2 Unix SMB/CIFS implementation.
4 trivial database library
6 Copyright (C) Andrew Tridgell 1999-2005
7 Copyright (C) Paul `Rusty' Russell 2000
8 Copyright (C) Jeremy Allison 2000-2003
10 ** NOTE! The following LGPL license applies to the tdb
11 ** library. This does NOT imply that all of Samba is released
14 This library is free software; you can redistribute it and/or
15 modify it under the terms of the GNU Lesser General Public
16 License as published by the Free Software Foundation; either
17 version 3 of the License, or (at your option) any later version.
19 This library is distributed in the hope that it will be useful,
20 but WITHOUT ANY WARRANTY; without even the implied warranty of
21 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
22 Lesser General Public License for more details.
24 You should have received a copy of the GNU Lesser General Public
25 License along with this library; if not, see <http://www.gnu.org/licenses/>.
28 #include "tdb_private.h"
30 /* 'right' merges can involve O(n^2) cost when combined with a
31 traverse, so they are disabled until we find a way to do them in
34 #define USE_RIGHT_MERGES 0
36 /* read a freelist record and check for simple errors */
37 int tdb_rec_free_read(struct tdb_context *tdb, tdb_off_t off, struct tdb_record *rec)
39 if (tdb->methods->tdb_read(tdb, off, rec, sizeof(*rec),DOCONV()) == -1)
42 if (rec->magic == TDB_MAGIC) {
43 /* this happens when a app is showdown while deleting a record - we should
44 not completely fail when this happens */
45 TDB_LOG((tdb, TDB_DEBUG_WARNING, "tdb_rec_free_read non-free magic 0x%x at offset=%u - fixing\n",
47 rec->magic = TDB_FREE_MAGIC;
48 if (tdb_rec_write(tdb, off, rec) == -1)
52 if (rec->magic != TDB_FREE_MAGIC) {
53 /* Ensure ecode is set for log fn. */
54 tdb->ecode = TDB_ERR_CORRUPT;
55 TDB_LOG((tdb, TDB_DEBUG_WARNING, "tdb_rec_free_read bad magic 0x%x at offset=%u\n",
59 if (tdb->methods->tdb_oob(tdb, rec->next, sizeof(*rec), 0) != 0)
66 /* Remove an element from the freelist. Must have alloc lock. */
67 static int remove_from_freelist(struct tdb_context *tdb, tdb_off_t off, tdb_off_t next)
69 tdb_off_t last_ptr, i;
71 /* read in the freelist top */
72 last_ptr = FREELIST_TOP;
73 while (tdb_ofs_read(tdb, last_ptr, &i) != -1 && i != 0) {
76 return tdb_ofs_write(tdb, last_ptr, &next);
78 /* Follow chain (next offset is at start of record) */
81 tdb->ecode = TDB_ERR_CORRUPT;
82 TDB_LOG((tdb, TDB_DEBUG_FATAL,"remove_from_freelist: not on list at off=%u\n", off));
88 /* update a record tailer (must hold allocation lock) */
89 static int update_tailer(struct tdb_context *tdb, tdb_off_t offset,
90 const struct tdb_record *rec)
94 /* Offset of tailer from record header */
95 totalsize = sizeof(*rec) + rec->rec_len;
96 return tdb_ofs_write(tdb, offset + totalsize - sizeof(tdb_off_t),
101 * Read the record directly on the left.
102 * Fail if there is no record on the left.
104 static int read_record_on_left(struct tdb_context *tdb, tdb_off_t rec_ptr,
106 struct tdb_record *left_r)
110 struct tdb_record left_rec;
113 left_ptr = rec_ptr - sizeof(tdb_off_t);
115 if (left_ptr <= TDB_DATA_START(tdb->hash_size)) {
116 /* no record on the left */
120 /* Read in tailer and jump back to header */
121 ret = tdb_ofs_read(tdb, left_ptr, &left_size);
123 TDB_LOG((tdb, TDB_DEBUG_FATAL,
124 "tdb_free: left offset read failed at %u\n", left_ptr));
128 /* it could be uninitialised data */
129 if (left_size == 0 || left_size == TDB_PAD_U32) {
133 if (left_size > rec_ptr) {
137 left_ptr = rec_ptr - left_size;
139 if (left_ptr < TDB_DATA_START(tdb->hash_size)) {
143 /* Now read in the left record */
144 ret = tdb->methods->tdb_read(tdb, left_ptr, &left_rec,
145 sizeof(left_rec), DOCONV());
147 TDB_LOG((tdb, TDB_DEBUG_FATAL,
148 "tdb_free: left read failed at %u (%u)\n",
149 left_ptr, left_size));
160 * Merge new freelist record with the direct left neighbour.
