X-Git-Url: https://git.ozlabs.org/?p=ccan;a=blobdiff_plain;f=ccan%2Ftdb2%2Ftdb.c;h=d19ceda239d50e71681607d39487d19187b61ac9;hp=43d1ef263d44d12cafbf23446e5868af2d1a9691;hb=cbfcfe53cb6184d76d2343b224409ee67841f458;hpb=ebdd6451e2d7aa185e62a59fa2c72ffe36772d9a

diff --git a/ccan/tdb2/tdb.c b/ccan/tdb2/tdb.c
index 43d1ef26..d19ceda2 100644
--- a/ccan/tdb2/tdb.c
+++ b/ccan/tdb2/tdb.c
@@ -21,29 +21,35 @@ null_log_fn(struct tdb_context *tdb,
 /* We do a lot of work assuming our copy of the header volatile area
  * is uptodate, and usually it is.  However, once we grab a lock, we have to
  * re-check it. */
-bool update_header(struct tdb_context *tdb)
+bool header_changed(struct tdb_context *tdb)
 {
-	struct tdb_header_volatile pad, *v;
+	uint64_t gen;
 
-	if (tdb->header_uptodate) {
+	if (!(tdb->flags & TDB_NOLOCK) && tdb->header_uptodate) {
 		tdb->log(tdb, TDB_DEBUG_WARNING, tdb->log_priv,
 			 "warning: header uptodate already\n");
 	}
 
 	/* We could get a partial update if we're not holding any locks. */
-	assert(tdb_has_locks(tdb));
+	assert((tdb->flags & TDB_NOLOCK) || tdb_has_locks(tdb));
 
-	v = tdb_get(tdb, offsetof(struct tdb_header, v), &pad, sizeof(*v));
-	if (!v) {
-		/* On failure, imply we updated header so they retry. */
-		return true;
-	}
 	tdb->header_uptodate = true;
-	if (likely(memcmp(&tdb->header.v, v, sizeof(*v)) == 0)) {
-		return false;
+	gen = tdb_read_off(tdb, offsetof(struct tdb_header, v.generation));
+	if (unlikely(gen != tdb->header.v.generation)) {
+		tdb_read_convert(tdb, offsetof(struct tdb_header, v),
+				 &tdb->header.v, sizeof(tdb->header.v));
+		return true;
 	}
-	tdb->header.v = *v;
-	return true;
+	return false;
+}
+
+int write_header(struct tdb_context *tdb)
+{
+	assert(tdb_read_off(tdb, offsetof(struct tdb_header, v.generation))
+	       == tdb->header.v.generation);
+	tdb->header.v.generation++;
+	return tdb_write_convert(tdb, offsetof(struct tdb_header, v),
+				 &tdb->header.v, sizeof(tdb->header.v));
 }
 
 static uint64_t jenkins_hash(const void *key, size_t length, uint64_t seed,
@@ -207,27 +213,47 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
 	uint64_t hash_test;
 	unsigned v;
 
-	if (!(tdb = (struct tdb_context *)calloc(1, sizeof *tdb))) {
+	tdb = malloc(sizeof(*tdb));
+	if (!tdb) {
 		/* Can't log this */
 		errno = ENOMEM;
 		goto fail;
 	}
-	tdb->fd = -1;
 	tdb->name = NULL;
 	tdb->map_ptr = NULL;
+	tdb->fd = -1;
+	/* map_size will be set below. */
+	tdb->ecode = TDB_SUCCESS;
+	/* header will be read in below. */
+	tdb->header_uptodate = false;
 	tdb->flags = tdb_flags;
 	tdb->log = null_log_fn;
 	tdb->log_priv = NULL;
 	tdb->khash = jenkins_hash;
 	tdb->hash_priv = NULL;
+	tdb->transaction = NULL;
+	/* last_zone will be set below. */
 	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) {
@@ -238,12 +264,13 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
 	}
 
 	if ((open_flags & O_ACCMODE) == O_RDONLY) {
-		tdb->read_only = 1;
+		tdb->read_only = true;
 		/* read only databases don't do locking */
 		tdb->flags |= TDB_NOLOCK;
-	}
+	} else
+		tdb->read_only = false;
 
