tdb2: add TDB_ATTRIBUTE_TDB1_HASHSIZE
[ccan] / ccan / tdb2 / open.c
1  /*
2    Trivial Database 2: opening and closing TDBs
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 <assert.h>
20
21 /* all tdbs, to detect double-opens (fcntl file don't nest!) */
22 static struct tdb_context *tdbs = NULL;
23
24 static struct tdb_file *find_file(dev_t device, ino_t ino)
25 {
26         struct tdb_context *i;
27
28         for (i = tdbs; i; i = i->next) {
29                 if (i->file->device == device && i->file->inode == ino) {
30                         i->file->refcnt++;
31                         return i->file;
32                 }
33         }
34         return NULL;
35 }
36
37 static bool read_all(int fd, void *buf, size_t len)
38 {
39         while (len) {
40                 ssize_t ret;
41                 ret = read(fd, buf, len);
42                 if (ret < 0)
43                         return false;
44                 if (ret == 0) {
45                         /* ETOOSHORT? */
46                         errno = EWOULDBLOCK;
47                         return false;
48                 }
49                 buf = (char *)buf + ret;
50                 len -= ret;
51         }
52         return true;
53 }
54
55 static uint64_t random_number(struct tdb_context *tdb)
56 {
57         int fd;
58         uint64_t ret = 0;
59         struct timeval now;
60
61         fd = open("/dev/urandom", O_RDONLY);
62         if (fd >= 0) {
63                 if (read_all(fd, &ret, sizeof(ret))) {
64                         close(fd);
65                         return ret;
66                 }
67                 close(fd);
68         }
69         /* FIXME: Untested!  Based on Wikipedia protocol description! */
70         fd = open("/dev/egd-pool", O_RDWR);
71         if (fd >= 0) {
72                 /* Command is 1, next byte is size we want to read. */
73                 char cmd[2] = { 1, sizeof(uint64_t) };
74                 if (write(fd, cmd, sizeof(cmd)) == sizeof(cmd)) {
75                         char reply[1 + sizeof(uint64_t)];
76                         int r = read(fd, reply, sizeof(reply));
77                         if (r > 1) {
78                                 /* Copy at least some bytes. */
79                                 memcpy(&ret, reply+1, r - 1);
80                                 if (reply[0] == sizeof(uint64_t)
81                                     && r == sizeof(reply)) {
82                                         close(fd);
83                                         return ret;
84                                 }
85                         }
86                 }
87                 close(fd);
88         }
89
90         /* Fallback: pid and time. */
91         gettimeofday(&now, NULL);
92         ret = getpid() * 100132289ULL + now.tv_sec * 1000000ULL + now.tv_usec;
93         tdb_logerr(tdb, TDB_SUCCESS, TDB_LOG_WARNING,
94                    "tdb_open: random from getpid and time");
95         return ret;
96 }
97
98 struct new_database {
99         struct tdb_header hdr;
100         struct tdb_freetable ftable;
101 };
102
103 /* initialise a new database */
104 static enum TDB_ERROR tdb_new_database(struct tdb_context *tdb,
105                                        struct tdb_attribute_seed *seed,
106                                        struct tdb_header *hdr)
107 {
108         /* We make it up in memory, then write it out if not internal */
109         struct new_database newdb;
110         unsigned int magic_len;
111         ssize_t rlen;
112         enum TDB_ERROR ecode;
113
114         /* Fill in the header */
115         newdb.hdr.version = TDB_VERSION;
116         if (seed)
117                 newdb.hdr.hash_seed = seed->seed;
118         else
119                 newdb.hdr.hash_seed = random_number(tdb);
120         newdb.hdr.hash_test = TDB_HASH_MAGIC;
121         newdb.hdr.hash_test = tdb->hash_fn(&newdb.hdr.hash_test,
122                                            sizeof(newdb.hdr.hash_test),
123                                            newdb.hdr.hash_seed,
124                                            tdb->hash_data);
125         newdb.hdr.recovery = 0;
126         newdb.hdr.features_used = newdb.hdr.features_offered = TDB_FEATURE_MASK;
