tdb2: make tests work in parallel.
[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->tdb2.ftable_off = off = first_ftable(tdb);
69         tdb->tdb2.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->tdb2.ftable_off = off;
79                         tdb->tdb2.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->tdb2.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->tdb2.ftable == ftable))
291                 return tdb->tdb2.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                         tdb->stats.alloc_coalesce_iterate_clash++;
375                         *protect = TDB_ERR_NOEXIST;
376                 }
377
378                 ecode = remove_from_list(tdb, nb_off, end, &rec);
379                 check_list(tdb, nb_off);
380                 if (ecode != TDB_SUCCESS) {
381                         tdb_unlock_free_bucket(tdb, nb_off);
382                         goto err;
383                 }
384
385                 end += sizeof(struct tdb_used_record) + frec_len(&rec);
386                 tdb_unlock_free_bucket(tdb, nb_off);
387                 tdb->stats.alloc_coalesce_num_merged++;
388         }
389
390         /* Didn't find any adjacent free? */
391         if (end == off + sizeof(struct tdb_used_record) + data_len)
392                 return 0;
393
394         /* Before we expand, check this isn't one you wanted protected? */
395         if (off == *protect) {
396                 *protect = TDB_ERR_EXISTS;
397                 tdb->stats.alloc_coalesce_iterate_clash++;
398         }
399
400         /* OK, expand initial record */
401         ecode = tdb_read_convert(tdb, off, &rec, sizeof(rec));
402         if (ecode != TDB_SUCCESS) {
403                 goto err;
404         }
405
406         if (frec_len(&rec) != data_len) {
407                 ecode = tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
408                                    "coalesce: expected data len %zu not %zu",
409                                    (size_t)data_len, (size_t)frec_len(&rec));
410                 goto err;
411         }
412
413         ecode = remove_from_list(tdb, b_off, off, &rec);
414         check_list(tdb, b_off);
415         if (ecode != TDB_SUCCESS) {
416                 goto err;
417         }
418
419         /* Try locking violation first.  We don't allow coalesce recursion! */
420         ecode = add_free_record(tdb, off, end - off, TDB_LOCK_NOWAIT, false);
421         if (ecode != TDB_SUCCESS) {
422                 /* Need to drop lock.  Can't rely on anything stable. */
423                 tdb->stats.alloc_coalesce_lockfail++;
424                 *protect = TDB_ERR_CORRUPT;
425
426                 /* We have to drop this to avoid deadlocks, so make sure record
427                  * doesn't get coalesced by someone else! */
428                 rec.ftable_and_len = (TDB_FTABLE_NONE
429                                       << (64 - TDB_OFF_UPPER_STEAL))
430                         | (end - off - sizeof(struct tdb_used_record));
431                 ecode = tdb_write_off(tdb,
432                                       off + offsetof(struct tdb_free_record,
433                                                      ftable_and_len),
434                                       rec.ftable_and_len);
435                 if (ecode != TDB_SUCCESS) {
436                         goto err;
437                 }
438
439                 tdb_unlock_free_bucket(tdb, b_off);
440
441                 ecode = add_free_record(tdb, off, end - off, TDB_LOCK_WAIT,
442                                         false);
443                 if (ecode != TDB_SUCCESS) {
444                         return ecode;
445                 }
446         } else if (TDB_OFF_IS_ERR(*protect)) {
447                 /* For simplicity, we always drop lock if they can't continue */
448                 tdb_unlock_free_bucket(tdb, b_off);
449         }
450         tdb->stats.alloc_coalesce_succeeded++;
451
452         /* Return usable length. */
453         return end - off - sizeof(struct tdb_used_record);
454
455 err:
456         /* To unify error paths, we *always* unlock bucket on error. */
457         tdb_unlock_free_bucket(tdb, b_off);
458         return ecode;
459 }
460
461 /* List is locked: we unlock it. */
462 static enum TDB_ERROR coalesce_list(struct tdb_context *tdb,
463                                     tdb_off_t ftable_off,
464                                     tdb_off_t b_off,
465                                     unsigned int limit)
466 {
467         enum TDB_ERROR ecode;
468         tdb_off_t off;
469
470         off = tdb_read_off(tdb, b_off);
471         if (TDB_OFF_IS_ERR(off)) {
472                 ecode = off;
473                 goto unlock_err;
474         }
475         /* A little bit of paranoia: counter should be 0. */
476         off &= TDB_OFF_MASK;
477
478         while (off && limit--) {
479                 struct tdb_free_record rec;
480                 tdb_len_t coal;
