*/
#include "private.h"
#include <ccan/likely/likely.h>
+#include <ccan/ilog/ilog.h>
#include <time.h>
#include <assert.h>
#include <limits.h>
-/* We have to be able to fit a free record here. */
-#define MIN_DATA_LEN \
- (sizeof(struct tdb_free_record) - sizeof(struct tdb_used_record))
-
static unsigned fls64(uint64_t val)
{
-#if HAVE_BUILTIN_CLZL
- if (val <= ULONG_MAX) {
- /* This is significantly faster! */
- return val ? sizeof(long) * CHAR_BIT - __builtin_clzl(val) : 0;
- } else {
-#endif
- uint64_t r = 64;
+ 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 & 0xffffffff00000000ull)) {
- val <<= 32;
- r -= 32;
+
+ if (!(val & 0xffffffff)) {
+ val >>= 32;
+ r += 32;
}
- if (!(val & 0xffff000000000000ull)) {
- val <<= 16;
- r -= 16;
+ if (!(val & 0xffff)) {
+ val >>= 16;
+ r += 16;
}
- if (!(val & 0xff00000000000000ull)) {
- val <<= 8;
- r -= 8;
+ if (!(val & 0xff)) {
+ val >>= 8;
+ r += 8;
}
- if (!(val & 0xf000000000000000ull)) {
- val <<= 4;
- r -= 4;
+ if (!(val & 0xf)) {
+ val >>= 4;
+ r += 4;
}
- if (!(val & 0xc000000000000000ull)) {
- val <<= 2;
- r -= 2;
+ if (!(val & 0x3)) {
+ val >>= 2;
+ r += 2;
}
- if (!(val & 0x8000000000000000ull)) {
- val <<= 1;
- r -= 1;
+ if (!(val & 0x1)) {
+ val >>= 1;
+ r += 1;
}
return r;
-#if HAVE_BUILTIN_CLZL
- }
#endif
}
unsigned int bucket;
/* We can't have records smaller than this. */
- assert(data_len >= MIN_DATA_LEN);
+ assert(data_len >= TDB_MIN_DATA_LEN);
/* Ignoring the header... */
- if (data_len - MIN_DATA_LEN <= 64) {
- /* 0 in bucket 0, 8 in bucket 1... 64 in bucket 6. */
- bucket = (data_len - MIN_DATA_LEN) / 8;
+ if (data_len - TDB_MIN_DATA_LEN <= 64) {
+ /* 0 in bucket 0, 8 in bucket 1... 64 in bucket 8. */
+ bucket = (data_len - TDB_MIN_DATA_LEN) / 8;
} else {
/* After that we go power of 2. */
- bucket = fls64(data_len - MIN_DATA_LEN) + 2;
+ bucket = fls64(data_len - TDB_MIN_DATA_LEN) + 2;
}
if (unlikely(bucket > BUCKETS_FOR_ZONE(zone_bits)))
return bucket;
}
-/* Subtract 1-byte tailer and header. Then round up to next power of 2. */
-static unsigned max_zone_bits(struct tdb_context *tdb)
-{
- return fls64(tdb->map_size-1-sizeof(struct tdb_header)-1) + 1;
-}
-
-/* Start by using a random zone to spread the load: returns the offset. */
-static uint64_t random_zone(struct tdb_context *tdb)
+/* 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)
{
- struct free_zone_header zhdr;
- tdb_off_t off = sizeof(struct tdb_header);
- tdb_len_t half_bits;
- uint64_t randbits = 0;
- unsigned int i;
+ tdb_off_t start, end;
- for (i = 0; i < 64; i += fls64(RAND_MAX))
- randbits ^= ((uint64_t)random()) << i;
-
- /* FIXME: Does this work? Test! */
- half_bits = max_zone_bits(tdb) - 1;
- do {
- /* Pick left or right side (not outside file) */
- if ((randbits & 1)
- && !tdb->methods->oob(tdb, off + (1ULL << half_bits)
- + sizeof(zhdr), true)) {
- off += 1ULL << half_bits;
- }
- randbits >>= 1;
+ start = sizeof(struct tdb_header);
+ end = start + (1ULL << fls64(tdb->map_size - start));
- if (tdb_read_convert(tdb, off, &zhdr, sizeof(zhdr)) == -1)
+ for (;;) {
+ if (tdb_read_convert(tdb, start, zhdr, sizeof(*zhdr)) == -1)
return TDB_OFF_ERR;
