ccan/ntdb: demote to junkcode.
[ccan] / junkcode / rusty@rustcorp.com.au-ntdb / 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 <limits.h>
23
24 static unsigned fls64(uint64_t val)
25 {
26         return ilog64(val);
27 }
28
29 /* In which bucket would we find a particular record size? (ignoring header) */
30 unsigned int size_to_bucket(ntdb_len_t data_len)
31 {
32         unsigned int bucket;
33
34         /* We can't have records smaller than this. */
35         assert(data_len >= NTDB_MIN_DATA_LEN);
36
37         /* Ignoring the header... */
38         if (data_len - NTDB_MIN_DATA_LEN <= 64) {
39                 /* 0 in bucket 0, 8 in bucket 1... 64 in bucket 8. */
40                 bucket = (data_len - NTDB_MIN_DATA_LEN) / 8;
41         } else {
42                 /* After that we go power of 2. */
43                 bucket = fls64(data_len - NTDB_MIN_DATA_LEN) + 2;
44         }
45
46         if (unlikely(bucket >= NTDB_FREE_BUCKETS))
47                 bucket = NTDB_FREE_BUCKETS - 1;
48         return bucket;
49 }
50
51 ntdb_off_t first_ftable(struct ntdb_context *ntdb)
52 {
53         return ntdb_read_off(ntdb, offsetof(struct ntdb_header, free_table));
54 }
55
56 ntdb_off_t next_ftable(struct ntdb_context *ntdb, ntdb_off_t ftable)
57 {
58         return ntdb_read_off(ntdb, ftable + offsetof(struct ntdb_freetable,next));
59 }
60
61 enum NTDB_ERROR ntdb_ftable_init(struct ntdb_context *ntdb)
62 {
63         /* Use reservoir sampling algorithm to select a free list at random. */
64         unsigned int rnd, max = 0, count = 0;
65         ntdb_off_t off;
66
67         ntdb->ftable_off = off = first_ftable(ntdb);
68         ntdb->ftable = 0;
69
70         while (off) {
71                 if (NTDB_OFF_IS_ERR(off)) {
72                         return NTDB_OFF_TO_ERR(off);
73                 }
74
75                 rnd = random();
76                 if (rnd >= max) {
77                         ntdb->ftable_off = off;
78                         ntdb->ftable = count;
79                         max = rnd;
80                 }
81
82                 off = next_ftable(ntdb, off);
83                 count++;
84         }
85         return NTDB_SUCCESS;
86 }
87
88 /* Offset of a given bucket. */
89 ntdb_off_t bucket_off(ntdb_off_t ftable_off, unsigned bucket)
90 {
91         return ftable_off + offsetof(struct ntdb_freetable, buckets)
92                 + bucket * sizeof(ntdb_off_t);
93 }
94
95 /* Returns free_buckets + 1, or list number to search, or -ve error. */
96 static ntdb_off_t find_free_head(struct ntdb_context *ntdb,
97                                 ntdb_off_t ftable_off,
98                                 ntdb_off_t bucket)
99 {
100         /* Speculatively search for a non-zero bucket. */
101         return ntdb_find_nonzero_off(ntdb, bucket_off(ftable_off, 0),
102                                     bucket, NTDB_FREE_BUCKETS);
103 }
104
105 static void check_list(struct ntdb_context *ntdb, ntdb_off_t b_off)
106 {
107 #ifdef CCAN_NTDB_DEBUG
108         ntdb_off_t off, prev = 0, first;
109         struct ntdb_free_record r;
110
111         first = off = (ntdb_read_off(ntdb, b_off) & NTDB_OFF_MASK);
112         while (off != 0) {
113                 ntdb_read_convert(ntdb, off, &r, sizeof(r));
114                 if (frec_magic(&r) != NTDB_FREE_MAGIC)
115                         abort();
116                 if (prev && frec_prev(&r) != prev)
117                         abort();
118                 prev = off;
119                 off = r.next;
120         }
121
122         if (first) {
123                 ntdb_read_convert(ntdb, first, &r, sizeof(r));
124                 if (frec_prev(&r) != prev)
125                         abort();
126         }
127 #endif
128 }
129
130 /* Remove from free bucket. */
131 static enum NTDB_ERROR remove_from_list(struct ntdb_context *ntdb,
132                                        ntdb_off_t b_off, ntdb_off_t r_off,
133                                        const struct ntdb_free_record *r)
134 {
135         ntdb_off_t off, prev_next, head;
136         enum NTDB_ERROR ecode;
137
138         /* Is this only element in list?  Zero out bucket, and we're done. */
139         if (frec_prev(r) == r_off)
140                 return ntdb_write_off(ntdb, b_off, 0);
141
142         /* off = &r->prev->next */
143         off = frec_prev(r) + offsetof(struct ntdb_free_record, next);
144
145         /* Get prev->next */
146         prev_next = ntdb_read_off(ntdb, off);
147         if (NTDB_OFF_IS_ERR(prev_next))
