tdb: defragment the freelist in tdb_allocate_from_freelist()
[Samba.git] / lib / tdb / common / freelist.c
blob86fac2ff078ba87c82969e7731f58ced7f713e8a
1 /*
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
12 ** under the LGPL
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
32 O(1) time
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)
40 return -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",
46 rec->magic, off));
47 rec->magic = TDB_FREE_MAGIC;
48 if (tdb_rec_write(tdb, off, rec) == -1)
49 return -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",
56 rec->magic, off));
57 return -1;
59 if (tdb->methods->tdb_oob(tdb, rec->next, sizeof(*rec), 0) != 0)
60 return -1;
61 return 0;
65 #if USE_RIGHT_MERGES
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) {
74 if (i == off) {
75 /* We've found it! */
76 return tdb_ofs_write(tdb, last_ptr, &next);
78 /* Follow chain (next offset is at start of record) */
79 last_ptr = i;
81 tdb->ecode = TDB_ERR_CORRUPT;
82 TDB_LOG((tdb, TDB_DEBUG_FATAL,"remove_from_freelist: not on list at off=%u\n", off));
83 return -1;
85 #endif
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)
92 tdb_off_t totalsize;
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),
97 &totalsize);
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,
105 tdb_off_t *left_p,
106 struct tdb_record *left_r)
108 tdb_off_t left_ptr;
109 tdb_off_t left_size;
110 struct tdb_record left_rec;
111 int ret;
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 */
117 return -1;
120 /* Read in tailer and jump back to header */
121 ret = tdb_ofs_read(tdb, left_ptr, &left_size);
122 if (ret == -1) {
123 TDB_LOG((tdb, TDB_DEBUG_FATAL,
124 "tdb_free: left offset read failed at %u\n", left_ptr));
125 return -1;
128 /* it could be uninitialised data */
129 if (left_size == 0 || left_size == TDB_PAD_U32) {
130 return -1;
133 if (left_size > rec_ptr) {
134 return -1;
137 left_ptr = rec_ptr - left_size;
139 if (left_ptr < TDB_DATA_START(tdb->hash_size)) {
140 return -1;
143 /* Now read in the left record */
144 ret = tdb->methods->tdb_read(tdb, left_ptr, &left_rec,
145 sizeof(left_rec), DOCONV());
146 if (ret == -1) {
147 TDB_LOG((tdb, TDB_DEBUG_FATAL,
148 "tdb_free: left read failed at %u (%u)\n",
149 left_ptr, left_size));
150 return -1;
153 *left_p = left_ptr;
154 *left_r = left_rec;
156 return 0;
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
163 * a freelist record.
165 static int merge_with_left_record(struct tdb_context *tdb,
166 tdb_off_t left_ptr,
167 struct tdb_record *left_rec,
168 struct tdb_record *right_rec)
170 int ret;
172 left_rec->rec_len += sizeof(*right_rec) + right_rec->rec_len;
174 ret = tdb_rec_write(tdb, left_ptr, left_rec);
175 if (ret == -1) {
176 TDB_LOG((tdb, TDB_DEBUG_FATAL,
177 "merge_with_left_record: update_left failed at %u\n",
178 left_ptr));
179 return -1;
182 ret = update_tailer(tdb, left_ptr, left_rec);
183 if (ret == -1) {
184 TDB_LOG((tdb, TDB_DEBUG_FATAL,
185 "merge_with_left_record: update_tailer failed at %u\n",
186 left_ptr));
187 return -1;
190 return 0;
194 * Check whether the record left of a given freelist record is
195 * also a freelist record, and if so, merge the two records.
