be9b18455342a76ee316f98eff4363ade141849e
[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 enum TDB_ERROR 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 (TDB_OFF_IS_ERR(off)) {
73                         return off;
74                 }
75
76                 rnd = random();
77                 if (rnd >= max) {
78                         tdb->ftable_off = off;
79                         tdb->ftable = count;
80                         max = rnd;
81                 }
82
83                 off = next_ftable(tdb, off);
84                 count++;
85         }
86         return TDB_SUCCESS;
87 }
88
89 /* Offset of a given bucket. */
90 tdb_off_t bucket_off(tdb_off_t ftable_off, unsigned bucket)
91 {
92         return ftable_off + offsetof(struct tdb_freetable, buckets)
93                 + bucket * sizeof(tdb_off_t);
94 }
95
96 /* Returns free_buckets + 1, or list number to search, or -ve error. */
97 static tdb_off_t find_free_head(struct tdb_context *tdb,
98                                 tdb_off_t ftable_off,
99                                 tdb_off_t bucket)
100 {
101         /* Speculatively search for a non-zero bucket. */
102         return tdb_find_nonzero_off(tdb, bucket_off(ftable_off, 0),
103                                     bucket, TDB_FREE_BUCKETS);
104 }
105
106 static void check_list(struct tdb_context *tdb, tdb_off_t b_off)
107 {
108 #ifdef CCAN_TDB2_DEBUG
109         tdb_off_t off, prev = 0, first;
110         struct tdb_free_record r;
111
112         first = off = (tdb_read_off(tdb, b_off) & TDB_OFF_MASK);
113         while (off != 0) {
114                 tdb_read_convert(tdb, off, &r, sizeof(r));
115                 if (frec_magic(&r) != TDB_FREE_MAGIC)
116                         abort();
117                 if (prev && frec_prev(&r) != prev)
118                         abort();
119                 prev = off;
120                 off = r.next;
121         }
122
123         if (first) {
124                 tdb_read_convert(tdb, first, &r, sizeof(r));
125                 if (frec_prev(&r) != prev)
126                         abort();
127         }
128 #endif
129 }
130
131 /* Remove from free bucket. */
132 static enum TDB_ERROR remove_from_list(struct tdb_context *tdb,
133                                        tdb_off_t b_off, tdb_off_t r_off,
134                                        const struct tdb_free_record *r)
135 {
136         tdb_off_t off, prev_next, head;
137         enum TDB_ERROR ecode;
138
139         /* Is this only element in list?  Zero out bucket, and we're done. */
140         if (frec_prev(r) == r_off)
141                 return tdb_write_off(tdb, b_off, 0);
142
143         /* off = &r->prev->next */
144         off = frec_prev(r) + offsetof(struct tdb_free_record, next);
145
146         /* Get prev->next */
147         prev_next = tdb_read_off(tdb, off);
148         if (TDB_OFF_IS_ERR(prev_next))
149                 return prev_next;
150
151         /* If prev->next == 0, we were head: update bucket to point to next. */
152         if (prev_next == 0) {
153                 /* We must preserve upper bits. */
154                 head = tdb_read_off(tdb, b_off);
155                 if (TDB_OFF_IS_ERR(head))
156                         return head;
157
158                 if ((head & TDB_OFF_MASK) != r_off) {
159                         return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
160                                           "remove_from_list:"
161                                           " %llu head %llu on list %llu",
162                                           (long long)r_off,
163                                           (long long)head,
164                                           (long long)b_off);
165                 }
166                 head = ((head & ~TDB_OFF_MASK) | r->next);
167                 ecode = tdb_write_off(tdb, b_off, head);
168                 if (ecode != TDB_SUCCESS)
169                         return ecode;
170         } else {
171                 /* r->prev->next = r->next */
172                 ecode = tdb_write_off(tdb, off, r->next);
173                 if (ecode != TDB_SUCCESS)
174                         return ecode;
175         }
176
177         /* If we were the tail, off = &head->prev. */
178         if (r->next == 0) {
179                 head = tdb_read_off(tdb, b_off);
180                 if (TDB_OFF_IS_ERR(head))
181                         return head;
182                 head &= TDB_OFF_MASK;
183                 off = head + offsetof(struct tdb_free_record, magic_and_prev);
184         } else {
