/* Remove from free bucket. */
static int remove_from_list(struct tdb_context *tdb,
tdb_off_t b_off, tdb_off_t r_off,
- struct tdb_free_record *r)
+ const struct tdb_free_record *r)
{
tdb_off_t off;
#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",
+ tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
+ "remove_from_list: %llu bad prev in list %llu",
(long long)r_off, (long long)b_off);
return -1;
}
#ifdef DEBUG
if (tdb_read_off(tdb, off) & TDB_OFF_MASK != 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);
+ tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
+ "remove_from_list: %llu bad list %llu",
+ (long long)r_off, (long long)b_off);
return -1;
}
#endif
new.next + offsetof(struct tdb_free_record,
magic_and_prev))
!= magic) {
- 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);
+ tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
+ "enqueue_in_free: %llu bad head"
+ " prev %llu",
+ (long long)new.next, (long long)b_off);
return -1;
}
#endif
return leftover;
}
-/* FIXME: Shortcut common case where tdb->flist == flist */
static tdb_off_t flist_offset(struct tdb_context *tdb, unsigned int flist)
{
- tdb_off_t off = first_flist(tdb);
+ tdb_off_t off;
unsigned int i;
+ if (likely(tdb->flist == flist))
+ return tdb->flist_off;
+
+ off = first_flist(tdb);
for (i = 0; i < flist; i++)
off = next_flist(tdb, off);
return off;
static int coalesce(struct tdb_context *tdb,
tdb_off_t off, tdb_off_t b_off, tdb_len_t data_len)
{
- struct tdb_free_record pad, *r;
tdb_off_t end;
+ struct tdb_free_record rec;
+ add_stat(tdb, alloc_coalesce_tried, 1);
end = off + sizeof(struct tdb_used_record) + data_len;
while (end < tdb->map_size) {
+ const struct tdb_free_record *r;
tdb_off_t nb_off;
unsigned flist, bucket;
- /* FIXME: do tdb_get here and below really win? */
- r = tdb_get(tdb, end, &pad, sizeof(pad));
+ r = tdb_access_read(tdb, end, sizeof(*r), true);
if (!r)
goto err;
- if (frec_magic(r) != TDB_FREE_MAGIC)
+ if (frec_magic(r) != TDB_FREE_MAGIC
+ || frec_flist(r) == TDB_FLIST_NONE) {
+ tdb_access_release(tdb, r);
break;
+ }
flist = frec_flist(r);
bucket = size_to_bucket(frec_len(r));
nb_off = bucket_off(flist_offset(tdb, flist), bucket);
+ tdb_access_release(tdb, r);
/* We may be violating lock order here, so best effort. */
- if (tdb_lock_free_bucket(tdb, nb_off, TDB_LOCK_NOWAIT) == -1)
+ if (tdb_lock_free_bucket(tdb, nb_off, TDB_LOCK_NOWAIT) == -1) {
+ add_stat(tdb, alloc_coalesce_lockfail, 1);
break;
+ }
/* Now we have lock, re-check. */
- r = tdb_get(tdb, end, &pad, sizeof(pad));
- if (!r) {
+ if (tdb_read_convert(tdb, end, &rec, sizeof(rec))) {
tdb_unlock_free_bucket(tdb, nb_off);
goto err;
}
- if (unlikely(frec_magic(r) != TDB_FREE_MAGIC)) {
+ if (unlikely(frec_magic(&rec) != TDB_FREE_MAGIC)) {
+ add_stat(tdb, alloc_coalesce_race, 1);
tdb_unlock_free_bucket(tdb, nb_off);
break;
}
- if (unlikely(frec_flist(r) != flist)
- || unlikely(size_to_bucket(frec_len(r)) != bucket)) {
+ if (unlikely(frec_flist(&rec) != flist)
+ || unlikely(size_to_bucket(frec_len(&rec)) != bucket)) {
+ add_stat(tdb, alloc_coalesce_race, 1);
tdb_unlock_free_bucket(tdb, nb_off);
break;
}
- if (remove_from_list(tdb, nb_off, end, r) == -1) {
+ if (remove_from_list(tdb, nb_off, end, &rec) == -1) {
tdb_unlock_free_bucket(tdb, nb_off);
goto err;
}
- end += sizeof(struct tdb_used_record) + frec_len(r);
+ end += sizeof(struct tdb_used_record) + frec_len(&rec);
tdb_unlock_free_bucket(tdb, nb_off);
+ add_stat(tdb, alloc_coalesce_num_merged, 1);
}
/* Didn't find any adjacent free? */
if (end == off + sizeof(struct tdb_used_record) + data_len)
return 0;
- /* OK, expand record */
- r = tdb_get(tdb, off, &pad, sizeof(pad));
- if (!r)
+ /* OK, expand initial record */
+ if (tdb_read_convert(tdb, off, &rec, sizeof(rec)))
goto err;
- if (frec_len(r) != data_len) {
- tdb->ecode = TDB_ERR_CORRUPT;
- tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
- "coalesce: expected data len %llu not %llu\n",
- (long long)data_len, (long long)frec_len(r));
+ if (frec_len(&rec) != data_len) {
+ tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
