attributes. If extended
2538                  * attributes are not supported we'll just skip this,
2539                  * and rely solely on mtime/atime/ctime of the file. */
2540
2541                 crtime = htole64((uint64_t) now(CLOCK_REALTIME));
2542                 fsetxattr(f->fd, "user.crtime_usec", &crtime, sizeof(crtime), XATTR_CREATE);
2543
2544 #ifdef HAVE_GCRYPT
2545                 /* Try to load the FSPRG state, and if we can't, then
2546                  * just don't do sealing */
2547                 if (f->seal) {
2548                         r = journal_file_fss_load(f);
2549                         if (r < 0)
2550                                 f->seal = false;
2551                 }
2552 #endif
2553
2554                 r = journal_file_init_header(f, template);
2555                 if (r < 0)
2556                         goto fail;
2557
2558                 if (fstat(f->fd, &f->last_stat) < 0) {
2559                         r = -errno;
2560                         goto fail;
2561                 }
2562
2563                 newly_created = true;
2564         }
2565
2566         if (f->last_stat.st_size < (off_t) HEADER_SIZE_MIN) {
2567                 r = -EIO;
2568                 goto fail;
2569         }
2570
2571         f->header = mmap(NULL, PAGE_ALIGN(sizeof(Header)), prot_from_flags(flags), MAP_SHARED, f->fd, 0);
2572         if (f->header == MAP_FAILED) {
2573                 f->header = NULL;
2574                 r = -errno;
2575                 goto fail;
2576         }
2577
2578         if (!newly_created) {
2579                 r = journal_file_verify_header(f);
2580                 if (r < 0)
2581                         goto fail;
2582         }
2583
2584 #ifdef HAVE_GCRYPT
2585         if (!newly_created && f->writable) {
2586                 r = journal_file_fss_load(f);
2587                 if (r < 0)
2588                         goto fail;
2589         }
2590 #endif
2591
2592         if (f->writable) {
2593                 if (metrics) {
2594                         journal_default_metrics(metrics, f->fd);
2595                         f->metrics = *metrics;
2596                 } else if (template)
2597                         f->metrics = template->metrics;
2598
2599                 r = journal_file_refresh_header(f);
2600                 if (r < 0)
2601                         goto fail;
2602         }
2603
2604 #ifdef HAVE_GCRYPT
2605         r = journal_file_hmac_setup(f);
2606         if (r < 0)
2607                 goto fail;
2608 #endif
2609
2610         if (newly_created) {
2611                 r = journal_file_setup_field_hash_table(f);
2612                 if (r < 0)
2613                         goto fail;
2614
2615                 r = journal_file_setup_data_hash_table(f);
2616                 if (r < 0)
2617                         goto fail;
2618
2619 #ifdef HAVE_GCRYPT
2620                 r = journal_file_append_first_tag(f);
2621                 if (r < 0)
2622                         goto fail;
2623 #endif
2624         }
2625
2626         r = journal_file_map_field_hash_table(f);
2627         if (r < 0)
2628                 goto fail;
2629
2630         r = journal_file_map_data_hash_table(f);
2631         if (r < 0)
2632                 goto fail;
2633
2634         *ret = f;
2635         return 0;
2636
2637 fail:
2638         journal_file_close(f);
2639
2640         return r;
2641 }
2642
2643 int journal_file_rotate(JournalFile **f, bool compress, bool seal) {
2644         _cleanup_free_ char *p = NULL;
2645         size_t l;
2646         JournalFile *old_file, *new_file = NULL;
2647         int r;
2648
2649         assert(f);
2650         assert(*f);
2651
2652         old_file = *f;
2653
2654         if (!old_file->writable)
2655                 return -EINVAL;
2656
2657         if (!endswith(old_file->path, ".journal"))
2658                 return -EINVAL;
2659
2660         l = strlen(old_file->path);
2661         r = asprintf(&p, "%.*s@" SD_ID128_FORMAT_STR "-%016"PRIx64"-%016"PRIx64".journal",
2662                      (int) l - 8, old_file->path,
2663                      SD_ID128_FORMAT_VAL(old_file->header->seqnum_id),
2664                      le64toh((*f)->header->head_entry_seqnum),
2665                      le64toh((*f)->header->head_entry_realtime));
2666         if (r < 0)
2667                 return -ENOMEM;
2668
2669         r = rename(old_file->path, p);
2670         if (r < 0)
2671                 return -errno;
2672
2673         old_file->header->state = STATE_ARCHIVED;
2674