185                 /* off = &r->next->prev */
186                 off = r->next + offsetof(struct tdb_free_record,
187                                          magic_and_prev);
188         }
189
190 #ifdef CCAN_TDB2_DEBUG
191         /* *off == r */
192         if ((tdb_read_off(tdb, off) & TDB_OFF_MASK) != r_off) {
193                 return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
194                                   "remove_from_list:"
195                                   " %llu bad prev in list %llu",
196                                   (long long)r_off, (long long)b_off);
197         }
198 #endif
199         /* r->next->prev = r->prev */
200         return tdb_write_off(tdb, off, r->magic_and_prev);
201 }
202
203 /* Enqueue in this free bucket: sets coalesce if we've added 128
204  * entries to it. */
205 static enum TDB_ERROR enqueue_in_free(struct tdb_context *tdb,
206                                       tdb_off_t b_off,
207                                       tdb_off_t off,
208                                       tdb_len_t len,
209                                       bool *coalesce)
210 {
211         struct tdb_free_record new;
212         enum TDB_ERROR ecode;
213         tdb_off_t prev, head;
214         uint64_t magic = (TDB_FREE_MAGIC << (64 - TDB_OFF_UPPER_STEAL));
215
216         head = tdb_read_off(tdb, b_off);
217         if (TDB_OFF_IS_ERR(head))
218                 return head;
219
220         /* We only need to set ftable_and_len; rest is set in enqueue_in_free */
221         new.ftable_and_len = ((uint64_t)tdb->ftable << (64 - TDB_OFF_UPPER_STEAL))
222                 | len;
223
224         /* new->next = head. */
225         new.next = (head & TDB_OFF_MASK);
226
227         /* First element?  Prev points to ourselves. */
228         if (!new.next) {
229                 new.magic_and_prev = (magic | off);
230         } else {
231                 /* new->prev = next->prev */
232                 prev = tdb_read_off(tdb,
233                                     new.next + offsetof(struct tdb_free_record,
234                                                         magic_and_prev));
235                 new.magic_and_prev = prev;
236                 if (frec_magic(&new) != TDB_FREE_MAGIC) {
237                         return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
238                                           "enqueue_in_free: %llu bad head"
239                                           " prev %llu",
240                                           (long long)new.next,
241                                           (long long)prev);
242                 }
243                 /* next->prev = new. */
244                 ecode = tdb_write_off(tdb, new.next
245                                       + offsetof(struct tdb_free_record,
246                                                  magic_and_prev),
247                                       off | magic);
248                 if (ecode != TDB_SUCCESS) {
249                         return ecode;
250                 }
251
252 #ifdef CCAN_TDB2_DEBUG
253                 prev = tdb_read_off(tdb, frec_prev(&new)
254                                     + offsetof(struct tdb_free_record, next));
255                 if (prev != 0) {
256                         return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
257                                           "enqueue_in_free:"
258                                           " %llu bad tail next ptr %llu",
259                                           (long long)frec_prev(&new)
260                                           + offsetof(struct tdb_free_record,
261                                                      next),
262                                           (long long)prev);
263                 }
264 #endif
265         }
266
267         /* Update enqueue count, but don't set high bit: see TDB_OFF_IS_ERR */
268         if (*coalesce)
269                 head += (1ULL << (64 - TDB_OFF_UPPER_STEAL));
270         head &= ~(TDB_OFF_MASK | (1ULL << 63));
271         head |= off;
272
273         ecode = tdb_write_off(tdb, b_off, head);
274         if (ecode != TDB_SUCCESS) {
275                 return ecode;
276         }
277
278         /* It's time to coalesce if counter wrapped. */
279         if (*coalesce)
280                 *coalesce = ((head & ~TDB_OFF_MASK) == 0);
281
282         return tdb_write_convert(tdb, off, &new, sizeof(new));
283 }
284
285 static tdb_off_t ftable_offset(struct tdb_context *tdb, unsigned int ftable)
286 {
287         tdb_off_t off;
288         unsigned int i;
