8ff5d74a3fd7a8a400b243b9dd8a22c1dfe44af6
[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_flist(struct tdb_context *tdb)
53 {
54         return tdb_read_off(tdb, offsetof(struct tdb_header, free_list));
55 }
56
57 tdb_off_t next_flist(struct tdb_context *tdb, tdb_off_t flist)
58 {
59         return tdb_read_off(tdb, flist + offsetof(struct tdb_freelist, next));
60 }
61
62 int tdb_flist_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->flist_off = off = first_flist(tdb);
69         tdb->flist = 0;
70
71         while (off) {
72                 if (off == TDB_OFF_ERR)
73                         return -1;
74
75                 rnd = random();
76                 if (rnd >= max) {
77                         tdb->flist_off = off;
78                         tdb->flist = count;
79                         max = rnd;
80                 }
81
82                 off = next_flist(tdb, off);
83                 count++;
84         }
85         return 0;
86 }
87
88 /* Offset of a given bucket. */
89 tdb_off_t bucket_off(tdb_off_t flist_off, unsigned bucket)
90 {
91         return flist_off + offsetof(struct tdb_freelist, buckets)
92                 + bucket * sizeof(tdb_off_t);
93 }
94
95 /* Returns free_buckets + 1, or list number to search. */
96 static tdb_off_t find_free_head(struct tdb_context *tdb,
97                                 tdb_off_t flist_off,
98                                 tdb_off_t bucket)
99 {
100         /* Speculatively search for a non-zero bucket. */
101         return tdb_find_nonzero_off(tdb, bucket_off(flist_off, 0),
102                                     bucket, TDB_FREE_BUCKETS);
103 }
104
105 /* Remove from free bucket. */
106 static int remove_from_list(struct tdb_context *tdb,
107                             tdb_off_t b_off, tdb_off_t r_off,
108                             struct tdb_free_record *r)
109 {
110         tdb_off_t off;
111
112         /* Front of list? */
113         if (frec_prev(r) == 0) {
114                 off = b_off;
115         } else {
116                 off = frec_prev(r) + offsetof(struct tdb_free_record, next);
117         }
118
119 #ifdef DEBUG
120         if (tdb_read_off(tdb, off) != r_off) {
121                 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
122                          "remove_from_list: %llu bad prev in list %llu\n",
123                          (long long)r_off, (long long)b_off);
124                 return -1;
125         }
126 #endif
127
128         /* r->prev->next = r->next */
129         if (tdb_write_off(tdb, off, r->next)) {
130                 return -1;
131         }
132
133         if (r->next != 0) {
134                 off = r->next + offsetof(struct tdb_free_record,magic_and_prev);
135                 /* r->next->prev = r->prev */
136
137 #ifdef DEBUG
138                 if (tdb_read_off(tdb, off) & TDB_OFF_MASK != r_off) {
139                         tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
140                                  "remove_from_list: %llu bad list %llu\n",
141                                  (long long)r_off, (long long)b_off);
142                         return -1;
143                 }
144 #endif
145
146                 if (tdb_write_off(tdb, off, r->magic_and_prev)) {
147                         return -1;
148                 }
149         }
150         return 0;
151 }
152
153 /* Enqueue in this free bucket. */
154 static int enqueue_in_free(struct tdb_context *tdb,
155                            tdb_off_t b_off,
156                            tdb_off_t off,
157                            tdb_len_t len)
158 {
159         struct tdb_free_record new;
160         uint64_t magic = (TDB_FREE_MAGIC << (64 - TDB_OFF_UPPER_STEAL));
161
162         /* We only need to set flist_and_len; rest is set in enqueue_in_free */
163         new.flist_and_len = ((uint64_t)tdb->flist << (64 - TDB_OFF_UPPER_STEAL))
164                 | len;
165         /* prev = 0. */
166         new.magic_and_prev = magic;
167
168         /* new->next = head. */
169         new.next = tdb_read_off(tdb, b_off);
170         if (new.next == TDB_OFF_ERR)
171                 return -1;
172
173         if (new.next) {
174 #ifdef DEBUG
175                 if (tdb_read_off(tdb,
176                                  new.next + offsetof(struct tdb_free_record,
