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_ftable(struct tdb_context *tdb)
54 return tdb_read_off(tdb, offsetof(struct tdb_header, free_table));
57 tdb_off_t next_ftable(struct tdb_context *tdb, tdb_off_t ftable)
59 return tdb_read_off(tdb, ftable + offsetof(struct tdb_freetable,next));
62 int tdb_ftable_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->ftable_off = off = first_ftable(tdb);
72 if (off == TDB_OFF_ERR)
77 tdb->ftable_off = off;
82 off = next_ftable(tdb, off);
88 /* Offset of a given bucket. */
89 tdb_off_t bucket_off(tdb_off_t ftable_off, unsigned bucket)
91 return ftable_off + offsetof(struct tdb_freetable, 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(ftable_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 const 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_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
122 "remove_from_list: %llu bad prev in list %llu",
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_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
140 "remove_from_list: %llu bad list %llu",
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 ftable_and_len; rest is set in enqueue_in_free */
163 new.ftable_and_len = ((uint64_t)tdb->ftable << (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_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
180 "enqueue_in_free: %llu bad head"
182 (long long)new.next, (long long)b_off);
186 /* next->prev = new. */
187 if (tdb_write_off(tdb, new.next
188 + offsetof(struct tdb_free_record,
194 if (tdb_write_off(tdb, b_off, off) != 0)
197 return tdb_write_convert(tdb, off, &new, sizeof(new));
200 /* List need not be locked. */
201 int add_free_record(struct tdb_context *tdb,
202 tdb_off_t off, tdb_len_t len_with_header)
208 assert(len_with_header >= sizeof(struct tdb_free_record));
210 len = len_with_header - sizeof(struct tdb_used_record);
212 b_off = bucket_off(tdb->ftable_off, size_to_bucket(len));
213 if (tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT) != 0)
216 ret = enqueue_in_free(tdb, b_off, off, len);
217 tdb_unlock_free_bucket(tdb, b_off);
221 static size_t adjust_size(size_t keylen, size_t datalen)
223 size_t size = keylen + datalen;
225 if (size < TDB_MIN_DATA_LEN)
226 size = TDB_MIN_DATA_LEN;
228 /* Round to next uint64_t boundary. */
229 return (size + (sizeof(uint64_t) - 1ULL)) & ~(sizeof(uint64_t) - 1ULL);
232 /* If we have enough left over to be useful, split that off. */
233 static size_t record_leftover(size_t keylen, size_t datalen,
234 bool want_extra, size_t total_len)
239 datalen += datalen / 2;
240 leftover = total_len - adjust_size(keylen, datalen);
242 if (leftover < (ssize_t)sizeof(struct tdb_free_record))
248 static tdb_off_t ftable_offset(struct tdb_context *tdb, unsigned int ftable)
253 if (likely(tdb->ftable == ftable))
254 return tdb->ftable_off;
256 off = first_ftable(tdb);
257 for (i = 0; i < ftable; i++)
258 off = next_ftable(tdb, off);
262 /* Note: we unlock the current bucket if we coalesce or fail. */
263 static int coalesce(struct tdb_context *tdb,
264 tdb_off_t off, tdb_off_t b_off, tdb_len_t data_len)
267 struct tdb_free_record rec;
269 add_stat(tdb, alloc_coalesce_tried, 1);
270 end = off + sizeof(struct tdb_used_record) + data_len;
272 while (end < tdb->map_size) {
273 const struct tdb_free_record *r;
275 unsigned ftable, bucket;
277 r = tdb_access_read(tdb, end, sizeof(*r), true);
281 if (frec_magic(r) != TDB_FREE_MAGIC
282 || frec_ftable(r) == TDB_FTABLE_NONE) {
283 tdb_access_release(tdb, r);
287 ftable = frec_ftable(r);
288 bucket = size_to_bucket(frec_len(r));
289 nb_off = bucket_off(ftable_offset(tdb, ftable), bucket);
290 tdb_access_release(tdb, r);
292 /* We may be violating lock order here, so best effort. */
293 if (tdb_lock_free_bucket(tdb, nb_off, TDB_LOCK_NOWAIT) == -1) {
294 add_stat(tdb, alloc_coalesce_lockfail, 1);
298 /* Now we have lock, re-check. */
299 if (tdb_read_convert(tdb, end, &rec, sizeof(rec))) {
300 tdb_unlock_free_bucket(tdb, nb_off);