-	/* internal databases don't mmap or lock */
+	/* internal databases don't need any of the rest. */
 	if (tdb->flags & TDB_INTERNAL) {
 		tdb->flags |= (TDB_NOLOCK | TDB_NOMMAP);
 		if (tdb_new_database(tdb) != 0) {
@@ -253,7 +280,9 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
 		}
 		TEST_IT(tdb->flags & TDB_CONVERT);
 		tdb_convert(tdb, &tdb->header, sizeof(tdb->header));
-		goto internal;
+		/* Zones don't matter for internal db. */
+		tdb->last_zone = 0;
+		return tdb;
 	}
 
 	if ((tdb->fd = open(name, open_flags, mode)) == -1) {
@@ -331,15 +360,10 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
 	tdb->map_size = st.st_size;
 	tdb->device = st.st_dev;
 	tdb->inode = st.st_ino;
-	tdb_io_init(tdb);
 	tdb_mmap(tdb);
-
- internal:
-	/* Internal (memory-only) databases skip all the code above to
-	 * do with disk files, and resume here by releasing their
-	 * open lock and hooking into the active list. */
 	tdb_unlock_open(tdb);
-	tdb->last_zone = random_free_zone(tdb);
+	tdb_zone_init(tdb);
+
 	tdb->next = tdbs;
 	tdbs = tdb;
 	return tdb;
@@ -370,140 +394,166 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
 	return NULL;
 }
 
-static int tdb_key_compare(TDB_DATA key, TDB_DATA data, void *private_data)
-{
-	return memcmp(data.dptr, key.dptr, data.dsize) == 0;
-}
-
-static void unlock_lists(struct tdb_context *tdb,
-			 uint64_t start, uint64_t end, int ltype)
+static tdb_off_t hash_off(struct tdb_context *tdb, uint64_t list)
 {
-	do {
-		tdb_unlock_list(tdb, start, ltype);
-		start = (start + ((1ULL << tdb->header.v.hash_bits) - 1))
-			& ((1ULL << tdb->header.v.hash_bits) - 1);
-	} while (start != end);
+	return tdb->header.v.hash_off
+		+ ((list & ((1ULL << tdb->header.v.hash_bits) - 1))
+		   * sizeof(tdb_off_t));
 }
 
-/* FIXME: Return header copy? */
-/* Returns -1 or offset of entry (0 if not found).
- * Locks hash entried from *start to *end (where the entry was found). */
-static tdb_off_t find_bucket_and_lock(struct tdb_context *tdb,
-				      const struct tdb_data *key,
-				      uint64_t hash,
-				      uint64_t *start,
-				      uint64_t *end,
-				      uint64_t *room,
-				      int ltype)
+/* Returns 0 if the entry is a zero (definitely not a match).
+ * Returns a valid entry offset if it's a match.  Fills in rec.
+ * Otherwise returns TDB_OFF_ERR: keep searching. */
+static tdb_off_t entry_matches(struct tdb_context *tdb,
+			       uint64_t list,
+			       uint64_t hash,
+			       const struct tdb_data *key,
+			       struct tdb_used_record *rec)
 {
-	uint64_t hextra;
 	tdb_off_t off;
+	uint64_t keylen;
+	const unsigned char *rkey;
 
-	/* hash_bits might be out of date... */
-again:
-	*start = *end = hash & ((1ULL << tdb->header.v.hash_bits) - 1);
-	hextra = hash >> tdb->header.v.hash_bits;
+	list &= ((1ULL << tdb->header.v.hash_bits) - 1);
 
