git.ozlabs.org Git - ccan/blob - ccan/tdb2/tdb.c

   1  /*
   2    Trivial Database 2: fetch, store and misc routines.
   3    Copyright (C) Rusty Russell 2010
   4
   5    This library is free software; you can redistribute it and/or
   6    modify it under the terms of the GNU Lesser General Public
   7    License as published by the Free Software Foundation; either
   8    version 3 of the License, or (at your option) any later version.
   9
  10    This library is distributed in the hope that it will be useful,
  11    but WITHOUT ANY WARRANTY; without even the implied warranty of
  12    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13    Lesser General Public License for more details.
  14
  15    You should have received a copy of the GNU Lesser General Public
  16    License along with this library; if not, see <http://www.gnu.org/licenses/>.
  17 */
  18 #include "private.h"
  19 #include <ccan/asprintf/asprintf.h>
  20 #include <stdarg.h>
  21
  22 static enum TDB_ERROR update_rec_hdr(struct tdb_context *tdb,
  23                                      tdb_off_t off,
  24                                      tdb_len_t keylen,
  25                                      tdb_len_t datalen,
  26                                      struct tdb_used_record *rec,
  27                                      uint64_t h)
  28 {
  29         uint64_t dataroom = rec_data_length(rec) + rec_extra_padding(rec);
  30         enum TDB_ERROR ecode;
  31
  32         ecode = set_header(tdb, rec, TDB_USED_MAGIC, keylen, datalen,
  33                            keylen + dataroom, h);
  34         if (ecode == TDB_SUCCESS) {
  35                 ecode = tdb_write_convert(tdb, off, rec, sizeof(*rec));
  36         }
  37         return ecode;
  38 }
  39
  40 static enum TDB_ERROR replace_data(struct tdb_context *tdb,
  41                                    struct hash_info *h,
  42                                    struct tdb_data key, struct tdb_data dbuf,
  43                                    tdb_off_t old_off, tdb_len_t old_room,
  44                                    bool growing)
  45 {
  46         tdb_off_t new_off;
  47         enum TDB_ERROR ecode;
  48
  49         /* Allocate a new record. */
  50         new_off = alloc(tdb, key.dsize, dbuf.dsize, h->h, TDB_USED_MAGIC,
  51                         growing);
  52         if (TDB_OFF_IS_ERR(new_off)) {
  53                 return new_off;
  54         }
  55
  56         /* We didn't like the existing one: remove it. */
  57         if (old_off) {
  58                 tdb->stats.frees++;
  59                 ecode = add_free_record(tdb, old_off,
  60                                         sizeof(struct tdb_used_record)
  61                                         + key.dsize + old_room,
  62                                         TDB_LOCK_WAIT, true);
  63                 if (ecode == TDB_SUCCESS)
  64                         ecode = replace_in_hash(tdb, h, new_off);
  65         } else {
  66                 ecode = add_to_hash(tdb, h, new_off);
  67         }
  68         if (ecode != TDB_SUCCESS) {
  69                 return ecode;
  70         }
  71
  72         new_off += sizeof(struct tdb_used_record);
  73         ecode = tdb->tdb2.io->twrite(tdb, new_off, key.dptr, key.dsize);
  74         if (ecode != TDB_SUCCESS) {
  75                 return ecode;
  76         }
  77
  78         new_off += key.dsize;
  79         ecode = tdb->tdb2.io->twrite(tdb, new_off, dbuf.dptr, dbuf.dsize);
  80         if (ecode != TDB_SUCCESS) {
  81                 return ecode;
  82         }
  83
  84         if (tdb->flags & TDB_SEQNUM)
  85                 tdb_inc_seqnum(tdb);
  86
  87         return TDB_SUCCESS;
  88 }
  89
  90 static enum TDB_ERROR update_data(struct tdb_context *tdb,
  91                                   tdb_off_t off,
  92                                   struct tdb_data dbuf,
  93                                   tdb_len_t extra)
  94 {
  95         enum TDB_ERROR ecode;
  96
  97         ecode = tdb->tdb2.io->twrite(tdb, off, dbuf.dptr, dbuf.dsize);
  98         if (ecode == TDB_SUCCESS && extra) {
  99                 /* Put a zero in; future versions may append other data. */
 100                 ecode = tdb->tdb2.io->twrite(tdb, off + dbuf.dsize, "", 1);
 101         }
 102         if (tdb->flags & TDB_SEQNUM)
 103                 tdb_inc_seqnum(tdb);
 104
 105         return ecode;
 106 }
 107
 108 enum TDB_ERROR tdb_store(struct tdb_context *tdb,