177                                                      magic_and_prev))
178                     != magic) {
179                         tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
180                                  "enqueue_in_free: %llu bad head prev %llu\n",
181                                  (long long)new.next, (long long)b_off);
182                         return -1;
183                 }
184 #endif
185                 /* next->prev = new. */
186                 if (tdb_write_off(tdb, new.next
187                                   + offsetof(struct tdb_free_record,
188                                              magic_and_prev),
189                                   off | magic) != 0)
190                         return -1;
191         }
192         /* head = new */
193         if (tdb_write_off(tdb, b_off, off) != 0)
194                 return -1;
195
196         return tdb_write_convert(tdb, off, &new, sizeof(new));
197 }
198
199 /* List need not be locked. */
200 int add_free_record(struct tdb_context *tdb,
201                     tdb_off_t off, tdb_len_t len_with_header)
202 {
203         tdb_off_t b_off;
204         tdb_len_t len;
205         int ret;
206
207         assert(len_with_header >= sizeof(struct tdb_free_record));
208
209         len = len_with_header - sizeof(struct tdb_used_record);
210
211         b_off = bucket_off(tdb->flist_off, size_to_bucket(len));
212         if (tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT) != 0)
213                 return -1;
214
215         ret = enqueue_in_free(tdb, b_off, off, len);
216         tdb_unlock_free_bucket(tdb, b_off);
217         return ret;
218 }
219
220 static size_t adjust_size(size_t keylen, size_t datalen)
221 {
222         size_t size = keylen + datalen;
223
224         if (size < TDB_MIN_DATA_LEN)
225                 size = TDB_MIN_DATA_LEN;
226
227         /* Round to next uint64_t boundary. */
228         return (size + (sizeof(uint64_t) - 1ULL)) & ~(sizeof(uint64_t) - 1ULL);
229 }
230
231 /* If we have enough left over to be useful, split that off. */
232 static size_t record_leftover(size_t keylen, size_t datalen,
233                               bool want_extra, size_t total_len)
234 {
235         ssize_t leftover;
236
237         if (want_extra)
238                 datalen += datalen / 2;
239         leftover = total_len - adjust_size(keylen, datalen);
240
241         if (leftover < (ssize_t)sizeof(struct tdb_free_record))
242                 return 0;
243
244         return leftover;
245 }
246
247 /* FIXME: Shortcut common case where tdb->flist == flist */
248 static tdb_off_t flist_offset(struct tdb_context *tdb, unsigned int flist)
249 {
250         tdb_off_t off = first_flist(tdb);
251         unsigned int i;
252
253         for (i = 0; i < flist; i++)
254                 off = next_flist(tdb, off);
255         return off;
256 }
257
258 /* Note: we unlock the current bucket if we coalesce or fail. */
259 static int coalesce(struct tdb_context *tdb,
260                     tdb_off_t off, tdb_off_t b_off, tdb_len_t data_len)
261 {
262         struct tdb_free_record pad, *r;
263         tdb_off_t end;
264
265         end = off + sizeof(struct tdb_used_record) + data_len;
266
267         while (end < tdb->map_size) {
268                 tdb_off_t nb_off;
269                 unsigned flist, bucket;
270
271                 /* FIXME: do tdb_get here and below really win? */
272                 r = tdb_get(tdb, end, &pad, sizeof(pad));
273                 if (!r)
274                         goto err;
275
276                 if (frec_magic(r) != TDB_FREE_MAGIC)
277                         break;
278
279                 flist = frec_flist(r);
280                 bucket = size_to_bucket(frec_len(r));
281                 nb_off = bucket_off(flist_offset(tdb, flist), bucket);
282
283                 /* We may be violating lock order here, so best effort. */
284                 if (tdb_lock_free_bucket(tdb, nb_off, TDB_LOCK_NOWAIT) == -1)
285                         break;
286
287                 /* Now we have lock, re-check. */
288                 r = tdb_get(tdb, end, &pad, sizeof(pad));
289                 if (!r) {
290                         tdb_unlock_free_bucket(tdb, nb_off);
291                         goto err;
292                 }
293
294                 if (unlikely(frec_magic(r) != TDB_FREE_MAGIC)) {
