fb45c80c7aecc65e6156180c3dc61e4db2750d17
[ccan] / ccan / tdb2 / free.c
1  /*
2    Trivial Database 2: free list/block handling
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/likely/likely.h>
20 #include <ccan/ilog/ilog.h>
21 #include <time.h>
22 #include <assert.h>
23 #include <limits.h>
24
25 static unsigned fls64(uint64_t val)
26 {
27         return ilog64(val);
28 }
29
30 /* In which bucket would we find a particular record size? (ignoring header) */
31 unsigned int size_to_bucket(tdb_len_t data_len)
32 {
33         unsigned int bucket;
34
35         /* We can't have records smaller than this. */
36         assert(data_len >= TDB_MIN_DATA_LEN);
37
38         /* Ignoring the header... */
39         if (data_len - TDB_MIN_DATA_LEN <= 64) {
40                 /* 0 in bucket 0, 8 in bucket 1... 64 in bucket 8. */
41                 bucket = (data_len - TDB_MIN_DATA_LEN) / 8;
42         } else {
43                 /* After that we go power of 2. */
44                 bucket = fls64(data_len - TDB_MIN_DATA_LEN) + 2;
45         }
46
47         if (unlikely(bucket >= TDB_FREE_BUCKETS))
48                 bucket = TDB_FREE_BUCKETS - 1;
49         return bucket;
50 }
51
52 tdb_off_t first_ftable(struct tdb_context *tdb)
53 {
54         return tdb_read_off(tdb, offsetof(struct tdb_header, free_table));
55 }
56
57 tdb_off_t next_ftable(struct tdb_context *tdb, tdb_off_t ftable)
58 {
59         return tdb_read_off(tdb, ftable + offsetof(struct tdb_freetable,next));
60 }
61
62 int tdb_ftable_init(struct tdb_context *tdb)
63 {
64         /* Use reservoir sampling algorithm to select a free list at random. */
65         unsigned int rnd, max = 0, count = 0;
66         tdb_off_t off;
67
68         tdb->ftable_off = off = first_ftable(tdb);
69         tdb->ftable = 0;
70
71         while (off) {
72                 if (off == TDB_OFF_ERR)
73                         return -1;
74
75                 rnd = random();
76                 if (rnd >= max) {
77                         tdb->ftable_off = off;
78                         tdb->ftable = count;
79                         max = rnd;
80                 }
81
82                 off = next_ftable(tdb, off);
83                 count++;
84         }
85         return 0;
86 }
87
88 /* Offset of a given bucket. */
89 tdb_off_t bucket_off(tdb_off_t ftable_off, unsigned bucket)
90 {
91         return ftable_off + offsetof(struct tdb_freetable, buckets)
92                 + bucket * sizeof(tdb_off_t);
93 }
94
95 /* Returns free_buckets + 1, or list number to search. */
96 static tdb_off_t find_free_head(struct tdb_context *tdb,
97                                 tdb_off_t ftable_off,
98                                 tdb_off_t bucket)
99 {
100         /* Speculatively search for a non-zero bucket. */
101         return tdb_find_nonzero_off(tdb, bucket_off(ftable_off, 0),
102                                     bucket, TDB_FREE_BUCKETS);
103 }
104
105 /* Remove from free bucket. */
106 static int remove_from_list(struct tdb_context *tdb,
107                             tdb_off_t b_off, tdb_off_t r_off,
108                             const struct tdb_free_record *r)
109 {
110         tdb_off_t off;
111         enum TDB_ERROR ecode;
112
113         /* Front of list? */
114         if (frec_prev(r) == 0) {
115                 off = b_off;
116         } else {
117                 off = frec_prev(r) + offsetof(struct tdb_free_record, next);
118         }
119
120 #ifdef CCAN_TDB2_DEBUG
121         if (tdb_read_off(tdb, off) != r_off) {
122                 tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
123                          "remove_from_list: %llu bad prev in list %llu",
124                          (long long)r_off, (long long)b_off);
125                 return -1;
126         }
127 #endif
128
129         /* r->prev->next = r->next */
130         ecode = tdb_write_off(tdb, off, r->next);
131         if (ecode != TDB_SUCCESS) {
132                 tdb->ecode = ecode;
133                 return -1;
134         }
135
136         if (r->next != 0) {
137                 off = r->next + offsetof(struct tdb_free_record,magic_and_prev);
138                 /* r->next->prev = r->prev */
139
140 #ifdef CCAN_TDB2_DEBUG
141                 if (tdb_read_off(tdb, off) & TDB_OFF_MASK != r_off) {
142                         tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