148                 return NTDB_OFF_TO_ERR(prev_next);
149
150         /* If prev->next == 0, we were head: update bucket to point to next. */
151         if (prev_next == 0) {
152                 /* We must preserve upper bits. */
153                 head = ntdb_read_off(ntdb, b_off);
154                 if (NTDB_OFF_IS_ERR(head))
155                         return NTDB_OFF_TO_ERR(head);
156
157                 if ((head & NTDB_OFF_MASK) != r_off) {
158                         return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
159                                           "remove_from_list:"
160                                           " %llu head %llu on list %llu",
161                                           (long long)r_off,
162                                           (long long)head,
163                                           (long long)b_off);
164                 }
165                 head = ((head & ~NTDB_OFF_MASK) | r->next);
166                 ecode = ntdb_write_off(ntdb, b_off, head);
167                 if (ecode != NTDB_SUCCESS)
168                         return ecode;
169         } else {
170                 /* r->prev->next = r->next */
171                 ecode = ntdb_write_off(ntdb, off, r->next);
172                 if (ecode != NTDB_SUCCESS)
173                         return ecode;
174         }
175
176         /* If we were the tail, off = &head->prev. */
177         if (r->next == 0) {
178                 head = ntdb_read_off(ntdb, b_off);
179                 if (NTDB_OFF_IS_ERR(head))
180                         return NTDB_OFF_TO_ERR(head);
181                 head &= NTDB_OFF_MASK;
182                 off = head + offsetof(struct ntdb_free_record, magic_and_prev);
183         } else {
184                 /* off = &r->next->prev */
185                 off = r->next + offsetof(struct ntdb_free_record,
186                                          magic_and_prev);
187         }
188
189 #ifdef CCAN_NTDB_DEBUG
190         /* *off == r */
191         if ((ntdb_read_off(ntdb, off) & NTDB_OFF_MASK) != r_off) {
192                 return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
193                                   "remove_from_list:"
194                                   " %llu bad prev in list %llu",
195                                   (long long)r_off, (long long)b_off);
196         }
197 #endif
198         /* r->next->prev = r->prev */
199         return ntdb_write_off(ntdb, off, r->magic_and_prev);
200 }
201
202 /* Enqueue in this free bucket: sets coalesce if we've added 128
203  * entries to it. */
204 static enum NTDB_ERROR enqueue_in_free(struct ntdb_context *ntdb,
205                                       ntdb_off_t b_off,
206                                       ntdb_off_t off,
207                                       ntdb_len_t len,
208                                       bool *coalesce)
209 {
210         struct ntdb_free_record new;
211         enum NTDB_ERROR ecode;
212         ntdb_off_t prev, head;
213         uint64_t magic = (NTDB_FREE_MAGIC << (64 - NTDB_OFF_UPPER_STEAL));
214
215         head = ntdb_read_off(ntdb, b_off);
216         if (NTDB_OFF_IS_ERR(head))
217                 return NTDB_OFF_TO_ERR(head);
218
219         /* We only need to set ftable_and_len; rest is set in enqueue_in_free */
220         new.ftable_and_len = ((uint64_t)ntdb->ftable
221                               << (64 - NTDB_OFF_UPPER_STEAL))
222                 | len;
223
224         /* new->next = head. */
225         new.next = (head & NTDB_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 = ntdb_read_off(ntdb,
233                                     new.next + offsetof(struct ntdb_free_record,
234                                                         magic_and_prev));
235                 new.magic_and_prev = prev;
236                 if (frec_magic(&new) != NTDB_FREE_MAGIC) {
237                         return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_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 = ntdb_write_off(ntdb, new.next
245                                       + offsetof(struct ntdb_free_record,
246                                                  magic_and_prev),
247                                       off | magic);
248                 if (ecode != NTDB_SUCCESS) {
249                         return ecode;
250                 }
251
252 #ifdef CCAN_NTDB_DEBUG
253                 prev = ntdb_read_off(ntdb, frec_prev(&new)