127         newdb.hdr.seqnum = 0;
128         memset(newdb.hdr.reserved, 0, sizeof(newdb.hdr.reserved));
129         /* Initial hashes are empty. */
130         memset(newdb.hdr.hashtable, 0, sizeof(newdb.hdr.hashtable));
131
132         /* Free is empty. */
133         newdb.hdr.free_table = offsetof(struct new_database, ftable);
134         memset(&newdb.ftable, 0, sizeof(newdb.ftable));
135         ecode = set_header(NULL, &newdb.ftable.hdr, TDB_FTABLE_MAGIC, 0,
136                            sizeof(newdb.ftable) - sizeof(newdb.ftable.hdr),
137                            sizeof(newdb.ftable) - sizeof(newdb.ftable.hdr),
138                            0);
139         if (ecode != TDB_SUCCESS) {
140                 return ecode;
141         }
142
143         /* Magic food */
144         memset(newdb.hdr.magic_food, 0, sizeof(newdb.hdr.magic_food));
145         strcpy(newdb.hdr.magic_food, TDB_MAGIC_FOOD);
146
147         /* This creates an endian-converted database, as if read from disk */
148         magic_len = sizeof(newdb.hdr.magic_food);
149         tdb_convert(tdb,
150                     (char *)&newdb.hdr + magic_len, sizeof(newdb) - magic_len);
151
152         *hdr = newdb.hdr;
153
154         if (tdb->flags & TDB_INTERNAL) {
155                 tdb->file->map_size = sizeof(newdb);
156                 tdb->file->map_ptr = malloc(tdb->file->map_size);
157                 if (!tdb->file->map_ptr) {
158                         return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
159                                           "tdb_new_database:"
160                                           " failed to allocate");
161                 }
162                 memcpy(tdb->file->map_ptr, &newdb, tdb->file->map_size);
163                 return TDB_SUCCESS;
164         }
165         if (lseek(tdb->file->fd, 0, SEEK_SET) == -1) {
166                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
167                                   "tdb_new_database:"
168                                   " failed to seek: %s", strerror(errno));
169         }
170
171         if (ftruncate(tdb->file->fd, 0) == -1) {
172                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
173                                   "tdb_new_database:"
174                                   " failed to truncate: %s", strerror(errno));
175         }
176
177         rlen = write(tdb->file->fd, &newdb, sizeof(newdb));
178         if (rlen != sizeof(newdb)) {
179                 if (rlen >= 0)
180                         errno = ENOSPC;
181                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
182                                   "tdb_new_database: %zi writing header: %s",
183                                   rlen, strerror(errno));
184         }
185         return TDB_SUCCESS;
186 }
187
188 static enum TDB_ERROR tdb_new_file(struct tdb_context *tdb)
189 {
190         tdb->file = malloc(sizeof(*tdb->file));
191         if (!tdb->file)
192                 return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
193                                   "tdb_open: cannot alloc tdb_file structure");
194         tdb->file->num_lockrecs = 0;
195         tdb->file->lockrecs = NULL;
196         tdb->file->allrecord_lock.count = 0;
197         tdb->file->refcnt = 1;
198         return TDB_SUCCESS;
199 }
200
201 enum TDB_ERROR tdb_set_attribute(struct tdb_context *tdb,
202                                  const union tdb_attribute *attr)
203 {
204         switch (attr->base.attr) {
205         case TDB_ATTRIBUTE_LOG:
206                 tdb->log_fn = attr->log.fn;
207                 tdb->log_data = attr->log.data;
208                 break;
209         case TDB_ATTRIBUTE_HASH:
210         case TDB_ATTRIBUTE_SEED:
211         case TDB_ATTRIBUTE_OPENHOOK:
212         case TDB_ATTRIBUTE_TDB1_HASHSIZE:
213                 return tdb->last_error
214                         = tdb_logerr(tdb, TDB_ERR_EINVAL,
215                                      TDB_LOG_USE_ERROR,
216                                      "tdb_set_attribute:"
217                                      " cannot set %s after opening",