- if (zhdr.zone_bits == half_bits)
- return off;
+ /* Is it inside this zone? */
+ if (off < start + (1ULL << zhdr->zone_bits))
+ return start;
- half_bits--;
- } while (half_bits >= INITIAL_ZONE_BITS);
+ /* In practice, start + end won't overflow. */
+ if (off >= (start + end) / 2)
+ start = (start + end) / 2;
+ else
+ end = (start + end) / 2;
+ }
+}
- tdb->ecode = TDB_ERR_CORRUPT;
- tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
- "random_zone: zone at %llu smaller than %u bits?",
- (long long)off, INITIAL_ZONE_BITS);
- return TDB_OFF_ERR;
+static tdb_off_t last_zone(struct tdb_context *tdb,
+ struct free_zone_header *zhdr)
+{
+ return off_to_zone(tdb, tdb->map_size - 1, zhdr);
}
int tdb_zone_init(struct tdb_context *tdb)
{
- tdb->zone_off = random_zone(tdb);
+ unsigned int i;
+ uint64_t randoff = 0;
+
+ /* 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->zone_off = off_to_zone(tdb, randoff, &tdb->zhdr);
if (tdb->zone_off == TDB_OFF_ERR)
return -1;
- if (tdb_read_convert(tdb, tdb->zone_off,
- &tdb->zhdr, sizeof(tdb->zhdr)) == -1)
- return -1;
return 0;
}
/* Returns free_buckets + 1, or list number to search. */
static tdb_off_t find_free_head(struct tdb_context *tdb, tdb_off_t bucket)
{
- tdb_off_t b;
-
/* Speculatively search for a non-zero bucket. */
- b = tdb_find_nonzero_off(tdb, bucket_off(tdb->zone_off, bucket),
- BUCKETS_FOR_ZONE(tdb->zhdr.zone_bits) + 1
- - bucket);
- return bucket + b;
+ return tdb_find_nonzero_off(tdb, bucket_off(tdb->zone_off, 0),
+ bucket,
+ BUCKETS_FOR_ZONE(tdb->zhdr.zone_bits) + 1);
}
/* Remove from free bucket. */
static int remove_from_list(struct tdb_context *tdb,
- tdb_off_t b_off, struct tdb_free_record *r)
+ tdb_off_t b_off, tdb_off_t r_off,
+ struct tdb_free_record *r)
{
tdb_off_t off;
} else {
off = r->prev + offsetof(struct tdb_free_record, next);
}
+
+#ifdef DEBUG
+ if (tdb_read_off(tdb, off) != r_off) {
+ tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
+ "remove_from_list: %llu bad prev in list %llu\n",
+ (long long)r_off, (long long)b_off);
+ return -1;
+ }
+#endif
+
/* r->prev->next = r->next */
if (tdb_write_off(tdb, off, r->next)) {
return -1;
if (r->next != 0) {
off = r->next + offsetof(struct tdb_free_record, prev);
/* r->next->prev = r->prev */
+
+#ifdef DEBUG
+ if (tdb_read_off(tdb, off) != r_off) {
+ tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
+ "remove_from_list: %llu bad list %llu\n",
+ (long long)r_off, (long long)b_off);
+ return -1;
+ }
+#endif
+
if (tdb_write_off(tdb, off, r->prev)) {
return -1;
}
return -1;
if (new->next) {
+#ifdef DEBUG
+ if (tdb_read_off(tdb,
+ new->next
+ + offsetof(struct tdb_free_record, prev))
+ != 0) {
+ tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
+ "enqueue_in_free: %llu bad head prev %llu\n",
+ (long long)new->next, (long long)b_off);
+ return -1;
+ }
+#endif
/* next->prev = new. */
if (tdb_write_off(tdb, new->next
+ offsetof(struct tdb_free_record, prev),
int ret;
assert(len_with_header >= sizeof(new));
- assert(zone_bits < (1 << 6));
+ assert(zone_bits < 64);
new.magic_and_meta = TDB_FREE_MAGIC | zone_bits;
new.data_len = len_with_header - sizeof(struct tdb_used_record);