289
290         if (likely(tdb->ftable == ftable))
291                 return tdb->ftable_off;
292
293         off = first_ftable(tdb);
294         for (i = 0; i < ftable; i++) {
295                 if (TDB_OFF_IS_ERR(off)) {
296                         break;
297                 }
298                 off = next_ftable(tdb, off);
299         }
300         return off;
301 }
302
303 /* Note: we unlock the current bucket if fail (-ve), or coalesce (+ve) and
304  * need to blatt the *protect record (which is set to an error). */
305 static tdb_len_t coalesce(struct tdb_context *tdb,
306                           tdb_off_t off, tdb_off_t b_off,
307                           tdb_len_t data_len,
308                           tdb_off_t *protect)
309 {
310         tdb_off_t end;
311         struct tdb_free_record rec;
312         enum TDB_ERROR ecode;
313
314         tdb->stats.alloc_coalesce_tried++;
315         end = off + sizeof(struct tdb_used_record) + data_len;
316
317         while (end < tdb->file->map_size) {
318                 const struct tdb_free_record *r;
319                 tdb_off_t nb_off;
320                 unsigned ftable, bucket;
321
322                 r = tdb_access_read(tdb, end, sizeof(*r), true);
323                 if (TDB_PTR_IS_ERR(r)) {
324                         ecode = TDB_PTR_ERR(r);
325                         goto err;
326                 }
327
328                 if (frec_magic(r) != TDB_FREE_MAGIC
329                     || frec_ftable(r) == TDB_FTABLE_NONE) {
330                         tdb_access_release(tdb, r);
331                         break;
332                 }
333
334                 ftable = frec_ftable(r);
335                 bucket = size_to_bucket(frec_len(r));
336                 nb_off = ftable_offset(tdb, ftable);
337                 if (TDB_OFF_IS_ERR(nb_off)) {
338                         tdb_access_release(tdb, r);
339                         ecode = nb_off;
340                         goto err;
341                 }
342                 nb_off = bucket_off(nb_off, bucket);
343                 tdb_access_release(tdb, r);
344
345                 /* We may be violating lock order here, so best effort. */
346                 if (tdb_lock_free_bucket(tdb, nb_off, TDB_LOCK_NOWAIT)
347                     != TDB_SUCCESS) {
348                         tdb->stats.alloc_coalesce_lockfail++;
349                         break;
350                 }
351
352                 /* Now we have lock, re-check. */
353                 ecode = tdb_read_convert(tdb, end, &rec, sizeof(rec));
354                 if (ecode != TDB_SUCCESS) {
355                         tdb_unlock_free_bucket(tdb, nb_off);
356                         goto err;
357                 }
358
359                 if (unlikely(frec_magic(&rec) != TDB_FREE_MAGIC)) {
360                         tdb->stats.alloc_coalesce_race++;
361                         tdb_unlock_free_bucket(tdb, nb_off);
362                         break;
363                 }
364
365                 if (unlikely(frec_ftable(&rec) != ftable)
366                     || unlikely(size_to_bucket(frec_len(&rec)) != bucket)) {
367                         tdb->stats.alloc_coalesce_race++;
368                         tdb_unlock_free_bucket(tdb, nb_off);
369                         break;
370                 }
371
372                 /* Did we just mess up a record you were hoping to use? */
373                 if (end == *protect)
374                         *protect = TDB_ERR_NOEXIST;
375
376                 ecode = remove_from_list(tdb, nb_off, end, &rec);
377                 check_list(tdb, nb_off);
378                 if (ecode != TDB_SUCCESS) {
379                         tdb_unlock_free_bucket(tdb, nb_off);
380                         goto err;
381                 }
382
383                 end += sizeof(struct tdb_used_record) + frec_len(&rec);
384                 tdb_unlock_free_bucket(tdb, nb_off);
385                 tdb->stats.alloc_coalesce_num_merged++;
386         }
387
388         /* Didn't find any adjacent free? */
389         if (end == off + sizeof(struct tdb_used_record) + data_len)
390                 return 0;
391
392         /* Before we expand, check this isn't one you wanted protected? */
393         if (off == *protect)
394                 *protect = TDB_ERR_EXISTS;
395
396         /* OK, expand initial record */
397         ecode = tdb_read_convert(tdb, off, &rec, sizeof(rec));
398         if (ecode != TDB_SUCCESS) {
399                 goto err;
400         }
401
402         if (frec_len(&rec) != data_len) {