218                                      attr->base.attr == TDB_ATTRIBUTE_HASH
219                                      ? "TDB_ATTRIBUTE_HASH"
220                                      : attr->base.attr == TDB_ATTRIBUTE_SEED
221                                      ? "TDB_ATTRIBUTE_SEED"
222                                      : attr->base.attr == TDB_ATTRIBUTE_OPENHOOK
223                                      ? "TDB_ATTRIBUTE_OPENHOOK"
224                                      : "TDB_ATTRIBUTE_TDB1_HASHSIZE");
225         case TDB_ATTRIBUTE_STATS:
226                 return tdb->last_error
227                         = tdb_logerr(tdb, TDB_ERR_EINVAL,
228                                      TDB_LOG_USE_ERROR,
229                                      "tdb_set_attribute:"
230                                      " cannot set TDB_ATTRIBUTE_STATS");
231         case TDB_ATTRIBUTE_FLOCK:
232                 tdb->lock_fn = attr->flock.lock;
233                 tdb->unlock_fn = attr->flock.unlock;
234                 tdb->lock_data = attr->flock.data;
235                 break;
236         default:
237                 return tdb->last_error
238                         = tdb_logerr(tdb, TDB_ERR_EINVAL,
239                                      TDB_LOG_USE_ERROR,
240                                      "tdb_set_attribute:"
241                                      " unknown attribute type %u",
242                                      attr->base.attr);
243         }
244         return TDB_SUCCESS;
245 }
246
247 enum TDB_ERROR tdb_get_attribute(struct tdb_context *tdb,
248                                  union tdb_attribute *attr)
249 {
250         switch (attr->base.attr) {
251         case TDB_ATTRIBUTE_LOG:
252                 if (!tdb->log_fn)
253                         return tdb->last_error = TDB_ERR_NOEXIST;
254                 attr->log.fn = tdb->log_fn;
255                 attr->log.data = tdb->log_data;
256                 break;
257         case TDB_ATTRIBUTE_HASH:
258                 attr->hash.fn = tdb->hash_fn;
259                 attr->hash.data = tdb->hash_data;
260                 break;
261         case TDB_ATTRIBUTE_SEED:
262                 attr->seed.seed = tdb->hash_seed;
263                 break;
264         case TDB_ATTRIBUTE_OPENHOOK:
265                 if (!tdb->openhook)
266                         return tdb->last_error = TDB_ERR_NOEXIST;
267                 attr->openhook.fn = tdb->openhook;
268                 attr->openhook.data = tdb->openhook_data;
269                 break;
270         case TDB_ATTRIBUTE_STATS: {
271                 size_t size = attr->stats.size;
272                 if (size > tdb->stats.size)
273                         size = tdb->stats.size;
274                 memcpy(&attr->stats, &tdb->stats, size);
275                 break;
276         }
277         case TDB_ATTRIBUTE_FLOCK:
278                 attr->flock.lock = tdb->lock_fn;
279                 attr->flock.unlock = tdb->unlock_fn;
280                 attr->flock.data = tdb->lock_data;
281                 break;
282         case TDB_ATTRIBUTE_TDB1_HASHSIZE:
283                 if (!(tdb->flags & TDB_VERSION1))
284                         return tdb->last_error
285                                 = tdb_logerr(tdb, TDB_ERR_EINVAL,
286                                              TDB_LOG_USE_ERROR,
287                                      "tdb_get_attribute:"
288                                      " cannot get TDB_ATTRIBUTE_TDB1_HASHSIZE"
289                                      " on TDB2 tdb.");
290                 attr->tdb1_hashsize.hsize = tdb->tdb1.header.hash_size;
291                 break;
292         default:
293                 return tdb->last_error
294                         = tdb_logerr(tdb, TDB_ERR_EINVAL,
295                                      TDB_LOG_USE_ERROR,
296                                      "tdb_get_attribute:"
297                                      " unknown attribute type %u",
298                                      attr->base.attr);
299         }
300         attr->base.next = NULL;
301         return TDB_SUCCESS;
302 }
303
304 void tdb_unset_attribute(struct tdb_context *tdb,
305                          enum tdb_attribute_type type)
306 {