return ret;
}
+static size_t adjust_size(size_t keylen, size_t datalen)
+{
+ size_t size = keylen + datalen;
+
+ if (size < TDB_MIN_DATA_LEN)
+ size = TDB_MIN_DATA_LEN;
+
+ /* Round to next uint64_t boundary. */
+ return (size + (sizeof(uint64_t) - 1ULL)) & ~(sizeof(uint64_t) - 1ULL);
+}
+
/* If we have enough left over to be useful, split that off. */
-static int to_used_record(struct tdb_context *tdb,
- unsigned int zone_bits,
- tdb_off_t off,
- tdb_len_t needed,
- tdb_len_t total_len,
- tdb_len_t *actual)
+static size_t record_leftover(size_t keylen, size_t datalen,
+ bool want_extra, size_t total_len)
{
- struct tdb_used_record used;
- tdb_len_t leftover;
+ ssize_t leftover;
- leftover = total_len - needed;
- if (leftover < sizeof(struct tdb_free_record))
- leftover = 0;
+ if (want_extra)
+ datalen += datalen / 2;
+ leftover = total_len - adjust_size(keylen, datalen);
- *actual = total_len - leftover;
+ if (leftover < (ssize_t)sizeof(struct tdb_free_record))
+ return 0;
- if (leftover) {
- if (add_free_record(tdb, zone_bits,
- off + sizeof(used) + *actual,
- total_len - needed))
- return -1;
- }
- return 0;
+ return leftover;
}
/* Note: we unlock the current bucket if we coalesce or fail. */
tdb_off_t off, tdb_off_t b_off, tdb_len_t data_len)
{
struct tdb_free_record pad, *r;
- tdb_off_t end = off + sizeof(struct tdb_used_record) + data_len;
+ tdb_off_t zone_end, 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_off + (1ULL << zone_bits))) {
+ while (end < zone_end) {
tdb_off_t nb_off;
/* FIXME: do tdb_get here and below really win? */
break;
}
- if (remove_from_list(tdb, nb_off, r) == -1) {
+ if (unlikely(bucket_off(zone_off,
+ size_to_bucket(zone_bits, r->data_len))
+ != nb_off)) {
+ tdb_unlock_free_bucket(tdb, nb_off);
+ break;
+ }
+
+ if (remove_from_list(tdb, nb_off, end, r) == -1) {
tdb_unlock_free_bucket(tdb, nb_off);
goto err;
}
goto err;
}
- if (remove_from_list(tdb, b_off, r) == -1)
+ if (remove_from_list(tdb, b_off, off, r) == -1)
+ goto err;
+
+ r = tdb_access_write(tdb, off, sizeof(*r), true);
+ if (!r)
+ goto err;
+
+ /* 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->data_len = end - off - sizeof(struct tdb_used_record);
+ if (tdb_access_commit(tdb, r) != 0)
goto err;
- /* We have to drop this to avoid deadlocks. */
tdb_unlock_free_bucket(tdb, b_off);
if (add_free_record(tdb, zone_bits, off, end - off) == -1)
tdb_off_t zone_off,
unsigned zone_bits,
tdb_off_t bucket,
- size_t size,
- tdb_len_t *actual)
+ size_t keylen, size_t datalen,
+ bool want_extra,
+ unsigned hashlow)
{
tdb_off_t off, b_off,best_off;
struct tdb_free_record pad, best = { 0 }, *r;
double multiplier;
+ size_t size = adjust_size(keylen, datalen);
again:
b_off = bucket_off(zone_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). */
/* Lock this bucket. */
if (tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT) == -1) {
return TDB_OFF_ERR;
best.data_len = -1ULL;
best_off = 0;
- /* FIXME: Start with larger multiplier if we're growing. */
- multiplier = 1.0;
+
+ /* Get slack if we're after extra. */
+ if (want_extra)
+ multiplier = 1.5;
+ else
+ multiplier = 1.0;