254                                     + offsetof(struct ntdb_free_record, next));
255                 if (prev != 0) {
256                         return ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
257                                           "enqueue_in_free:"
258                                           " %llu bad tail next ptr %llu",
259                                           (long long)frec_prev(&new)
260                                           + offsetof(struct ntdb_free_record,
261                                                      next),
262                                           (long long)prev);
263                 }
264 #endif
265         }
266
267         /* Update enqueue count, but don't set high bit: see NTDB_OFF_IS_ERR */
268         if (*coalesce)
269                 head += (1ULL << (64 - NTDB_OFF_UPPER_STEAL));
270         head &= ~(NTDB_OFF_MASK | (1ULL << 63));
271         head |= off;
272
273         ecode = ntdb_write_off(ntdb, b_off, head);
274         if (ecode != NTDB_SUCCESS) {
275                 return ecode;
276         }
277
278         /* It's time to coalesce if counter wrapped. */
279         if (*coalesce)
280                 *coalesce = ((head & ~NTDB_OFF_MASK) == 0);
281
282         return ntdb_write_convert(ntdb, off, &new, sizeof(new));
283 }
284
285 static ntdb_off_t ftable_offset(struct ntdb_context *ntdb, unsigned int ftable)
286 {
287         ntdb_off_t off;
288         unsigned int i;
289
290         if (likely(ntdb->ftable == ftable))
291                 return ntdb->ftable_off;
292
293         off = first_ftable(ntdb);
294         for (i = 0; i < ftable; i++) {
295                 if (NTDB_OFF_IS_ERR(off)) {
296                         break;
297                 }
298                 off = next_ftable(ntdb, 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 ntdb_len_t coalesce(struct ntdb_context *ntdb,
306                           ntdb_off_t off, ntdb_off_t b_off,
307                           ntdb_len_t data_len,
308                           ntdb_off_t *protect)
309 {
310         ntdb_off_t end;
311         struct ntdb_free_record rec;
312         enum NTDB_ERROR ecode;
313
314         ntdb->stats.alloc_coalesce_tried++;
315         end = off + sizeof(struct ntdb_used_record) + data_len;
316
317         while (end < ntdb->file->map_size) {
318                 const struct ntdb_free_record *r;
319                 ntdb_off_t nb_off;
320                 unsigned ftable, bucket;
321
322                 r = ntdb_access_read(ntdb, end, sizeof(*r), true);
323                 if (NTDB_PTR_IS_ERR(r)) {
324                         ecode = NTDB_PTR_ERR(r);
325                         goto err;
326                 }
327
328                 if (frec_magic(r) != NTDB_FREE_MAGIC
329                     || frec_ftable(r) == NTDB_FTABLE_NONE) {
330                         ntdb_access_release(ntdb, r);
331                         break;
332                 }
333
334                 ftable = frec_ftable(r);
335                 bucket = size_to_bucket(frec_len(r));
336                 nb_off = ftable_offset(ntdb, ftable);
337                 if (NTDB_OFF_IS_ERR(nb_off)) {
338                         ntdb_access_release(ntdb, r);
339                         ecode = NTDB_OFF_TO_ERR(nb_off);
340                         goto err;
341                 }
342                 nb_off = bucket_off(nb_off, bucket);
343                 ntdb_access_release(ntdb, r);
344
345                 /* We may be violating lock order here, so best effort. */
346                 if (ntdb_lock_free_bucket(ntdb, nb_off, NTDB_LOCK_NOWAIT)
347                     != NTDB_SUCCESS) {
348                         ntdb->stats.alloc_coalesce_lockfail++;
349                         break;
350                 }
351
352                 /* Now we have lock, re-check. */
353                 ecode = ntdb_read_convert(ntdb, end, &rec, sizeof(rec));
354                 if (ecode != NTDB_SUCCESS) {
355                         ntdb_unlock_free_bucket(ntdb, nb_off);
356                         goto err;
357                 }
358
359                 if (unlikely(frec_magic(&rec) != NTDB_FREE_MAGIC)) {
360                         ntdb->stats.alloc_coalesce_race++;
361                         ntdb_unlock_free_bucket(ntdb, nb_off);
362                         break;
363                 }
364
365                 if (unlikely(frec_ftable(&rec) != ftable)
366                     || unlikely(size_to_bucket(frec_len(&rec)) != bucket)) {