 109                          struct tdb_data key, struct tdb_data dbuf, int flag)
 110 {
 111         struct hash_info h;
 112         tdb_off_t off;
 113         tdb_len_t old_room = 0;
 114         struct tdb_used_record rec;
 115         enum TDB_ERROR ecode;
 116
 117         if (tdb->flags & TDB_VERSION1) {
 118                 if (tdb1_store(tdb, key, dbuf, flag) == -1)
 119                         return tdb->last_error;
 120                 return TDB_SUCCESS;
 121         }
 122
 123         off = find_and_lock(tdb, key, F_WRLCK, &h, &rec, NULL);
 124         if (TDB_OFF_IS_ERR(off)) {
 125                 return tdb->last_error = off;
 126         }
 127
 128         /* Now we have lock on this hash bucket. */
 129         if (flag == TDB_INSERT) {
 130                 if (off) {
 131                         ecode = TDB_ERR_EXISTS;
 132                         goto out;
 133                 }
 134         } else {
 135                 if (off) {
 136                         old_room = rec_data_length(&rec)
 137                                 + rec_extra_padding(&rec);
 138                         if (old_room >= dbuf.dsize) {
 139                                 /* Can modify in-place.  Easy! */
 140                                 ecode = update_rec_hdr(tdb, off,
 141                                                        key.dsize, dbuf.dsize,
 142                                                        &rec, h.h);
 143                                 if (ecode != TDB_SUCCESS) {
 144                                         goto out;
 145                                 }
 146                                 ecode = update_data(tdb,
 147                                                     off + sizeof(rec)
 148                                                     + key.dsize, dbuf,
 149                                                     old_room - dbuf.dsize);
 150                                 if (ecode != TDB_SUCCESS) {
 151                                         goto out;
 152                                 }
 153                                 tdb_unlock_hashes(tdb, h.hlock_start,
 154                                                   h.hlock_range, F_WRLCK);
 155                                 return tdb->last_error = TDB_SUCCESS;
 156                         }
 157                 } else {
 158                         if (flag == TDB_MODIFY) {
 159                                 /* if the record doesn't exist and we
 160                                    are in TDB_MODIFY mode then we should fail
 161                                    the store */
 162                                 ecode = TDB_ERR_NOEXIST;
 163                                 goto out;
 164                         }
 165                 }
 166         }
 167
 168         /* If we didn't use the old record, this implies we're growing. */
 169         ecode = replace_data(tdb, &h, key, dbuf, off, old_room, off);
 170 out:
 171         tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range, F_WRLCK);
 172         return tdb->last_error = ecode;
 173 }
 174
 175 enum TDB_ERROR tdb_append(struct tdb_context *tdb,
 176                           struct tdb_data key, struct tdb_data dbuf)
 177 {
 178         struct hash_info h;
 179         tdb_off_t off;
 180         struct tdb_used_record rec;
 181         tdb_len_t old_room = 0, old_dlen;
 182         unsigned char *newdata;
 183         struct tdb_data new_dbuf;
 184         enum TDB_ERROR ecode;
 185
 186         off = find_and_lock(tdb, key, F_WRLCK, &h, &rec, NULL);
 187         if (TDB_OFF_IS_ERR(off)) {
 188                 return tdb->last_error = off;
 189         }
 190
 191         if (off) {
 192                 old_dlen = rec_data_length(&rec);
 193                 old_room = old_dlen + rec_extra_padding(&rec);
 194
 195                 /* Fast path: can append in place. */
 196                 if (rec_extra_padding(&rec) >= dbuf.dsize) {
 197                         ecode = update_rec_hdr(tdb, off, key.dsize,
 198                                                old_dlen + dbuf.dsize, &rec,
 199                                                h.h);
 200                         if (ecode != TDB_SUCCESS) {
 201                                 goto out;
 202                         }
 203
 204                         off += sizeof(rec) + key.dsize + old_dlen;
 205                         ecode = update_data(tdb, off, dbuf,
 206                                             rec_extra_padding(&rec));
 207                         goto out;