403                 ecode = tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
404                                    "coalesce: expected data len %zu not %zu",
405                                    (size_t)data_len, (size_t)frec_len(&rec));
406                 goto err;
407         }
408
409         ecode = remove_from_list(tdb, b_off, off, &rec);
410         check_list(tdb, b_off);
411         if (ecode != TDB_SUCCESS) {
412                 goto err;
413         }
414
415         /* Try locking violation first.  We don't allow coalesce recursion! */
416         ecode = add_free_record(tdb, off, end - off, TDB_LOCK_NOWAIT, false);
417         if (ecode != TDB_SUCCESS) {
418                 /* Need to drop lock.  Can't rely on anything stable. */
419                 *protect = TDB_ERR_CORRUPT;
420
421                 /* We have to drop this to avoid deadlocks, so make sure record
422                  * doesn't get coalesced by someone else! */
423                 rec.ftable_and_len = (TDB_FTABLE_NONE
424                                       << (64 - TDB_OFF_UPPER_STEAL))
425                         | (end - off - sizeof(struct tdb_used_record));
426                 ecode = tdb_write_off(tdb,
427                                       off + offsetof(struct tdb_free_record,
428                                                      ftable_and_len),
429                                       rec.ftable_and_len);
430                 if (ecode != TDB_SUCCESS) {
431                         goto err;
432                 }
433
434                 tdb_unlock_free_bucket(tdb, b_off);
435
436                 ecode = add_free_record(tdb, off, end - off, TDB_LOCK_WAIT,
437                                         false);
438                 if (ecode != TDB_SUCCESS) {
439                         return ecode;
440                 }
441         } else if (TDB_OFF_IS_ERR(*protect)) {
442                 /* For simplicity, we always drop lock if they can't continue */
443                 tdb_unlock_free_bucket(tdb, b_off);
444         }
445         tdb->stats.alloc_coalesce_succeeded++;
446
447         /* Return usable length. */
448         return end - off - sizeof(struct tdb_used_record);
449
450 err:
451         /* To unify error paths, we *always* unlock bucket on error. */
452         tdb_unlock_free_bucket(tdb, b_off);
453         return ecode;
454 }
455
456 /* List is locked: we unlock it. */
457 static enum TDB_ERROR coalesce_list(struct tdb_context *tdb,
458                                     tdb_off_t ftable_off,
459                                     tdb_off_t b_off,
460                                     unsigned int limit)
461 {
462         enum TDB_ERROR ecode;
463         tdb_off_t off;
464
465         off = tdb_read_off(tdb, b_off);
466         if (TDB_OFF_IS_ERR(off)) {
467                 ecode = off;
468                 goto unlock_err;
469         }
470         /* A little bit of paranoia: counter should be 0. */
471         off &= TDB_OFF_MASK;
472
473         while (off && limit--) {
474                 struct tdb_free_record rec;
475                 tdb_len_t coal;
476                 tdb_off_t next;
477
478                 ecode = tdb_read_convert(tdb, off, &rec, sizeof(rec));
479                 if (ecode != TDB_SUCCESS)
480                         goto unlock_err;
481
482                 next = rec.next;
483                 coal = coalesce(tdb, off, b_off, frec_len(&rec), &next);
484                 if (TDB_OFF_IS_ERR(coal)) {
485                         /* This has already unlocked on error. */
486                         return coal;
487                 }
488                 if (TDB_OFF_IS_ERR(next)) {
489                         /* Coalescing had to unlock, so stop. */
490                         return TDB_SUCCESS;
491                 }
492                 /* Keep going if we're doing well... */
493                 limit += size_to_bucket(coal / 16 + TDB_MIN_DATA_LEN);
494                 off = next;
495         }
496
497         /* Now, move those elements to the tail of the list so we get something
498          * else next time. */
499         if (off) {
500                 struct tdb_free_record oldhrec, newhrec, oldtrec, newtrec;
501                 tdb_off_t oldhoff, oldtoff, newtoff;
502
503                 /* The record we were up to is the new head. */
504                 ecode = tdb_read_convert(tdb, off, &newhrec, sizeof(newhrec));
505                 if (ecode != TDB_SUCCESS)
506                         goto unlock_err;
507
508                 /* Get the new tail. */
509                 newtoff = frec_prev(&newhrec);