197 * Return code:
198 * -1 upon error
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
205 * in lp and lr;
207 static int check_merge_with_left_record(struct tdb_context *tdb,
208 tdb_off_t rec_ptr,
209 struct tdb_record *rec,
210 tdb_off_t *lp,
211 struct tdb_record *lr)
213 tdb_off_t left_ptr;
214 struct tdb_record left_rec;
215 int ret;
217 ret = read_record_on_left(tdb, rec_ptr, &left_ptr, &left_rec);
218 if (ret != 0) {
219 return 0;
222 if (left_rec.magic != TDB_FREE_MAGIC) {
223 return 0;
226 /* It's free - expand to include it. */
227 ret = merge_with_left_record(tdb, left_ptr, &left_rec, rec);
228 if (ret != 0) {
229 return -1;
232 if (lp != NULL) {
233 *lp = left_ptr;
236 if (lr != NULL) {
237 *lr = left_rec;
240 return 1;
244 * Check whether the record left of a given freelist record is
245 * also a freelist record, and if so, merge the two records.
247 * Return code:
248 * -1 upon error
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,
260 tdb_off_t rec_ptr,
261 tdb_off_t *next_ptr)
263 tdb_off_t left_ptr;
264 struct tdb_record rec, left_rec;
265 int ret;
267 ret = read_record_on_left(tdb, rec_ptr, &left_ptr, &left_rec);
268 if (ret != 0) {
269 return 0;
272 if (left_rec.magic != TDB_FREE_MAGIC) {
273 return 0;
276 /* It's free - expand to include it. */
278 ret = tdb->methods->tdb_read(tdb, rec_ptr, &rec,
279 sizeof(rec), DOCONV());
280 if (ret != 0) {
281 return -1;
284 ret = merge_with_left_record(tdb, left_ptr, &left_rec, &rec);
285 if (ret != 0) {
286 return -1;
289 if (next_ptr != NULL) {
290 *next_ptr = rec.next;
293 return 1;
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)
309 int ret;
311 /* Allocation and tailer lock */
312 if (tdb_lock(tdb, -1, F_WRLCK) != 0)
313 return -1;
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"));
318 goto fail;
321 #if USE_RIGHT_MERGES
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;
325 struct tdb_record r;
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));
329 goto left;
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));
336 goto left;
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));
341 goto fail;
345 left:
346 #endif
348 ret = check_merge_with_left_record(tdb, offset, rec, NULL, NULL);
349 if (ret == -1) {
350 goto fail;
352 if (ret == 1) {
353 /* merged */
354 goto done;
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));
365 goto fail;
368 done:
369 /* And we're done. */
370 tdb_unlock(tdb, -1, F_WRLCK);
371 return 0;
373 fail:
374 tdb_unlock(tdb, -1, F_WRLCK);
375 return -1;
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) {
399 return 0;
402 /* mark it not free */
403 rec->magic = TDB_MAGIC;
404 if (tdb_rec_write(tdb, rec_ptr, rec) == -1) {
405 return 0;
407 return rec_ptr;
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) {
413 return 0;
415 if (update_tailer(tdb, rec_ptr, rec) == -1) {
416 return 0;
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) {
427 return 0;
430 if (update_tailer(tdb, rec_ptr, rec) == -1) {
431 return 0;
434 return rec_ptr;
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 {
448 tdb_off_t rec_ptr, last_ptr;
449 tdb_len_t rec_len;
450 } bestfit;
451 float multiplier = 1.0;
452 bool merge_created_candidate;
454 /* over-allocate to reduce fragmentation */
455 length *= 1.25;
457 /* Extra bytes required for tailer */
458 length += sizeof(tdb_off_t);
459 length = TDB_ALIGN(length, TDB_ALIGNMENT);
461 again:
462 merge_created_candidate = false;
463 last_ptr = FREELIST_TOP;
465 /* read in the freelist top */
466 if (tdb_ofs_read(tdb, FREELIST_TOP, &rec_ptr) == -1)
467 return 0;
469 bestfit.rec_ptr = 0;
470 bestfit.last_ptr = 0;
471 bestfit.rec_len = 0;
474 this is a best fit allocation strategy. Originally we used
475 a first fit strategy, but it suffered from massive fragmentation
476 issues when faced with a slowly increasing record size.
