tdb_io_init(tdb);
tdb_lock_init(tdb);
- /* FIXME */
- if (attr) {
- tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
- "tdb_open: attributes not yet supported\n");
- errno = EINVAL;
- goto fail;
+ while (attr) {
+ switch (attr->base.attr) {
+ case TDB_ATTRIBUTE_LOG:
+ tdb->log = attr->log.log_fn;
+ tdb->log_priv = attr->log.log_private;
+ break;
+ case TDB_ATTRIBUTE_HASH:
+ tdb->khash = attr->hash.hash_fn;
+ tdb->hash_priv = attr->hash.hash_private;
+ break;
+ default:
+ tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
+ "tdb_open: unknown attribute type %u\n",
+ attr->base.attr);
+ errno = EINVAL;
+ goto fail;
+ }
+ attr = attr->base.next;
}
if ((open_flags & O_ACCMODE) == O_WRONLY) {
}
TEST_IT(tdb->flags & TDB_CONVERT);
tdb_convert(tdb, &tdb->header, sizeof(tdb->header));
+ /* Zones don't matter for internal db. */
+ tdb->last_zone = 0;
return tdb;
}
uint64_t keylen;
const unsigned char *rkey;
+ list &= ((1ULL << tdb->header.v.hash_bits) - 1);
+
off = tdb_read_off(tdb, tdb->header.v.hash_off
+ list * sizeof(tdb_off_t));
if (off == 0 || off == TDB_OFF_ERR)
if (keylen != key->dsize)
return TDB_OFF_ERR;
- rkey = tdb_access_read(tdb, off + sizeof(*rec), keylen);
+ rkey = tdb_access_read(tdb, off + sizeof(*rec), keylen, false);
if (!rkey)
return TDB_OFF_ERR;
if (memcmp(rkey, key->dptr, keylen) != 0)
}
/* FIXME: Optimize? */
-static void unlock_range(struct tdb_context *tdb,
+static void unlock_lists(struct tdb_context *tdb,
tdb_off_t list, tdb_len_t num,
int ltype)
{
}
/* FIXME: Optimize? */
-static int lock_range(struct tdb_context *tdb,
+static int lock_lists(struct tdb_context *tdb,
tdb_off_t list, tdb_len_t num,
int ltype)
{
tdb_off_t i;
for (i = list; i < list + num; i++) {
- if (tdb_lock_list(tdb, i, ltype, TDB_LOCK_WAIT) != 0) {
- unlock_range(tdb, list, i - list, ltype);
+ if (tdb_lock_list(tdb, i, ltype, TDB_LOCK_WAIT)
+ == TDB_OFF_ERR) {
+ unlock_lists(tdb, list, i - list, ltype);
return -1;
}
}
static tdb_len_t relock_hash_to_zero(struct tdb_context *tdb,
tdb_off_t start, int ltype)
{
- tdb_len_t num, len, pre_locks;
+ tdb_len_t num, len;
again:
num = 1ULL << tdb->header.v.hash_bits;
if (unlikely(len == num - start)) {
/* We hit the end of the hash range. Drop lock: we have
to lock start of hash first. */
+ tdb_len_t pre_locks;
+
tdb_unlock_list(tdb, start, ltype);
+
/* Grab something, so header is stable. */
if (tdb_lock_list(tdb, 0, ltype, TDB_LOCK_WAIT))
return TDB_OFF_ERR;
- len = tdb_find_zero_off(tdb, hash_off(tdb, 0), num);
- if (lock_range(tdb, 1, len, ltype) == -1) {
+ pre_locks = tdb_find_zero_off(tdb, hash_off(tdb, 0), num);
+ /* We want to lock the zero entry as well. */
+ pre_locks++;
+ if (lock_lists(tdb, 1, pre_locks - 1, ltype) == -1) {
tdb_unlock_list(tdb, 0, ltype);
return TDB_OFF_ERR;
}
- pre_locks = len;
- len = num - start;
+
+ /* Now lock later ones. */
+ if (unlikely(lock_lists(tdb, start, len, ltype) == -1)) {
+ unlock_lists(tdb, 0, pre_locks, ltype);
+ return TDB_OFF_ERR;
+ }
+ len += pre_locks;
} else {
- /* We already have lock on start. */
- start++;
- pre_locks = 0;
- }