510                 ecode = tdb_read_convert(tdb, newtoff, &newtrec,
511                                          sizeof(newtrec));
512                 if (ecode != TDB_SUCCESS)
513                         goto unlock_err;
514
515                 /* Get the old head. */
516                 oldhoff = tdb_read_off(tdb, b_off);
517                 if (TDB_OFF_IS_ERR(oldhoff)) {
518                         ecode = oldhoff;
519                         goto unlock_err;
520                 }
521
522                 /* This could happen if they all coalesced away. */
523                 if (oldhoff == off)
524                         goto out;
525
526                 ecode = tdb_read_convert(tdb, oldhoff, &oldhrec,
527                                          sizeof(oldhrec));
528                 if (ecode != TDB_SUCCESS)
529                         goto unlock_err;
530
531                 /* Get the old tail. */
532                 oldtoff = frec_prev(&oldhrec);
533                 ecode = tdb_read_convert(tdb, oldtoff, &oldtrec,
534                                          sizeof(oldtrec));
535                 if (ecode != TDB_SUCCESS)
536                         goto unlock_err;
537
538                 /* Old tail's next points to old head. */
539                 oldtrec.next = oldhoff;
540
541                 /* Old head's prev points to old tail. */
542                 oldhrec.magic_and_prev
543                         = (TDB_FREE_MAGIC << (64 - TDB_OFF_UPPER_STEAL))
544                         | oldtoff;
545
546                 /* New tail's next is 0. */
547                 newtrec.next = 0;
548
549                 /* Write out the modified versions. */
550                 ecode = tdb_write_convert(tdb, oldtoff, &oldtrec,
551                                           sizeof(oldtrec));
552                 if (ecode != TDB_SUCCESS)
553                         goto unlock_err;
554
555                 ecode = tdb_write_convert(tdb, oldhoff, &oldhrec,
556                                           sizeof(oldhrec));
557                 if (ecode != TDB_SUCCESS)
558                         goto unlock_err;
559
560                 ecode = tdb_write_convert(tdb, newtoff, &newtrec,
561                                           sizeof(newtrec));
562                 if (ecode != TDB_SUCCESS)
563                         goto unlock_err;
564                 
565                 /* And finally link in new head. */
566                 ecode = tdb_write_off(tdb, b_off, off);
567                 if (ecode != TDB_SUCCESS)
568                         goto unlock_err;
569         }
570 out:
571         tdb_unlock_free_bucket(tdb, b_off);
572         return TDB_SUCCESS;
573
574 unlock_err:
575         tdb_unlock_free_bucket(tdb, b_off);
576         return ecode;
577 }
578
579 /* List must not be locked if coalesce_ok is set. */
580 enum TDB_ERROR add_free_record(struct tdb_context *tdb,
581                                tdb_off_t off, tdb_len_t len_with_header,
582                                enum tdb_lock_flags waitflag,
583                                bool coalesce)
584 {
585         tdb_off_t b_off;
586         tdb_len_t len;
587         enum TDB_ERROR ecode;
588
589         assert(len_with_header >= sizeof(struct tdb_free_record));
590
591         len = len_with_header - sizeof(struct tdb_used_record);
592
593         b_off = bucket_off(tdb->ftable_off, size_to_bucket(len));
594         ecode = tdb_lock_free_bucket(tdb, b_off, waitflag);
595         if (ecode != TDB_SUCCESS) {
596                 return ecode;
597         }
598
599         ecode = enqueue_in_free(tdb, b_off, off, len, &coalesce);
600         check_list(tdb, b_off);
601
602         /* Coalescing unlocks free list. */
603         if (!ecode && coalesce)
604                 ecode = coalesce_list(tdb, tdb->ftable_off, b_off, 2);
605         else
606                 tdb_unlock_free_bucket(tdb, b_off);
607         return ecode;
608 }
609
610 static size_t adjust_size(size_t keylen, size_t datalen)
611 {
612         size_t size = keylen + datalen;
613
614         if (size < TDB_MIN_DATA_LEN)
615                 size = TDB_MIN_DATA_LEN;
616
617         /* Round to next uint64_t boundary. */
618         return (size + (sizeof(uint64_t) - 1ULL)) & ~(sizeof(uint64_t) - 1ULL);
619 }
620
621 /* If we have enough left over to be useful, split that off. */
622 static size_t record_leftover(size_t keylen, size_t datalen,
623                               bool want_extra, size_t total_len)
624 {
625         ssize_t leftover;
626
627         if (want_extra)
