return ilog64(val);
}
-static unsigned ffs64(uint64_t val)
-{
-#if HAVE_BUILTIN_FFSLL
- return __builtin_ffsll(val);
-#else
- unsigned r = 0;
-
- if (!val)
- return 0;
-
- if (!(val & 0xffffffff)) {
- val >>= 32;
- r += 32;
- }
- if (!(val & 0xffff)) {
- val >>= 16;
- r += 16;
- }
- if (!(val & 0xff)) {
- val >>= 8;
- r += 8;
- }
- if (!(val & 0xf)) {
- val >>= 4;
- r += 4;
- }
- if (!(val & 0x3)) {
- val >>= 2;
- r += 2;
- }
- if (!(val & 0x1)) {
- val >>= 1;
- r += 1;
- }
- return r;
-#endif
-}
-
/* In which bucket would we find a particular record size? (ignoring header) */
-unsigned int size_to_bucket(unsigned int zone_bits, tdb_len_t data_len)
+unsigned int size_to_bucket(tdb_len_t data_len)
{
unsigned int bucket;
bucket = fls64(data_len - TDB_MIN_DATA_LEN) + 2;
}
- if (unlikely(bucket > BUCKETS_FOR_ZONE(zone_bits)))
- bucket = BUCKETS_FOR_ZONE(zone_bits);
+ if (unlikely(bucket >= TDB_FREE_BUCKETS))
+ bucket = TDB_FREE_BUCKETS - 1;
return bucket;
}
-/* Binary search for the zone for this offset. */
-static tdb_off_t off_to_zone(struct tdb_context *tdb, tdb_off_t off,
- struct free_zone_header *zhdr)
+tdb_off_t first_flist(struct tdb_context *tdb)
{
- tdb_off_t start, end;
-
- start = sizeof(struct tdb_header);
- end = start + (1ULL << fls64(tdb->map_size - start));
-
- for (;;) {
- if (tdb_read_convert(tdb, start, zhdr, sizeof(*zhdr)) == -1)
- return TDB_OFF_ERR;
-
- /* Is it inside this zone? */
- if (off < start + (1ULL << zhdr->zone_bits))
- return start;
-
- /* In practice, start + end won't overflow. */
- if (off >= (start + end) / 2)
- start = (start + end) / 2;
- else
- end = (start + end) / 2;
- }
+ return tdb_read_off(tdb, offsetof(struct tdb_header, free_list));
}
-static tdb_off_t last_zone(struct tdb_context *tdb,
- struct free_zone_header *zhdr)
+tdb_off_t next_flist(struct tdb_context *tdb, tdb_off_t flist)
{
- return off_to_zone(tdb, tdb->map_size - 1, zhdr);
+ return tdb_read_off(tdb, flist + offsetof(struct tdb_freelist, next));
}
-int tdb_zone_init(struct tdb_context *tdb)
+int tdb_flist_init(struct tdb_context *tdb)
{
- unsigned int i;
- uint64_t randoff = 0;
+ /* Use reservoir sampling algorithm to select a free list at random. */
+ unsigned int rnd, max = 0;
+ tdb_off_t off;
- /* We start in a random zone, to spread the load. */
- for (i = 0; i < 64; i += fls64(RAND_MAX))
- randoff ^= ((uint64_t)random()) << i;
- randoff = sizeof(struct tdb_header)
- + (randoff % (tdb->map_size - sizeof(struct tdb_header)));
+ tdb->flist_off = off = first_flist(tdb);
- tdb->zone_off = off_to_zone(tdb, randoff, &tdb->zhdr);
- if (tdb->zone_off == TDB_OFF_ERR)
- return -1;
- return 0;
-}
+ while (off) {
+ if (off == TDB_OFF_ERR)
+ return -1;
-/* Where's the header, given a zone size of 1 << zone_bits? */
-static tdb_off_t zone_off(tdb_off_t off, unsigned int zone_bits)
-{
- off -= sizeof(struct tdb_header);
- return (off & ~((1ULL << zone_bits) - 1)) + sizeof(struct tdb_header);
+ rnd = random();
+ if (rnd >= max) {
+ tdb->flist_off = off;
+ max = rnd;
+ }
+
+ off = next_flist(tdb, off);
+ }
+ return 0;
}
/* Offset of a given bucket. */
-/* FIXME: bucket can be "unsigned" everywhere, or even uint8/16. */
-tdb_off_t bucket_off(tdb_off_t zone_off, tdb_off_t bucket)
+tdb_off_t bucket_off(tdb_off_t flist_off, unsigned bucket)
{