-	/* FIXME: can we avoid locks for some fast paths? */
-	if (tdb_lock_list(tdb, *end, ltype, TDB_LOCK_WAIT) == -1)
+	off = tdb_read_off(tdb, tdb->header.v.hash_off
+			   + list * sizeof(tdb_off_t));
+	if (off == 0 || off == TDB_OFF_ERR)
+		return off;
+
+#if 0 /* FIXME: Check other bits. */
+	unsigned int bits, bitmask, hoffextra;
+	/* Bottom three bits show how many extra hash bits. */
+	bits = (off & ((1 << TDB_EXTRA_HASHBITS_NUM) - 1)) + 1;
+	bitmask = (1 << bits)-1;
+	hoffextra = ((off >> TDB_EXTRA_HASHBITS_NUM) & bitmask);
+	uint64_t hextra = hash >> tdb->header.v.hash_bits;
+	if ((hextra & bitmask) != hoffextra) 
 		return TDB_OFF_ERR;
+	off &= ~...;
+#endif
 
-	/* We only need to check this for first lock. */
-	if (unlikely(update_header(tdb))) {
-		tdb_unlock_list(tdb, *end, ltype);
-		goto again;
-	}
+	if (tdb_read_convert(tdb, off, rec, sizeof(*rec)) == -1)
+		return TDB_OFF_ERR;
+
+	/* FIXME: check extra bits in header! */
+	keylen = rec_key_length(rec);
+	if (keylen != key->dsize)
+		return TDB_OFF_ERR;
 
-	while ((off = tdb_read_off(tdb, tdb->header.v.hash_off
-				   + *end * sizeof(tdb_off_t)))
-	       != TDB_OFF_ERR) {
-		struct tdb_used_record pad, *r;
-		uint64_t keylen, next;
+	rkey = tdb_access_read(tdb, off + sizeof(*rec), keylen, false);
+	if (!rkey)
+		return TDB_OFF_ERR;
+	if (memcmp(rkey, key->dptr, keylen) != 0)
+		off = TDB_OFF_ERR;
+	tdb_access_release(tdb, rkey);
+	return off;
+}
 
-		/* Didn't find it? */
-		if (!off)
-			return 0;
+/* FIXME: Optimize? */
+static void unlock_lists(struct tdb_context *tdb,
+			 tdb_off_t list, tdb_len_t num,
+			 int ltype)
+{
+	tdb_off_t i;
 
-#if 0 /* FIXME: Check other bits. */
-		unsigned int bits, bitmask, hoffextra;
-		/* Bottom three bits show how many extra hash bits. */
-		bits = (off & ((1 << TDB_EXTRA_HASHBITS_NUM) - 1)) + 1;
-		bitmask = (1 << bits)-1;
-		hoffextra = ((off >> TDB_EXTRA_HASHBITS_NUM) & bitmask);
-		if ((hextra & bitmask) != hoffextra) 
-			goto lock_next;
-#endif
+	for (i = list; i < list + num; i++)
+		tdb_unlock_list(tdb, i, ltype);
+}
 
-		r = tdb_get(tdb, off, &pad, sizeof(*r));
-		if (!r)
-			goto unlock_err;
+/* FIXME: Optimize? */
+static int lock_lists(struct tdb_context *tdb,
+		      tdb_off_t list, tdb_len_t num,
+		      int ltype)
+{
+	tdb_off_t i;
 
-		if (rec_magic(r) != TDB_MAGIC) {
-			tdb->ecode = TDB_ERR_CORRUPT;
-			tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
-				 "find_bucket_and_lock: bad magic 0x%llx"
-				 " at offset %llu!\n",
-				 (long long)rec_magic(r), (long long)off);
-			goto unlock_err;
+	for (i = list; i < list + num; i++) {
+		if (tdb_lock_list(tdb, i, ltype, TDB_LOCK_WAIT)
+		    == TDB_OFF_ERR) {
+			unlock_lists(tdb, list, i - list, ltype);
+			return -1;
 		}
+	}
+	return 0;
+}
 
-		/* FIXME: check extra bits in header! */
-		keylen = rec_key_length(r);
-		if (keylen != key->dsize)
-			goto lock_next;
-
-		switch (tdb_parse_data(tdb, *key, off + sizeof(*r), key->dsize,
-				       tdb_key_compare, NULL)) {
-		case 1:
-			/* Match! */
-			*room = rec_data_length(r) + rec_extra_padding(r);
-			return off >> TDB_EXTRA_HASHBITS_NUM;
-		case 0:
-			break;
-		default:
-			goto unlock_err;
+/* We lock hashes up to the next empty offset.  We already hold the
+ * lock on the start bucket, but we may need to release and re-grab
+ * it.  If we fail, we hold no locks at all! */
+static tdb_len_t relock_hash_to_zero(struct tdb_context *tdb,
+				     tdb_off_t start, int ltype)
+{
+	tdb_len_t num, len;
+
+again:
+	num = 1ULL << tdb->header.v.hash_bits;
+	len = tdb_find_zero_off(tdb, hash_off(tdb, start), num - start);
+	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;
+		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;
 		}
 