143                                    "remove_from_list: %llu bad list %llu",
144                                    (long long)r_off, (long long)b_off);
145                         return -1;
146                 }
147 #endif
148
149                 ecode = tdb_write_off(tdb, off, r->magic_and_prev);
150                 if (ecode != TDB_SUCCESS) {
151                         tdb->ecode = ecode;
152                         return -1;
153                 }
154         }
155         return 0;
156 }
157
158 /* Enqueue in this free bucket. */
159 static int enqueue_in_free(struct tdb_context *tdb,
160                            tdb_off_t b_off,
161                            tdb_off_t off,
162                            tdb_len_t len)
163 {
164         struct tdb_free_record new;
165         enum TDB_ERROR ecode;
166         uint64_t magic = (TDB_FREE_MAGIC << (64 - TDB_OFF_UPPER_STEAL));
167
168         /* We only need to set ftable_and_len; rest is set in enqueue_in_free */
169         new.ftable_and_len = ((uint64_t)tdb->ftable << (64 - TDB_OFF_UPPER_STEAL))
170                 | len;
171         /* prev = 0. */
172         new.magic_and_prev = magic;
173
174         /* new->next = head. */
175         new.next = tdb_read_off(tdb, b_off);
176         if (new.next == TDB_OFF_ERR)
177                 return -1;
178
179         if (new.next) {
180 #ifdef CCAN_TDB2_DEBUG
181                 if (tdb_read_off(tdb,
182                                  new.next + offsetof(struct tdb_free_record,
183                                                      magic_and_prev))
184                     != magic) {
185                         tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
186                                    "enqueue_in_free: %llu bad head"
187                                    " prev %llu",
188                                    (long long)new.next, (long long)b_off);
189                         return -1;
190                 }
191 #endif
192                 /* next->prev = new. */
193                 ecode = tdb_write_off(tdb, new.next
194                                       + offsetof(struct tdb_free_record,
195                                                  magic_and_prev),
196                                       off | magic);
197                 if (ecode != TDB_SUCCESS) {
198                         tdb->ecode = ecode;
199                         return -1;
200                 }
201         }
202         /* head = new */
203         ecode = tdb_write_off(tdb, b_off, off);
204         if (ecode != TDB_SUCCESS) {
205                 tdb->ecode = ecode;
206                 return -1;
207         }
208
209         ecode = tdb_write_convert(tdb, off, &new, sizeof(new));
210         if (ecode != TDB_SUCCESS) {
211                 tdb->ecode = ecode;
212                 return -1;
213         }
214         return 0;
215 }
216
217 /* List need not be locked. */
218 int add_free_record(struct tdb_context *tdb,
219                     tdb_off_t off, tdb_len_t len_with_header)
220 {
221         tdb_off_t b_off;
222         tdb_len_t len;
223         int ret;
224         enum TDB_ERROR ecode;
225
226         assert(len_with_header >= sizeof(struct tdb_free_record));
227
228         len = len_with_header - sizeof(struct tdb_used_record);
229
230         b_off = bucket_off(tdb->ftable_off, size_to_bucket(len));
231         ecode = tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT);
232         if (ecode != TDB_SUCCESS) {
233                 tdb->ecode = ecode;
234                 return -1;
235         }
236
237         ret = enqueue_in_free(tdb, b_off, off, len);
238         tdb_unlock_free_bucket(tdb, b_off);
239         return ret;
240 }
241
242 static size_t adjust_size(size_t keylen, size_t datalen)
243 {
244         size_t size = keylen + datalen;
245
246         if (size < TDB_MIN_DATA_LEN)
247                 size = TDB_MIN_DATA_LEN;
248
249         /* Round to next uint64_t boundary. */
250         return (size + (sizeof(uint64_t) - 1ULL)) & ~(sizeof(uint64_t) - 1ULL);
251 }
252
253 /* If we have enough left over to be useful, split that off. */
254 static size_t record_leftover(size_t keylen, size_t datalen,
255                               bool want_extra, size_t total_len)
256 {
257         ssize_t leftover;
258
259         if (want_extra)