- return zone_off
- + sizeof(struct free_zone_header)
+ return flist_off + offsetof(struct tdb_freelist, buckets)
+ bucket * sizeof(tdb_off_t);
}
/* Returns free_buckets + 1, or list number to search. */
-static tdb_off_t find_free_head(struct tdb_context *tdb, tdb_off_t bucket)
+static tdb_off_t find_free_head(struct tdb_context *tdb,
+ tdb_off_t flist_off,
+ tdb_off_t bucket)
{
/* Speculatively search for a non-zero bucket. */
- return tdb_find_nonzero_off(tdb, bucket_off(tdb->zone_off, 0),
- bucket,
- BUCKETS_FOR_ZONE(tdb->zhdr.zone_bits) + 1);
+ return tdb_find_nonzero_off(tdb, bucket_off(flist_off, 0),
+ bucket, TDB_FREE_BUCKETS);
}
/* Remove from free bucket. */
/* List need not be locked. */
int add_free_record(struct tdb_context *tdb,
- unsigned int zone_bits,
tdb_off_t off, tdb_len_t len_with_header)
{
struct tdb_free_record new;
int ret;
assert(len_with_header >= sizeof(new));
- assert(zone_bits < 64);
- new.magic_and_meta = TDB_FREE_MAGIC | zone_bits;
+ new.magic_and_meta = TDB_FREE_MAGIC << (64 - TDB_OFF_UPPER_STEAL)
+ | tdb->flist_off;
new.data_len = len_with_header - sizeof(struct tdb_used_record);
- b_off = bucket_off(zone_off(off, zone_bits),
- size_to_bucket(zone_bits, new.data_len));
+ b_off = bucket_off(tdb->flist_off, size_to_bucket(new.data_len));
if (tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT) != 0)
return -1;
return ret;
}
-static size_t adjust_size(size_t keylen, size_t datalen, bool want_extra)
+static size_t adjust_size(size_t keylen, size_t datalen)
{
size_t size = keylen + datalen;
- /* We want at least 50% growth for data. */
- if (want_extra)
- size += datalen/2;
-
if (size < TDB_MIN_DATA_LEN)
size = TDB_MIN_DATA_LEN;
{
ssize_t leftover;
- /* We might *want* extra, but not have it, so leftover is negative. */
- leftover = total_len - adjust_size(keylen, datalen, want_extra);
- if (leftover < (ssize_t)sizeof(struct tdb_free_record))
- return 0;
+ if (want_extra)
+ datalen += datalen / 2;
+ leftover = total_len - adjust_size(keylen, datalen);
- /* If we want extra anwyay, don't split unless we have 2x size. */
- if (want_extra && leftover <= datalen / 2)
+ if (leftover < (ssize_t)sizeof(struct tdb_free_record))
return 0;
return leftover;
/* Note: we unlock the current bucket if we coalesce or fail. */
static int coalesce(struct tdb_context *tdb,
- tdb_off_t zone_off, unsigned zone_bits,
tdb_off_t off, tdb_off_t b_off, tdb_len_t data_len)
{
struct tdb_free_record pad, *r;
- tdb_off_t zone_end, end;
+ tdb_off_t end;
end = off + sizeof(struct tdb_used_record) + data_len;
- zone_end = zone_off + (1ULL << zone_bits);
- if (tdb->methods->oob(tdb, zone_end, true))
- zone_end = tdb->map_size;
-
- while (end < zone_end) {
+ while (end < tdb->map_size) {
tdb_off_t nb_off;
/* FIXME: do tdb_get here and below really win? */
if (frec_magic(r) != TDB_FREE_MAGIC)
break;
- nb_off = bucket_off(zone_off,
- size_to_bucket(zone_bits, r->data_len));
+ nb_off = bucket_off(frec_flist(r), size_to_bucket(r->data_len));
/* We may be violating lock order here, so best effort. */
if (tdb_lock_free_bucket(tdb, nb_off, TDB_LOCK_NOWAIT) == -1)
break;
}
- if (unlikely(bucket_off(zone_off,
- size_to_bucket(zone_bits, r->data_len))
+ if (unlikely(bucket_off(frec_flist(r),
+ size_to_bucket(r->data_len))
!= nb_off)) {
tdb_unlock_free_bucket(tdb, nb_off);