-	lock_next:
-		/* Lock next bucket. */
-		/* FIXME: We can deadlock if this wraps! */
-		next = (*end + 1) & ((1ULL << tdb->header.v.hash_bits) - 1);
-		if (next == *start) {
-			tdb->ecode = TDB_ERR_CORRUPT;
-			tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
-				 "find_bucket_and_lock: full hash table!\n");
-			goto unlock_err;
+		/* 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 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;
 		}
-		if (tdb_lock_list(tdb, next, ltype, TDB_LOCK_WAIT) == -1)
-			goto unlock_err;
-		*end = next;
 	}
 
-unlock_err:
-	TEST_IT(*end < *start);
-	unlock_lists(tdb, *start, *end, ltype);
-	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, start + len - 1)) != 0)) {
+		/* Leave the start locked, as expected. */
+		unlock_lists(tdb, start + 1, len - 1, ltype);
+		goto again;
+	}
+
+	return len;
 }
 
+/* FIXME: modify, don't rewrite! */
 static int update_rec_hdr(struct tdb_context *tdb,
 			  tdb_off_t off,
 			  tdb_len_t keylen,
 			  tdb_len_t datalen,
-			  tdb_len_t room,
+			  struct tdb_used_record *rec,
 			  uint64_t h)
 {
-	struct tdb_used_record 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));
+	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;
 
@@ -511,32 +561,36 @@ static void enlarge_hash(struct tdb_context *tdb)
 	if (tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_WAIT, false) == -1)
 		return;
 
-	if (unlikely(update_header(tdb))) {
-		/* Someone else enlarged for us?  Nothing to do. */
-		if ((1ULL << tdb->header.v.hash_bits) != num)
-			goto unlock;
-	}
+	/* Someone else enlarged for us?  Nothing to do. */
+	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)) {
@@ -545,27 +599,32 @@ static void enlarge_hash(struct tdb_context *tdb)
 				 "find_bucket_and_lock: 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);
 }
@@ -573,15 +632,43 @@ unlock:
 int tdb_store(struct tdb_context *tdb,
 	      struct tdb_data key, struct tdb_data dbuf, int flag)
 {
-	tdb_off_t new_off, off, start, end, room;
+	tdb_off_t new_off, off, old_bucket, start, num_locks = 1;
+	struct tdb_used_record rec;
 	uint64_t h;
 	bool growing = false;
 
 	h = tdb_hash(tdb, key.dptr, key.dsize);
-	off = find_bucket_and_lock(tdb, &key, h, &start, &end, &room, F_WRLCK);
-	if (off == TDB_OFF_ERR)
+
+	/* 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)
 		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)
+				break;
+		}
+		if (i == start + num_locks)
+			off = 0;
+
+		/* Even if not found, this is where we put the new entry. */
+		old_bucket = i;
+	}
+
 	/* Now we have lock on this hash bucket. */
 	if (flag == TDB_INSERT) {
 		if (off) {
@@ -590,11 +677,12 @@ int tdb_store(struct tdb_context *tdb,
 		}
 	} else {
 		if (off) {
-			if (room >= key.dsize + dbuf.dsize) {
+			if (rec_data_length(&rec) + rec_extra_padding(&rec)
+			    >= dbuf.dsize) {
 				new_off = off;
 				if (update_rec_hdr(tdb, off,
 						   key.dsize, dbuf.dsize,
-						   room, h))
+						   &rec, h))
 					goto fail;
 				goto write;
 			}
@@ -615,7 +703,7 @@ int tdb_store(struct tdb_context *tdb,
 	/* Allocate a new record. */
 	new_off = alloc(tdb, key.dsize, dbuf.dsize, h, growing);
 	if (new_off == 0) {
-		unlock_lists(tdb, start, end, F_WRLCK);
+		unlock_lists(tdb, start, num_locks, F_WRLCK);
 		/* Expand, then try again... */
 		if (tdb_expand(tdb, key.dsize, dbuf.dsize, growing) == -1)
 			return -1;
@@ -625,15 +713,16 @@ int tdb_store(struct tdb_context *tdb,
 	/* We didn't like the existing one: remove it. */
 	if (off) {
 		add_free_record(tdb, off, sizeof(struct tdb_used_record)
-				+ key.dsize + room);
+				+ rec_key_length(&rec)
+				+ rec_data_length(&rec)
+				+ rec_extra_padding(&rec));
 	}
 