367                         ntdb->stats.alloc_coalesce_race++;
368                         ntdb_unlock_free_bucket(ntdb, nb_off);
369                         break;
370                 }
371
372                 /* Did we just mess up a record you were hoping to use? */
373                 if (end == *protect) {
374                         ntdb->stats.alloc_coalesce_iterate_clash++;
375                         *protect = NTDB_ERR_TO_OFF(NTDB_ERR_NOEXIST);
376                 }
377
378                 ecode = remove_from_list(ntdb, nb_off, end, &rec);
379                 check_list(ntdb, nb_off);
380                 if (ecode != NTDB_SUCCESS) {
381                         ntdb_unlock_free_bucket(ntdb, nb_off);
382                         goto err;
383                 }
384
385                 end += sizeof(struct ntdb_used_record) + frec_len(&rec);
386                 ntdb_unlock_free_bucket(ntdb, nb_off);
387                 ntdb->stats.alloc_coalesce_num_merged++;
388         }
389
390         /* Didn't find any adjacent free? */
391         if (end == off + sizeof(struct ntdb_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 = NTDB_ERR_TO_OFF(NTDB_ERR_EXISTS);
397                 ntdb->stats.alloc_coalesce_iterate_clash++;
398         }
399
400         /* OK, expand initial record */
401         ecode = ntdb_read_convert(ntdb, off, &rec, sizeof(rec));
402         if (ecode != NTDB_SUCCESS) {
403                 goto err;
404         }
405
406         if (frec_len(&rec) != data_len) {
407                 ecode = ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_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(ntdb, b_off, off, &rec);
414         check_list(ntdb, b_off);
415         if (ecode != NTDB_SUCCESS) {
416                 goto err;
417         }
418
419         /* Try locking violation first.  We don't allow coalesce recursion! */
420         ecode = add_free_record(ntdb, off, end - off, NTDB_LOCK_NOWAIT, false);
421         if (ecode != NTDB_SUCCESS) {
422                 /* Need to drop lock.  Can't rely on anything stable. */
423                 ntdb->stats.alloc_coalesce_lockfail++;
424                 *protect = NTDB_ERR_TO_OFF(NTDB_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 = (NTDB_FTABLE_NONE
429                                       << (64 - NTDB_OFF_UPPER_STEAL))
430                         | (end - off - sizeof(struct ntdb_used_record));
431                 ecode = ntdb_write_off(ntdb,
432                                       off + offsetof(struct ntdb_free_record,
433                                                      ftable_and_len),
434                                       rec.ftable_and_len);
435                 if (ecode != NTDB_SUCCESS) {
436                         goto err;
437                 }
438
439                 ntdb_unlock_free_bucket(ntdb, b_off);
440
441                 ecode = add_free_record(ntdb, off, end - off, NTDB_LOCK_WAIT,
442                                         false);
443                 if (ecode != NTDB_SUCCESS) {
444                         return NTDB_ERR_TO_OFF(ecode);
445                 }
446         } else if (NTDB_OFF_IS_ERR(*protect)) {
447                 /* For simplicity, we always drop lock if they can't continue */
448                 ntdb_unlock_free_bucket(ntdb, b_off);
449         }
450         ntdb->stats.alloc_coalesce_succeeded++;
451
452         /* Return usable length. */
453         return end - off - sizeof(struct ntdb_used_record);
454
455 err:
456         /* To unify error paths, we *always* unlock bucket on error. */
457         ntdb_unlock_free_bucket(ntdb, b_off);
458         return NTDB_ERR_TO_OFF(ecode);
459 }
460
461 /* List is locked: we unlock it. */
462 static enum NTDB_ERROR coalesce_list(struct ntdb_context *ntdb,
463                                     ntdb_off_t ftable_off,
464                                     ntdb_off_t b_off,
465                                     unsigned int limit)
466 {
467         enum NTDB_ERROR ecode;
468         ntdb_off_t off;
469
470         off = ntdb_read_off(ntdb, b_off);
471         if (NTDB_OFF_IS_ERR(off)) {
472                 ecode = NTDB_OFF_TO_ERR(off);
473                 goto unlock_err;
474         }
475         /* A little bit of paranoia: counter should be 0. */
476         off &= NTDB_OFF_MASK;
477
478         while (off && limit--) {
479                 struct ntdb_free_record rec;
480                 ntdb_len_t coal;
481                 ntdb_off_t next;
482