break;
/* We have to drop this to avoid deadlocks, so make sure record
* doesn't get coalesced by someone else! */
- r->magic_and_meta = TDB_COALESCING_MAGIC | zone_bits;
+ r->magic_and_meta = TDB_COALESCING_MAGIC << (64 - TDB_OFF_UPPER_STEAL);
r->data_len = end - off - sizeof(struct tdb_used_record);
if (tdb_access_commit(tdb, r) != 0)
goto err;
tdb_unlock_free_bucket(tdb, b_off);
- if (add_free_record(tdb, zone_bits, off, end - off) == -1)
+ if (add_free_record(tdb, off, end - off) == -1)
return -1;
return 1;
/* We need size bytes to put our key and data in. */
static tdb_off_t lock_and_alloc(struct tdb_context *tdb,
- tdb_off_t zone_off,
- unsigned zone_bits,
+ tdb_off_t flist_off,
tdb_off_t bucket,
size_t keylen, size_t datalen,
bool want_extra,
tdb_off_t off, b_off,best_off;
struct tdb_free_record pad, best = { 0 }, *r;
double multiplier;
- size_t size = keylen + datalen;
+ size_t size = adjust_size(keylen, datalen);
again:
- b_off = bucket_off(zone_off, bucket);
+ b_off = bucket_off(flist_off, bucket);
- /* FIXME: Try non-blocking wait first, to measure contention.
- * If we're contented, try switching zones, and don't enlarge zone
- * next time (we want more zones). */
+ /* FIXME: Try non-blocking wait first, to measure contention. */
/* Lock this bucket. */
if (tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT) == -1) {
return TDB_OFF_ERR;
multiplier *= 1.01;
/* Since we're going slow anyway, try coalescing here. */
- switch (coalesce(tdb, zone_off, zone_bits, off, b_off,
- r->data_len)) {
+ switch (coalesce(tdb, off, b_off, r->data_len)) {
case -1:
/* This has already unlocked on error. */
return -1;
leftover = record_leftover(keylen, datalen, want_extra,
best.data_len);
+ assert(keylen + datalen + leftover <= best.data_len);
/* We need to mark non-free before we drop lock, otherwise
* coalesce() could try to merge it! */
- if (set_header(tdb, &rec, keylen, datalen,
- best.data_len - leftover,
- hashlow, zone_bits) != 0)
+ if (set_used_header(tdb, &rec, keylen, datalen,
+ best.data_len - leftover,
+ hashlow) != 0)
goto unlock_err;
if (tdb_write_convert(tdb, best_off, &rec, sizeof(rec)) != 0)
tdb_unlock_free_bucket(tdb, b_off);
if (leftover) {
- if (add_free_record(tdb, zone_bits,
+ if (add_free_record(tdb,
best_off + sizeof(rec)
+ best.data_len - leftover,
leftover))
return TDB_OFF_ERR;
}
-static bool next_zone(struct tdb_context *tdb)
-{
- tdb_off_t next = tdb->zone_off + (1ULL << tdb->zhdr.zone_bits);
-
- /* We must have a header. */
- if (tdb->methods->oob(tdb, next + sizeof(tdb->zhdr), true))
- return false;
-
- tdb->zone_off = next;
- return tdb_read_convert(tdb, next, &tdb->zhdr, sizeof(tdb->zhdr)) == 0;
-}
-
-/* Offset returned is within current zone (which it may alter). */
+/* Get a free block from current free list, or 0 if none. */
static tdb_off_t get_free(struct tdb_context *tdb,
size_t keylen, size_t datalen, bool want_extra,
unsigned hashlow)
{
- tdb_off_t start_zone = tdb->zone_off, off;
+ tdb_off_t off, flist;
+ unsigned start_b, b;
bool wrapped = false;
- size_t size = adjust_size(keylen, datalen, want_extra);
/* If they are growing, add 50% to get to higher bucket. */
if (want_extra)
- size += datalen / 2;
-
- /* FIXME: If we don't get a hit in the first bucket we want,