2675         r = journal_file_open(old_file->path, old_file->flags, old_file->mode, compress, seal, NULL, old_file->mmap, old_file, &new_file);
2676         journal_file_close(old_file);
2677
2678         *f = new_file;
2679         return r;
2680 }
2681
2682 int journal_file_open_reliably(
2683                 const char *fname,
2684                 int flags,
2685                 mode_t mode,
2686                 bool compress,
2687                 bool seal,
2688                 JournalMetrics *metrics,
2689                 MMapCache *mmap_cache,
2690                 JournalFile *template,
2691                 JournalFile **ret) {
2692
2693         int r;
2694         size_t l;
2695         _cleanup_free_ char *p = NULL;
2696
2697         r = journal_file_open(fname, flags, mode, compress, seal,
2698                               metrics, mmap_cache, template, ret);
2699         if (r != -EBADMSG && /* corrupted */
2700             r != -ENODATA && /* truncated */
2701             r != -EHOSTDOWN && /* other machine */
2702             r != -EPROTONOSUPPORT && /* incompatible feature */
2703             r != -EBUSY && /* unclean shutdown */
2704             r != -ESHUTDOWN /* already archived */)
2705                 return r;
2706
2707         if ((flags & O_ACCMODE) == O_RDONLY)
2708                 return r;
2709
2710         if (!(flags & O_CREAT))
2711                 return r;
2712
2713         if (!endswith(fname, ".journal"))
2714                 return r;
2715
2716         /* The file is corrupted. Rotate it away and try it again (but only once) */
2717
2718         l = strlen(fname);
2719         if (asprintf(&p, "%.*s@%016llx-%016" PRIx64 ".journal~",
2720                      (int) l - 8, fname,
2721                      (unsigned long long) now(CLOCK_REALTIME),
2722                      random_u64()) < 0)
2723                 return -ENOMEM;
2724
2725         r = rename(fname, p);
2726         if (r < 0)
2727                 return -errno;
2728
2729         log_warning("File %s corrupted or uncleanly shut down, renaming and replacing.", fname);
2730
2731         return journal_file_open(fname, flags, mode, compress, seal,
2732                                  metrics, mmap_cache, template, ret);
2733 }
2734
2735 int journal_file_copy_entry(JournalFile *from, JournalFile *to, Object *o, uint64_t p, uint64_t *seqnum, Object **ret, uint64_t *offset) {
2736         uint64_t i, n;
2737         uint64_t q, xor_hash = 0;
2738         int r;
2739         EntryItem *items;
2740         dual_timestamp ts;
2741
2742         assert(from);
2743         assert(to);
2744         assert(o);
2745         assert(p);
2746
2747         if (!to->writable)
2748                 return -EPERM;
2749
2750         ts.monotonic = le64toh(o->entry.monotonic);
2751         ts.realtime = le64toh(o->entry.realtime);
2752
2753         n = journal_file_entry_n_items(o);
2754         /* alloca() can't take 0, hence let's allocate at least one */
2755         items = alloca(sizeof(EntryItem) * MAX(1u, n));
2756
2757         for (i = 0; i < n; i++) {
2758                 uint64_t l, h;
2759                 le64_t le_hash;
2760                 size_t t;
2761                 void *data;
2762                 Object *u;
2763
2764                 q = le64toh(o->entry.items[i].object_offset);
2765                 le_hash = o->entry.items[i].hash;
2766
2767                 r = journal_file_move_to_object(from, OBJECT_DATA, q, &o);
2768                 if (r < 0)
2769                         return r;
2770
2771                 if (le_hash != o->data.hash)
2772                         return -EBADMSG;
2773
2774                 l = le64toh(o->object.size) - offsetof(Object, data.payload);
2775                 t = (size_t) l;
2776
2777                 /* We hit the limit on 32bit machines */
2778                 if ((uint64_t) t != l)
2779                         return -E2BIG;
2780
2781                 if (o->object.flags & OBJECT_COMPRESSION_MASK) {
2782 #if defined(HAVE_XZ) || defined(HAVE_LZ4)
2783                         size_t rsize;
2784
2785                         r = decompress_blob(o->object.flags & OBJECT_COMPRESSION_MASK,
2786                                             o->data.payload, l, &from->compress_buffer, &from->compress_buffer_size, &rsize, 0);
2787                         if (r < 0)
2788                                 return r;
2789
2790                         data = from->compress_buffer;
2791                         l = rsize;
2792 #else
2793                         return -EPROTONOSUPPORT;
2794 #endif