- if (unlikely(lock_range(tdb, start, len, ltype) == -1)) {
- if (pre_locks)
- unlock_range(tdb, 0, pre_locks, ltype);
- else
+ /* We want to lock the zero entry as well. */
+ len++;
+ /* But we already have lock on start. */
+ if (unlikely(lock_lists(tdb, start+1, len-1, ltype) == -1)) {
tdb_unlock_list(tdb, start, ltype);
- return TDB_OFF_ERR;
+ return TDB_OFF_ERR;
+ }
}
/* Now, did we lose the race, and it's not zero any more? */
- if (unlikely(tdb_read_off(tdb, hash_off(tdb, pre_locks + len)) != 0)) {
- unlock_range(tdb, 0, pre_locks, ltype);
+ if (unlikely(tdb_read_off(tdb, hash_off(tdb, start + len - 1)) != 0)) {
/* Leave the start locked, as expected. */
- unlock_range(tdb, start + 1, len - 1, ltype);
+ unlock_lists(tdb, start + 1, len - 1, ltype);
goto again;
}
- return pre_locks + len;
+ return len;
}
/* FIXME: modify, don't rewrite! */
struct tdb_used_record *rec,
uint64_t h)
{
- uint64_t room = rec_data_length(rec) + rec_extra_padding(rec);
+ uint64_t dataroom = rec_data_length(rec) + rec_extra_padding(rec);
- if (set_header(tdb, rec, keylen, datalen, room - datalen, h))
+ if (set_header(tdb, rec, keylen, datalen, keylen + dataroom, h))
return -1;
return tdb_write_convert(tdb, off, rec, sizeof(*rec));
/* If we fail, others will try after us. */
static void enlarge_hash(struct tdb_context *tdb)
{
- tdb_off_t newoff, i;
+ tdb_off_t newoff, oldoff, i;
+ tdb_len_t hlen;
uint64_t h, num = 1ULL << tdb->header.v.hash_bits;
struct tdb_used_record pad, *r;
if ((1ULL << tdb->header.v.hash_bits) != num)
goto unlock;
- newoff = alloc(tdb, 0, num * 2, 0, false);
+ /* Allocate our new array. */
+ hlen = num * sizeof(tdb_off_t) * 2;
+ newoff = alloc(tdb, 0, hlen, 0, false);
if (unlikely(newoff == TDB_OFF_ERR))
goto unlock;
if (unlikely(newoff == 0)) {
- if (tdb_expand(tdb, 0, num * 2, false) == -1)
+ if (tdb_expand(tdb, 0, hlen, false) == -1)
goto unlock;
- newoff = alloc(tdb, 0, num * 2, 0, false);
+ newoff = alloc(tdb, 0, hlen, 0, false);
if (newoff == TDB_OFF_ERR || newoff == 0)
goto unlock;
}
+ /* Step over record header! */
+ newoff += sizeof(struct tdb_used_record);
+
+ /* Starts all zero. */
+ if (zero_out(tdb, newoff, hlen) == -1)
+ goto unlock;
/* FIXME: If the space before is empty, we know this is in its ideal
- * location. We can steal a bit from the pointer to avoid rehash. */
- for (i = tdb_find_nonzero_off(tdb, tdb->header.v.hash_off, num);
+ * location. Or steal a bit from the pointer to avoid rehash. */
+ for (i = tdb_find_nonzero_off(tdb, hash_off(tdb, 0), num);
i < num;
- i += tdb_find_nonzero_off(tdb, tdb->header.v.hash_off
- + i*sizeof(tdb_off_t), num - i)) {
+ i += tdb_find_nonzero_off(tdb, hash_off(tdb, i), num - i)) {
tdb_off_t off;
- off = tdb_read_off(tdb, tdb->header.v.hash_off
- + i*sizeof(tdb_off_t));
+ off = tdb_read_off(tdb, hash_off(tdb, i));
if (unlikely(off == TDB_OFF_ERR))
goto unlock;
if (unlikely(!off)) {
tdb->ecode = TDB_ERR_CORRUPT;
tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
- "find_bucket_and_lock: zero hash bucket!\n");
+ "enlarge_hash: zero hash bucket!\n");
goto unlock;
}
- h = hash_record(tdb, off);
+