+ "coalesce: expected data len %zu not %zu",
+ (size_t)data_len, (size_t)frec_len(&rec));
goto err;
}
- if (remove_from_list(tdb, b_off, off, r) == -1)
- goto err;
-
- r = tdb_access_write(tdb, off, sizeof(*r), true);
- if (!r)
+ if (remove_from_list(tdb, b_off, off, &rec) == -1)
goto err;
/* We have to drop this to avoid deadlocks, so make sure record
* doesn't get coalesced by someone else! */
- r->magic_and_prev = TDB_COALESCING_MAGIC << (64 - TDB_OFF_UPPER_STEAL);
- /* FIXME: Use 255 as invalid free list? */
- r->flist_and_len = end - off - sizeof(struct tdb_used_record);
- if (tdb_access_commit(tdb, r) != 0)
+ rec.flist_and_len = (TDB_FLIST_NONE << (64 - TDB_OFF_UPPER_STEAL))
+ | (end - off - sizeof(struct tdb_used_record));
+ if (tdb_write_off(tdb, off + offsetof(struct tdb_free_record,
+ flist_and_len),
+ rec.flist_and_len) != 0)
goto err;
+ add_stat(tdb, alloc_coalesce_succeeded, 1);
tdb_unlock_free_bucket(tdb, b_off);
if (add_free_record(tdb, off, end - off) == -1)
unsigned hashlow)
{
tdb_off_t off, b_off,best_off;
- struct tdb_free_record pad, best = { 0 }, *r;
+ struct tdb_free_record best = { 0 };
double multiplier;
size_t size = adjust_size(keylen, datalen);
+ add_stat(tdb, allocs, 1);
again:
b_off = bucket_off(flist_off, bucket);
goto unlock_err;
while (off) {
- /* FIXME: Does tdb_get win anything here? */
- r = tdb_get(tdb, off, &pad, sizeof(*r));
+ const struct tdb_free_record *r;
+ tdb_len_t len;
+ tdb_off_t next;
+
+ r = tdb_access_read(tdb, off, sizeof(*r), true);
if (!r)
goto unlock_err;
if (frec_magic(r) != TDB_FREE_MAGIC) {
- tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
- "lock_and_alloc: %llu non-free 0x%llx\n",
+ tdb_access_release(tdb, r);
+ tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
+ "lock_and_alloc: %llu non-free 0x%llx",
(long long)off, (long long)r->magic_and_prev);
goto unlock_err;
}
best = *r;
}
- if (frec_len(&best) < size * multiplier && best_off)
+ if (frec_len(&best) < size * multiplier && best_off) {
+ tdb_access_release(tdb, r);
break;
+ }
multiplier *= 1.01;
+ next = r->next;
+ len = frec_len(r);
+ tdb_access_release(tdb, r);
+
/* Since we're going slow anyway, try coalescing here. */
- switch (coalesce(tdb, off, b_off, frec_len(r))) {
+ switch (coalesce(tdb, off, b_off, len)) {
case -1:
/* This has already unlocked on error. */
return -1;
/* This has unlocked list, restart. */
goto again;
}
- off = r->next;
+ off = next;
}
/* If we found anything at all, use it. */
if (tdb_write_convert(tdb, best_off, &rec, sizeof(rec)) != 0)
goto unlock_err;
- tdb_unlock_free_bucket(tdb, b_off);
-
+ /* Bucket of leftover will be <= current bucket, so nested
+ * locking is allowed. */
if (leftover) {
+ add_stat(tdb, alloc_leftover, 1);
if (add_free_record(tdb,
best_off + sizeof(rec)
+ frec_len(&best) - leftover,
leftover))
- return TDB_OFF_ERR;
+ best_off = TDB_OFF_ERR;
}
+ tdb_unlock_free_bucket(tdb, b_off);
+
return best_off;
}
if (off == TDB_OFF_ERR)
return TDB_OFF_ERR;
if (off != 0) {
+ if (b == start_b)
+ add_stat(tdb, alloc_bucket_exact, 1);
+ if (b == TDB_FREE_BUCKETS - 1)
+ add_stat(tdb, alloc_bucket_max, 1);
/* Worked? Stay using this list. */
tdb->flist_off = flist_off;
tdb->flist = flist;
/* Hmm, try next list. */
flist_off = next_flist(tdb, flist_off);
flist++;
+
if (flist_off == 0) {
wrapped = true;
flist_off = first_flist(tdb);
if (rec_key_length(rec) != keylen
|| rec_data_length(rec) != datalen
|| rec_extra_padding(rec) != actuallen - (keylen + datalen)) {
- tdb->ecode = TDB_ERR_IO;
- tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
- "Could not encode k=%llu,d=%llu,a=%llu\n",
+ tdb_logerr(tdb, TDB_ERR_IO, TDB_DEBUG_ERROR,
+ "Could not encode k=%llu,d=%llu,a=%llu",
(long long)keylen, (long long)datalen,
(long long)actuallen);
return -1;
/* 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");
+ tdb_logerr(tdb, TDB_ERR_LOCK, TDB_DEBUG_ERROR,
+ "tdb_expand: must hold lock during expand");
return -1;
}
/* We need to drop this lock before adding free record. */
tdb_unlock_expand(tdb, F_WRLCK);
+ add_stat(tdb, expands, 1);
return add_free_record(tdb, old_size, wanted);
}