483                 ecode = ntdb_read_convert(ntdb, off, &rec, sizeof(rec));
484                 if (ecode != NTDB_SUCCESS)
485                         goto unlock_err;
486
487                 next = rec.next;
488                 coal = coalesce(ntdb, off, b_off, frec_len(&rec), &next);
489                 if (NTDB_OFF_IS_ERR(coal)) {
490                         /* This has already unlocked on error. */
491                         return NTDB_OFF_TO_ERR(coal);
492                 }
493                 if (NTDB_OFF_IS_ERR(next)) {
494                         /* Coalescing had to unlock, so stop. */
495                         return NTDB_SUCCESS;
496                 }
497                 /* Keep going if we're doing well... */
498                 limit += size_to_bucket(coal / 16 + NTDB_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 ntdb_free_record oldhrec, newhrec, oldtrec, newtrec;
506                 ntdb_off_t oldhoff, oldtoff, newtoff;
507
508                 /* The record we were up to is the new head. */
509                 ecode = ntdb_read_convert(ntdb, off, &newhrec, sizeof(newhrec));
510                 if (ecode != NTDB_SUCCESS)
511                         goto unlock_err;
512
513                 /* Get the new tail. */
514                 newtoff = frec_prev(&newhrec);
515                 ecode = ntdb_read_convert(ntdb, newtoff, &newtrec,
516                                          sizeof(newtrec));
517                 if (ecode != NTDB_SUCCESS)
518                         goto unlock_err;
519
520                 /* Get the old head. */
521                 oldhoff = ntdb_read_off(ntdb, b_off);
522                 if (NTDB_OFF_IS_ERR(oldhoff)) {
523                         ecode = NTDB_OFF_TO_ERR(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 = ntdb_read_convert(ntdb, oldhoff, &oldhrec,
532                                          sizeof(oldhrec));
533                 if (ecode != NTDB_SUCCESS)
534                         goto unlock_err;
535
536                 /* Get the old tail. */
537                 oldtoff = frec_prev(&oldhrec);
538                 ecode = ntdb_read_convert(ntdb, oldtoff, &oldtrec,
539                                          sizeof(oldtrec));
540                 if (ecode != NTDB_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                         = (NTDB_FREE_MAGIC << (64 - NTDB_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 = ntdb_write_convert(ntdb, oldtoff, &oldtrec,
556                                           sizeof(oldtrec));
557                 if (ecode != NTDB_SUCCESS)
558                         goto unlock_err;
559
560                 ecode = ntdb_write_convert(ntdb, oldhoff, &oldhrec,
561                                           sizeof(oldhrec));
562                 if (ecode != NTDB_SUCCESS)
563                         goto unlock_err;
564
565                 ecode = ntdb_write_convert(ntdb, newtoff, &newtrec,
566                                           sizeof(newtrec));
567                 if (ecode != NTDB_SUCCESS)
568                         goto unlock_err;
569
570                 /* And finally link in new head. */
571                 ecode = ntdb_write_off(ntdb, b_off, off);
572                 if (ecode != NTDB_SUCCESS)
573                         goto unlock_err;
574         }
575 out:
576         ntdb_unlock_free_bucket(ntdb, b_off);
577         return NTDB_SUCCESS;
578
579 unlock_err:
580         ntdb_unlock_free_bucket(ntdb, b_off);
581         return ecode;
582 }
583
584 /* List must not be locked if coalesce_ok is set. */
585 enum NTDB_ERROR add_free_record(struct ntdb_context *ntdb,
586                                ntdb_off_t off, ntdb_len_t len_with_header,
587                                enum ntdb_lock_flags waitflag,
588                                bool coalesce_ok)
589 {
590         ntdb_off_t b_off;
591         ntdb_len_t len;
592         enum NTDB_ERROR ecode;
593
594         assert(len_with_header >= sizeof(struct ntdb_free_record));
595
596         len = len_with_header - sizeof(struct ntdb_used_record);
597
598         b_off = bucket_off(ntdb->ftable_off, size_to_bucket(len));
599         ecode = ntdb_lock_free_bucket(ntdb, b_off, waitflag);
600         if (ecode != NTDB_SUCCESS) {
601                 return ecode;
602         }
603
604         ecode = enqueue_in_free(ntdb, b_off, off, len, &coalesce_ok);
605         check_list(ntdb, b_off);
606
607         /* Coalescing unlocks free list. */
608         if (!ecode && coalesce_ok)