 208                 }
 209
 210                 /* Slow path. */
 211                 newdata = malloc(key.dsize + old_dlen + dbuf.dsize);
 212                 if (!newdata) {
 213                         ecode = tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
 214                                            "tdb_append:"
 215                                            " failed to allocate %zu bytes",
 216                                            (size_t)(key.dsize + old_dlen
 217                                                     + dbuf.dsize));
 218                         goto out;
 219                 }
 220                 ecode = tdb->tdb2.io->tread(tdb, off + sizeof(rec) + key.dsize,
 221                                             newdata, old_dlen);
 222                 if (ecode != TDB_SUCCESS) {
 223                         goto out_free_newdata;
 224                 }
 225                 memcpy(newdata + old_dlen, dbuf.dptr, dbuf.dsize);
 226                 new_dbuf.dptr = newdata;
 227                 new_dbuf.dsize = old_dlen + dbuf.dsize;
 228         } else {
 229                 newdata = NULL;
 230                 new_dbuf = dbuf;
 231         }
 232
 233         /* If they're using tdb_append(), it implies they're growing record. */
 234         ecode = replace_data(tdb, &h, key, new_dbuf, off, old_room, true);
 235
 236 out_free_newdata:
 237         free(newdata);
 238 out:
 239         tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range, F_WRLCK);
 240         return tdb->last_error = ecode;
 241 }
 242
 243 enum TDB_ERROR tdb_fetch(struct tdb_context *tdb, struct tdb_data key,
 244                          struct tdb_data *data)
 245 {
 246         tdb_off_t off;
 247         struct tdb_used_record rec;
 248         struct hash_info h;
 249         enum TDB_ERROR ecode;
 250
 251         off = find_and_lock(tdb, key, F_RDLCK, &h, &rec, NULL);
 252         if (TDB_OFF_IS_ERR(off)) {
 253                 return tdb->last_error = off;
 254         }
 255
 256         if (!off) {
 257                 ecode = TDB_ERR_NOEXIST;
 258         } else {
 259                 data->dsize = rec_data_length(&rec);
 260                 data->dptr = tdb_alloc_read(tdb, off + sizeof(rec) + key.dsize,
 261                                             data->dsize);
 262                 if (TDB_PTR_IS_ERR(data->dptr)) {
 263                         ecode = TDB_PTR_ERR(data->dptr);
 264                 } else
 265                         ecode = TDB_SUCCESS;
 266         }
 267
 268         tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range, F_RDLCK);
 269         return tdb->last_error = ecode;
 270 }
 271
 272 bool tdb_exists(struct tdb_context *tdb, TDB_DATA key)
 273 {
 274         tdb_off_t off;
 275         struct tdb_used_record rec;
 276         struct hash_info h;
 277
 278         off = find_and_lock(tdb, key, F_RDLCK, &h, &rec, NULL);
 279         if (TDB_OFF_IS_ERR(off)) {
 280                 tdb->last_error = off;
 281                 return false;
 282         }
 283         tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range, F_RDLCK);
 284
 285         tdb->last_error = TDB_SUCCESS;
 286         return off ? true : false;
 287 }
 288
 289 enum TDB_ERROR tdb_delete(struct tdb_context *tdb, struct tdb_data key)
 290 {
 291         tdb_off_t off;
 292         struct tdb_used_record rec;
 293         struct hash_info h;
 294         enum TDB_ERROR ecode;
 295
 296         off = find_and_lock(tdb, key, F_WRLCK, &h, &rec, NULL);
 297         if (TDB_OFF_IS_ERR(off)) {
 298                 return tdb->last_error = off;
 299         }
 300
 301         if (!off) {
 302                 ecode = TDB_ERR_NOEXIST;
 303                 goto unlock;
 304         }
 305
 306         ecode = delete_from_hash(tdb, &h);
 307         if (ecode != TDB_SUCCESS) {
 308                 goto unlock;
 309         }
 310
 311         /* Free the deleted entry. */
 312         tdb->stats.frees++;
 313         ecode = add_free_record(tdb, off,
 314                                 sizeof(struct tdb_used_record)
 315                                 + rec_key_length(&rec)
 316                                 + rec_data_length(&rec)
 317                                 + rec_extra_padding(&rec),
 318                                 TDB_LOCK_WAIT, true);
 319
 320         if (tdb->flags & TDB_SEQNUM)
 321                 tdb_inc_seqnum(tdb);
 322
 323 unlock:
 324         tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range, F_WRLCK);