161 * This assumes that left_rec represents the record
162 * directly to the left of right_rec and that this is
165 static int merge_with_left_record(struct tdb_context *tdb,
167 struct tdb_record *left_rec,
168 struct tdb_record *right_rec)
172 left_rec->rec_len += sizeof(*right_rec) + right_rec->rec_len;
174 ret = tdb_rec_write(tdb, left_ptr, left_rec);
176 TDB_LOG((tdb, TDB_DEBUG_FATAL,
177 "merge_with_left_record: update_left failed at %u\n",
182 ret = update_tailer(tdb, left_ptr, left_rec);
184 TDB_LOG((tdb, TDB_DEBUG_FATAL,
185 "merge_with_left_record: update_tailer failed at %u\n",
194 * Check whether the record left of a given freelist record is
195 * also a freelist record, and if so, merge the two records.
199 * 0 if left was not a free record
200 * 1 if left was free and successfully merged.
202 * The currend record is handed in with pointer and fully read record.
204 * The left record pointer and struct can be retrieved as result
207 static int check_merge_with_left_record(struct tdb_context *tdb,
209 struct tdb_record *rec,
211 struct tdb_record *lr)
214 struct tdb_record left_rec;
217 ret = read_record_on_left(tdb, rec_ptr, &left_ptr, &left_rec);
222 if (left_rec.magic != TDB_FREE_MAGIC) {
226 /* It's free - expand to include it. */
227 ret = merge_with_left_record(tdb, left_ptr, &left_rec, rec);
244 * Check whether the record left of a given freelist record is
245 * also a freelist record, and if so, merge the two records.
249 * 0 if left was not a free record
250 * 1 if left was free and successfully merged.
252 * In this variant, the input record is specified just as the pointer
253 * and is read from the database if needed.
255 * next_ptr will contain the original record's next pointer after
256 * successful merging (which will be lost after merging), so that
257 * the caller can update the last pointer.
259 static int check_merge_ptr_with_left_record(struct tdb_context *tdb,
264 struct tdb_record rec, left_rec;
267 ret = read_record_on_left(tdb, rec_ptr, &left_ptr, &left_rec);
272 if (left_rec.magic != TDB_FREE_MAGIC) {
276 /* It's free - expand to include it. */
278 ret = tdb->methods->tdb_read(tdb, rec_ptr, &rec,
279 sizeof(rec), DOCONV());
284 ret = merge_with_left_record(tdb, left_ptr, &left_rec, &rec);
289 if (next_ptr != NULL) {
290 *next_ptr = rec.next;
297 * Add an element into the freelist.
299 * We merge the new record into the left record if it is also a
300 * free record, but not with the right one. This makes the
301 * operation O(1) instead of O(n): merging with the right record
302 * requires a traverse of the freelist to find the previous
303 * record in the free list.
305 * This prevents db traverses from being O(n^2) after a lot of deletes.
307 int tdb_free(struct tdb_context *tdb, tdb_off_t offset, struct tdb_record *rec)
311 /* Allocation and tailer lock */
312 if (tdb_lock(tdb, -1, F_WRLCK) != 0)
315 /* set an initial tailer, so if we fail we don't leave a bogus record */
316 if (update_tailer(tdb, offset, rec) != 0) {
317 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: update_tailer failed!\n"));
322 /* Look right first (I'm an Australian, dammit) */
323 if (offset + sizeof(*rec) + rec->rec_len + sizeof(*rec) <= tdb->map_size) {
324 tdb_off_t right = offset + sizeof(*rec) + rec->rec_len;
327 if (tdb->methods->tdb_read(tdb, right, &r, sizeof(r), DOCONV()) == -1) {
328 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: right read failed at %u\n", right));
332 /* If it's free, expand to include it. */
333 if (r.magic == TDB_FREE_MAGIC) {
334 if (remove_from_freelist(tdb, right, r.next) == -1) {
335 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: right free failed at %u\n", right));
338 rec->rec_len += sizeof(r) + r.rec_len;
339 if (update_tailer(tdb, offset, rec) == -1) {
340 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: update_tailer failed at %u\n", offset));
348 ret = check_merge_with_left_record(tdb, offset, rec, NULL, NULL);