481                 tdb_off_t next;
482
483                 ecode = tdb_read_convert(tdb, off, &rec, sizeof(rec));
484                 if (ecode != TDB_SUCCESS)
485                         goto unlock_err;
486
487                 next = rec.next;
488                 coal = coalesce(tdb, off, b_off, frec_len(&rec), &next);
489                 if (TDB_OFF_IS_ERR(coal)) {
490                         /* This has already unlocked on error. */
491                         return coal;
492                 }
493                 if (TDB_OFF_IS_ERR(next)) {
494                         /* Coalescing had to unlock, so stop. */
495                         return TDB_SUCCESS;
496                 }
497                 /* Keep going if we're doing well... */
498                 limit += size_to_bucket(coal / 16 + TDB_MIN_DATA_LEN);
499                 off = next;
500         }
501
502         /* Now, move those elements to the tail of the list so we get something
503          * else next time. */
504         if (off) {
505                 struct tdb_free_record oldhrec, newhrec, oldtrec, newtrec;
506                 tdb_off_t oldhoff, oldtoff, newtoff;
507
508                 /* The record we were up to is the new head. */
509                 ecode = tdb_read_convert(tdb, off, &newhrec, sizeof(newhrec));
510                 if (ecode != TDB_SUCCESS)
511                         goto unlock_err;
512
513                 /* Get the new tail. */
514                 newtoff = frec_prev(&newhrec);
515                 ecode = tdb_read_convert(tdb, newtoff, &newtrec,
516                                          sizeof(newtrec));
517                 if (ecode != TDB_SUCCESS)
518                         goto unlock_err;
519
520                 /* Get the old head. */
521                 oldhoff = tdb_read_off(tdb, b_off);
522                 if (TDB_OFF_IS_ERR(oldhoff)) {
523                         ecode = oldhoff;
524                         goto unlock_err;
525                 }
526
527                 /* This could happen if they all coalesced away. */
528                 if (oldhoff == off)
529                         goto out;
530
531                 ecode = tdb_read_convert(tdb, oldhoff, &oldhrec,
532                                          sizeof(oldhrec));
533                 if (ecode != TDB_SUCCESS)
534                         goto unlock_err;
535
536                 /* Get the old tail. */
537                 oldtoff = frec_prev(&oldhrec);
538                 ecode = tdb_read_convert(tdb, oldtoff, &oldtrec,
539                                          sizeof(oldtrec));
540                 if (ecode != TDB_SUCCESS)
541                         goto unlock_err;
542
543                 /* Old tail's next points to old head. */
544                 oldtrec.next = oldhoff;
545
546                 /* Old head's prev points to old tail. */
547                 oldhrec.magic_and_prev
548                         = (TDB_FREE_MAGIC << (64 - TDB_OFF_UPPER_STEAL))
549                         | oldtoff;
550
551                 /* New tail's next is 0. */
552                 newtrec.next = 0;
553
554                 /* Write out the modified versions. */
555                 ecode = tdb_write_convert(tdb, oldtoff, &oldtrec,
556                                           sizeof(oldtrec));
557                 if (ecode != TDB_SUCCESS)
558                         goto unlock_err;
559
560                 ecode = tdb_write_convert(tdb, oldhoff, &oldhrec,
561                                           sizeof(oldhrec));
562                 if (ecode != TDB_SUCCESS)
563                         goto unlock_err;
564
565                 ecode = tdb_write_convert(tdb, newtoff, &newtrec,
566                                           sizeof(newtrec));
567                 if (ecode != TDB_SUCCESS)
568                         goto unlock_err;
569                 
570                 /* And finally link in new head. */
571                 ecode = tdb_write_off(tdb, b_off, off);
572                 if (ecode != TDB_SUCCESS)
573                         goto unlock_err;
574         }
575 out:
576         tdb_unlock_free_bucket(tdb, b_off);
577         return TDB_SUCCESS;
578
579 unlock_err:
580         tdb_unlock_free_bucket(tdb, b_off);
581         return ecode;
582 }
583
584 /* List must not be locked if coalesce_ok is set. */
585 enum TDB_ERROR add_free_record(struct tdb_context *tdb,
586                                tdb_off_t off, tdb_len_t len_with_header,
587                                enum tdb_lock_flags waitflag,
588                                bool coalesce)
589 {
590         tdb_off_t b_off;
591         tdb_len_t len;
592         enum TDB_ERROR ecode;
593
594         assert(len_with_header >= sizeof(struct tdb_free_record));
595
596         len = len_with_header - sizeof(struct tdb_used_record);
597
598         b_off = bucket_off(tdb->tdb2.ftable_off, size_to_bucket(len));
599         ecode = tdb_lock_free_bucket(tdb, b_off, waitflag);