/* Walk the list to see if any are large enough, getting less fussy
* as we go. */
}
if (best.data_len < size * multiplier && best_off)
- goto use_best;
+ break;
multiplier *= 1.01;
/* If we found anything at all, use it. */
if (best_off) {
- use_best:
+ struct tdb_used_record rec;
+ size_t leftover;
+
/* We're happy with this size: take it. */
- if (remove_from_list(tdb, b_off, &best) != 0)
+ if (remove_from_list(tdb, b_off, best_off, &best) != 0)
+ goto unlock_err;
+
+ 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)
goto unlock_err;
+
+ if (tdb_write_convert(tdb, best_off, &rec, sizeof(rec)) != 0)
+ goto unlock_err;
+
tdb_unlock_free_bucket(tdb, b_off);
- if (to_used_record(tdb, zone_bits, best_off, size,
- best.data_len, actual)) {
- return -1;
+ if (leftover) {
+ if (add_free_record(tdb, zone_bits,
+ best_off + sizeof(rec)
+ + best.data_len - leftover,
+ leftover))
+ return TDB_OFF_ERR;
}
return best_off;
}
}
/* Offset returned is within current zone (which it may alter). */
-static tdb_off_t get_free(struct tdb_context *tdb, size_t size,
- tdb_len_t *actual)
+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;
bool wrapped = false;
+ size_t size = adjust_size(keylen, datalen);
+
+ /* 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)
- continue;
+ goto next;
/* 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)) {
+ b = find_free_head(tdb, b + 1)) {
/* Try getting one from list. */
off = lock_and_alloc(tdb, tdb->zone_off,
tdb->zhdr.zone_bits,
- b, size, actual);
+ b, keylen, datalen, want_extra,
+ hashlow);
if (off == TDB_OFF_ERR)
return TDB_OFF_ERR;
if (off != 0)
/* Didn't work. Try next bucket. */
}
+ next:
/* Didn't work, try next zone, if it exists. */
if (!next_zone(tdb)) {
wrapped = true;
int set_header(struct tdb_context *tdb,
struct tdb_used_record *rec,
uint64_t keylen, uint64_t datalen,
- uint64_t actuallen, uint64_t hash,
+ uint64_t actuallen, unsigned hashlow,
unsigned int zone_bits)
{
uint64_t keybits = (fls64(keylen) + 1) / 2;
- /* Use top bits of hash, so it's independent of hash table size. */
+ /* Use bottom bits of hash, so it's independent of hash table size. */
rec->magic_and_meta
= zone_bits
- | ((hash >> 59) << 6)
+ | ((hashlow & ((1 << 5)-1)) << 6)
| ((actuallen - (keylen + datalen)) << 11)
| (keybits << 43)
| (TDB_MAGIC << 48);
return 0;
}
-static bool zones_happy(struct tdb_context *tdb)
+static bool zones_contended(struct tdb_context *tdb)
{
- /* FIXME: look at distribution of zones. */
- return true;
+ return false;
}
/* Assume we want buckets up to the comfort factor. */
{
uint64_t old_size;
tdb_off_t off;
- uint8_t zone_bits;
- unsigned int num_buckets;
- tdb_len_t wanted;
+ unsigned int num_buckets, zone_bits;
+ tdb_len_t wanted, expand;
struct free_zone_header zhdr;
- bool enlarge_zone;
/* We need room for the record header too. */
wanted = sizeof(struct tdb_used_record) + size;
if (tdb->map_size != old_size)
goto success;
- /* zone bits tailer char is protected by EXPAND lock. */
- if (tdb->methods->read(tdb, old_size - 1, &zone_bits, 1) == -1)
+ /* Treat last zone as minimum reasonable zone size. */
+ off = last_zone(tdb, &zhdr);