260                 datalen += datalen / 2;
261         leftover = total_len - adjust_size(keylen, datalen);
262
263         if (leftover < (ssize_t)sizeof(struct tdb_free_record))
264                 return 0;
265
266         return leftover;
267 }
268
269 static tdb_off_t ftable_offset(struct tdb_context *tdb, unsigned int ftable)
270 {
271         tdb_off_t off;
272         unsigned int i;
273
274         if (likely(tdb->ftable == ftable))
275                 return tdb->ftable_off;
276
277         off = first_ftable(tdb);
278         for (i = 0; i < ftable; i++)
279                 off = next_ftable(tdb, off);
280         return off;
281 }
282
283 /* Note: we unlock the current bucket if we coalesce or fail. */
284 static int coalesce(struct tdb_context *tdb,
285                     tdb_off_t off, tdb_off_t b_off, tdb_len_t data_len)
286 {
287         tdb_off_t end;
288         struct tdb_free_record rec;
289         enum TDB_ERROR ecode;
290
291         add_stat(tdb, alloc_coalesce_tried, 1);
292         end = off + sizeof(struct tdb_used_record) + data_len;
293
294         while (end < tdb->map_size) {
295                 const struct tdb_free_record *r;
296                 tdb_off_t nb_off;
297                 unsigned ftable, bucket;
298
299                 r = tdb_access_read(tdb, end, sizeof(*r), true);
300                 if (!r)
301                         goto err;
302
303                 if (frec_magic(r) != TDB_FREE_MAGIC
304                     || frec_ftable(r) == TDB_FTABLE_NONE) {
305                         tdb_access_release(tdb, r);
306                         break;
307                 }
308
309                 ftable = frec_ftable(r);
310                 bucket = size_to_bucket(frec_len(r));
311                 nb_off = bucket_off(ftable_offset(tdb, ftable), bucket);
312                 tdb_access_release(tdb, r);
313
314                 /* We may be violating lock order here, so best effort. */
315                 if (tdb_lock_free_bucket(tdb, nb_off, TDB_LOCK_NOWAIT)
316                     != TDB_SUCCESS) {
317                         add_stat(tdb, alloc_coalesce_lockfail, 1);
318                         break;
319                 }
320
321                 /* Now we have lock, re-check. */
322                 ecode = tdb_read_convert(tdb, end, &rec, sizeof(rec));
323                 if (ecode != TDB_SUCCESS) {
324                         tdb->ecode = ecode;
325                         tdb_unlock_free_bucket(tdb, nb_off);
326                         goto err;
327                 }
328
329                 if (unlikely(frec_magic(&rec) != TDB_FREE_MAGIC)) {
330                         add_stat(tdb, alloc_coalesce_race, 1);
331                         tdb_unlock_free_bucket(tdb, nb_off);
332                         break;
333                 }
334
335                 if (unlikely(frec_ftable(&rec) != ftable)
336                     || unlikely(size_to_bucket(frec_len(&rec)) != bucket)) {
337                         add_stat(tdb, alloc_coalesce_race, 1);
338                         tdb_unlock_free_bucket(tdb, nb_off);
339                         break;
340                 }
341
342                 if (remove_from_list(tdb, nb_off, end, &rec) == -1) {
343                         tdb_unlock_free_bucket(tdb, nb_off);
344                         goto err;
345                 }
346
347                 end += sizeof(struct tdb_used_record) + frec_len(&rec);
348                 tdb_unlock_free_bucket(tdb, nb_off);
349                 add_stat(tdb, alloc_coalesce_num_merged, 1);
350         }
351
352         /* Didn't find any adjacent free? */
353         if (end == off + sizeof(struct tdb_used_record) + data_len)
354                 return 0;
355
356         /* OK, expand initial record */
357         ecode = tdb_read_convert(tdb, off, &rec, sizeof(rec));
358         if (ecode != TDB_SUCCESS) {
359                 tdb->ecode = ecode;
360                 goto err;
361         }
362
363         if (frec_len(&rec) != data_len) {
364                 tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
365                            "coalesce: expected data len %zu not %zu",
366                            (size_t)data_len, (size_t)frec_len(&rec));
367                 goto err;
368         }
369
370         if (remove_from_list(tdb, b_off, off, &rec) == -1)
