2 Trivial Database 2: free list/block handling
3 Copyright (C) Rusty Russell 2010
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.
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.
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/>.
19 #include <ccan/likely/likely.h>
20 #include <ccan/ilog/ilog.h>
25 static unsigned fls64(uint64_t val)
30 /* In which bucket would we find a particular record size? (ignoring header) */
31 unsigned int size_to_bucket(tdb_len_t data_len)
35 /* We can't have records smaller than this. */
36 assert(data_len >= TDB_MIN_DATA_LEN);
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;
43 /* After that we go power of 2. */
44 bucket = fls64(data_len - TDB_MIN_DATA_LEN) + 2;
47 if (unlikely(bucket >= TDB_FREE_BUCKETS))
48 bucket = TDB_FREE_BUCKETS - 1;
52 tdb_off_t first_flist(struct tdb_context *tdb)
54 return tdb_read_off(tdb, offsetof(struct tdb_header, free_list));
57 tdb_off_t next_flist(struct tdb_context *tdb, tdb_off_t flist)
59 return tdb_read_off(tdb, flist + offsetof(struct tdb_freelist, next));
62 int tdb_flist_init(struct tdb_context *tdb)
64 /* Use reservoir sampling algorithm to select a free list at random. */
65 unsigned int rnd, max = 0, count = 0;
68 tdb->flist_off = off = first_flist(tdb);
72 if (off == TDB_OFF_ERR)
82 off = next_flist(tdb, off);
88 /* Offset of a given bucket. */
89 tdb_off_t bucket_off(tdb_off_t flist_off, unsigned bucket)
91 return flist_off + offsetof(struct tdb_freelist, buckets)
92 + bucket * sizeof(tdb_off_t);
95 /* Returns free_buckets + 1, or list number to search. */
96 static tdb_off_t find_free_head(struct tdb_context *tdb,
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);
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)
113 if (frec_prev(r) == 0) {
116 off = frec_prev(r) + offsetof(struct tdb_free_record, next);
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);
128 /* r->prev->next = r->next */
129 if (tdb_write_off(tdb, off, r->next)) {
134 off = r->next + offsetof(struct tdb_free_record,magic_and_prev);
135 /* r->next->prev = r->prev */
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);
146 if (tdb_write_off(tdb, off, r->magic_and_prev)) {
153 /* Enqueue in this free bucket. */
154 static int enqueue_in_free(struct tdb_context *tdb,
159 struct tdb_free_record new;
160 uint64_t magic = (TDB_FREE_MAGIC << (64 - TDB_OFF_UPPER_STEAL));
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))
166 new.magic_and_prev = magic;
168 /* new->next = head. */
169 new.next = tdb_read_off(tdb, b_off);
170 if (new.next == TDB_OFF_ERR)
175 if (tdb_read_off(tdb,
176 new.next + offsetof(struct tdb_free_record,
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);
185 /* next->prev = new. */
186 if (tdb_write_off(tdb, new.next
187 + offsetof(struct tdb_free_record,
193 if (tdb_write_off(tdb, b_off, off) != 0)
196 return tdb_write_convert(tdb, off, &new, sizeof(new));
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)
207 assert(len_with_header >= sizeof(struct tdb_free_record));
209 len = len_with_header - sizeof(struct tdb_used_record);
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)
215 ret = enqueue_in_free(tdb, b_off, off, len);
216 tdb_unlock_free_bucket(tdb, b_off);
220 static size_t adjust_size(size_t keylen, size_t datalen)
222 size_t size = keylen + datalen;
224 if (size < TDB_MIN_DATA_LEN)
225 size = TDB_MIN_DATA_LEN;
227 /* Round to next uint64_t boundary. */
228 return (size + (sizeof(uint64_t) - 1ULL)) & ~(sizeof(uint64_t) - 1ULL);
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)
238 datalen += datalen / 2;
239 leftover = total_len - adjust_size(keylen, datalen);
241 if (leftover < (ssize_t)sizeof(struct tdb_free_record))
247 /* FIXME: Shortcut common case where tdb->flist == flist */
248 static tdb_off_t flist_offset(struct tdb_context *tdb, unsigned int flist)
250 tdb_off_t off = first_flist(tdb);
253 for (i = 0; i < flist; i++)
254 off = next_flist(tdb, off);
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)
262 struct tdb_free_record pad, *r;
265 add_stat(tdb, alloc_coalesce_tried, 1);
266 end = off + sizeof(struct tdb_used_record) + data_len;
268 while (end < tdb->map_size) {