478 while (rec_ptr) {
479 int ret;
480 tdb_off_t left_ptr;
481 struct tdb_record left_rec;
483 if (tdb_rec_free_read(tdb, rec_ptr, rec) == -1) {
484 return 0;
487 ret = check_merge_with_left_record(tdb, rec_ptr, rec,
488 &left_ptr, &left_rec);
489 if (ret == -1) {
490 return 0;
492 if (ret == 1) {
493 /* merged */
494 rec_ptr = rec->next;
495 ret = tdb_ofs_write(tdb, last_ptr, &rec->next);
496 if (ret == -1) {
497 return 0;
501 * We have merged the current record into the left
502 * neighbour. So our traverse of the freelist will
503 * skip it and consider the next record in the chain.
505 * But the enlarged left neighbour may be a candidate.
506 * If it is, we can not directly use it, though.
507 * The only thing we can do and have to do here is to
508 * update the current best fit size in the chain if the
509 * current best fit is the left record. (By that we may
510 * worsen the best fit we already had, bit this is not a
511 * problem.)
513 * If the current best fit is not the left record,
514 * all we can do is remember the fact that a merge
515 * created a new candidate so that we can trigger
516 * a second walk of the freelist if at the end of
517 * the first walk we have not found any fit.
518 * This way we can avoid expanding the database.
521 if (bestfit.rec_ptr == left_ptr) {
522 bestfit.rec_len = left_rec.rec_len;
525 if (left_rec.rec_len > length) {
526 merge_created_candidate = true;
529 continue;
532 if (rec->rec_len >= length) {
533 if (bestfit.rec_ptr == 0 ||
534 rec->rec_len < bestfit.rec_len) {
535 bestfit.rec_len = rec->rec_len;
536 bestfit.rec_ptr = rec_ptr;
537 bestfit.last_ptr = last_ptr;
541 /* move to the next record */
542 last_ptr = rec_ptr;
543 rec_ptr = rec->next;
545 /* if we've found a record that is big enough, then
546 stop searching if its also not too big. The
547 definition of 'too big' changes as we scan
548 through */
549 if (bestfit.rec_len > 0 &&
550 bestfit.rec_len < length * multiplier) {
551 break;
554 /* this multiplier means we only extremely rarely
555 search more than 50 or so records. At 50 records we
556 accept records up to 11 times larger than what we
557 want */
558 multiplier *= 1.05;
561 if (bestfit.rec_ptr != 0) {
562 if (tdb_rec_free_read(tdb, bestfit.rec_ptr, rec) == -1) {
563 return 0;
566 newrec_ptr = tdb_allocate_ofs(tdb, length, bestfit.rec_ptr,
567 rec, bestfit.last_ptr);
568 return newrec_ptr;
571 if (merge_created_candidate) {
572 goto again;
575 /* we didn't find enough space. See if we can expand the
576 database and if we can then try again */
577 if (tdb_expand(tdb, length + sizeof(*rec)) == 0)
578 goto again;
580 return 0;
583 static bool tdb_alloc_dead(
584 struct tdb_context *tdb, int hash, tdb_len_t length,
585 tdb_off_t *rec_ptr, struct tdb_record *rec)
587 tdb_off_t last_ptr;
589 *rec_ptr = tdb_find_dead(tdb, hash, rec, length, &last_ptr);
590 if (*rec_ptr == 0) {
591 return false;
594 * Unlink the record from the hash chain, it's about to be moved into
595 * another one.
597 return (tdb_ofs_write(tdb, last_ptr, &rec->next) == 0);
601 * Chain "hash" is assumed to be locked
604 tdb_off_t tdb_allocate(struct tdb_context *tdb, int hash, tdb_len_t length,
605 struct tdb_record *rec)
607 tdb_off_t ret;
608 int i;
610 if (tdb->max_dead_records == 0) {
612 * No dead records to expect anywhere. Do the blocking
613 * freelist lock without trying to steal from others
615 goto blocking_freelist_allocate;
619 * The following loop tries to get the freelist lock nonblocking. If
620 * it gets the lock, allocate from there. If the freelist is busy,
621 * instead of waiting we try to steal dead records from other hash
622 * chains.