295                         tdb_unlock_free_bucket(tdb, nb_off);
296                         break;
297                 }
298
299                 if (unlikely(frec_flist(r) != flist)
300                     || unlikely(size_to_bucket(frec_len(r)) != bucket)) {
301                         tdb_unlock_free_bucket(tdb, nb_off);
302                         break;
303                 }
304
305                 if (remove_from_list(tdb, nb_off, end, r) == -1) {
306                         tdb_unlock_free_bucket(tdb, nb_off);
307                         goto err;
308                 }
309
310                 end += sizeof(struct tdb_used_record) + frec_len(r);
311                 tdb_unlock_free_bucket(tdb, nb_off);
312         }
313
314         /* Didn't find any adjacent free? */
315         if (end == off + sizeof(struct tdb_used_record) + data_len)
316                 return 0;
317
318         /* OK, expand record */
319         r = tdb_get(tdb, off, &pad, sizeof(pad));
320         if (!r)
321                 goto err;
322
323         if (frec_len(r) != data_len) {
324                 tdb->ecode = TDB_ERR_CORRUPT;
325                 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
326                          "coalesce: expected data len %llu not %llu\n",
327                          (long long)data_len, (long long)frec_len(r));
328                 goto err;
329         }
330
331         if (remove_from_list(tdb, b_off, off, r) == -1)
332                 goto err;
333
334         r = tdb_access_write(tdb, off, sizeof(*r), true);
335         if (!r)
336                 goto err;
337
338         /* We have to drop this to avoid deadlocks, so make sure record
339          * doesn't get coalesced by someone else! */
340         r->magic_and_prev = TDB_COALESCING_MAGIC << (64 - TDB_OFF_UPPER_STEAL);
341         /* FIXME: Use 255 as invalid free list? */
342         r->flist_and_len = end - off - sizeof(struct tdb_used_record);
343         if (tdb_access_commit(tdb, r) != 0)
344                 goto err;
345
346         tdb_unlock_free_bucket(tdb, b_off);
347
348         if (add_free_record(tdb, off, end - off) == -1)
349                 return -1;
350         return 1;
351
352 err:
353         /* To unify error paths, we *always* unlock bucket on error. */
354         tdb_unlock_free_bucket(tdb, b_off);
355         return -1;
356 }
357
358 /* We need size bytes to put our key and data in. */
359 static tdb_off_t lock_and_alloc(struct tdb_context *tdb,
360                                 tdb_off_t flist_off,
361                                 tdb_off_t bucket,
362                                 size_t keylen, size_t datalen,
363                                 bool want_extra,
364                                 unsigned hashlow)
365 {
366         tdb_off_t off, b_off,best_off;
367         struct tdb_free_record pad, best = { 0 }, *r;
368         double multiplier;
369         size_t size = adjust_size(keylen, datalen);
370
371 again:
372         b_off = bucket_off(flist_off, bucket);
373
374         /* FIXME: Try non-blocking wait first, to measure contention. */
375         /* Lock this bucket. */
376         if (tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT) == -1) {
377                 return TDB_OFF_ERR;
378         }
379
380         best.flist_and_len = -1ULL;
381         best_off = 0;
382
383         /* Get slack if we're after extra. */
384         if (want_extra)
385                 multiplier = 1.5;
386         else
387                 multiplier = 1.0;
388
389         /* Walk the list to see if any are large enough, getting less fussy
390          * as we go. */
391         off = tdb_read_off(tdb, b_off);
392         if (unlikely(off == TDB_OFF_ERR))
393                 goto unlock_err;
394
395         while (off) {
396                 /* FIXME: Does tdb_get win anything here? */
397                 r = tdb_get(tdb, off, &pad, sizeof(*r));
398                 if (!r)
399                         goto unlock_err;
400
401                 if (frec_magic(r) != TDB_FREE_MAGIC) {
402                         tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
403                                  "lock_and_alloc: %llu non-free 0x%llx\n",
404                                  (long long)off, (long long)r->magic_and_prev);
405                         goto unlock_err;