371                 goto err;
372
373         /* We have to drop this to avoid deadlocks, so make sure record
374          * doesn't get coalesced by someone else! */
375         rec.ftable_and_len = (TDB_FTABLE_NONE << (64 - TDB_OFF_UPPER_STEAL))
376                 | (end - off - sizeof(struct tdb_used_record));
377         ecode = tdb_write_off(tdb, off + offsetof(struct tdb_free_record,
378                                                   ftable_and_len),
379                               rec.ftable_and_len);
380         if (ecode != TDB_SUCCESS) {
381                 tdb->ecode = ecode;
382                 goto err;
383         }
384
385         add_stat(tdb, alloc_coalesce_succeeded, 1);
386         tdb_unlock_free_bucket(tdb, b_off);
387
388         if (add_free_record(tdb, off, end - off) == -1)
389                 return -1;
390         return 1;
391
392 err:
393         /* To unify error paths, we *always* unlock bucket on error. */
394         tdb_unlock_free_bucket(tdb, b_off);
395         return -1;
396 }
397
398 /* We need size bytes to put our key and data in. */
399 static tdb_off_t lock_and_alloc(struct tdb_context *tdb,
400                                 tdb_off_t ftable_off,
401                                 tdb_off_t bucket,
402                                 size_t keylen, size_t datalen,
403                                 bool want_extra,
404                                 unsigned magic,
405                                 unsigned hashlow)
406 {
407         tdb_off_t off, b_off,best_off;
408         struct tdb_free_record best = { 0 };
409         double multiplier;
410         size_t size = adjust_size(keylen, datalen);
411         enum TDB_ERROR ecode;
412
413         add_stat(tdb, allocs, 1);
414 again:
415         b_off = bucket_off(ftable_off, bucket);
416
417         /* FIXME: Try non-blocking wait first, to measure contention. */
418         /* Lock this bucket. */
419         ecode = tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT);
420         if (ecode != TDB_SUCCESS) {
421                 tdb->ecode = ecode;
422                 return TDB_OFF_ERR;
423         }
424
425         best.ftable_and_len = -1ULL;
426         best_off = 0;
427
428         /* Get slack if we're after extra. */
429         if (want_extra)
430                 multiplier = 1.5;
431         else
432                 multiplier = 1.0;
433
434         /* Walk the list to see if any are large enough, getting less fussy
435          * as we go. */
436         off = tdb_read_off(tdb, b_off);
437         if (unlikely(off == TDB_OFF_ERR))
438                 goto unlock_err;
439
440         while (off) {
441                 const struct tdb_free_record *r;
442                 tdb_len_t len;
443                 tdb_off_t next;
444
445                 r = tdb_access_read(tdb, off, sizeof(*r), true);
446                 if (!r)
447                         goto unlock_err;
448
449                 if (frec_magic(r) != TDB_FREE_MAGIC) {
450                         tdb_access_release(tdb, r);
451                         tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
452                                  "lock_and_alloc: %llu non-free 0x%llx",
453                                  (long long)off, (long long)r->magic_and_prev);
454                         goto unlock_err;
455                 }
456
457                 if (frec_len(r) >= size && frec_len(r) < frec_len(&best)) {
458                         best_off = off;
459                         best = *r;
460                 }
461
462                 if (frec_len(&best) < size * multiplier && best_off) {
463                         tdb_access_release(tdb, r);
464                         break;
465                 }
466
467                 multiplier *= 1.01;
468
469                 next = r->next;
470                 len = frec_len(r);
471                 tdb_access_release(tdb, r);
472
473                 /* Since we're going slow anyway, try coalescing here. */
474                 switch (coalesce(tdb, off, b_off, len)) {
475                 case -1:
476                         /* This has already unlocked on error. */
477                         return -1;
478                 case 1:
479                         /* This has unlocked list, restart. */
480                         goto again;
481                 }