307         switch (type) {
308         case TDB_ATTRIBUTE_LOG:
309                 tdb->log_fn = NULL;
310                 break;
311         case TDB_ATTRIBUTE_OPENHOOK:
312                 tdb->openhook = NULL;
313                 break;
314         case TDB_ATTRIBUTE_HASH:
315         case TDB_ATTRIBUTE_SEED:
316         case TDB_ATTRIBUTE_TDB1_HASHSIZE:
317                 tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
318                            "tdb_unset_attribute: cannot unset %s after opening",
319                            type == TDB_ATTRIBUTE_HASH
320                            ? "TDB_ATTRIBUTE_HASH"
321                            : type == TDB_ATTRIBUTE_SEED
322                            ? "TDB_ATTRIBUTE_SEED"
323                            : "TDB_ATTRIBUTE_TDB1_HASHSIZE");
324                 break;
325         case TDB_ATTRIBUTE_STATS:
326                 tdb_logerr(tdb, TDB_ERR_EINVAL,
327                            TDB_LOG_USE_ERROR,
328                            "tdb_unset_attribute:"
329                            "cannot unset TDB_ATTRIBUTE_STATS");
330                 break;
331         case TDB_ATTRIBUTE_FLOCK:
332                 tdb->lock_fn = tdb_fcntl_lock;
333                 tdb->unlock_fn = tdb_fcntl_unlock;
334                 break;
335         default:
336                 tdb_logerr(tdb, TDB_ERR_EINVAL,
337                            TDB_LOG_USE_ERROR,
338                            "tdb_unset_attribute: unknown attribute type %u",
339                            type);
340         }
341 }
342
343 struct tdb_context *tdb_open(const char *name, int tdb_flags,
344                              int open_flags, mode_t mode,
345                              union tdb_attribute *attr)
346 {
347         struct tdb_context *tdb;
348         struct stat st;
349         int saved_errno = 0;
350         uint64_t hash_test;
351         unsigned v;
352         ssize_t rlen;
353         struct tdb_header hdr;
354         struct tdb_attribute_seed *seed = NULL;
355         struct tdb_attribute_tdb1_hashsize *hsize_attr = NULL;
356         tdb_bool_err berr;
357         enum TDB_ERROR ecode;
358         int openlock;
359
360         tdb = malloc(sizeof(*tdb) + (name ? strlen(name) + 1 : 0));
361         if (!tdb) {
362                 /* Can't log this */
363                 errno = ENOMEM;
364                 return NULL;
365         }
366         /* Set name immediately for logging functions. */
367         if (name) {
368                 tdb->name = strcpy((char *)(tdb + 1), name);
369         } else {
370                 tdb->name = NULL;
371         }
372         tdb->flags = tdb_flags;
373         tdb->log_fn = NULL;
374         tdb->open_flags = open_flags;
375         tdb->last_error = TDB_SUCCESS;
376         tdb->file = NULL;
377         tdb->openhook = NULL;
378         tdb->lock_fn = tdb_fcntl_lock;
379         tdb->unlock_fn = tdb_fcntl_unlock;
380         tdb->hash_fn = tdb_jenkins_hash;
381         memset(&tdb->stats, 0, sizeof(tdb->stats));
382         tdb->stats.base.attr = TDB_ATTRIBUTE_STATS;
383         tdb->stats.size = sizeof(tdb->stats);
384         tdb_io_init(tdb);
385         tdb->tdb2.direct_access = 0;
386         tdb->tdb2.transaction = NULL;
387         tdb->tdb2.access = NULL;
388
389         while (attr) {
390                 switch (attr->base.attr) {
391                 case TDB_ATTRIBUTE_HASH:
392                         tdb->hash_fn = attr->hash.fn;
393                         tdb->hash_data = attr->hash.data;
394                         break;
395                 case TDB_ATTRIBUTE_SEED:
396                         seed = &attr->seed;
397                         break;
398                 case TDB_ATTRIBUTE_OPENHOOK:
399                         tdb->openhook = attr->openhook.fn;
400                         tdb->openhook_data = attr->openhook.data;
401                         break;
402                 case TDB_ATTRIBUTE_TDB1_HASHSIZE:
403                         hsize_attr = &attr->tdb1_hashsize;
404                         break;
405                 default:
406                         /* These are set as normal. */
407                         ecode = tdb_set_attribute(tdb, attr);