-write:
-	off = tdb->header.v.hash_off + end * sizeof(tdb_off_t);
 	/* FIXME: Encode extra hash bits! */
-	if (tdb_write_off(tdb, off, new_off) == -1)
+	if (tdb_write_off(tdb, hash_off(tdb, old_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;
@@ -642,48 +731,65 @@ write:
 		goto fail;
 
 	/* FIXME: tdb_increment_seqnum(tdb); */
-	unlock_lists(tdb, start, end, F_WRLCK);
+	unlock_lists(tdb, start, num_locks, F_WRLCK);
 
-	/* By simple trial and error, this roughly approximates a 60%
-	 * full measure. */
-	if (unlikely(end - start > 4 * tdb->header.v.hash_bits - 32))
+	/* FIXME: by simple simulation, this approximated 60% full.
+	 * Check in real case! */
+	if (unlikely(num_locks > 4 * tdb->header.v.hash_bits - 30))
 		enlarge_hash(tdb);
 
 	return 0;
 
 fail:
-	unlock_lists(tdb, start, end, F_WRLCK);
+	unlock_lists(tdb, start, num_locks, F_WRLCK);
 	return -1;
 }
 
 struct tdb_data tdb_fetch(struct tdb_context *tdb, struct tdb_data key)
 {
-	tdb_off_t off, start, end, room;
+	tdb_off_t off, start, num_locks = 1;
+	struct tdb_used_record rec;
 	uint64_t h;
-	struct tdb_used_record pad, *r;
 	struct tdb_data ret;
 
 	h = tdb_hash(tdb, key.dptr, key.dsize);
-	off = find_bucket_and_lock(tdb, &key, h, &start, &end, &room, F_RDLCK);
-	if (off == TDB_OFF_ERR)
-		return tdb_null;
 
-	if (!off) {
-		unlock_lists(tdb, start, end, F_RDLCK);
-		tdb->ecode = TDB_SUCCESS;
+	/* 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)
 		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;
 	}
 
-	r = tdb_get(tdb, off, &pad, sizeof(*r));
-	if (!r) {
-		unlock_lists(tdb, start, end, F_RDLCK);
+	if (!off) {
+		unlock_lists(tdb, start, num_locks, F_RDLCK);
+		tdb->ecode = TDB_ERR_NOEXIST;
 		return tdb_null;
 	}
 
-	ret.dsize = rec_data_length(r);
-	ret.dptr = tdb_alloc_read(tdb, off + sizeof(*r) + key.dsize,
+	ret.dsize = rec_data_length(&rec);
+	ret.dptr = tdb_alloc_read(tdb, off + sizeof(rec) + key.dsize,
 				  ret.dsize);
-	unlock_lists(tdb, start, end, F_RDLCK);
+	unlock_lists(tdb, start, num_locks, F_RDLCK);
 	return ret;
 }
 
@@ -691,154 +797,128 @@ static int hash_add(struct tdb_context *tdb, uint64_t h, tdb_off_t off)
 {
 	tdb_off_t i, hoff, len, num;
 
+	/* Look for next space. */
 	i = (h & ((1ULL << tdb->header.v.hash_bits) - 1));
-	hoff = tdb->header.v.hash_off + i * sizeof(tdb_off_t);
 	len = (1ULL << tdb->header.v.hash_bits) - i;
+	num = tdb_find_zero_off(tdb, hash_off(tdb, i), len);
 