 325         return tdb->last_error = ecode;
 326 }
 327
 328 unsigned int tdb_get_flags(struct tdb_context *tdb)
 329 {
 330         return tdb->flags;
 331 }
 332
 333 static bool inside_transaction(const struct tdb_context *tdb)
 334 {
 335         if (tdb->flags & TDB_VERSION1)
 336                 return tdb->tdb1.transaction != NULL;
 337         else
 338                 return tdb->tdb2.transaction != NULL;
 339 }
 340
 341 static bool readonly_changable(struct tdb_context *tdb, const char *caller)
 342 {
 343         if (inside_transaction(tdb)) {
 344                 tdb->last_error = tdb_logerr(tdb, TDB_ERR_EINVAL,
 345                                              TDB_LOG_USE_ERROR,
 346                                              "%s: can't change"
 347                                              " TDB_RDONLY inside transaction",
 348                                              caller);
 349                 return false;
 350         }
 351
 352         if (tdb->file->allrecord_lock.count != 0
 353             || tdb->file->num_lockrecs != 0) {
 354                 tdb->last_error = tdb_logerr(tdb, TDB_ERR_EINVAL,
 355                                              TDB_LOG_USE_ERROR,
 356                                              "%s: can't change"
 357                                              " TDB_RDONLY holding locks",
 358                                              caller);
 359                 return false;
 360         }
 361         return true;
 362 }
 363
 364 void tdb_add_flag(struct tdb_context *tdb, unsigned flag)
 365 {
 366         if (tdb->flags & TDB_INTERNAL) {
 367                 tdb->last_error = tdb_logerr(tdb, TDB_ERR_EINVAL,
 368                                              TDB_LOG_USE_ERROR,
 369                                              "tdb_add_flag: internal db");
 370                 return;
 371         }
 372         switch (flag) {
 373         case TDB_NOLOCK:
 374                 tdb->flags |= TDB_NOLOCK;
 375                 break;
 376         case TDB_NOMMAP:
 377                 tdb->flags |= TDB_NOMMAP;
 378                 tdb_munmap(tdb->file);
 379                 break;
 380         case TDB_NOSYNC:
 381                 tdb->flags |= TDB_NOSYNC;
 382                 break;
 383         case TDB_SEQNUM:
 384                 tdb->flags |= TDB_SEQNUM;
 385                 break;
 386         case TDB_ALLOW_NESTING:
 387                 tdb->flags |= TDB_ALLOW_NESTING;
 388                 break;
 389         case TDB_RDONLY:
 390                 if (readonly_changable(tdb, "tdb_add_flag"))
 391                         tdb->flags |= TDB_RDONLY;
 392                 break;
 393         default:
 394                 tdb->last_error = tdb_logerr(tdb, TDB_ERR_EINVAL,
 395                                              TDB_LOG_USE_ERROR,
 396                                              "tdb_add_flag: Unknown flag %u",
 397                                              flag);
 398         }
 399 }
 400
 401 void tdb_remove_flag(struct tdb_context *tdb, unsigned flag)
 402 {
 403         if (tdb->flags & TDB_INTERNAL) {
 404                 tdb->last_error = tdb_logerr(tdb, TDB_ERR_EINVAL,
 405                                              TDB_LOG_USE_ERROR,
 406                                              "tdb_remove_flag: internal db");
 407                 return;
 408         }
 409         switch (flag) {
 410         case TDB_NOLOCK:
 411                 tdb->flags &= ~TDB_NOLOCK;
 412                 break;
 413         case TDB_NOMMAP:
 414                 tdb->flags &= ~TDB_NOMMAP;
 415                 tdb_mmap(tdb);
 416                 break;
 417         case TDB_NOSYNC:
 418                 tdb->flags &= ~TDB_NOSYNC;
 419                 break;
 420         case TDB_SEQNUM:
 421                 tdb->flags &= ~TDB_SEQNUM;
 422                 break;
 423         case TDB_ALLOW_NESTING:
 424                 tdb->flags &= ~TDB_ALLOW_NESTING;
 425                 break;
 426         case TDB_RDONLY:
 427                 if ((tdb->open_flags & O_ACCMODE) == O_RDONLY) {
 428                         tdb->last_error = tdb_logerr(tdb, TDB_ERR_EINVAL,
 429                                                      TDB_LOG_USE_ERROR,
 430                                                      "tdb_remove_flag: can't"
 431                                                      " remove TDB_RDONLY on tdb"
 432                                                      " opened with O_RDONLY");
 433                         break;