600         if (ecode != TDB_SUCCESS) {
601                 return ecode;
602         }
603
604         ecode = enqueue_in_free(tdb, b_off, off, len, &coalesce);
605         check_list(tdb, b_off);
606
607         /* Coalescing unlocks free list. */
608         if (!ecode && coalesce)
609                 ecode = coalesce_list(tdb, tdb->tdb2.ftable_off, b_off, 2);
610         else
611                 tdb_unlock_free_bucket(tdb, b_off);
612         return ecode;
613 }
614
615 static size_t adjust_size(size_t keylen, size_t datalen)
616 {
617         size_t size = keylen + datalen;
618
619         if (size < TDB_MIN_DATA_LEN)
620                 size = TDB_MIN_DATA_LEN;
621
622         /* Round to next uint64_t boundary. */
623         return (size + (sizeof(uint64_t) - 1ULL)) & ~(sizeof(uint64_t) - 1ULL);
624 }
625
626 /* If we have enough left over to be useful, split that off. */
627 static size_t record_leftover(size_t keylen, size_t datalen,
628                               bool want_extra, size_t total_len)
629 {
630         ssize_t leftover;
631
632         if (want_extra)
633                 datalen += datalen / 2;
634         leftover = total_len - adjust_size(keylen, datalen);
635
636         if (leftover < (ssize_t)sizeof(struct tdb_free_record))
637                 return 0;
638
639         return leftover;
640 }
641
642 /* We need size bytes to put our key and data in. */
643 static tdb_off_t lock_and_alloc(struct tdb_context *tdb,
644                                 tdb_off_t ftable_off,
645                                 tdb_off_t bucket,
646                                 size_t keylen, size_t datalen,
647                                 bool want_extra,
648                                 unsigned magic,
649                                 unsigned hashlow)
650 {
651         tdb_off_t off, b_off,best_off;
652         struct tdb_free_record best = { 0 };
653         double multiplier;
654         size_t size = adjust_size(keylen, datalen);
655         enum TDB_ERROR ecode;
656
657         tdb->stats.allocs++;
658         b_off = bucket_off(ftable_off, bucket);
659
660         /* FIXME: Try non-blocking wait first, to measure contention. */
661         /* Lock this bucket. */
662         ecode = tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT);
663         if (ecode != TDB_SUCCESS) {
664                 return ecode;
665         }
666
667         best.ftable_and_len = -1ULL;
668         best_off = 0;
669
670         /* Get slack if we're after extra. */
671         if (want_extra)
672                 multiplier = 1.5;
673         else
674                 multiplier = 1.0;
675
676         /* Walk the list to see if any are large enough, getting less fussy
677          * as we go. */
678         off = tdb_read_off(tdb, b_off);
679         if (TDB_OFF_IS_ERR(off)) {
680                 ecode = off;
681                 goto unlock_err;
682         }
683         off &= TDB_OFF_MASK;
684
685         while (off) {
686                 const struct tdb_free_record *r;
687                 tdb_len_t len;
688                 tdb_off_t next;
689
690                 r = tdb_access_read(tdb, off, sizeof(*r), true);
691                 if (TDB_PTR_IS_ERR(r)) {
692                         ecode = TDB_PTR_ERR(r);
693                         goto unlock_err;
694                 }
695
696                 if (frec_magic(r) != TDB_FREE_MAGIC) {
697                         ecode = tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
698                                            "lock_and_alloc:"
699                                            " %llu non-free 0x%llx",
700                                            (long long)off,
701                                            (long long)r->magic_and_prev);
702                         tdb_access_release(tdb, r);
703                         goto unlock_err;
704                 }
705
706                 if (frec_len(r) >= size && frec_len(r) < frec_len(&best)) {
707                         best_off = off;
708                         best = *r;
709                 }
710
711                 if (frec_len(&best) <= size * multiplier && best_off) {
712                         tdb_access_release(tdb, r);
713                         break;
714                 }
715
716                 multiplier *= 1.01;
717
718                 next = r->next;
719                 len = frec_len(r);
720                 tdb_access_release(tdb, r);
721                 off = next;
722         }
723
724         /* If we found anything at all, use it. */
725         if (best_off) {
726                 struct tdb_used_record rec;
727                 size_t leftover;
728
729                 /* We're happy with this size: take it. */
730                 ecode = remove_from_list(tdb, b_off, best_off, &best);
731                 check_list(tdb, b_off);
732                 if (ecode != TDB_SUCCESS) {
733                         goto unlock_err;
734                 }