628                 datalen += datalen / 2;
629         leftover = total_len - adjust_size(keylen, datalen);
630
631         if (leftover < (ssize_t)sizeof(struct tdb_free_record))
632                 return 0;
633
634         return leftover;
635 }
636
637 /* We need size bytes to put our key and data in. */
638 static tdb_off_t lock_and_alloc(struct tdb_context *tdb,
639                                 tdb_off_t ftable_off,
640                                 tdb_off_t bucket,
641                                 size_t keylen, size_t datalen,
642                                 bool want_extra,
643                                 unsigned magic,
644                                 unsigned hashlow)
645 {
646         tdb_off_t off, b_off,best_off;
647         struct tdb_free_record best = { 0 };
648         double multiplier;
649         size_t size = adjust_size(keylen, datalen);
650         enum TDB_ERROR ecode;
651
652         tdb->stats.allocs++;
653         b_off = bucket_off(ftable_off, bucket);
654
655         /* FIXME: Try non-blocking wait first, to measure contention. */
656         /* Lock this bucket. */
657         ecode = tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT);
658         if (ecode != TDB_SUCCESS) {
659                 return ecode;
660         }
661
662         best.ftable_and_len = -1ULL;
663         best_off = 0;
664
665         /* Get slack if we're after extra. */
666         if (want_extra)
667                 multiplier = 1.5;
668         else
669                 multiplier = 1.0;
670
671         /* Walk the list to see if any are large enough, getting less fussy
672          * as we go. */
673         off = tdb_read_off(tdb, b_off);
674         if (TDB_OFF_IS_ERR(off)) {
675                 ecode = off;
676                 goto unlock_err;
677         }
678         off &= TDB_OFF_MASK;
679
680         while (off) {
681                 const struct tdb_free_record *r;
682                 tdb_len_t len;
683                 tdb_off_t next;
684
685                 r = tdb_access_read(tdb, off, sizeof(*r), true);
686                 if (TDB_PTR_IS_ERR(r)) {
687                         ecode = TDB_PTR_ERR(r);
688                         goto unlock_err;
689                 }
690
691                 if (frec_magic(r) != TDB_FREE_MAGIC) {
692                         ecode = tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
693                                            "lock_and_alloc:"
694                                            " %llu non-free 0x%llx",
695                                            (long long)off,
696                                            (long long)r->magic_and_prev);
697                         tdb_access_release(tdb, r);
698                         goto unlock_err;
699                 }
700
701                 if (frec_len(r) >= size && frec_len(r) < frec_len(&best)) {
702                         best_off = off;
703                         best = *r;
704                 }
705
706                 if (frec_len(&best) <= size * multiplier && best_off) {
707                         tdb_access_release(tdb, r);
708                         break;
709                 }
710
711                 multiplier *= 1.01;
712
713                 next = r->next;
714                 len = frec_len(r);
715                 tdb_access_release(tdb, r);
716                 off = next;
717         }
718
719         /* If we found anything at all, use it. */
720         if (best_off) {
721                 struct tdb_used_record rec;
722                 size_t leftover;
723
724                 /* We're happy with this size: take it. */
725                 ecode = remove_from_list(tdb, b_off, best_off, &best);
726                 check_list(tdb, b_off);
727                 if (ecode != TDB_SUCCESS) {
728                         goto unlock_err;
729                 }
730
731                 leftover = record_leftover(keylen, datalen, want_extra,
732                                            frec_len(&best));
733
734                 assert(keylen + datalen + leftover <= frec_len(&best));
735                 /* We need to mark non-free before we drop lock, otherwise
736                  * coalesce() could try to merge it! */
737                 ecode = set_header(tdb, &rec, magic, keylen, datalen,
738                                    frec_len(&best) - leftover, hashlow);
739                 if (ecode != TDB_SUCCESS) {
740                         goto unlock_err;
741                 }
742
743                 ecode = tdb_write_convert(tdb, best_off, &rec, sizeof(rec));