408                         if (ecode != TDB_SUCCESS)
409                                 goto fail;
410                 }
411                 attr = attr->base.next;
412         }
413
414         if (tdb_flags & ~(TDB_INTERNAL | TDB_NOLOCK | TDB_NOMMAP | TDB_CONVERT
415                           | TDB_NOSYNC | TDB_SEQNUM | TDB_ALLOW_NESTING
416                           | TDB_RDONLY)) {
417                 ecode = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
418                                    "tdb_open: unknown flags %u", tdb_flags);
419                 goto fail;
420         }
421
422         if (hsize_attr) {
423                 if (!(tdb_flags & TDB_VERSION1) ||
424                     (!(tdb_flags & TDB_INTERNAL) && !(open_flags & O_CREAT))) {
425                         ecode = tdb_logerr(tdb, TDB_ERR_EINVAL,
426                                            TDB_LOG_USE_ERROR,
427                                            "tdb_open: can only use"
428                                            " TDB_ATTRIBUTE_TDB1_HASHSIZE when"
429                                            " creating a TDB_VERSION1 tdb");
430                         goto fail;
431                 }
432         }
433
434         if ((open_flags & O_ACCMODE) == O_WRONLY) {
435                 ecode = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
436                                    "tdb_open: can't open tdb %s write-only",
437                                    name);
438                 goto fail;
439         }
440
441         if ((open_flags & O_ACCMODE) == O_RDONLY) {
442                 openlock = F_RDLCK;
443                 tdb->flags |= TDB_RDONLY;
444         } else {
445                 if (tdb_flags & TDB_RDONLY) {
446                         ecode = tdb_logerr(tdb, TDB_ERR_EINVAL,
447                                            TDB_LOG_USE_ERROR,
448                                            "tdb_open: can't use TDB_RDONLY"
449                                            " without O_RDONLY");
450                         goto fail;
451                 }
452                 openlock = F_WRLCK;
453         }
454
455         /* internal databases don't need any of the rest. */
456         if (tdb->flags & TDB_INTERNAL) {
457                 tdb->flags |= (TDB_NOLOCK | TDB_NOMMAP);
458                 ecode = tdb_new_file(tdb);
459                 if (ecode != TDB_SUCCESS) {
460                         goto fail;
461                 }
462                 tdb->file->fd = -1;
463                 ecode = tdb_new_database(tdb, seed, &hdr);
464                 if (ecode != TDB_SUCCESS) {
465                         goto fail;
466                 }
467                 tdb_convert(tdb, &hdr.hash_seed, sizeof(hdr.hash_seed));
468                 tdb->hash_seed = hdr.hash_seed;
469                 tdb_ftable_init(tdb);
470                 return tdb;
471         }
472
473         if (stat(name, &st) != -1)
474                 tdb->file = find_file(st.st_dev, st.st_ino);
475
476         if (!tdb->file) {
477                 int fd;
478
479                 if ((fd = open(name, open_flags, mode)) == -1) {
480                         /* errno set by open(2) */
481                         saved_errno = errno;
482                         tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
483                                    "tdb_open: could not open file %s: %s",
484                                    name, strerror(errno));
485                         goto fail_errno;
486                 }
487
488                 /* on exec, don't inherit the fd */
489                 v = fcntl(fd, F_GETFD, 0);
490                 fcntl(fd, F_SETFD, v | FD_CLOEXEC);
491
492                 if (fstat(fd, &st) == -1) {
493                         saved_errno = errno;
494                         tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
495                                    "tdb_open: could not stat open %s: %s",
496                                    name, strerror(errno));
497                         close(fd);
498                         goto fail_errno;
499                 }
500
501                 ecode = tdb_new_file(tdb);
502                 if (ecode != TDB_SUCCESS) {