609                 ecode = coalesce_list(ntdb, ntdb->ftable_off, b_off, 2);
610         else
611                 ntdb_unlock_free_bucket(ntdb, 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 < NTDB_MIN_DATA_LEN)
620                 size = NTDB_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 ntdb_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 ntdb_off_t lock_and_alloc(struct ntdb_context *ntdb,
644                                 ntdb_off_t ftable_off,
645                                 ntdb_off_t bucket,
646                                 size_t keylen, size_t datalen,
647                                 bool want_extra,
648                                 unsigned magic)
649 {
650         ntdb_off_t off, b_off,best_off;
651         struct ntdb_free_record best = { 0 };
652         double multiplier;
653         size_t size = adjust_size(keylen, datalen);
654         enum NTDB_ERROR ecode;
655
656         ntdb->stats.allocs++;
657         b_off = bucket_off(ftable_off, bucket);
658
659         /* FIXME: Try non-blocking wait first, to measure contention. */
660         /* Lock this bucket. */
661         ecode = ntdb_lock_free_bucket(ntdb, b_off, NTDB_LOCK_WAIT);
662         if (ecode != NTDB_SUCCESS) {
663                 return NTDB_ERR_TO_OFF(ecode);
664         }
665
666         best.ftable_and_len = -1ULL;
667         best_off = 0;
668
669         /* Get slack if we're after extra. */
670         if (want_extra)
671                 multiplier = 1.5;
672         else
673                 multiplier = 1.0;
674
675         /* Walk the list to see if any are large enough, getting less fussy
676          * as we go. */
677         off = ntdb_read_off(ntdb, b_off);
678         if (NTDB_OFF_IS_ERR(off)) {
679                 ecode = NTDB_OFF_TO_ERR(off);
680                 goto unlock_err;
681         }
682         off &= NTDB_OFF_MASK;
683
684         while (off) {
685                 const struct ntdb_free_record *r;
686                 ntdb_off_t next;
687
688                 r = ntdb_access_read(ntdb, off, sizeof(*r), true);
689                 if (NTDB_PTR_IS_ERR(r)) {
690                         ecode = NTDB_PTR_ERR(r);
691                         goto unlock_err;
692                 }
693
694                 if (frec_magic(r) != NTDB_FREE_MAGIC) {
695                         ecode = ntdb_logerr(ntdb, NTDB_ERR_CORRUPT, NTDB_LOG_ERROR,
696                                            "lock_and_alloc:"
697                                            " %llu non-free 0x%llx",
698                                            (long long)off,
699                                            (long long)r->magic_and_prev);
700                         ntdb_access_release(ntdb, r);
701                         goto unlock_err;
702                 }
703
704                 if (frec_len(r) >= size && frec_len(r) < frec_len(&best)) {
705                         best_off = off;
706                         best = *r;
707                 }
708
709                 if (frec_len(&best) <= size * multiplier && best_off) {
710                         ntdb_access_release(ntdb, r);
711                         break;
712                 }
713
714                 multiplier *= 1.01;
715
716                 next = r->next;
717                 ntdb_access_release(ntdb, r);
718                 off = next;
719         }
720
721         /* If we found anything at all, use it. */
722         if (best_off) {
723                 struct ntdb_used_record rec;
724                 size_t leftover;
725
726                 /* We're happy with this size: take it. */
727                 ecode = remove_from_list(ntdb, b_off, best_off, &best);
728                 check_list(ntdb, b_off);
729                 if (ecode != NTDB_SUCCESS) {
730                         goto unlock_err;
731                 }
732
733                 leftover = record_leftover(keylen, datalen, want_extra,
734                                            frec_len(&best));
735
736                 assert(keylen + datalen + leftover <= frec_len(&best));
737                 /* We need to mark non-free before we drop lock, otherwise
738                  * coalesce() could try to merge it! */
739                 ecode = set_header(ntdb, &rec, magic, keylen, datalen,
740                                    frec_len(&best) - leftover);
741                 if (ecode != NTDB_SUCCESS) {
742                         goto unlock_err;
743                 }
744
745                 ecode = ntdb_write_convert(ntdb, best_off, &rec, sizeof(rec));