735
736                 leftover = record_leftover(keylen, datalen, want_extra,
737                                            frec_len(&best));
738
739                 assert(keylen + datalen + leftover <= frec_len(&best));
740                 /* We need to mark non-free before we drop lock, otherwise
741                  * coalesce() could try to merge it! */
742                 ecode = set_header(tdb, &rec, magic, keylen, datalen,
743                                    frec_len(&best) - leftover, hashlow);
744                 if (ecode != TDB_SUCCESS) {
745                         goto unlock_err;
746                 }
747
748                 ecode = tdb_write_convert(tdb, best_off, &rec, sizeof(rec));
749                 if (ecode != TDB_SUCCESS) {
750                         goto unlock_err;
751                 }
752
753                 /* For futureproofing, we put a 0 in any unused space. */
754                 if (rec_extra_padding(&rec)) {
755                         ecode = tdb->tdb2.io->twrite(tdb, best_off + sizeof(rec)
756                                                      + keylen + datalen, "", 1);
757                         if (ecode != TDB_SUCCESS) {
758                                 goto unlock_err;
759                         }
760                 }
761
762                 /* Bucket of leftover will be <= current bucket, so nested
763                  * locking is allowed. */
764                 if (leftover) {
765                         tdb->stats.alloc_leftover++;
766                         ecode = add_free_record(tdb,
767                                                 best_off + sizeof(rec)
768                                                 + frec_len(&best) - leftover,
769                                                 leftover, TDB_LOCK_WAIT, false);
770                         if (ecode != TDB_SUCCESS) {
771                                 best_off = ecode;
772                         }
773                 }
774                 tdb_unlock_free_bucket(tdb, b_off);
775
776                 return best_off;
777         }
778
779         tdb_unlock_free_bucket(tdb, b_off);
780         return 0;
781
782 unlock_err:
783         tdb_unlock_free_bucket(tdb, b_off);
784         return ecode;
785 }
786
787 /* Get a free block from current free list, or 0 if none, -ve on error. */
788 static tdb_off_t get_free(struct tdb_context *tdb,
789                           size_t keylen, size_t datalen, bool want_extra,
790                           unsigned magic, unsigned hashlow)
791 {
792         tdb_off_t off, ftable_off;
793         tdb_off_t start_b, b, ftable;
794         bool wrapped = false;
795
796         /* If they are growing, add 50% to get to higher bucket. */
797         if (want_extra)
798                 start_b = size_to_bucket(adjust_size(keylen,
799                                                      datalen + datalen / 2));
800         else
801                 start_b = size_to_bucket(adjust_size(keylen, datalen));
802
803         ftable_off = tdb->tdb2.ftable_off;
804         ftable = tdb->tdb2.ftable;
805         while (!wrapped || ftable_off != tdb->tdb2.ftable_off) {
806                 /* Start at exact size bucket, and search up... */
807                 for (b = find_free_head(tdb, ftable_off, start_b);
808                      b < TDB_FREE_BUCKETS;
809                      b = find_free_head(tdb, ftable_off, b + 1)) {
810                         /* Try getting one from list. */
811                         off = lock_and_alloc(tdb, ftable_off,
812                                              b, keylen, datalen, want_extra,
813                                              magic, hashlow);
814                         if (TDB_OFF_IS_ERR(off))
815                                 return off;
816                         if (off != 0) {
817                                 if (b == start_b)
818                                         tdb->stats.alloc_bucket_exact++;
819                                 if (b == TDB_FREE_BUCKETS - 1)
820                                         tdb->stats.alloc_bucket_max++;
821                                 /* Worked?  Stay using this list. */
822                                 tdb->tdb2.ftable_off = ftable_off;
823                                 tdb->tdb2.ftable = ftable;
824                                 return off;
825                         }
826                         /* Didn't work.  Try next bucket. */
827                 }
828
829                 if (TDB_OFF_IS_ERR(b)) {
830                         return b;
831                 }
832
833                 /* Hmm, try next table. */
834                 ftable_off = next_ftable(tdb, ftable_off);
835                 if (TDB_OFF_IS_ERR(ftable_off)) {
836                         return ftable_off;
837                 }
838                 ftable++;
839
840                 if (ftable_off == 0) {
841                         wrapped = true;
