X-Git-Url: https://git.ozlabs.org/?a=blobdiff_plain;f=ccan%2Ftdb2%2Ftdb.c;h=40811c2eff83609ca2c55d757815e2a0cae9f9d1;hb=4f73f6a6dfc0d65aa9a5055683bf7baa5a7b622e;hp=19ebfb4961677f3b0740fa0f07c6f66b5aef1785;hpb=cb399d268c839dd75a59678268f3ba666147cc20;p=ccan

diff --git a/ccan/tdb2/tdb.c b/ccan/tdb2/tdb.c
index 19ebfb49..40811c2e 100644
--- a/ccan/tdb2/tdb.c
+++ b/ccan/tdb2/tdb.c
@@ -124,12 +124,22 @@ static uint64_t random_number(struct tdb_context *tdb)
 	return ret;
 }
 
-struct new_database {
+struct new_db_head {
 	struct tdb_header hdr;
+	struct free_zone_header zhdr;
+	tdb_off_t free[BUCKETS_FOR_ZONE(INITIAL_ZONE_BITS) + 1];
 	struct tdb_used_record hrec;
 	tdb_off_t hash[1ULL << INITIAL_HASH_BITS];
-	struct tdb_used_record frec;
-	tdb_off_t free[INITIAL_FREE_BUCKETS + 1]; /* One overflow bucket */
+	struct tdb_free_record frec;
+};
+
+struct new_database {
+	struct new_db_head h;
+	/* Rest up to 1 << INITIAL_ZONE_BITS is empty. */
+	char space[(1 << INITIAL_ZONE_BITS)
+		   - (sizeof(struct new_db_head) - sizeof(struct tdb_header))];
+	uint8_t tailer;
+	/* Don't count final padding! */
 };
 
 /* initialise a new database */
@@ -137,51 +147,64 @@ static int tdb_new_database(struct tdb_context *tdb)
 {
 	/* We make it up in memory, then write it out if not internal */
 	struct new_database newdb;
-	unsigned int magic_off = offsetof(struct tdb_header, magic_food);
+	unsigned int bucket, magic_off, dbsize;
 
-	/* Fill in the header */
-	newdb.hdr.version = TDB_VERSION;
-	newdb.hdr.hash_seed = random_number(tdb);
-	newdb.hdr.hash_test = TDB_HASH_MAGIC;
-	newdb.hdr.hash_test = tdb->khash(&newdb.hdr.hash_test,
-					 sizeof(newdb.hdr.hash_test),
-					 newdb.hdr.hash_seed,
-					 tdb->hash_priv);
-
-	newdb.hdr.v.generation = 0;
-
-	/* The initial zone must cover the initial database size! */
-	BUILD_ASSERT((1ULL << INITIAL_ZONE_BITS) >= sizeof(newdb));
-
-	/* Free array has 1 zone, 10 buckets.  All buckets empty. */
-	newdb.hdr.v.num_zones = 1;
-	newdb.hdr.v.zone_bits = INITIAL_ZONE_BITS;
-	newdb.hdr.v.free_buckets = INITIAL_FREE_BUCKETS;
-	newdb.hdr.v.free_off = offsetof(struct new_database, free);
-	set_header(tdb, &newdb.frec, 0,
-		   sizeof(newdb.free), sizeof(newdb.free), 0);
-	memset(newdb.free, 0, sizeof(newdb.free));
+	/* Don't want any extra padding! */
+	dbsize = offsetof(struct new_database, tailer) + sizeof(newdb.tailer);
 