+ /* Find next empty hash slot. */
+ for (h = hash_record(tdb, off);
+ tdb_read_off(tdb, newoff + (h & ((num * 2)-1))
+ * sizeof(tdb_off_t)) != 0;
+ h++);
+
/* FIXME: Encode extra hash bits! */
- if (tdb_write_off(tdb, newoff
- + (h & ((num * 2) - 1)) * sizeof(uint64_t),
- off) == -1)
+ if (tdb_write_off(tdb, newoff + (h & ((num * 2)-1))
+ * sizeof(tdb_off_t), off) == -1)
goto unlock;
+ i++;
}
/* Free up old hash. */
- r = tdb_get(tdb, tdb->header.v.hash_off, &pad, sizeof(*r));
+ oldoff = tdb->header.v.hash_off - sizeof(*r);
+ r = tdb_get(tdb, oldoff, &pad, sizeof(*r));
if (!r)
goto unlock;
- add_free_record(tdb, tdb->header.v.hash_off,
- rec_data_length(r) + rec_extra_padding(r));
+ add_free_record(tdb, oldoff,
+ sizeof(*r)+rec_data_length(r)+rec_extra_padding(r));
/* Now we write the modified header. */
- tdb->header.v.generation++;
tdb->header.v.hash_bits++;
tdb->header.v.hash_off = newoff;
- tdb_write_convert(tdb, offsetof(struct tdb_header, v),
- &tdb->header.v, sizeof(tdb->header.v));
+ write_header(tdb);
unlock:
tdb_allrecord_unlock(tdb, F_WRLCK);
}
+/* This is the core routine which searches the hashtable for an entry.
+ * On error, no locks are held and TDB_OFF_ERR is returned.
+ * Otherwise, *num_locks locks of type ltype from *start_lock are held.
+ * The bucket where the entry is (or would be) is in *bucket.
+ * If not found, the return value is 0.
+ * If found, the return value is the offset, and *rec is the record. */
+static tdb_off_t find_and_lock(struct tdb_context *tdb,
+ struct tdb_data key,
+ uint64_t h,
+ int ltype,
+ tdb_off_t *start_lock,
+ tdb_len_t *num_locks,
+ tdb_off_t *bucket,
+ struct tdb_used_record *rec)
+{
+ tdb_off_t off;
+
+ /* FIXME: can we avoid locks for some fast paths? */
+ *start_lock = tdb_lock_list(tdb, h, ltype, TDB_LOCK_WAIT);
+ if (*start_lock == TDB_OFF_ERR)
+ return TDB_OFF_ERR;
+
+ /* Fast path. */
+ off = entry_matches(tdb, *start_lock, h, &key, rec);
+ if (likely(off != TDB_OFF_ERR)) {
+ *bucket = *start_lock;
+ *num_locks = 1;
+ return off;
+ }
+
+ /* Slow path, need to grab more locks and search. */
+
+ /* Warning: this may drop the lock on *bucket! */
+ *num_locks = relock_hash_to_zero(tdb, *start_lock, ltype);
+ if (*num_locks == TDB_OFF_ERR)
+ return TDB_OFF_ERR;
+
+ for (*bucket = *start_lock;
+ *bucket < *start_lock + *num_locks;
+ (*bucket)++) {
+ off = entry_matches(tdb, *bucket, h, &key, rec);
+ /* Empty entry or we found it? */
+ if (off == 0 || off != TDB_OFF_ERR)
+ return off;
+ }
+
+ /* We didn't find a zero entry? Something went badly wrong... */
+ unlock_lists(tdb, *start_lock, *start_lock + *num_locks, ltype);
+ tdb->ecode = TDB_ERR_CORRUPT;
+ tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
+ "find_and_lock: expected to find an empty hash bucket!\n");
+ return TDB_OFF_ERR;
+}
+
int tdb_store(struct tdb_context *tdb,
struct tdb_data key, struct tdb_data dbuf, int flag)
{
- tdb_off_t new_off, off, old_bucket, start, num_locks = 1;
+ tdb_off_t new_off, off, bucket, start, num;
struct tdb_used_record rec;
uint64_t h;
bool growing = false;
h = tdb_hash(tdb, key.dptr, key.dsize);