842                         ftable_off = first_ftable(tdb);
843                         if (TDB_OFF_IS_ERR(ftable_off)) {
844                                 return ftable_off;
845                         }
846                         ftable = 0;
847                 }
848         }
849
850         return 0;
851 }
852
853 enum TDB_ERROR set_header(struct tdb_context *tdb,
854                           struct tdb_used_record *rec,
855                           unsigned magic, uint64_t keylen, uint64_t datalen,
856                           uint64_t actuallen, unsigned hashlow)
857 {
858         uint64_t keybits = (fls64(keylen) + 1) / 2;
859
860         /* Use bottom bits of hash, so it's independent of hash table size. */
861         rec->magic_and_meta = (hashlow & ((1 << 11)-1))
862                 | ((actuallen - (keylen + datalen)) << 11)
863                 | (keybits << 43)
864                 | ((uint64_t)magic << 48);
865         rec->key_and_data_len = (keylen | (datalen << (keybits*2)));
866
867         /* Encoding can fail on big values. */
868         if (rec_key_length(rec) != keylen
869             || rec_data_length(rec) != datalen
870             || rec_extra_padding(rec) != actuallen - (keylen + datalen)) {
871                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
872                                   "Could not encode k=%llu,d=%llu,a=%llu",
873                                   (long long)keylen, (long long)datalen,
874                                   (long long)actuallen);
875         }
876         return TDB_SUCCESS;
877 }
878
879 /* Expand the database. */
880 static enum TDB_ERROR tdb_expand(struct tdb_context *tdb, tdb_len_t size)
881 {
882         uint64_t old_size, rec_size, map_size;
883         tdb_len_t wanted;
884         enum TDB_ERROR ecode;
885
886         /* Need to hold a hash lock to expand DB: transactions rely on it. */
887         if (!(tdb->flags & TDB_NOLOCK)
888             && !tdb->file->allrecord_lock.count && !tdb_has_hash_locks(tdb)) {
889                 return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
890                                   "tdb_expand: must hold lock during expand");
891         }
892
893         /* Only one person can expand file at a time. */
894         ecode = tdb_lock_expand(tdb, F_WRLCK);
895         if (ecode != TDB_SUCCESS) {
896                 return ecode;
897         }
898
899         /* Someone else may have expanded the file, so retry. */
900         old_size = tdb->file->map_size;
901         tdb->tdb2.io->oob(tdb, tdb->file->map_size + 1, true);
902         if (tdb->file->map_size != old_size) {
903                 tdb_unlock_expand(tdb, F_WRLCK);
904                 return TDB_SUCCESS;
905         }
906
907         /* limit size in order to avoid using up huge amounts of memory for
908          * in memory tdbs if an oddball huge record creeps in */
909         if (size > 100 * 1024) {
910                 rec_size = size * 2;
911         } else {
912                 rec_size = size * 100;
913         }
914
915         /* always make room for at least rec_size more records, and at
916            least 25% more space. if the DB is smaller than 100MiB,
917            otherwise grow it by 10% only. */
918         if (old_size > 100 * 1024 * 1024) {
919                 map_size = old_size / 10;
920         } else {
921                 map_size = old_size / 4;
922         }
923
924         if (map_size > rec_size) {
925                 wanted = map_size;
926         } else {
927                 wanted = rec_size;
928         }
929
930         /* We need room for the record header too. */
931         wanted = adjust_size(0, sizeof(struct tdb_used_record) + wanted);
932
933         ecode = tdb->tdb2.io->expand_file(tdb, wanted);
934         if (ecode != TDB_SUCCESS) {
935                 tdb_unlock_expand(tdb, F_WRLCK);
936                 return ecode;
937         }
938
939         /* We need to drop this lock before adding free record. */
940         tdb_unlock_expand(tdb, F_WRLCK);
941
942         tdb->stats.expands++;
943         return add_free_record(tdb, old_size, wanted, TDB_LOCK_WAIT, true);
944 }
945
946 /* This won't fail: it will expand the database if it has to. */
947 tdb_off_t alloc(struct tdb_context *tdb, size_t keylen, size_t datalen,
948                 uint64_t hash, unsigned magic, bool growing)
949 {
950         tdb_off_t off;
951
952         /* We can't hold pointers during this: we could unmap! */
953         assert(!tdb->tdb2.direct_access);
954
955         for (;;) {
956                 enum TDB_ERROR ecode;
957                 off = get_free(tdb, keylen, datalen, growing, magic, hash);
958                 if (likely(off != 0))
959                         break;
960
961                 ecode = tdb_expand(tdb, adjust_size(keylen, datalen));
962                 if (ecode != TDB_SUCCESS) {
963                         return ecode;
964                 }
965         }
966
967         return off;
968 }