2795                 } else
2796                         data = o->data.payload;
2797
2798                 r = journal_file_append_data(to, data, l, &u, &h);
2799                 if (r < 0)
2800                         return r;
2801
2802                 xor_hash ^= le64toh(u->data.hash);
2803                 items[i].object_offset = htole64(h);
2804                 items[i].hash = u->data.hash;
2805
2806                 r = journal_file_move_to_object(from, OBJECT_ENTRY, p, &o);
2807                 if (r < 0)
2808                         return r;
2809         }
2810
2811         return journal_file_append_entry_internal(to, &ts, xor_hash, items, n, seqnum, ret, offset);
2812 }
2813
2814 void journal_default_metrics(JournalMetrics *m, int fd) {
2815         uint64_t fs_size = 0;
2816         struct statvfs ss;
2817         char a[FORMAT_BYTES_MAX], b[FORMAT_BYTES_MAX], c[FORMAT_BYTES_MAX], d[FORMAT_BYTES_MAX];
2818
2819         assert(m);
2820         assert(fd >= 0);
2821
2822         if (fstatvfs(fd, &ss) >= 0)
2823                 fs_size = ss.f_frsize * ss.f_blocks;
2824
2825         if (m->max_use == (uint64_t) -1) {
2826
2827                 if (fs_size > 0) {
2828                         m->max_use = PAGE_ALIGN(fs_size / 10); /* 10% of file system size */
2829
2830                         if (m->max_use > DEFAULT_MAX_USE_UPPER)
2831                                 m->max_use = DEFAULT_MAX_USE_UPPER;
2832
2833                         if (m->max_use < DEFAULT_MAX_USE_LOWER)
2834                                 m->max_use = DEFAULT_MAX_USE_LOWER;
2835                 } else
2836                         m->max_use = DEFAULT_MAX_USE_LOWER;
2837         } else {
2838                 m->max_use = PAGE_ALIGN(m->max_use);
2839
2840                 if (m->max_use < JOURNAL_FILE_SIZE_MIN*2)
2841                         m->max_use = JOURNAL_FILE_SIZE_MIN*2;
2842         }
2843
2844         if (m->max_size == (uint64_t) -1) {
2845                 m->max_size = PAGE_ALIGN(m->max_use / 8); /* 8 chunks */
2846
2847                 if (m->max_size > DEFAULT_MAX_SIZE_UPPER)
2848                         m->max_size = DEFAULT_MAX_SIZE_UPPER;
2849         } else
2850                 m->max_size = PAGE_ALIGN(m->max_size);
2851
2852         if (m->max_size < JOURNAL_FILE_SIZE_MIN)
2853                 m->max_size = JOURNAL_FILE_SIZE_MIN;
2854
2855         if (m->max_size*2 > m->max_use)
2856                 m->max_use = m->max_size*2;
2857
2858         if (m->min_size == (uint64_t) -1)
2859                 m->min_size = JOURNAL_FILE_SIZE_MIN;
2860         else {
2861                 m->min_size = PAGE_ALIGN(m->min_size);
2862
2863                 if (m->min_size < JOURNAL_FILE_SIZE_MIN)
2864                         m->min_size = JOURNAL_FILE_SIZE_MIN;
2865
2866                 if (m->min_size > m->max_size)
2867                         m->max_size = m->min_size;
2868         }
2869
2870         if (m->keep_free == (uint64_t) -1) {
2871
2872                 if (fs_size > 0) {
2873                         m->keep_free = PAGE_ALIGN(fs_size * 3 / 20); /* 15% of file system size */
2874
2875                         if (m->keep_free > DEFAULT_KEEP_FREE_UPPER)
2876                                 m->keep_free = DEFAULT_KEEP_FREE_UPPER;
2877
2878                 } else
2879                         m->keep_free = DEFAULT_KEEP_FREE;
2880         }
2881
2882         log_debug("Fixed max_use=%s max_size=%s min_size=%s keep_free=%s",
2883                   format_bytes(a, sizeof(a), m->max_use),
2884                   format_bytes(b, sizeof(b), m->max_size),
2885                   format_bytes(c, sizeof(c), m->min_size),
2886                   format_bytes(d, sizeof(d), m->keep_free));
2887 }
2888
2889 int journal_file_get_cutoff_realtime_usec(JournalFile *f, usec_t *from, usec_t *to) {
2890         assert(f);
2891         assert(from || to);
2892
2893         if (from) {
2894                 if (f->header->head_entry_realtime == 0)
2895                         return -ENOENT;
2896
2897                 *from = le64toh(f->header->head_entry_realtime);
2898         }
2899
2900         if (to) {
2901                 if (f->header->tail_entry_realtime == 0)
2902                         return -ENOENT;
2903
2904                 *to = le64toh(f->header->tail_entry_realtime);
2905         }
2906
2907         return 1;
2908 }
2909
2910 int journal_file_get_cutoff_monotonic_usec(JournalFile *f, sd_id128_t boot_id, usec_t *from, usec_t *to) {
2911         Object *o;
2912         uint64_t p;
2913         int r;
2914