503                         close(fd);
504                         goto fail;
505                 }
506
507                 tdb->file->fd = fd;
508                 tdb->file->device = st.st_dev;
509                 tdb->file->inode = st.st_ino;
510                 tdb->file->map_ptr = NULL;
511                 tdb->file->map_size = sizeof(struct tdb_header);
512         }
513
514         /* ensure there is only one process initialising at once */
515         ecode = tdb_lock_open(tdb, openlock, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK);
516         if (ecode != TDB_SUCCESS) {
517                 saved_errno = errno;
518                 goto fail_errno;
519         }
520
521         /* call their open hook if they gave us one. */
522         if (tdb->openhook) {
523                 ecode = tdb->openhook(tdb->file->fd, tdb->openhook_data);
524                 if (ecode != TDB_SUCCESS) {
525                         tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
526                                    "tdb_open: open hook failed");
527                         goto fail;
528                 }
529                 open_flags |= O_CREAT;
530         }
531
532         /* If they used O_TRUNC, read will return 0. */
533         rlen = pread(tdb->file->fd, &hdr, sizeof(hdr), 0);
534         if (rlen == 0 && (open_flags & O_CREAT)) {
535                 ecode = tdb_new_database(tdb, seed, &hdr);
536                 if (ecode != TDB_SUCCESS) {
537                         goto fail;
538                 }
539         } else if (rlen < 0) {
540                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
541                                    "tdb_open: error %s reading %s",
542                                    strerror(errno), name);
543                 goto fail;
544         } else if (rlen < sizeof(hdr)
545                    || strcmp(hdr.magic_food, TDB_MAGIC_FOOD) != 0) {
546                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
547                                    "tdb_open: %s is not a tdb file", name);
548                 goto fail;
549         }
550
551         if (hdr.version != TDB_VERSION) {
552                 if (hdr.version == bswap_64(TDB_VERSION))
553                         tdb->flags |= TDB_CONVERT;
554                 else {
555                         /* wrong version */
556                         ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
557                                            "tdb_open:"
558                                            " %s is unknown version 0x%llx",
559                                            name, (long long)hdr.version);
560                         goto fail;
561                 }
562         } else if (tdb->flags & TDB_CONVERT) {
563                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
564                                    "tdb_open:"
565                                    " %s does not need TDB_CONVERT",
566                                    name);
567                 goto fail;
568         }
569
570         tdb_convert(tdb, &hdr, sizeof(hdr));
571         tdb->hash_seed = hdr.hash_seed;
572         hash_test = TDB_HASH_MAGIC;
573         hash_test = tdb_hash(tdb, &hash_test, sizeof(hash_test));
574         if (hdr.hash_test != hash_test) {
575                 /* wrong hash variant */
576                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
577                                    "tdb_open:"
578                                    " %s uses a different hash function",
579                                    name);
580                 goto fail;
581         }
582
583         /* Clear any features we don't understand. */
584         if ((open_flags & O_ACCMODE) != O_RDONLY) {
585                 hdr.features_used &= TDB_FEATURE_MASK;
586                 ecode = tdb_write_convert(tdb, offsetof(struct tdb_header,
587                                                         features_used),
588                                           &hdr.features_used,
589                                           sizeof(hdr.features_used));
590                 if (ecode != TDB_SUCCESS)
591                         goto fail;