482                 off = next;
483         }
484
485         /* If we found anything at all, use it. */
486         if (best_off) {
487                 struct tdb_used_record rec;
488                 size_t leftover;
489
490                 /* We're happy with this size: take it. */
491                 if (remove_from_list(tdb, b_off, best_off, &best) != 0)
492                         goto unlock_err;
493
494                 leftover = record_leftover(keylen, datalen, want_extra,
495                                            frec_len(&best));
496
497                 assert(keylen + datalen + leftover <= frec_len(&best));
498                 /* We need to mark non-free before we drop lock, otherwise
499                  * coalesce() could try to merge it! */
500                 if (set_header(tdb, &rec, magic, keylen, datalen,
501                                frec_len(&best) - leftover, hashlow) != 0)
502                         goto unlock_err;
503
504                 ecode = tdb_write_convert(tdb, best_off, &rec, sizeof(rec));
505                 if (ecode != TDB_SUCCESS) {
506                         tdb->ecode = ecode;
507                         goto unlock_err;
508                 }
509
510                 /* Bucket of leftover will be <= current bucket, so nested
511                  * locking is allowed. */
512                 if (leftover) {
513                         add_stat(tdb, alloc_leftover, 1);
514                         if (add_free_record(tdb,
515                                             best_off + sizeof(rec)
516                                             + frec_len(&best) - leftover,
517                                             leftover))
518                                 best_off = TDB_OFF_ERR;
519                 }
520                 tdb_unlock_free_bucket(tdb, b_off);
521
522                 return best_off;
523         }
524
525         tdb_unlock_free_bucket(tdb, b_off);
526         return 0;
527
528 unlock_err:
529         tdb_unlock_free_bucket(tdb, b_off);
530         return TDB_OFF_ERR;
531 }
532
533 /* Get a free block from current free list, or 0 if none. */
534 static tdb_off_t get_free(struct tdb_context *tdb,
535                           size_t keylen, size_t datalen, bool want_extra,
536                           unsigned magic, unsigned hashlow)
537 {
538         tdb_off_t off, ftable_off;
539         unsigned start_b, b, ftable;
540         bool wrapped = false;
541
542         /* If they are growing, add 50% to get to higher bucket. */
543         if (want_extra)
544                 start_b = size_to_bucket(adjust_size(keylen,
545                                                      datalen + datalen / 2));
546         else
547                 start_b = size_to_bucket(adjust_size(keylen, datalen));
548
549         ftable_off = tdb->ftable_off;
550         ftable = tdb->ftable;
551         while (!wrapped || ftable_off != tdb->ftable_off) {
552                 /* Start at exact size bucket, and search up... */
553                 for (b = find_free_head(tdb, ftable_off, start_b);
554                      b < TDB_FREE_BUCKETS;
555                      b = find_free_head(tdb, ftable_off, b + 1)) {
556                         /* Try getting one from list. */
557                         off = lock_and_alloc(tdb, ftable_off,
558                                              b, keylen, datalen, want_extra,
559                                              magic, hashlow);
560                         if (off == TDB_OFF_ERR)
561                                 return TDB_OFF_ERR;
562                         if (off != 0) {
563                                 if (b == start_b)
564                                         add_stat(tdb, alloc_bucket_exact, 1);
565                                 if (b == TDB_FREE_BUCKETS - 1)
566                                         add_stat(tdb, alloc_bucket_max, 1);
567                                 /* Worked?  Stay using this list. */
568                                 tdb->ftable_off = ftable_off;
569                                 tdb->ftable = ftable;
570                                 return off;
571                         }
572                         /* Didn't work.  Try next bucket. */
573                 }
574
575                 /* Hmm, try next table. */