2915         assert(f);
2916         assert(from || to);
2917
2918         r = find_data_object_by_boot_id(f, boot_id, &o, &p);
2919         if (r <= 0)
2920                 return r;
2921
2922         if (le64toh(o->data.n_entries) <= 0)
2923                 return 0;
2924
2925         if (from) {
2926                 r = journal_file_move_to_object(f, OBJECT_ENTRY, le64toh(o->data.entry_offset), &o);
2927                 if (r < 0)
2928                         return r;
2929
2930                 *from = le64toh(o->entry.monotonic);
2931         }
2932
2933         if (to) {
2934                 r = journal_file_move_to_object(f, OBJECT_DATA, p, &o);
2935                 if (r < 0)
2936                         return r;
2937
2938                 r = generic_array_get_plus_one(f,
2939                                                le64toh(o->data.entry_offset),
2940                                                le64toh(o->data.entry_array_offset),
2941                                                le64toh(o->data.n_entries)-1,
2942                                                &o, NULL);
2943                 if (r <= 0)
2944                         return r;
2945
2946                 *to = le64toh(o->entry.monotonic);
2947         }
2948
2949         return 1;
2950 }
2951
2952 bool journal_file_rotate_suggested(JournalFile *f, usec_t max_file_usec) {
2953         assert(f);
2954
2955         /* If we gained new header fields we gained new features,
2956          * hence suggest a rotation */
2957         if (le64toh(f->header->header_size) < sizeof(Header)) {
2958                 log_debug("%s uses an outdated header, suggesting rotation.", f->path);
2959                 return true;
2960         }
2961
2962         /* Let's check if the hash tables grew over a certain fill
2963          * level (75%, borrowing this value from Java's hash table
2964          * implementation), and if so suggest a rotation. To calculate
2965          * the fill level we need the n_data field, which only exists
2966          * in newer versions. */
2967
2968         if (JOURNAL_HEADER_CONTAINS(f->header, n_data))
2969                 if (le64toh(f->header->n_data) * 4ULL > (le64toh(f->header->data_hash_table_size) / sizeof(HashItem)) * 3ULL) {
2970                         log_debug("Data hash table of %s has a fill level at %.1f (%"PRIu64" of %"PRIu64" items, %llu file size, %"PRIu64" bytes per hash table item), suggesting rotation.",
2971                                   f->path,
2972                                   100.0 * (double) le64toh(f->header->n_data) / ((double) (le64toh(f->header->data_hash_table_size) / sizeof(HashItem))),
2973                                   le64toh(f->header->n_data),
2974                                   le64toh(f->header->data_hash_table_size) / sizeof(HashItem),
2975                                   (unsigned long long) f->last_stat.st_size,
2976                                   f->last_stat.st_size / le64toh(f->header->n_data));
2977                         return true;
2978                 }
2979
2980         if (JOURNAL_HEADER_CONTAINS(f->header, n_fields))
2981                 if (le64toh(f->header->n_fields) * 4ULL > (le64toh(f->header->field_hash_table_size) / sizeof(HashItem)) * 3ULL) {
2982                         log_debug("Field hash table of %s has a fill level at %.1f (%"PRIu64" of %"PRIu64" items), suggesting rotation.",
2983                                   f->path,
2984                                   100.0 * (double) le64toh(f->header->n_fields) / ((double) (le64toh(f->header->field_hash_table_size) / sizeof(HashItem))),
2985                                   le64toh(f->header->n_fields),
2986                                   le64toh(f->header->field_hash_table_size) / sizeof(HashItem));
2987                         return true;
2988                 }
2989
2990         /* Are the data objects properly indexed by field objects? */
2991         if (JOURNAL_HEADER_CONTAINS(f->header, n_data) &&
2992             JOURNAL_HEADER_CONTAINS(f->header, n_fields) &&
2993             le64toh(f->header->n_data) > 0 &&
2994             le64toh(f->header->n_fields) == 0)
2995                 return true;
2996
2997         if (max_file_usec > 0) {
2998                 usec_t t, h;
2999
3000                 h = le64toh(f->header->head_entry_realtime);
3001                 t = now(CLOCK_REALTIME);
3002
3003                 if (h > 0 && t > h + max_file_usec)
3004                         return true;
3005         }
3006
3007         return false;
3008 }