406                 }
407
408                 if (frec_len(r) >= size && frec_len(r) < frec_len(&best)) {
409                         best_off = off;
410                         best = *r;
411                 }
412
413                 if (frec_len(&best) < size * multiplier && best_off)
414                         break;
415
416                 multiplier *= 1.01;
417
418                 /* Since we're going slow anyway, try coalescing here. */
419                 switch (coalesce(tdb, off, b_off, frec_len(r))) {
420                 case -1:
421                         /* This has already unlocked on error. */
422                         return -1;
423                 case 1:
424                         /* This has unlocked list, restart. */
425                         goto again;
426                 }
427                 off = r->next;
428         }
429
430         /* If we found anything at all, use it. */
431         if (best_off) {
432                 struct tdb_used_record rec;
433                 size_t leftover;
434
435                 /* We're happy with this size: take it. */
436                 if (remove_from_list(tdb, b_off, best_off, &best) != 0)
437                         goto unlock_err;
438
439                 leftover = record_leftover(keylen, datalen, want_extra,
440                                            frec_len(&best));
441
442                 assert(keylen + datalen + leftover <= frec_len(&best));
443                 /* We need to mark non-free before we drop lock, otherwise
444                  * coalesce() could try to merge it! */
445                 if (set_used_header(tdb, &rec, keylen, datalen,
446                                     frec_len(&best) - leftover,
447                                     hashlow) != 0)
448                         goto unlock_err;
449
450                 if (tdb_write_convert(tdb, best_off, &rec, sizeof(rec)) != 0)
451                         goto unlock_err;
452
453                 tdb_unlock_free_bucket(tdb, b_off);
454
455                 if (leftover) {
456                         if (add_free_record(tdb,
457                                             best_off + sizeof(rec)
458                                             + frec_len(&best) - leftover,
459                                             leftover))
460                                 return TDB_OFF_ERR;
461                 }
462                 return best_off;
463         }
464
465         tdb_unlock_free_bucket(tdb, b_off);
466         return 0;
467
468 unlock_err:
469         tdb_unlock_free_bucket(tdb, b_off);
470         return TDB_OFF_ERR;
471 }
472
473 /* Get a free block from current free list, or 0 if none. */
474 static tdb_off_t get_free(struct tdb_context *tdb,
475                           size_t keylen, size_t datalen, bool want_extra,
476                           unsigned hashlow)
477 {
478         tdb_off_t off, flist_off;
479         unsigned start_b, b, flist;
480         bool wrapped = false;
481
482         /* If they are growing, add 50% to get to higher bucket. */
483         if (want_extra)
484                 start_b = size_to_bucket(adjust_size(keylen,
485                                                      datalen + datalen / 2));
486         else
487                 start_b = size_to_bucket(adjust_size(keylen, datalen));
488
489         flist_off = tdb->flist_off;
490         flist = tdb->flist;
491         while (!wrapped || flist_off != tdb->flist_off) {
492                 /* Start at exact size bucket, and search up... */
493                 for (b = find_free_head(tdb, flist_off, start_b);
494                      b < TDB_FREE_BUCKETS;
495                      b = find_free_head(tdb, flist_off, b + 1)) {
496                         /* Try getting one from list. */
497                         off = lock_and_alloc(tdb, flist_off,
498                                              b, keylen, datalen, want_extra,
499                                              hashlow);
500                         if (off == TDB_OFF_ERR)
501                                 return TDB_OFF_ERR;
502                         if (off != 0) {
503                                 /* Worked?  Stay using this list. */
504                                 tdb->flist_off = flist_off;
505                                 tdb->flist = flist;