-
- /* FIXME: can we avoid locks for some fast paths? */
- start = tdb_lock_list(tdb, h, F_WRLCK, TDB_LOCK_WAIT);
- if (start == TDB_OFF_ERR)
+ off = find_and_lock(tdb, key, h, F_WRLCK, &start, &num, &bucket, &rec);
+ if (unlikely(off == TDB_OFF_ERR))
return -1;
- /* Fast path. */
- old_bucket = start;
- off = entry_matches(tdb, start, h, &key, &rec);
- if (unlikely(off == TDB_OFF_ERR)) {
- /* Slow path, need to grab more locks and search. */
- tdb_off_t i;
-
- /* Warning: this may drop the lock! Does that on error. */
- num_locks = relock_hash_to_zero(tdb, start, F_WRLCK);
- if (num_locks == TDB_OFF_ERR)
- return -1;
-
- for (i = start; i < start + num_locks; i++) {
- off = entry_matches(tdb, i, h, &key, &rec);
- /* Empty entry or we found it? */
- if (off == 0 || off != TDB_OFF_ERR) {
- old_bucket = i;
- break;
- }
- }
- if (i == start + num_locks)
- off = 0;
- }
-
/* Now we have lock on this hash bucket. */
if (flag == TDB_INSERT) {
if (off) {
/* Allocate a new record. */
new_off = alloc(tdb, key.dsize, dbuf.dsize, h, growing);
if (new_off == 0) {
- unlock_range(tdb, start, num_locks, F_WRLCK);
+ unlock_lists(tdb, start, num, F_WRLCK);
/* Expand, then try again... */
if (tdb_expand(tdb, key.dsize, dbuf.dsize, growing) == -1)
return -1;
+ rec_extra_padding(&rec));
}
-write:
/* FIXME: Encode extra hash bits! */
- if (tdb_write_off(tdb, hash_off(tdb, old_bucket), new_off) == -1)
+ if (tdb_write_off(tdb, hash_off(tdb, bucket), new_off) == -1)
goto fail;
+write:
off = new_off + sizeof(struct tdb_used_record);
if (tdb->methods->write(tdb, off, key.dptr, key.dsize) == -1)
goto fail;
goto fail;
/* FIXME: tdb_increment_seqnum(tdb); */
- unlock_range(tdb, start, num_locks, F_WRLCK);
+ unlock_lists(tdb, start, num, F_WRLCK);
/* FIXME: by simple simulation, this approximated 60% full.
* Check in real case! */
- if (unlikely(num_locks > 4 * tdb->header.v.hash_bits - 31))
+ if (unlikely(num > 4 * tdb->header.v.hash_bits - 30))
enlarge_hash(tdb);
return 0;
fail:
- unlock_range(tdb, start, num_locks, F_WRLCK);
+ unlock_lists(tdb, start, num, F_WRLCK);
return -1;
}
struct tdb_data tdb_fetch(struct tdb_context *tdb, struct tdb_data key)
{
- tdb_off_t off, start, num_locks = 1;
+ tdb_off_t off, start, num, bucket;
struct tdb_used_record rec;
uint64_t h;
struct tdb_data ret;
h = tdb_hash(tdb, key.dptr, key.dsize);
-
- /* FIXME: can we avoid locks for some fast paths? */
- start = tdb_lock_list(tdb, h, F_RDLCK, TDB_LOCK_WAIT);
- if (start == TDB_OFF_ERR)
+ off = find_and_lock(tdb, key, h, F_RDLCK, &start, &num, &bucket, &rec);
+ if (unlikely(off == TDB_OFF_ERR))
return tdb_null;
- /* Fast path. */
- off = entry_matches(tdb, start, h, &key, &rec);
- if (unlikely(off == TDB_OFF_ERR)) {
- /* Slow path, need to grab more locks and search. */
- tdb_off_t i;
-
- /* Warning: this may drop the lock! Does that on error. */
- num_locks = relock_hash_to_zero(tdb, start, F_RDLCK);
- if (num_locks == TDB_OFF_ERR)
- return tdb_null;
-
- for (i = start; i < start + num_locks; i++) {
- off = entry_matches(tdb, i, h, &key, &rec);
- /* Empty entry or we found it? */