576                 ftable_off = next_ftable(tdb, ftable_off);
577                 ftable++;
578
579                 if (ftable_off == 0) {
580                         wrapped = true;
581                         ftable_off = first_ftable(tdb);
582                         ftable = 0;
583                 }
584         }
585
586         return 0;
587 }
588
589 int set_header(struct tdb_context *tdb,
590                struct tdb_used_record *rec,
591                unsigned magic, uint64_t keylen, uint64_t datalen,
592                uint64_t actuallen, unsigned hashlow)
593 {
594         uint64_t keybits = (fls64(keylen) + 1) / 2;
595
596         /* Use bottom bits of hash, so it's independent of hash table size. */
597         rec->magic_and_meta = (hashlow & ((1 << 11)-1))
598                 | ((actuallen - (keylen + datalen)) << 11)
599                 | (keybits << 43)
600                 | ((uint64_t)magic << 48);
601         rec->key_and_data_len = (keylen | (datalen << (keybits*2)));
602
603         /* Encoding can fail on big values. */
604         if (rec_key_length(rec) != keylen
605             || rec_data_length(rec) != datalen
606             || rec_extra_padding(rec) != actuallen - (keylen + datalen)) {
607                 tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
608                          "Could not encode k=%llu,d=%llu,a=%llu",
609                          (long long)keylen, (long long)datalen,
610                          (long long)actuallen);
611                 return -1;
612         }
613         return 0;
614 }
615
616 /* Expand the database. */
617 static int tdb_expand(struct tdb_context *tdb, tdb_len_t size)
618 {
619         uint64_t old_size;
620         tdb_len_t wanted;
621         enum TDB_ERROR ecode;
622
623         /* We need room for the record header too. */
624         wanted = sizeof(struct tdb_used_record) + size;
625
626         /* Need to hold a hash lock to expand DB: transactions rely on it. */
627         if (!(tdb->flags & TDB_NOLOCK)
628             && !tdb->allrecord_lock.count && !tdb_has_hash_locks(tdb)) {
629                 tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
630                            "tdb_expand: must hold lock during expand");
631                 return -1;
632         }
633
634         /* always make room for at least 100 more records, and at
635            least 25% more space. */
636         if (size * TDB_EXTENSION_FACTOR > tdb->map_size / 4)
637                 wanted = size * TDB_EXTENSION_FACTOR;
638         else
639                 wanted = tdb->map_size / 4;
640         wanted = adjust_size(0, wanted);
641
642         /* Only one person can expand file at a time. */
643         ecode = tdb_lock_expand(tdb, F_WRLCK);
644         if (ecode != TDB_SUCCESS) {
645                 tdb->ecode = ecode;
646                 return -1;
647         }
648
649         /* Someone else may have expanded the file, so retry. */
650         old_size = tdb->map_size;
651         tdb->methods->oob(tdb, tdb->map_size + 1, true);
652         if (tdb->map_size != old_size) {
653                 tdb_unlock_expand(tdb, F_WRLCK);
654                 return 0;
655         }
656
657         ecode = tdb->methods->expand_file(tdb, wanted);
658         if (ecode != TDB_SUCCESS) {
659                 tdb->ecode = ecode;
660                 tdb_unlock_expand(tdb, F_WRLCK);
661                 return -1;
662         }
663
664         /* We need to drop this lock before adding free record. */
665         tdb_unlock_expand(tdb, F_WRLCK);
666
667         add_stat(tdb, expands, 1);
668         return add_free_record(tdb, old_size, wanted);
669 }
670
671 /* This won't fail: it will expand the database if it has to. */
672 tdb_off_t alloc(struct tdb_context *tdb, size_t keylen, size_t datalen,
673                 uint64_t hash, unsigned magic, bool growing)
674 {
675         tdb_off_t off;
676
677         /* We can't hold pointers during this: we could unmap! */
678         assert(!tdb->direct_access);
679
680         for (;;) {
681                 off = get_free(tdb, keylen, datalen, growing, magic, hash);
682                 if (likely(off != 0))
683                         break;
684
685                 if (tdb_expand(tdb, adjust_size(keylen, datalen)))
686                         return TDB_OFF_ERR;
687         }
688
689         return off;
690 }