- * try changing zones for next time. That should help wear
- * zones evenly, so we don't need to search all of them before
- * expanding. */
- while (!wrapped || tdb->zone_off != start_zone) {
- tdb_off_t b;
-
- /* Shortcut for really huge allocations... */
- if ((size >> tdb->zhdr.zone_bits) != 0)
- goto next;
+ start_b = size_to_bucket(adjust_size(keylen,
+ datalen + datalen / 2));
+ else
+ start_b = size_to_bucket(adjust_size(keylen, datalen));
+ flist = tdb->flist_off;
+ while (!wrapped || flist != tdb->flist_off) {
/* Start at exact size bucket, and search up... */
- b = size_to_bucket(tdb->zhdr.zone_bits, size);
- for (b = find_free_head(tdb, b);
- b <= BUCKETS_FOR_ZONE(tdb->zhdr.zone_bits);
- b += find_free_head(tdb, b + 1)) {
+ for (b = find_free_head(tdb, flist, start_b);
+ b < TDB_FREE_BUCKETS;
+ b = find_free_head(tdb, flist, b + 1)) {
/* Try getting one from list. */
- off = lock_and_alloc(tdb, tdb->zone_off,
- tdb->zhdr.zone_bits,
+ off = lock_and_alloc(tdb, flist,
b, keylen, datalen, want_extra,
hashlow);
if (off == TDB_OFF_ERR)
return TDB_OFF_ERR;
- if (off != 0)
+ if (off != 0) {
+ /* Worked? Stay using this list. */
+ tdb->flist_off = flist;
return off;
+ }
/* Didn't work. Try next bucket. */
}
- next:
- /* Didn't work, try next zone, if it exists. */
- if (!next_zone(tdb)) {
+ /* Hmm, try next list. */
+ flist = next_flist(tdb, flist);
+ if (flist == 0) {
wrapped = true;
- tdb->zone_off = sizeof(struct tdb_header);
- if (tdb_read_convert(tdb, tdb->zone_off,
- &tdb->zhdr, sizeof(tdb->zhdr))) {
- return TDB_OFF_ERR;
- }
+ flist = first_flist(tdb);
}
}
+
return 0;
}
-int set_header(struct tdb_context *tdb,
- struct tdb_used_record *rec,
- uint64_t keylen, uint64_t datalen,
- uint64_t actuallen, unsigned hashlow,
- unsigned int zone_bits)
+int set_used_header(struct tdb_context *tdb,
+ struct tdb_used_record *rec,
+ uint64_t keylen, uint64_t datalen,
+ uint64_t actuallen, unsigned hashlow)
{
uint64_t keybits = (fls64(keylen) + 1) / 2;
/* Use bottom bits of hash, so it's independent of hash table size. */
- rec->magic_and_meta
- = zone_bits
- | ((hashlow & ((1 << 5)-1)) << 6)
+ rec->magic_and_meta = (hashlow & ((1 << 11)-1))
| ((actuallen - (keylen + datalen)) << 11)
| (keybits << 43)
| (TDB_MAGIC << 48);
return 0;
}
-static bool zones_contended(struct tdb_context *tdb)
-{
- return false;
-}
-
-/* Assume we want buckets up to the comfort factor. */
-static tdb_len_t overhead(unsigned int zone_bits)
-{
- return sizeof(struct free_zone_header)
- + (BUCKETS_FOR_ZONE(zone_bits) + 1) * sizeof(tdb_off_t);
-}
-
-/* Expand the database (by adding a zone). */
+/* Expand the database. */
static int tdb_expand(struct tdb_context *tdb, tdb_len_t size)
{
uint64_t old_size;
- tdb_off_t off;
- unsigned int num_buckets, zone_bits;
- tdb_len_t wanted, expand;
- struct free_zone_header zhdr;
+ tdb_len_t wanted;
/* We need room for the record header too. */
wanted = sizeof(struct tdb_used_record) + size;
+ /* Need to hold a hash lock to expand DB: transactions rely on it. */
+ if (!(tdb->flags & TDB_NOLOCK)
+ && !tdb->allrecord_lock.count && !tdb_has_hash_locks(tdb)) {
+ tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
+ "tdb_expand: must hold lock during expand\n");
+ return -1;
+ }
+
/* Only one person can expand file at a time. */
if (tdb_lock_expand(tdb, F_WRLCK) != 0)