746                 if (ecode != NTDB_SUCCESS) {
747                         goto unlock_err;
748                 }
749
750                 /* For futureproofing, we put a 0 in any unused space. */
751                 if (rec_extra_padding(&rec)) {
752                         ecode = ntdb->io->twrite(ntdb, best_off + sizeof(rec)
753                                                 + keylen + datalen, "", 1);
754                         if (ecode != NTDB_SUCCESS) {
755                                 goto unlock_err;
756                         }
757                 }
758
759                 /* Bucket of leftover will be <= current bucket, so nested
760                  * locking is allowed. */
761                 if (leftover) {
762                         ntdb->stats.alloc_leftover++;
763                         ecode = add_free_record(ntdb,
764                                                 best_off + sizeof(rec)
765                                                 + frec_len(&best) - leftover,
766                                                 leftover, NTDB_LOCK_WAIT, false);
767                         if (ecode != NTDB_SUCCESS) {
768                                 best_off = NTDB_ERR_TO_OFF(ecode);
769                         }
770                 }
771                 ntdb_unlock_free_bucket(ntdb, b_off);
772
773                 return best_off;
774         }
775
776         ntdb_unlock_free_bucket(ntdb, b_off);
777         return 0;
778
779 unlock_err:
780         ntdb_unlock_free_bucket(ntdb, b_off);
781         return NTDB_ERR_TO_OFF(ecode);
782 }
783
784 /* Get a free block from current free list, or 0 if none, -ve on error. */
785 static ntdb_off_t get_free(struct ntdb_context *ntdb,
786                           size_t keylen, size_t datalen, bool want_extra,
787                           unsigned magic)
788 {
789         ntdb_off_t off, ftable_off;
790         ntdb_off_t start_b, b, ftable;
791         bool wrapped = false;
792
793         /* If they are growing, add 50% to get to higher bucket. */
794         if (want_extra)
795                 start_b = size_to_bucket(adjust_size(keylen,
796                                                      datalen + datalen / 2));
797         else
798                 start_b = size_to_bucket(adjust_size(keylen, datalen));
799
800         ftable_off = ntdb->ftable_off;
801         ftable = ntdb->ftable;
802         while (!wrapped || ftable_off != ntdb->ftable_off) {
803                 /* Start at exact size bucket, and search up... */
804                 for (b = find_free_head(ntdb, ftable_off, start_b);
805                      b < NTDB_FREE_BUCKETS;
806                      b = find_free_head(ntdb, ftable_off, b + 1)) {
807                         /* Try getting one from list. */
808                         off = lock_and_alloc(ntdb, ftable_off,
809                                              b, keylen, datalen, want_extra,
810                                              magic);
811                         if (NTDB_OFF_IS_ERR(off))
812                                 return off;
813                         if (off != 0) {
814                                 if (b == start_b)
815                                         ntdb->stats.alloc_bucket_exact++;
816                                 if (b == NTDB_FREE_BUCKETS - 1)
817                                         ntdb->stats.alloc_bucket_max++;
818                                 /* Worked?  Stay using this list. */
819                                 ntdb->ftable_off = ftable_off;
820                                 ntdb->ftable = ftable;
821                                 return off;
822                         }
823                         /* Didn't work.  Try next bucket. */
824                 }
825
826                 if (NTDB_OFF_IS_ERR(b)) {
827                         return b;
828                 }
829
830                 /* Hmm, try next table. */
831                 ftable_off = next_ftable(ntdb, ftable_off);
832                 if (NTDB_OFF_IS_ERR(ftable_off)) {
833                         return ftable_off;
834                 }
835                 ftable++;
836
837                 if (ftable_off == 0) {
838                         wrapped = true;
839                         ftable_off = first_ftable(ntdb);
840                         if (NTDB_OFF_IS_ERR(ftable_off)) {
841                                 return ftable_off;
842                         }
843                         ftable = 0;