-	/* Look for next space. */
-	num = tdb_find_zero_off(tdb, hoff, len);
 	if (unlikely(num == len)) {
-		hoff = tdb->header.v.hash_off;
+		/* We wrapped.  Look through start of hash table. */
+		hoff = hash_off(tdb, 0);
 		len = (1ULL << tdb->header.v.hash_bits);
 		num = tdb_find_zero_off(tdb, hoff, len);
-		if (i == len)
+		if (i == len) {
+			tdb->ecode = TDB_ERR_CORRUPT;
+			tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
+				 "hash_add: full hash table!\n");
 			return -1;
+		}
 	}
 	/* FIXME: Encode extra hash bits! */
-	return tdb_write_off(tdb, hoff + num * sizeof(tdb_off_t), off);
+	return tdb_write_off(tdb, hash_off(tdb, i + num), off);
 }
 
-static int unlink_used_record(struct tdb_context *tdb, tdb_off_t chain,
-			      uint64_t *extra_locks)
+int tdb_delete(struct tdb_context *tdb, struct tdb_data key)
 {
-	tdb_off_t num, len, i, hoff;
+	tdb_off_t i, old_bucket, off, start, num_locks = 1;
+	struct tdb_used_record rec;
+	uint64_t h;
 
-	/* FIXME: Maybe lock more in search?  Maybe don't lock if scan
-	 * finds none? */
-again:
-	len = (1ULL << tdb->header.v.hash_bits) - (chain + 1);
-	hoff = tdb->header.v.hash_off + (chain + 1) * sizeof(tdb_off_t);
-	num = tdb_find_zero_off(tdb, hoff, len);
-
-	/* We want to lock the zero entry, too.  In the wrap case,
-	 * this locks one extra.  That's harmless. */
-	num++;
-
-	for (i = chain + 1; i < chain + 1 + num; i++) {
-		if (tdb_lock_list(tdb, i, F_WRLCK, TDB_LOCK_WAIT) == -1) {
-			if (i != chain + 1)
-				unlock_lists(tdb, chain + 1, i-1, F_WRLCK);
+	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)
+		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;
-		}
-	}
 
-	/* The wrap case: we need those locks out of order! */
-	if (unlikely(num == len + 1)) {
-		*extra_locks = tdb_find_zero_off(tdb, tdb->header.v.hash_off,
-						 1ULL << tdb->header.v.hash_bits);
-		(*extra_locks)++;
-		for (i = 0; i < *extra_locks; i++) {
-			if (tdb_lock_list(tdb, i, F_WRLCK, TDB_LOCK_NOWAIT)) {
-				/* Failed.  Caller must lock in order. */
-				if (i)
-					unlock_lists(tdb, 0, i-1, F_WRLCK);
-				unlock_lists(tdb, chain + 1, chain + num,
-					     F_WRLCK);
-				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;
 			}
 		}
-		num += *extra_locks;
+		if (i == start + num_locks)
+			off = 0;
 	}
 
-	/* Now we have the locks, be certain that offset is still 0! */
-	hoff = tdb->header.v.hash_off
-		+ (((chain + num) * sizeof(tdb_off_t))
-		   & ((1ULL << tdb->header.v.hash_bits) - 1));
-
-	if (unlikely(tdb_read_off(tdb, hoff) != 0)) {
-		unlock_lists(tdb, chain + 1, chain + num, F_WRLCK);
-		goto again;
+	if (!off) {
+		unlock_lists(tdb, start, num_locks, F_WRLCK);
+		tdb->ecode = TDB_ERR_NOEXIST;
+		return -1;
 	}
 
-	/* OK, all locked.  Unlink first one. */
-	hoff = tdb->header.v.hash_off + chain * sizeof(tdb_off_t);
-	if (tdb_write_off(tdb, hoff, 0) == -1)
+delete:
+	/* This actually unlinks it. */
+	if (tdb_write_off(tdb, hash_off(tdb, old_bucket), 0) == -1)
 		goto unlock_err;
 