270 unsigned flist, bucket;
272 /* FIXME: do tdb_get here and below really win? */
273 r = tdb_get(tdb, end, &pad, sizeof(pad));
277 if (frec_magic(r) != TDB_FREE_MAGIC)
280 flist = frec_flist(r);
281 bucket = size_to_bucket(frec_len(r));
282 nb_off = bucket_off(flist_offset(tdb, flist), bucket);
284 /* We may be violating lock order here, so best effort. */
285 if (tdb_lock_free_bucket(tdb, nb_off, TDB_LOCK_NOWAIT) == -1) {
286 add_stat(tdb, alloc_coalesce_lockfail, 1);
290 /* Now we have lock, re-check. */
291 r = tdb_get(tdb, end, &pad, sizeof(pad));
293 tdb_unlock_free_bucket(tdb, nb_off);
297 if (unlikely(frec_magic(r) != TDB_FREE_MAGIC)) {
298 add_stat(tdb, alloc_coalesce_race, 1);
299 tdb_unlock_free_bucket(tdb, nb_off);
303 if (unlikely(frec_flist(r) != flist)
304 || unlikely(size_to_bucket(frec_len(r)) != bucket)) {
305 add_stat(tdb, alloc_coalesce_race, 1);
306 tdb_unlock_free_bucket(tdb, nb_off);
310 if (remove_from_list(tdb, nb_off, end, r) == -1) {
311 tdb_unlock_free_bucket(tdb, nb_off);
315 end += sizeof(struct tdb_used_record) + frec_len(r);
316 tdb_unlock_free_bucket(tdb, nb_off);
317 add_stat(tdb, alloc_coalesce_num_merged, 1);
320 /* Didn't find any adjacent free? */
321 if (end == off + sizeof(struct tdb_used_record) + data_len)
324 /* OK, expand record */
325 r = tdb_get(tdb, off, &pad, sizeof(pad));
329 if (frec_len(r) != data_len) {
330 tdb->ecode = TDB_ERR_CORRUPT;
331 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
332 "coalesce: expected data len %llu not %llu\n",
333 (long long)data_len, (long long)frec_len(r));
337 if (remove_from_list(tdb, b_off, off, r) == -1)
340 r = tdb_access_write(tdb, off, sizeof(*r), true);
344 /* We have to drop this to avoid deadlocks, so make sure record
345 * doesn't get coalesced by someone else! */
346 r->magic_and_prev = TDB_COALESCING_MAGIC << (64 - TDB_OFF_UPPER_STEAL);
347 /* FIXME: Use 255 as invalid free list? */
348 r->flist_and_len = end - off - sizeof(struct tdb_used_record);
349 if (tdb_access_commit(tdb, r) != 0)
352 add_stat(tdb, alloc_coalesce_succeeded, 1);
353 tdb_unlock_free_bucket(tdb, b_off);
355 if (add_free_record(tdb, off, end - off) == -1)
360 /* To unify error paths, we *always* unlock bucket on error. */
361 tdb_unlock_free_bucket(tdb, b_off);
365 /* We need size bytes to put our key and data in. */
366 static tdb_off_t lock_and_alloc(struct tdb_context *tdb,
369 size_t keylen, size_t datalen,
373 tdb_off_t off, b_off,best_off;
374 struct tdb_free_record pad, best = { 0 }, *r;
376 size_t size = adjust_size(keylen, datalen);
378 add_stat(tdb, allocs, 1);
380 b_off = bucket_off(flist_off, bucket);
382 /* FIXME: Try non-blocking wait first, to measure contention. */
383 /* Lock this bucket. */
384 if (tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT) == -1) {
388 best.flist_and_len = -1ULL;
391 /* Get slack if we're after extra. */
397 /* Walk the list to see if any are large enough, getting less fussy
399 off = tdb_read_off(tdb, b_off);
400 if (unlikely(off == TDB_OFF_ERR))
404 /* FIXME: Does tdb_get win anything here? */
405 r = tdb_get(tdb, off, &pad, sizeof(*r));
409 if (frec_magic(r) != TDB_FREE_MAGIC) {
410 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
411 "lock_and_alloc: %llu non-free 0x%llx\n",
412 (long long)off, (long long)r->magic_and_prev);
416 if (frec_len(r) >= size && frec_len(r) < frec_len(&best)) {
421 if (frec_len(&best) < size * multiplier && best_off)
426 /* Since we're going slow anyway, try coalescing here. */
427 switch (coalesce(tdb, off, b_off, frec_len(r))) {
429 /* This has already unlocked on error. */
432 /* This has unlocked list, restart. */
438 /* If we found anything at all, use it. */
440 struct tdb_used_record rec;
443 /* We're happy with this size: take it. */
444 if (remove_from_list(tdb, b_off, best_off, &best) != 0)
447 leftover = record_leftover(keylen, datalen, want_extra,
450 assert(keylen + datalen + leftover <= frec_len(&best));
451 /* We need to mark non-free before we drop lock, otherwise
452 * coalesce() could try to merge it! */
453 if (set_used_header(tdb, &rec, keylen, datalen,
454 frec_len(&best) - leftover,
458 if (tdb_write_convert(tdb, best_off, &rec, sizeof(rec)) != 0)