506                                 return off;
507                         }
508                         /* Didn't work.  Try next bucket. */
509                 }
510
511                 /* Hmm, try next list. */
512                 flist_off = next_flist(tdb, flist_off);
513                 flist++;
514                 if (flist_off == 0) {
515                         wrapped = true;
516                         flist_off = first_flist(tdb);
517                         flist = 0;
518                 }
519         }
520
521         return 0;
522 }
523
524 int set_used_header(struct tdb_context *tdb,
525                     struct tdb_used_record *rec,
526                     uint64_t keylen, uint64_t datalen,
527                     uint64_t actuallen, unsigned hashlow)
528 {
529         uint64_t keybits = (fls64(keylen) + 1) / 2;
530
531         /* Use bottom bits of hash, so it's independent of hash table size. */
532         rec->magic_and_meta = (hashlow & ((1 << 11)-1))
533                 | ((actuallen - (keylen + datalen)) << 11)
534                 | (keybits << 43)
535                 | (TDB_MAGIC << 48);
536         rec->key_and_data_len = (keylen | (datalen << (keybits*2)));
537
538         /* Encoding can fail on big values. */
539         if (rec_key_length(rec) != keylen
540             || rec_data_length(rec) != datalen
541             || rec_extra_padding(rec) != actuallen - (keylen + datalen)) {
542                 tdb->ecode = TDB_ERR_IO;
543                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
544                          "Could not encode k=%llu,d=%llu,a=%llu\n",
545                          (long long)keylen, (long long)datalen,
546                          (long long)actuallen);
547                 return -1;
548         }
549         return 0;
550 }
551
552 /* Expand the database. */
553 static int tdb_expand(struct tdb_context *tdb, tdb_len_t size)
554 {
555         uint64_t old_size;
556         tdb_len_t wanted;
557
558         /* We need room for the record header too. */
559         wanted = sizeof(struct tdb_used_record) + size;
560
561         /* Need to hold a hash lock to expand DB: transactions rely on it. */
562         if (!(tdb->flags & TDB_NOLOCK)
563             && !tdb->allrecord_lock.count && !tdb_has_hash_locks(tdb)) {
564                 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
565                          "tdb_expand: must hold lock during expand\n");
566                 return -1;
567         }
568
569         /* always make room for at least 100 more records, and at
570            least 25% more space. */
571         if (size * TDB_EXTENSION_FACTOR > tdb->map_size / 4)
572                 wanted = size * TDB_EXTENSION_FACTOR;
573         else
574                 wanted = tdb->map_size / 4;
575         wanted = adjust_size(0, wanted);
576
577         /* Only one person can expand file at a time. */
578         if (tdb_lock_expand(tdb, F_WRLCK) != 0)
579                 return -1;
580
581         /* Someone else may have expanded the file, so retry. */
582         old_size = tdb->map_size;
583         tdb->methods->oob(tdb, tdb->map_size + 1, true);
584         if (tdb->map_size != old_size) {
585                 tdb_unlock_expand(tdb, F_WRLCK);
586                 return 0;
587         }
588
589         if (tdb->methods->expand_file(tdb, wanted) == -1) {
590                 tdb_unlock_expand(tdb, F_WRLCK);
591                 return -1;
592         }
593
594         /* We need to drop this lock before adding free record. */
595         tdb_unlock_expand(tdb, F_WRLCK);
596
597         return add_free_record(tdb, old_size, wanted);
598 }
599
600 /* This won't fail: it will expand the database if it has to. */
601 tdb_off_t alloc(struct tdb_context *tdb, size_t keylen, size_t datalen,
602                 uint64_t hash, bool growing)
603 {
604         tdb_off_t off;
605
606         /* We can't hold pointers during this: we could unmap! */
607         assert(!tdb->direct_access);
608
609         for (;;) {
610                 off = get_free(tdb, keylen, datalen, growing, hash);
611                 if (likely(off != 0))
612                         break;
613
614                 if (tdb_expand(tdb, adjust_size(keylen, datalen)))
615                         return TDB_OFF_ERR;
616         }
617
618         return off;
619 }