357 /* Nothing to merge, prepend to free list */
359 rec->magic = TDB_FREE_MAGIC;
361 if (tdb_ofs_read(tdb, FREELIST_TOP, &rec->next) == -1 ||
362 tdb_rec_write(tdb, offset, rec) == -1 ||
363 tdb_ofs_write(tdb, FREELIST_TOP, &offset) == -1) {
364 TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free record write failed at offset=%u\n", offset));
369 /* And we're done. */
370 tdb_unlock(tdb, -1, F_WRLCK);
374 tdb_unlock(tdb, -1, F_WRLCK);
381 the core of tdb_allocate - called when we have decided which
382 free list entry to use
384 Note that we try to allocate by grabbing data from the end of an existing record,
385 not the beginning. This is so the left merge in a free is more likely to be
386 able to free up the record without fragmentation
388 static tdb_off_t tdb_allocate_ofs(struct tdb_context *tdb,
389 tdb_len_t length, tdb_off_t rec_ptr,
390 struct tdb_record *rec, tdb_off_t last_ptr)
392 #define MIN_REC_SIZE (sizeof(struct tdb_record) + sizeof(tdb_off_t) + 8)
394 if (rec->rec_len < length + MIN_REC_SIZE) {
395 /* we have to grab the whole record */
397 /* unlink it from the previous record */
398 if (tdb_ofs_write(tdb, last_ptr, &rec->next) == -1) {
402 /* mark it not free */
403 rec->magic = TDB_MAGIC;
404 if (tdb_rec_write(tdb, rec_ptr, rec) == -1) {
410 /* we're going to just shorten the existing record */
411 rec->rec_len -= (length + sizeof(*rec));
412 if (tdb_rec_write(tdb, rec_ptr, rec) == -1) {
415 if (update_tailer(tdb, rec_ptr, rec) == -1) {
419 /* and setup the new record */
420 rec_ptr += sizeof(*rec) + rec->rec_len;
422 memset(rec, '\0', sizeof(*rec));
423 rec->rec_len = length;
424 rec->magic = TDB_MAGIC;
426 if (tdb_rec_write(tdb, rec_ptr, rec) == -1) {
430 if (update_tailer(tdb, rec_ptr, rec) == -1) {
437 /* allocate some space from the free list. The offset returned points
438 to a unconnected tdb_record within the database with room for at
439 least length bytes of total data
441 0 is returned if the space could not be allocated
443 static tdb_off_t tdb_allocate_from_freelist(
444 struct tdb_context *tdb, tdb_len_t length, struct tdb_record *rec)
446 tdb_off_t rec_ptr, last_ptr, newrec_ptr;
447 struct tdb_chainwalk_ctx chainwalk;
450 tdb_off_t rec_ptr, last_ptr;
453 float multiplier = 1.0;
454 bool merge_created_candidate;
456 /* over-allocate to reduce fragmentation */
459 /* Extra bytes required for tailer */
460 length += sizeof(tdb_off_t);
461 length = TDB_ALIGN(length, TDB_ALIGNMENT);
464 merge_created_candidate = false;
465 last_ptr = FREELIST_TOP;
467 /* read in the freelist top */
468 if (tdb_ofs_read(tdb, FREELIST_TOP, &rec_ptr) == -1)
472 tdb_chainwalk_init(&chainwalk, rec_ptr);
475 bestfit.last_ptr = 0;
479 this is a best fit allocation strategy. Originally we used
480 a first fit strategy, but it suffered from massive fragmentation
481 issues when faced with a slowly increasing record size.
486 struct tdb_record left_rec;
488 if (tdb_rec_free_read(tdb, rec_ptr, rec) == -1) {
492 ret = check_merge_with_left_record(tdb, rec_ptr, rec,
493 &left_ptr, &left_rec);
500 ret = tdb_ofs_write(tdb, last_ptr, &rec->next);
506 * We have merged the current record into the left
507 * neighbour. So our traverse of the freelist will
508 * skip it and consider the next record in the chain.
510 * But the enlarged left neighbour may be a candidate.
511 * If it is, we can not directly use it, though.
512 * The only thing we can do and have to do here is to
513 * update the current best fit size in the chain if the
514 * current best fit is the left record. (By that we may
515 * worsen the best fit we already had, bit this is not a
518 * If the current best fit is not the left record,
519 * all we can do is remember the fact that a merge
520 * created a new candidate so that we can trigger
521 * a second walk of the freelist if at the end of
522 * the first walk we have not found any fit.
523 * This way we can avoid expanding the database.