- if (off == 0 || off != TDB_OFF_ERR)
- break;
- }
- if (i == start + num_locks)
- off = 0;
- }
-
if (!off) {
- unlock_range(tdb, start, num_locks, F_RDLCK);
tdb->ecode = TDB_ERR_NOEXIST;
- return tdb_null;
+ ret = tdb_null;
+ } else {
+ ret.dsize = rec_data_length(&rec);
+ ret.dptr = tdb_alloc_read(tdb, off + sizeof(rec) + key.dsize,
+ ret.dsize);
}
- ret.dsize = rec_data_length(&rec);
- ret.dptr = tdb_alloc_read(tdb, off + sizeof(rec) + key.dsize,
- ret.dsize);
- unlock_range(tdb, start, num_locks, F_RDLCK);
+ unlock_lists(tdb, start, num, F_RDLCK);
return ret;
}
int tdb_delete(struct tdb_context *tdb, struct tdb_data key)
{
- tdb_off_t i, old_bucket, off, start, num_locks = 1;
+ tdb_off_t i, bucket, off, start, num;
struct tdb_used_record rec;
uint64_t h;
h = tdb_hash(tdb, key.dptr, key.dsize);
-
- /* FIXME: can we avoid locks for some fast paths? */
- start = tdb_lock_list(tdb, h, F_WRLCK, TDB_LOCK_WAIT);
- if (start == TDB_OFF_ERR)
+ off = find_and_lock(tdb, key, h, F_WRLCK, &start, &num, &bucket, &rec);
+ if (unlikely(off == TDB_OFF_ERR))
return -1;
- /* Fast path. */
- old_bucket = start;
- off = entry_matches(tdb, start, h, &key, &rec);
- if (off && off != TDB_OFF_ERR) {
- /* We can only really fastpath delete if next bucket
- * is 0. Note that we haven't locked it, but our lock
- * on this bucket stops anyone overflowing into it
- * while we look. */
- if (tdb_read_off(tdb, hash_off(tdb, h+1)) == 0)
- goto delete;
- /* Slow path. */
- off = TDB_OFF_ERR;
- }
-
- if (unlikely(off == TDB_OFF_ERR)) {
- /* Slow path, need to grab more locks and search. */
- tdb_off_t i;
-
- /* Warning: this may drop the lock! Does that on error. */
- num_locks = relock_hash_to_zero(tdb, start, F_WRLCK);
- if (num_locks == TDB_OFF_ERR)
- return -1;
-
- for (i = start; i < start + num_locks; i++) {
- off = entry_matches(tdb, i, h, &key, &rec);
- /* Empty entry or we found it? */
- if (off == 0 || off != TDB_OFF_ERR) {
- old_bucket = i;
- break;
- }
- }
- if (i == start + num_locks)
- off = 0;
- }
-
if (!off) {
- unlock_range(tdb, start, num_locks, F_WRLCK);
+ unlock_lists(tdb, start, num, F_WRLCK);
tdb->ecode = TDB_ERR_NOEXIST;
return -1;
}
-delete:
/* This actually unlinks it. */
- if (tdb_write_off(tdb, hash_off(tdb, old_bucket), 0) == -1)
+ if (tdb_write_off(tdb, hash_off(tdb, bucket), 0) == -1)
goto unlock_err;
/* Rehash anything following. */
- for (i = hash_off(tdb, old_bucket+1);
- i != hash_off(tdb, h + num_locks);
+ for (i = hash_off(tdb, bucket+1);
+ i != hash_off(tdb, h + num - 1);
i += sizeof(tdb_off_t)) {
tdb_off_t off2;
uint64_t h2;
if (unlikely(off2 == TDB_OFF_ERR))
goto unlock_err;
+ /* This can happen if we raced. */
+ if (unlikely(off2 == 0))
+ break;
+
/* Maybe use a bit to indicate it is in ideal place? */
h2 = hash_record(tdb, off2);
/* Is it happy where it is? */
+ rec_extra_padding(&rec)) != 0)
goto unlock_err;
- unlock_range(tdb, start, num_locks, F_WRLCK);
+ unlock_lists(tdb, start, num, F_WRLCK);
return 0;
unlock_err:
- unlock_range(tdb, start, num_locks, F_WRLCK);
+ unlock_lists(tdb, start, num, F_WRLCK);
return -1;
}