744                 if (ecode != TDB_SUCCESS) {
745                         goto unlock_err;
746                 }
747
748                 /* For futureproofing, we put a 0 in any unused space. */
749                 if (rec_extra_padding(&rec)) {
750                         ecode = tdb->methods->twrite(tdb, best_off + sizeof(rec)
751                                                      + keylen + datalen, "", 1);
752                         if (ecode != TDB_SUCCESS) {
753                                 goto unlock_err;
754                         }
755                 }
756
757                 /* Bucket of leftover will be <= current bucket, so nested
758                  * locking is allowed. */
759                 if (leftover) {
760                         tdb->stats.alloc_leftover++;
761                         ecode = add_free_record(tdb,
762                                                 best_off + sizeof(rec)
763                                                 + frec_len(&best) - leftover,
764                                                 leftover, TDB_LOCK_WAIT, false);
765                         if (ecode != TDB_SUCCESS) {
766                                 best_off = ecode;
767                         }
768                 }
769                 tdb_unlock_free_bucket(tdb, b_off);
770
771                 return best_off;
772         }
773
774         tdb_unlock_free_bucket(tdb, b_off);
775         return 0;
776
777 unlock_err:
778         tdb_unlock_free_bucket(tdb, b_off);
779         return ecode;
780 }
781
782 /* Get a free block from current free list, or 0 if none, -ve on error. */
783 static tdb_off_t get_free(struct tdb_context *tdb,
784                           size_t keylen, size_t datalen, bool want_extra,
785                           unsigned magic, unsigned hashlow)
786 {
787         tdb_off_t off, ftable_off;
788         tdb_off_t start_b, b, ftable;
789         bool wrapped = false;
790
791         /* If they are growing, add 50% to get to higher bucket. */
792         if (want_extra)
793                 start_b = size_to_bucket(adjust_size(keylen,
794                                                      datalen + datalen / 2));
795         else
796                 start_b = size_to_bucket(adjust_size(keylen, datalen));
797
798         ftable_off = tdb->ftable_off;
799         ftable = tdb->ftable;
800         while (!wrapped || ftable_off != tdb->ftable_off) {
801                 /* Start at exact size bucket, and search up... */
802                 for (b = find_free_head(tdb, ftable_off, start_b);
803                      b < TDB_FREE_BUCKETS;
804                      b = find_free_head(tdb, ftable_off, b + 1)) {
805                         /* Try getting one from list. */
806                         off = lock_and_alloc(tdb, ftable_off,
807                                              b, keylen, datalen, want_extra,
808                                              magic, hashlow);
809                         if (TDB_OFF_IS_ERR(off))
810                                 return off;
811                         if (off != 0) {
812                                 if (b == start_b)
813                                         tdb->stats.alloc_bucket_exact++;
814                                 if (b == TDB_FREE_BUCKETS - 1)
815                                         tdb->stats.alloc_bucket_max++;
816                                 /* Worked?  Stay using this list. */
817                                 tdb->ftable_off = ftable_off;
818                                 tdb->ftable = ftable;
819                                 return off;
820                         }
821                         /* Didn't work.  Try next bucket. */
822                 }
823
824                 if (TDB_OFF_IS_ERR(b)) {
825                         return b;
826                 }
827
828                 /* Hmm, try next table. */
829                 ftable_off = next_ftable(tdb, ftable_off);
830                 if (TDB_OFF_IS_ERR(ftable_off)) {
831                         return ftable_off;
832                 }
833                 ftable++;
834
835                 if (ftable_off == 0) {
836                         wrapped = true;
837                         ftable_off = first_ftable(tdb);
838                         if (TDB_OFF_IS_ERR(ftable_off)) {
839                                 return ftable_off;
840                         }
841                         ftable = 0;
842                 }
843         }
844
845         return 0;
846 }
847
848 enum TDB_ERROR set_header(struct tdb_context *tdb,