 434                 }
 435                 if (readonly_changable(tdb, "tdb_remove_flag"))
 436                         tdb->flags &= ~TDB_RDONLY;
 437                 break;
 438         default:
 439                 tdb->last_error = tdb_logerr(tdb, TDB_ERR_EINVAL,
 440                                              TDB_LOG_USE_ERROR,
 441                                              "tdb_remove_flag: Unknown flag %u",
 442                                              flag);
 443         }
 444 }
 445
 446 const char *tdb_errorstr(enum TDB_ERROR ecode)
 447 {
 448         /* Gcc warns if you miss a case in the switch, so use that. */
 449         switch (ecode) {
 450         case TDB_SUCCESS: return "Success";
 451         case TDB_ERR_CORRUPT: return "Corrupt database";
 452         case TDB_ERR_IO: return "IO Error";
 453         case TDB_ERR_LOCK: return "Locking error";
 454         case TDB_ERR_OOM: return "Out of memory";
 455         case TDB_ERR_EXISTS: return "Record exists";
 456         case TDB_ERR_EINVAL: return "Invalid parameter";
 457         case TDB_ERR_NOEXIST: return "Record does not exist";
 458         case TDB_ERR_RDONLY: return "write not permitted";
 459         }
 460         return "Invalid error code";
 461 }
 462
 463 enum TDB_ERROR tdb_error(struct tdb_context *tdb)
 464 {
 465         return tdb->last_error;
 466 }
 467
 468 enum TDB_ERROR COLD tdb_logerr(struct tdb_context *tdb,
 469                                enum TDB_ERROR ecode,
 470                                enum tdb_log_level level,
 471                                const char *fmt, ...)
 472 {
 473         char *message;
 474         va_list ap;
 475         size_t len;
 476         /* tdb_open paths care about errno, so save it. */
 477         int saved_errno = errno;
 478
 479         if (!tdb->log_fn)
 480                 return ecode;
 481
 482         va_start(ap, fmt);
 483         len = vasprintf(&message, fmt, ap);
 484         va_end(ap);
 485
 486         if (len < 0) {
 487                 tdb->log_fn(tdb, TDB_LOG_ERROR, TDB_ERR_OOM,
 488                             "out of memory formatting message:", tdb->log_data);
 489                 tdb->log_fn(tdb, level, ecode, fmt, tdb->log_data);
 490         } else {
 491                 tdb->log_fn(tdb, level, ecode, message, tdb->log_data);
 492                 free(message);
 493         }
 494         errno = saved_errno;
 495         return ecode;
 496 }
 497
 498 enum TDB_ERROR tdb_parse_record_(struct tdb_context *tdb,
 499                                  TDB_DATA key,
 500                                  enum TDB_ERROR (*parse)(TDB_DATA k,
 501                                                          TDB_DATA d,
 502                                                          void *data),
 503                                  void *data)
 504 {
 505         tdb_off_t off;
 506         struct tdb_used_record rec;
 507         struct hash_info h;
 508         enum TDB_ERROR ecode;
 509
 510         off = find_and_lock(tdb, key, F_RDLCK, &h, &rec, NULL);
 511         if (TDB_OFF_IS_ERR(off)) {
 512                 return tdb->last_error = off;
 513         }
 514
 515         if (!off) {
 516                 ecode = TDB_ERR_NOEXIST;
 517         } else {
 518                 const void *dptr;
 519                 dptr = tdb_access_read(tdb, off + sizeof(rec) + key.dsize,
 520                                        rec_data_length(&rec), false);
 521                 if (TDB_PTR_IS_ERR(dptr)) {
 522                         ecode = TDB_PTR_ERR(dptr);
 523                 } else {
 524                         TDB_DATA d = tdb_mkdata(dptr, rec_data_length(&rec));
 525
 526                         ecode = parse(key, d, data);
 527                         tdb_access_release(tdb, dptr);
 528                 }
 529         }
 530
 531         tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range, F_RDLCK);
 532         return tdb->last_error = ecode;
 533 }
 534
 535 const char *tdb_name(const struct tdb_context *tdb)
 536 {
 537         return tdb->name;
 538 }
 539
 540 int64_t tdb_get_seqnum(struct tdb_context *tdb)
 541 {
 542         tdb_off_t off = tdb_read_off(tdb, offsetof(struct tdb_header, seqnum));
 543         if (TDB_OFF_IS_ERR(off))
 544                 tdb->last_error = off;
 545         else
 546                 tdb->last_error = TDB_SUCCESS;
 547         return off;
 548 }
 549
 550
 551 int tdb_fd(const struct tdb_context *tdb)
 552 {
 553         return tdb->file->fd;
 554 }