304 if (unlikely(frec_magic(&rec) != TDB_FREE_MAGIC)) {
305 add_stat(tdb, alloc_coalesce_race, 1);
306 tdb_unlock_free_bucket(tdb, nb_off);
310 if (unlikely(frec_ftable(&rec) != ftable)
311 || unlikely(size_to_bucket(frec_len(&rec)) != bucket)) {
312 add_stat(tdb, alloc_coalesce_race, 1);
313 tdb_unlock_free_bucket(tdb, nb_off);
317 if (remove_from_list(tdb, nb_off, end, &rec) == -1) {
318 tdb_unlock_free_bucket(tdb, nb_off);
322 end += sizeof(struct tdb_used_record) + frec_len(&rec);
323 tdb_unlock_free_bucket(tdb, nb_off);
324 add_stat(tdb, alloc_coalesce_num_merged, 1);
327 /* Didn't find any adjacent free? */
328 if (end == off + sizeof(struct tdb_used_record) + data_len)
331 /* OK, expand initial record */
332 if (tdb_read_convert(tdb, off, &rec, sizeof(rec)))
335 if (frec_len(&rec) != data_len) {
336 tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
337 "coalesce: expected data len %zu not %zu",
338 (size_t)data_len, (size_t)frec_len(&rec));
342 if (remove_from_list(tdb, b_off, off, &rec) == -1)
345 /* We have to drop this to avoid deadlocks, so make sure record
346 * doesn't get coalesced by someone else! */
347 rec.ftable_and_len = (TDB_FTABLE_NONE << (64 - TDB_OFF_UPPER_STEAL))
348 | (end - off - sizeof(struct tdb_used_record));
349 if (tdb_write_off(tdb, off + offsetof(struct tdb_free_record,
351 rec.ftable_and_len) != 0)
354 add_stat(tdb, alloc_coalesce_succeeded, 1);
355 tdb_unlock_free_bucket(tdb, b_off);
357 if (add_free_record(tdb, off, end - off) == -1)
362 /* To unify error paths, we *always* unlock bucket on error. */
363 tdb_unlock_free_bucket(tdb, b_off);
367 /* We need size bytes to put our key and data in. */
368 static tdb_off_t lock_and_alloc(struct tdb_context *tdb,
369 tdb_off_t ftable_off,
371 size_t keylen, size_t datalen,
376 tdb_off_t off, b_off,best_off;
377 struct tdb_free_record best = { 0 };
379 size_t size = adjust_size(keylen, datalen);
381 add_stat(tdb, allocs, 1);
383 b_off = bucket_off(ftable_off, bucket);
385 /* FIXME: Try non-blocking wait first, to measure contention. */
386 /* Lock this bucket. */
387 if (tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT) == -1) {
391 best.ftable_and_len = -1ULL;
394 /* Get slack if we're after extra. */
400 /* Walk the list to see if any are large enough, getting less fussy
402 off = tdb_read_off(tdb, b_off);
403 if (unlikely(off == TDB_OFF_ERR))
407 const struct tdb_free_record *r;
411 r = tdb_access_read(tdb, off, sizeof(*r), true);
415 if (frec_magic(r) != TDB_FREE_MAGIC) {
416 tdb_access_release(tdb, r);
417 tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
418 "lock_and_alloc: %llu non-free 0x%llx",
419 (long long)off, (long long)r->magic_and_prev);
423 if (frec_len(r) >= size && frec_len(r) < frec_len(&best)) {
428 if (frec_len(&best) < size * multiplier && best_off) {
429 tdb_access_release(tdb, r);
437 tdb_access_release(tdb, r);
439 /* Since we're going slow anyway, try coalescing here. */
440 switch (coalesce(tdb, off, b_off, len)) {
442 /* This has already unlocked on error. */
445 /* This has unlocked list, restart. */
451 /* If we found anything at all, use it. */
453 struct tdb_used_record rec;
456 /* We're happy with this size: take it. */
457 if (remove_from_list(tdb, b_off, best_off, &best) != 0)
460 leftover = record_leftover(keylen, datalen, want_extra,
463 assert(keylen + datalen + leftover <= frec_len(&best));
464 /* We need to mark non-free before we drop lock, otherwise
465 * coalesce() could try to merge it! */
466 if (set_header(tdb, &rec, magic, keylen, datalen,
467 frec_len(&best) - leftover, hashlow) != 0)
470 if (tdb_write_convert(tdb, best_off, &rec, sizeof(rec)) != 0)
473 /* Bucket of leftover will be <= current bucket, so nested
474 * locking is allowed. */
476 add_stat(tdb, alloc_leftover, 1);
477 if (add_free_record(tdb,
478 best_off + sizeof(rec)
479 + frec_len(&best) - leftover,
481 best_off = TDB_OFF_ERR;
483 tdb_unlock_free_bucket(tdb, b_off);
488 tdb_unlock_free_bucket(tdb, b_off);