-	/* Rehash the rest. */
-	for (i = 1; i < num; i++) {
-		tdb_off_t off;
-		uint64_t h;
+	/* Rehash anything following. */
+	for (i = hash_off(tdb, old_bucket+1);
+	     i != hash_off(tdb, h + num_locks);
+	     i += sizeof(tdb_off_t)) {
+		tdb_off_t off2;
+		uint64_t h2;
 
-		hoff = tdb->header.v.hash_off
-			+ (((chain + i) * sizeof(tdb_off_t))
-			   & ((1ULL << tdb->header.v.hash_bits) - 1));
-		off = tdb_read_off(tdb, hoff);
-		if (unlikely(off == TDB_OFF_ERR))
+		off2 = tdb_read_off(tdb, i);
+		if (unlikely(off2 == TDB_OFF_ERR))
 			goto unlock_err;
 
 		/* Maybe use a bit to indicate it is in ideal place? */
-		h = hash_record(tdb, off);
+		h2 = hash_record(tdb, off2);
 		/* Is it happy where it is? */
-		if ((h & ((1ULL << tdb->header.v.hash_bits)-1)) == (chain + i))
+		if (hash_off(tdb, h2) == i)
 			continue;
 
 		/* Remove it. */
-		if (tdb_write_off(tdb, hoff, 0) == -1)
+		if (tdb_write_off(tdb, i, 0) == -1)
 			goto unlock_err;
 
 		/* Rehash it. */
-		if (hash_add(tdb, h, off) == -1)
+		if (hash_add(tdb, h2, off2) == -1)
 			goto unlock_err;
 	}
-	unlock_lists(tdb, chain + 1, chain + num, F_WRLCK);
+
+	/* Free the deleted entry. */
+	if (add_free_record(tdb, off,
+			    sizeof(struct tdb_used_record)
+			    + rec_key_length(&rec)
+			    + rec_data_length(&rec)
+			    + rec_extra_padding(&rec)) != 0)
+		goto unlock_err;
+
+	unlock_lists(tdb, start, num_locks, F_WRLCK);
 	return 0;
 
 unlock_err:
-	unlock_lists(tdb, chain + 1, chain + num, F_WRLCK);
+	unlock_lists(tdb, start, num_locks, F_WRLCK);
 	return -1;
 }
 
-int tdb_delete(struct tdb_context *tdb, struct tdb_data key)
-{
-	tdb_off_t off, start, end, room, extra_locks = 0;
-	uint64_t h;
-	int ret;
-
-	h = tdb_hash(tdb, key.dptr, key.dsize);
-	off = find_bucket_and_lock(tdb, &key, h, &start, &end, &room, F_WRLCK);
-	if (off == TDB_OFF_ERR)
-		return -1;
-
-	if (off == 0) {
-		unlock_lists(tdb, start, end, F_WRLCK);
-		tdb->ecode = TDB_ERR_NOEXIST;
-		return -1;
-	}
-
-	ret = unlink_used_record(tdb, end, &extra_locks);
-	if (unlikely(ret == 1)) {
-		unsigned int i;
-
-		unlock_lists(tdb, start, end, F_WRLCK);
-
-		/* We need extra locks at the start. */
-		for (i = 0; i < extra_locks; i++) {
-			if (tdb_lock_list(tdb, i, F_WRLCK, TDB_LOCK_WAIT)) {
-				if (i)
-					unlock_lists(tdb, 0, i-1, F_WRLCK);
-				return -1;
-			}
-		}
-		/* Try again now we're holding more locks. */
-		ret = tdb_delete(tdb, key);
-		unlock_lists(tdb, 0, i, F_WRLCK);
-		return ret;
-	}
-	unlock_lists(tdb, start, end, F_WRLCK);
-	return ret;
-}
-
 int tdb_close(struct tdb_context *tdb)
 {
 	struct tdb_context **i;
@@ -879,3 +959,8 @@ int tdb_close(struct tdb_context *tdb)
 
 	return ret;
 }
+
+enum TDB_ERROR tdb_error(struct tdb_context *tdb)
+{
+	return tdb->ecode;
+}