624 * Be aware that we do nonblocking locks on the other hash chains as
625 * well and fail gracefully. This way we avoid deadlocks (we block two
626 * hash chains, something which is pretty bad normally)
629 for (i=0; i<tdb->hash_size; i++) {
631 int list;
633 list = BUCKET(hash+i);
635 if (tdb_lock_nonblock(tdb, list, F_WRLCK) == 0) {
636 bool got_dead;
638 got_dead = tdb_alloc_dead(tdb, list, length, &ret, rec);
639 tdb_unlock(tdb, list, F_WRLCK);
641 if (got_dead) {
642 return ret;
646 if (tdb_lock_nonblock(tdb, -1, F_WRLCK) == 0) {
648 * Under the freelist lock take the chance to give
649 * back our dead records.
651 tdb_purge_dead(tdb, hash);
653 ret = tdb_allocate_from_freelist(tdb, length, rec);
654 tdb_unlock(tdb, -1, F_WRLCK);
655 return ret;
659 blocking_freelist_allocate:
661 if (tdb_lock(tdb, -1, F_WRLCK) == -1) {
662 return 0;
664 ret = tdb_allocate_from_freelist(tdb, length, rec);
665 tdb_unlock(tdb, -1, F_WRLCK);
666 return ret;
670 * Merge adjacent records in the freelist.
672 static int tdb_freelist_merge_adjacent(struct tdb_context *tdb,
673 int *count_records, int *count_merged)
675 tdb_off_t cur, next;
676 int count = 0;
677 int merged = 0;
678 int ret;
680 ret = tdb_lock(tdb, -1, F_RDLCK);
681 if (ret == -1) {
682 return -1;
685 cur = FREELIST_TOP;
686 while (tdb_ofs_read(tdb, cur, &next) == 0 && next != 0) {
687 tdb_off_t next2;
689 count++;
691 ret = check_merge_ptr_with_left_record(tdb, next, &next2);
692 if (ret == -1) {
693 goto done;
695 if (ret == 1) {
697 * merged:
698 * now let cur->next point to next2 instead of next
701 ret = tdb_ofs_write(tdb, cur, &next2);
702 if (ret != 0) {
703 goto done;
706 next = next2;
707 merged++;
710 cur = next;
713 if (count_records != NULL) {
714 *count_records = count;
717 if (count_merged != NULL) {
718 *count_merged = merged;
721 ret = 0;
723 done:
724 tdb_unlock(tdb, -1, F_RDLCK);
725 return ret;
729 * return the size of the freelist - no merging done
731 static int tdb_freelist_size_no_merge(struct tdb_context *tdb)
733 tdb_off_t ptr;
734 int count=0;
736 if (tdb_lock(tdb, -1, F_RDLCK) == -1) {
737 return -1;
740 ptr = FREELIST_TOP;
741 while (tdb_ofs_read(tdb, ptr, &ptr) == 0 && ptr != 0) {
742 count++;
745 tdb_unlock(tdb, -1, F_RDLCK);
746 return count;
750 * return the size of the freelist - used to decide if we should repack
752 * As a side effect, adjacent records are merged unless the
753 * database is read-only, in order to reduce the fragmentation
754 * without repacking.
756 _PUBLIC_ int tdb_freelist_size(struct tdb_context *tdb)
759 int count = 0;
761 if (tdb->read_only) {
762 count = tdb_freelist_size_no_merge(tdb);
763 } else {
764 int ret;
765 ret = tdb_freelist_merge_adjacent(tdb, &count, NULL);
766 if (ret != 0) {
767 return -1;
771 return count;