849                           struct tdb_used_record *rec,
850                           unsigned magic, uint64_t keylen, uint64_t datalen,
851                           uint64_t actuallen, unsigned hashlow)
852 {
853         uint64_t keybits = (fls64(keylen) + 1) / 2;
854
855         /* Use bottom bits of hash, so it's independent of hash table size. */
856         rec->magic_and_meta = (hashlow & ((1 << 11)-1))
857                 | ((actuallen - (keylen + datalen)) << 11)
858                 | (keybits << 43)
859                 | ((uint64_t)magic << 48);
860         rec->key_and_data_len = (keylen | (datalen << (keybits*2)));
861
862         /* Encoding can fail on big values. */
863         if (rec_key_length(rec) != keylen
864             || rec_data_length(rec) != datalen
865             || rec_extra_padding(rec) != actuallen - (keylen + datalen)) {
866                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
867                                   "Could not encode k=%llu,d=%llu,a=%llu",
868                                   (long long)keylen, (long long)datalen,
869                                   (long long)actuallen);
870         }
871         return TDB_SUCCESS;
872 }
873
874 /* Expand the database. */
875 static enum TDB_ERROR tdb_expand(struct tdb_context *tdb, tdb_len_t size)
876 {
877         uint64_t old_size, rec_size, map_size;
878         tdb_len_t wanted;
879         enum TDB_ERROR ecode;
880
881         /* Need to hold a hash lock to expand DB: transactions rely on it. */
882         if (!(tdb->flags & TDB_NOLOCK)
883             && !tdb->file->allrecord_lock.count && !tdb_has_hash_locks(tdb)) {
884                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
885                                   "tdb_expand: must hold lock during expand");
886         }
887
888         /* Only one person can expand file at a time. */
889         ecode = tdb_lock_expand(tdb, F_WRLCK);
890         if (ecode != TDB_SUCCESS) {
891                 return ecode;
892         }
893
894         /* Someone else may have expanded the file, so retry. */
895         old_size = tdb->file->map_size;
896         tdb->methods->oob(tdb, tdb->file->map_size + 1, true);
897         if (tdb->file->map_size != old_size) {
898                 tdb_unlock_expand(tdb, F_WRLCK);
899                 return TDB_SUCCESS;
900         }
901
902         /* limit size in order to avoid using up huge amounts of memory for
903          * in memory tdbs if an oddball huge record creeps in */
904         if (size > 100 * 1024) {
905                 rec_size = size * 2;
906         } else {
907                 rec_size = size * 100;
908         }
909
910         /* always make room for at least rec_size more records, and at
911            least 25% more space. if the DB is smaller than 100MiB,
912            otherwise grow it by 10% only. */
913         if (old_size > 100 * 1024 * 1024) {
914                 map_size = old_size / 10;
915         } else {
916                 map_size = old_size / 4;
917         }
918
919         if (map_size > rec_size) {
920                 wanted = map_size;
921         } else {
922                 wanted = rec_size;
923         }
924
925         /* We need room for the record header too. */
926         wanted = adjust_size(0, sizeof(struct tdb_used_record) + wanted);
927
928         ecode = tdb->methods->expand_file(tdb, wanted);
929         if (ecode != TDB_SUCCESS) {
930                 tdb_unlock_expand(tdb, F_WRLCK);
931                 return ecode;
932         }
933
934         /* We need to drop this lock before adding free record. */
935         tdb_unlock_expand(tdb, F_WRLCK);
936
937         tdb->stats.expands++;
938         return add_free_record(tdb, old_size, wanted, TDB_LOCK_WAIT, true);
939 }
940
941 /* This won't fail: it will expand the database if it has to. */
942 tdb_off_t alloc(struct tdb_context *tdb, size_t keylen, size_t datalen,
943                 uint64_t hash, unsigned magic, bool growing)
944 {
945         tdb_off_t off;
946
947         /* We can't hold pointers during this: we could unmap! */
948         assert(!tdb->direct_access);
949
950         for (;;) {
951                 enum TDB_ERROR ecode;
952                 off = get_free(tdb, keylen, datalen, growing, magic, hash);
953                 if (likely(off != 0))
954                         break;
955
956                 ecode = tdb_expand(tdb, adjust_size(keylen, datalen));
957                 if (ecode != TDB_SUCCESS) {
958                         return ecode;
959                 }
960         }
961
962         return off;
963 }