592         }
593
594         tdb_unlock_open(tdb, openlock);
595
596         /* This make sure we have current map_size and mmap. */
597         ecode = tdb->tdb2.io->oob(tdb, tdb->file->map_size + 1, true);
598         if (unlikely(ecode != TDB_SUCCESS))
599                 goto fail;
600
601         /* Now it's fully formed, recover if necessary. */
602         berr = tdb_needs_recovery(tdb);
603         if (unlikely(berr != false)) {
604                 if (berr < 0) {
605                         ecode = berr;
606                         goto fail;
607                 }
608                 ecode = tdb_lock_and_recover(tdb);
609                 if (ecode != TDB_SUCCESS) {
610                         goto fail;
611                 }
612         }
613
614         ecode = tdb_ftable_init(tdb);
615         if (ecode != TDB_SUCCESS) {
616                 goto fail;
617         }
618
619         tdb->next = tdbs;
620         tdbs = tdb;
621         return tdb;
622
623  fail:
624         /* Map ecode to some logical errno. */
625         switch (ecode) {
626         case TDB_ERR_CORRUPT:
627         case TDB_ERR_IO:
628                 saved_errno = EIO;
629                 break;
630         case TDB_ERR_LOCK:
631                 saved_errno = EWOULDBLOCK;
632                 break;
633         case TDB_ERR_OOM:
634                 saved_errno = ENOMEM;
635                 break;
636         case TDB_ERR_EINVAL:
637                 saved_errno = EINVAL;
638                 break;
639         default:
640                 saved_errno = EINVAL;
641                 break;
642         }
643
644 fail_errno:
645 #ifdef TDB_TRACE
646         close(tdb->tracefd);
647 #endif
648         if (tdb->file) {
649                 tdb_lock_cleanup(tdb);
650                 if (--tdb->file->refcnt == 0) {
651                         assert(tdb->file->num_lockrecs == 0);
652                         if (tdb->file->map_ptr) {
653                                 if (tdb->flags & TDB_INTERNAL) {
654                                         free(tdb->file->map_ptr);
655                                 } else
656                                         tdb_munmap(tdb->file);
657                         }
658                         if (close(tdb->file->fd) != 0)
659                                 tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
660                                            "tdb_open: failed to close tdb fd"
661                                            " on error: %s", strerror(errno));
662                         free(tdb->file->lockrecs);
663                         free(tdb->file);
664                 }
665         }
666
667         free(tdb);
668         errno = saved_errno;
669         return NULL;
670 }
671
672 int tdb_close(struct tdb_context *tdb)
673 {
674         int ret = 0;
675         struct tdb_context **i;
676
677         tdb_trace(tdb, "tdb_close");
678
679         if (tdb->tdb2.transaction) {
680                 tdb_transaction_cancel(tdb);
681         }
682
683         if (tdb->file->map_ptr) {
684                 if (tdb->flags & TDB_INTERNAL)
685                         free(tdb->file->map_ptr);
686                 else
687                         tdb_munmap(tdb->file);
688         }
689         if (tdb->file) {
690                 tdb_lock_cleanup(tdb);
691                 if (--tdb->file->refcnt == 0) {
692                         ret = close(tdb->file->fd);
693                         free(tdb->file->lockrecs);
694                         free(tdb->file);
695                 }
696         }
697
698         /* Remove from tdbs list */
699         for (i = &tdbs; *i; i = &(*i)->next) {
700                 if (*i == tdb) {
701                         *i = tdb->next;
702                         break;
703                 }
704         }
705
706 #ifdef TDB_TRACE
707         close(tdb->tracefd);
708 #endif
709         free(tdb);
710
711         return ret;
712 }
713
714 void tdb_foreach_(int (*fn)(struct tdb_context *, void *), void *p)
715 {
716         struct tdb_context *i;
717
718         for (i = tdbs; i; i = i->next) {
719                 if (fn(i, p) != 0)
720                         break;
721         }
722 }