+ if (off == TDB_OFF_ERR)
goto fail;
- /* If zones aren't working well, add larger zone if possible. */
- enlarge_zone = !zones_happy(tdb);
+ /* 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. */
+ 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);
+ }
- /* New zone can be between zone_bits or larger if we're on the right
- * boundary. */
- for (;;) {
- /* Does this fit the allocation comfortably? */
- if ((1ULL << zone_bits) >= overhead(zone_bits) + wanted) {
- /* Only let enlarge_zone enlarge us once. */
- if (!enlarge_zone)
- break;
- enlarge_zone = false;
- }
- if ((old_size - 1 - sizeof(struct tdb_header))
- & (1 << zone_bits))
- break;
- zone_bits++;
+ /* 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;
}
zhdr.zone_bits = zone_bits;
num_buckets = BUCKETS_FOR_ZONE(zone_bits);
- if (tdb->methods->expand_file(tdb, 1ULL << zone_bits) == -1)
- goto fail;
+ /* 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);
- /* Write new tailer. */
- if (tdb->methods->write(tdb, tdb->map_size - 1, &zone_bits, 1) == -1)
+ if (tdb->methods->expand_file(tdb, expand) == -1)
goto fail;
- /* Write new zone header (just before old tailer). */
- off = old_size - 1;
+ /* Write new zone header (at old end). */
+ off = old_size;
if (tdb_write_convert(tdb, off, &zhdr, sizeof(zhdr)) == -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, tdb->map_size-1-off) == -1)
+ 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 - 1;
+ tdb->zone_off = old_size;
tdb->zhdr = zhdr;
success:
return -1;
}
-static tdb_len_t adjust_size(size_t keylen, size_t datalen, bool growing)
-{
- tdb_len_t size = keylen + datalen;
-
- if (size < MIN_DATA_LEN)
- size = MIN_DATA_LEN;
-
- /* Overallocate if this is coming from an enlarging store. */
- if (growing)
- size += datalen / 2;
-
- /* Round to next uint64_t boundary. */
- return (size + (sizeof(uint64_t) - 1ULL)) & ~(sizeof(uint64_t) - 1ULL);
-}
-
/* This won't fail: it will expand the database if it has to. */
tdb_off_t alloc(struct tdb_context *tdb, size_t keylen, size_t datalen,
uint64_t hash, bool growing)
{
tdb_off_t off;
- tdb_len_t size, actual;
- struct tdb_used_record rec;
- /* We don't want header to change during this! */
- assert(tdb->header_uptodate);
+ /* We can't hold pointers during this: we could unmap! */
+ assert(!tdb->direct_access);
- size = adjust_size(keylen, datalen, growing);
-
-again:
- off = get_free(tdb, size, &actual);
- if (unlikely(off == TDB_OFF_ERR))
- return off;
+ for (;;) {
+ off = get_free(tdb, keylen, datalen, growing, hash);
+ if (likely(off != 0))
+ break;
- if (unlikely(off == 0)) {
- if (tdb_expand(tdb, size) == -1)
+ if (tdb_expand(tdb, adjust_size(keylen, datalen)))
return TDB_OFF_ERR;
- goto again;
- }
-
- /* Some supergiant values can't be encoded. */
- /* FIXME: Check before, and limit actual in get_free. */
- if (set_header(tdb, &rec, keylen, datalen, actual, hash,
- tdb->zhdr.zone_bits) != 0) {
- add_free_record(tdb, tdb->zhdr.zone_bits, off,
- sizeof(rec) + actual);
- return TDB_OFF_ERR;
}
- if (tdb_write_convert(tdb, off, &rec, sizeof(rec)) != 0)
- return TDB_OFF_ERR;
-
return off;
}
projects / ccan / blobdiff