844                 }
845         }
846
847         return 0;
848 }
849
850 enum NTDB_ERROR set_header(struct ntdb_context *ntdb,
851                           struct ntdb_used_record *rec,
852                           unsigned magic, uint64_t keylen, uint64_t datalen,
853                           uint64_t actuallen)
854 {
855         uint64_t keybits = (fls64(keylen) + 1) / 2;
856
857         rec->magic_and_meta = ((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 ntdb_logerr(ntdb, NTDB_ERR_IO, NTDB_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 NTDB_SUCCESS;
872 }
873
874 /* You need 'size', this tells you how much you should expand by. */
875 ntdb_off_t ntdb_expand_adjust(ntdb_off_t map_size, ntdb_off_t size)
876 {
877         ntdb_off_t new_size, top_size;
878
879         /* limit size in order to avoid using up huge amounts of memory for
880          * in memory tdbs if an oddball huge record creeps in */
881         if (size > 100 * 1024) {
882                 top_size = map_size + size * 2;
883         } else {
884                 top_size = map_size + size * 100;
885         }
886
887         /* always make room for at least top_size more records, and at
888            least 25% more space. if the DB is smaller than 100MiB,
889            otherwise grow it by 10% only. */
890         if (map_size > 100 * 1024 * 1024) {
891                 new_size = map_size * 1.10;
892         } else {
893                 new_size = map_size * 1.25;
894         }
895
896         if (new_size < top_size)
897                 new_size = top_size;
898
899         /* We always make the file a multiple of transaction page
900          * size.  This guarantees that the transaction recovery area
901          * is always aligned, otherwise the transaction code can overwrite
902          * itself. */
903         new_size = (new_size + NTDB_PGSIZE-1) & ~(NTDB_PGSIZE-1);
904         return new_size - map_size;
905 }
906
907 /* Expand the database. */
908 static enum NTDB_ERROR ntdb_expand(struct ntdb_context *ntdb, ntdb_len_t size)
909 {
910         uint64_t old_size;
911         ntdb_len_t wanted;
912         enum NTDB_ERROR ecode;
913
914         /* Need to hold a hash lock to expand DB: transactions rely on it. */
915         if (!(ntdb->flags & NTDB_NOLOCK)
916             && !ntdb->file->allrecord_lock.count && !ntdb_has_hash_locks(ntdb)) {
917                 return ntdb_logerr(ntdb, NTDB_ERR_LOCK, NTDB_LOG_ERROR,
918                                   "ntdb_expand: must hold lock during expand");
919         }
920
921         /* Only one person can expand file at a time. */
922         ecode = ntdb_lock_expand(ntdb, F_WRLCK);
923         if (ecode != NTDB_SUCCESS) {
924                 return ecode;
925         }
926
927         /* Someone else may have expanded the file, so retry. */
928         old_size = ntdb->file->map_size;
929         ntdb_oob(ntdb, ntdb->file->map_size, 1, true);
930         if (ntdb->file->map_size != old_size) {
931                 ntdb_unlock_expand(ntdb, F_WRLCK);
932                 return NTDB_SUCCESS;
933         }
934
935         /* We need room for the record header too. */
936         size = adjust_size(0, sizeof(struct ntdb_used_record) + size);
937         /* Overallocate. */
938         wanted = ntdb_expand_adjust(old_size, size);
939
940         ecode = ntdb->io->expand_file(ntdb, wanted);
941         if (ecode != NTDB_SUCCESS) {
942                 ntdb_unlock_expand(ntdb, F_WRLCK);
943                 return ecode;
944         }
945
946         /* We need to drop this lock before adding free record. */
947         ntdb_unlock_expand(ntdb, F_WRLCK);
948
949         ntdb->stats.expands++;
950         return add_free_record(ntdb, old_size, wanted, NTDB_LOCK_WAIT, true);
951 }
952
953 /* This won't fail: it will expand the database if it has to. */
954 ntdb_off_t alloc(struct ntdb_context *ntdb, size_t keylen, size_t datalen,
955                  unsigned magic, bool growing)
956 {
957         ntdb_off_t off;
958
959         for (;;) {
960                 enum NTDB_ERROR ecode;
961                 off = get_free(ntdb, keylen, datalen, growing, magic);
962                 if (likely(off != 0))
963                         break;
964
965                 ecode = ntdb_expand(ntdb, adjust_size(keylen, datalen));
966                 if (ecode != NTDB_SUCCESS) {
967                         return NTDB_ERR_TO_OFF(ecode);
968                 }
969         }
970
971         return off;
972 }