492 tdb_unlock_free_bucket(tdb, b_off);
496 /* Get a free block from current free list, or 0 if none. */
497 static tdb_off_t get_free(struct tdb_context *tdb,
498 size_t keylen, size_t datalen, bool want_extra,
499 unsigned magic, unsigned hashlow)
501 tdb_off_t off, ftable_off;
502 unsigned start_b, b, ftable;
503 bool wrapped = false;
505 /* If they are growing, add 50% to get to higher bucket. */
507 start_b = size_to_bucket(adjust_size(keylen,
508 datalen + datalen / 2));
510 start_b = size_to_bucket(adjust_size(keylen, datalen));
512 ftable_off = tdb->ftable_off;
513 ftable = tdb->ftable;
514 while (!wrapped || ftable_off != tdb->ftable_off) {
515 /* Start at exact size bucket, and search up... */
516 for (b = find_free_head(tdb, ftable_off, start_b);
517 b < TDB_FREE_BUCKETS;
518 b = find_free_head(tdb, ftable_off, b + 1)) {
519 /* Try getting one from list. */
520 off = lock_and_alloc(tdb, ftable_off,
521 b, keylen, datalen, want_extra,
523 if (off == TDB_OFF_ERR)
527 add_stat(tdb, alloc_bucket_exact, 1);
528 if (b == TDB_FREE_BUCKETS - 1)
529 add_stat(tdb, alloc_bucket_max, 1);
530 /* Worked? Stay using this list. */
531 tdb->ftable_off = ftable_off;
532 tdb->ftable = ftable;
535 /* Didn't work. Try next bucket. */
538 /* Hmm, try next table. */
539 ftable_off = next_ftable(tdb, ftable_off);
542 if (ftable_off == 0) {
544 ftable_off = first_ftable(tdb);
552 int set_header(struct tdb_context *tdb,
553 struct tdb_used_record *rec,
554 unsigned magic, uint64_t keylen, uint64_t datalen,
555 uint64_t actuallen, unsigned hashlow)
557 uint64_t keybits = (fls64(keylen) + 1) / 2;
559 /* Use bottom bits of hash, so it's independent of hash table size. */
560 rec->magic_and_meta = (hashlow & ((1 << 11)-1))
561 | ((actuallen - (keylen + datalen)) << 11)
563 | ((uint64_t)magic << 48);
564 rec->key_and_data_len = (keylen | (datalen << (keybits*2)));
566 /* Encoding can fail on big values. */
567 if (rec_key_length(rec) != keylen
568 || rec_data_length(rec) != datalen
569 || rec_extra_padding(rec) != actuallen - (keylen + datalen)) {
570 tdb_logerr(tdb, TDB_ERR_IO, TDB_DEBUG_ERROR,
571 "Could not encode k=%llu,d=%llu,a=%llu",
572 (long long)keylen, (long long)datalen,
573 (long long)actuallen);
579 /* Expand the database. */
580 static int tdb_expand(struct tdb_context *tdb, tdb_len_t size)
585 /* We need room for the record header too. */
586 wanted = sizeof(struct tdb_used_record) + size;
588 /* Need to hold a hash lock to expand DB: transactions rely on it. */
589 if (!(tdb->flags & TDB_NOLOCK)
590 && !tdb->allrecord_lock.count && !tdb_has_hash_locks(tdb)) {
591 tdb_logerr(tdb, TDB_ERR_LOCK, TDB_DEBUG_ERROR,
592 "tdb_expand: must hold lock during expand");
596 /* always make room for at least 100 more records, and at
597 least 25% more space. */
598 if (size * TDB_EXTENSION_FACTOR > tdb->map_size / 4)
599 wanted = size * TDB_EXTENSION_FACTOR;
601 wanted = tdb->map_size / 4;
602 wanted = adjust_size(0, wanted);
604 /* Only one person can expand file at a time. */
605 if (tdb_lock_expand(tdb, F_WRLCK) != 0)
608 /* Someone else may have expanded the file, so retry. */
609 old_size = tdb->map_size;
610 tdb->methods->oob(tdb, tdb->map_size + 1, true);
611 if (tdb->map_size != old_size) {
612 tdb_unlock_expand(tdb, F_WRLCK);
616 if (tdb->methods->expand_file(tdb, wanted) == -1) {
617 tdb_unlock_expand(tdb, F_WRLCK);
621 /* We need to drop this lock before adding free record. */
622 tdb_unlock_expand(tdb, F_WRLCK);
624 add_stat(tdb, expands, 1);
625 return add_free_record(tdb, old_size, wanted);
628 /* This won't fail: it will expand the database if it has to. */
629 tdb_off_t alloc(struct tdb_context *tdb, size_t keylen, size_t datalen,
630 uint64_t hash, unsigned magic, bool growing)
634 /* We can't hold pointers during this: we could unmap! */
635 assert(!tdb->direct_access);
638 off = get_free(tdb, keylen, datalen, growing, magic, hash);
639 if (likely(off != 0))
642 if (tdb_expand(tdb, adjust_size(keylen, datalen)))