526 if (bestfit.rec_ptr == left_ptr) {
527 bestfit.rec_len = left_rec.rec_len;
530 if (left_rec.rec_len > length) {
531 merge_created_candidate = true;
539 if (rec->rec_len >= length) {
540 if (bestfit.rec_ptr == 0 ||
541 rec->rec_len < bestfit.rec_len) {
542 bestfit.rec_len = rec->rec_len;
543 bestfit.rec_ptr = rec_ptr;
544 bestfit.last_ptr = last_ptr;
548 /* move to the next record */
554 ok = tdb_chainwalk_check(tdb, &chainwalk, rec_ptr);
560 /* if we've found a record that is big enough, then
561 stop searching if its also not too big. The
562 definition of 'too big' changes as we scan
564 if (bestfit.rec_len > 0 &&
565 bestfit.rec_len < length * multiplier) {
569 /* this multiplier means we only extremely rarely
570 search more than 50 or so records. At 50 records we
571 accept records up to 11 times larger than what we
576 if (bestfit.rec_ptr != 0) {
577 if (tdb_rec_free_read(tdb, bestfit.rec_ptr, rec) == -1) {
581 newrec_ptr = tdb_allocate_ofs(tdb, length, bestfit.rec_ptr,
582 rec, bestfit.last_ptr);
586 if (merge_created_candidate) {
590 /* we didn't find enough space. See if we can expand the
591 database and if we can then try again */
592 if (tdb_expand(tdb, length + sizeof(*rec)) == 0)
598 static bool tdb_alloc_dead(
599 struct tdb_context *tdb, int hash, tdb_len_t length,
600 tdb_off_t *rec_ptr, struct tdb_record *rec)
604 *rec_ptr = tdb_find_dead(tdb, hash, rec, length, &last_ptr);
609 * Unlink the record from the hash chain, it's about to be moved into
612 return (tdb_ofs_write(tdb, last_ptr, &rec->next) == 0);
616 * Chain "hash" is assumed to be locked
619 tdb_off_t tdb_allocate(struct tdb_context *tdb, int hash, tdb_len_t length,
620 struct tdb_record *rec)
625 if (tdb->max_dead_records == 0) {
627 * No dead records to expect anywhere. Do the blocking
628 * freelist lock without trying to steal from others
630 goto blocking_freelist_allocate;
634 * The following loop tries to get the freelist lock nonblocking. If
635 * it gets the lock, allocate from there. If the freelist is busy,
636 * instead of waiting we try to steal dead records from other hash
639 * Be aware that we do nonblocking locks on the other hash chains as
640 * well and fail gracefully. This way we avoid deadlocks (we block two
641 * hash chains, something which is pretty bad normally)
644 for (i=0; i<tdb->hash_size; i++) {
648 list = BUCKET(hash+i);
650 if (tdb_lock_nonblock(tdb, list, F_WRLCK) == 0) {
653 got_dead = tdb_alloc_dead(tdb, list, length, &ret, rec);
654 tdb_unlock(tdb, list, F_WRLCK);
661 if (tdb_lock_nonblock(tdb, -1, F_WRLCK) == 0) {
663 * Under the freelist lock take the chance to give
664 * back our dead records.
666 tdb_purge_dead(tdb, hash);
668 ret = tdb_allocate_from_freelist(tdb, length, rec);
669 tdb_unlock(tdb, -1, F_WRLCK);
674 blocking_freelist_allocate:
676 if (tdb_lock(tdb, -1, F_WRLCK) == -1) {
679 ret = tdb_allocate_from_freelist(tdb, length, rec);
680 tdb_unlock(tdb, -1, F_WRLCK);
685 * Merge adjacent records in the freelist.
687 static int tdb_freelist_merge_adjacent(struct tdb_context *tdb,
688 int *count_records, int *count_merged)
695 ret = tdb_lock(tdb, -1, F_RDLCK);
701 while (tdb_ofs_read(tdb, cur, &next) == 0 && next != 0) {
706 ret = check_merge_ptr_with_left_record(tdb, next, &next2);
713 * now let cur->next point to next2 instead of next
716 ret = tdb_ofs_write(tdb, cur, &next2);
728 if (count_records != NULL) {
729 *count_records = count;
732 if (count_merged != NULL) {
733 *count_merged = merged;
739 tdb_unlock(tdb, -1, F_RDLCK);
744 * return the size of the freelist - no merging done
746 static int tdb_freelist_size_no_merge(struct tdb_context *tdb)
751 if (tdb_lock(tdb, -1, F_RDLCK) == -1) {
756 while (tdb_ofs_read(tdb, ptr, &ptr) == 0 && ptr != 0) {
760 tdb_unlock(tdb, -1, F_RDLCK);
765 * return the size of the freelist - used to decide if we should repack
767 * As a side effect, adjacent records are merged unless the
768 * database is read-only, in order to reduce the fragmentation
771 _PUBLIC_ int tdb_freelist_size(struct tdb_context *tdb)
776 if (tdb->read_only) {
777 count = tdb_freelist_size_no_merge(tdb);
780 ret = tdb_freelist_merge_adjacent(tdb, &count, NULL);