return -1;
/* Someone else may have expanded the file, so retry. */
old_size = tdb->map_size;
tdb->methods->oob(tdb, tdb->map_size + 1, true);
- if (tdb->map_size != old_size)
- goto success;
-
- /* Treat last zone as minimum reasonable zone size. */
- off = last_zone(tdb, &zhdr);
- if (off == TDB_OFF_ERR)
- goto fail;
-
- /* Zone isn't fully expanded? */
- if (tdb->map_size < off + (1ULL << zhdr.zone_bits)) {
- expand = off + (1ULL << zhdr.zone_bits) - tdb->map_size;
- /* Expand more than we want. */
- if (expand > (wanted << TDB_COMFORT_FACTOR_BITS))
- expand = (wanted << TDB_COMFORT_FACTOR_BITS);
- if (tdb->methods->expand_file(tdb, expand) == -1)
- goto fail;
- /* We need to drop this lock before adding free record. */
+ if (tdb->map_size != old_size) {
tdb_unlock_expand(tdb, F_WRLCK);
-
- /* Allocate from here. */
- tdb->zone_off = off;
- tdb->zhdr = zhdr;
-
- /* FIXME: If this isn't sufficient, we search again... */
- return add_free_record(tdb, zhdr.zone_bits,
- tdb->map_size - expand, expand);
+ return 0;
}
- /* We are never allowed to cross a power-of-two boundary, and our
- * minimum zone size is 1 << INITIAL_ZONE_BITS.
- *
- * If our filesize is 128k, we can add a 64k or a 128k zone. If it's
- * 192k, we can only add a 64k zone.
- *
- * In other words, our max zone size is (1 << (ffs(filesize) - 1)) */
- zone_bits = ffs64(old_size - sizeof(struct tdb_header)) - 1;
- assert(zone_bits >= INITIAL_ZONE_BITS);
-
- /* Big zones generally good, but more zones wanted if contended. */
- if (zones_contended(tdb)) {
- /* If it suffices, make zone same size as last one. */
- if (zhdr.zone_bits < zone_bits
- && (1ULL << zhdr.zone_bits) >= overhead(zone_bits)+wanted)
- zone_bits = zhdr.zone_bits;
+ if (tdb->methods->expand_file(tdb, wanted*TDB_EXTENSION_FACTOR) == -1) {
+ tdb_unlock_expand(tdb, F_WRLCK);
+ return -1;
}
- zhdr.zone_bits = zone_bits;
- num_buckets = BUCKETS_FOR_ZONE(zone_bits);
-
- /* Expand the file by more than we need right now. */
- expand = 1ULL << zone_bits;
- if (expand > overhead(zone_bits) + (wanted << TDB_COMFORT_FACTOR_BITS))
- expand = overhead(zone_bits)
- + (wanted << TDB_COMFORT_FACTOR_BITS);
-
- if (tdb->methods->expand_file(tdb, expand) == -1)
- goto fail;
-
- /* Write new zone header (at old end). */
- off = old_size;
- if (tdb_write_convert(tdb, off, &zhdr, sizeof(zhdr)) == -1)
- goto fail;
-
- /* Now write empty buckets. */
- off += sizeof(zhdr);
- if (zero_out(tdb, off, (num_buckets+1) * sizeof(tdb_off_t)) == -1)
- goto fail;
- off += (num_buckets+1) * sizeof(tdb_off_t);
-
- /* Now add the rest as our free record. */
- if (add_free_record(tdb, zone_bits, off, expand - overhead(zone_bits))
- == -1)
- goto fail;
-
- /* Try allocating from this zone now. */
- tdb->zone_off = old_size;
- tdb->zhdr = zhdr;
-
-success:
+ /* We need to drop this lock before adding free record. */
tdb_unlock_expand(tdb, F_WRLCK);
- return 0;
-fail:
- tdb_unlock_expand(tdb, F_WRLCK);
- return -1;
+ return add_free_record(tdb, old_size, wanted * TDB_EXTENSION_FACTOR);
}
/* This won't fail: it will expand the database if it has to. */
if (likely(off != 0))
break;
- if (tdb_expand(tdb, adjust_size(keylen, datalen, growing)))
+ if (tdb_expand(tdb, adjust_size(keylen, datalen)))
return TDB_OFF_ERR;
}
projects / ccan / blobdiff