461 /* Bucket of leftover will be <= current bucket, so nested
462 * locking is allowed. */
464 add_stat(tdb, alloc_leftover, 1);
465 if (add_free_record(tdb,
466 best_off + sizeof(rec)
467 + frec_len(&best) - leftover,
469 best_off = TDB_OFF_ERR;
471 tdb_unlock_free_bucket(tdb, b_off);
476 tdb_unlock_free_bucket(tdb, b_off);
480 tdb_unlock_free_bucket(tdb, b_off);
484 /* Get a free block from current free list, or 0 if none. */
485 static tdb_off_t get_free(struct tdb_context *tdb,
486 size_t keylen, size_t datalen, bool want_extra,
489 tdb_off_t off, flist_off;
490 unsigned start_b, b, flist;
491 bool wrapped = false;
493 /* If they are growing, add 50% to get to higher bucket. */
495 start_b = size_to_bucket(adjust_size(keylen,
496 datalen + datalen / 2));
498 start_b = size_to_bucket(adjust_size(keylen, datalen));
500 flist_off = tdb->flist_off;
502 while (!wrapped || flist_off != tdb->flist_off) {
503 /* Start at exact size bucket, and search up... */
504 for (b = find_free_head(tdb, flist_off, start_b);
505 b < TDB_FREE_BUCKETS;
506 b = find_free_head(tdb, flist_off, b + 1)) {
507 /* Try getting one from list. */
508 off = lock_and_alloc(tdb, flist_off,
509 b, keylen, datalen, want_extra,
511 if (off == TDB_OFF_ERR)
515 add_stat(tdb, alloc_bucket_exact, 1);
516 if (b == TDB_FREE_BUCKETS - 1)
517 add_stat(tdb, alloc_bucket_max, 1);
518 /* Worked? Stay using this list. */
519 tdb->flist_off = flist_off;
523 /* Didn't work. Try next bucket. */
526 /* Hmm, try next list. */
527 flist_off = next_flist(tdb, flist_off);
530 if (flist_off == 0) {
532 flist_off = first_flist(tdb);
540 int set_used_header(struct tdb_context *tdb,
541 struct tdb_used_record *rec,
542 uint64_t keylen, uint64_t datalen,
543 uint64_t actuallen, unsigned hashlow)
545 uint64_t keybits = (fls64(keylen) + 1) / 2;
547 /* Use bottom bits of hash, so it's independent of hash table size. */
548 rec->magic_and_meta = (hashlow & ((1 << 11)-1))
549 | ((actuallen - (keylen + datalen)) << 11)
552 rec->key_and_data_len = (keylen | (datalen << (keybits*2)));
554 /* Encoding can fail on big values. */
555 if (rec_key_length(rec) != keylen
556 || rec_data_length(rec) != datalen
557 || rec_extra_padding(rec) != actuallen - (keylen + datalen)) {
558 tdb->ecode = TDB_ERR_IO;
559 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
560 "Could not encode k=%llu,d=%llu,a=%llu\n",
561 (long long)keylen, (long long)datalen,
562 (long long)actuallen);
568 /* Expand the database. */
569 static int tdb_expand(struct tdb_context *tdb, tdb_len_t size)
574 /* We need room for the record header too. */
575 wanted = sizeof(struct tdb_used_record) + size;
577 /* Need to hold a hash lock to expand DB: transactions rely on it. */
578 if (!(tdb->flags & TDB_NOLOCK)
579 && !tdb->allrecord_lock.count && !tdb_has_hash_locks(tdb)) {
580 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
581 "tdb_expand: must hold lock during expand\n");
585 /* always make room for at least 100 more records, and at
586 least 25% more space. */
587 if (size * TDB_EXTENSION_FACTOR > tdb->map_size / 4)
588 wanted = size * TDB_EXTENSION_FACTOR;
590 wanted = tdb->map_size / 4;
591 wanted = adjust_size(0, wanted);
593 /* Only one person can expand file at a time. */
594 if (tdb_lock_expand(tdb, F_WRLCK) != 0)
597 /* Someone else may have expanded the file, so retry. */
598 old_size = tdb->map_size;
599 tdb->methods->oob(tdb, tdb->map_size + 1, true);
600 if (tdb->map_size != old_size) {
601 tdb_unlock_expand(tdb, F_WRLCK);
605 if (tdb->methods->expand_file(tdb, wanted) == -1) {
606 tdb_unlock_expand(tdb, F_WRLCK);
610 /* We need to drop this lock before adding free record. */
611 tdb_unlock_expand(tdb, F_WRLCK);
613 add_stat(tdb, expands, 1);
614 return add_free_record(tdb, old_size, wanted);
617 /* This won't fail: it will expand the database if it has to. */
618 tdb_off_t alloc(struct tdb_context *tdb, size_t keylen, size_t datalen,
619 uint64_t hash, bool growing)
623 /* We can't hold pointers during this: we could unmap! */
624 assert(!tdb->direct_access);
627 off = get_free(tdb, keylen, datalen, growing, hash);
628 if (likely(off != 0))
631 if (tdb_expand(tdb, adjust_size(keylen, datalen)))