+	/* Fill in the header */
+	newdb.h.hdr.version = TDB_VERSION;
+	newdb.h.hdr.hash_seed = random_number(tdb);
+	newdb.h.hdr.hash_test = TDB_HASH_MAGIC;
+	newdb.h.hdr.hash_test = tdb->khash(&newdb.h.hdr.hash_test,
+					   sizeof(newdb.h.hdr.hash_test),
+					   newdb.h.hdr.hash_seed,
+					   tdb->hash_priv);
+	memset(newdb.h.hdr.reserved, 0, sizeof(newdb.h.hdr.reserved));
+	newdb.h.hdr.v.generation = 0;
 	/* Initial hashes are empty. */
-	newdb.hdr.v.hash_bits = INITIAL_HASH_BITS;
-	newdb.hdr.v.hash_off = offsetof(struct new_database, hash);
-	set_header(tdb, &newdb.hrec, 0,
-		   sizeof(newdb.hash), sizeof(newdb.hash), 0);
-	memset(newdb.hash, 0, sizeof(newdb.hash));
+	newdb.h.hdr.v.hash_bits = INITIAL_HASH_BITS;
+	newdb.h.hdr.v.hash_off = offsetof(struct new_database, h.hash);
+	set_header(tdb, &newdb.h.hrec, 0,
+		   sizeof(newdb.h.hash), sizeof(newdb.h.hash), 0,
+		   INITIAL_ZONE_BITS);
+	memset(newdb.h.hash, 0, sizeof(newdb.h.hash));
+
+	/* Create the single free entry. */
+	newdb.h.frec.magic_and_meta = TDB_FREE_MAGIC | INITIAL_ZONE_BITS;
+	newdb.h.frec.data_len = (sizeof(newdb.h.frec)
+				 - sizeof(struct tdb_used_record)
+				 + sizeof(newdb.space));
+
+	/* Free is mostly empty... */
+	newdb.h.zhdr.zone_bits = INITIAL_ZONE_BITS;
+	memset(newdb.h.free, 0, sizeof(newdb.h.free));
+
+	/* ... except for this one bucket. */
+	bucket = size_to_bucket(INITIAL_ZONE_BITS, newdb.h.frec.data_len);
+	newdb.h.free[bucket] = offsetof(struct new_database, h.frec);
+	newdb.h.frec.next = newdb.h.frec.prev = 0;
+
+	/* Clear free space to keep valgrind happy, and avoid leaking stack. */
+	memset(newdb.space, 0, sizeof(newdb.space));
+
+	/* Tailer contains maximum number of free_zone bits. */
+	newdb.tailer = INITIAL_ZONE_BITS;
 
 	/* Magic food */
-	memset(newdb.hdr.magic_food, 0, sizeof(newdb.hdr.magic_food));
-	strcpy(newdb.hdr.magic_food, TDB_MAGIC_FOOD);
+	memset(newdb.h.hdr.magic_food, 0, sizeof(newdb.h.hdr.magic_food));
+	strcpy(newdb.h.hdr.magic_food, TDB_MAGIC_FOOD);
 
 	/* This creates an endian-converted database, as if read from disk */
+	magic_off = offsetof(struct tdb_header, magic_food);
 	tdb_convert(tdb,
-		    (char *)&newdb.hdr + magic_off,
-		    sizeof(newdb) - magic_off);
+		    (char *)&newdb.h.hdr + magic_off,
+		    dbsize - 1 - magic_off);
 
-	tdb->header = newdb.hdr;
+	tdb->header = newdb.h.hdr;
 
 	if (tdb->flags & TDB_INTERNAL) {
-		tdb->map_size = sizeof(newdb);
+		tdb->map_size = dbsize;
 		tdb->map_ptr = malloc(tdb->map_size);
 		if (!tdb->map_ptr) {
 			tdb->ecode = TDB_ERR_OOM;
@@ -196,7 +219,7 @@ static int tdb_new_database(struct tdb_context *tdb)
 	if (ftruncate(tdb->fd, 0) == -1)
 		return -1;
 
-	if (!tdb_pwrite_all(tdb->fd, &newdb, sizeof(newdb), 0)) {
+	if (!tdb_pwrite_all(tdb->fd, &newdb, dbsize, 0)) {
 		tdb->ecode = TDB_ERR_IO;
 		return -1;
 	}
@@ -222,7 +245,7 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
 	tdb->name = NULL;
 	tdb->map_ptr = NULL;
 	tdb->fd = -1;
-	/* map_size will be set below. */
+	tdb->map_size = sizeof(struct tdb_header);
 	tdb->ecode = TDB_SUCCESS;
 	/* header will be read in below. */
 	tdb->header_uptodate = false;
@@ -236,12 +259,24 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
 	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) {
@@ -268,8 +303,7 @@ 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));
-		/* Zones don't matter for internal db. */
-		tdb->last_zone = 0;
+		tdb_zone_init(tdb);
 		return tdb;
 	}
 
@@ -345,12 +379,16 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
 		goto fail;
 	}
 
-	tdb->map_size = st.st_size;
 	tdb->device = st.st_dev;
 	tdb->inode = st.st_ino;
-	tdb_mmap(tdb);
 	tdb_unlock_open(tdb);
-	tdb_zone_init(tdb);
+
+	/* This make sure we have current map_size and mmap. */
+	tdb->methods->oob(tdb, tdb->map_size + 1, true);
+
+	/* Now we can pick a random free zone to start from. */
+	if (tdb_zone_init(tdb) == -1)
+		goto fail;
 
 	tdb->next = tdbs;
 	tdbs = tdb;
@@ -382,7 +420,7 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
 	return NULL;
 }
 
-static tdb_off_t hash_off(struct tdb_context *tdb, uint64_t list)
+tdb_off_t hash_off(struct tdb_context *tdb, uint64_t list)
 {
 	return tdb->header.v.hash_off
 		+ ((list & ((1ULL << tdb->header.v.hash_bits) - 1))
@@ -402,6 +440,8 @@ static tdb_off_t entry_matches(struct tdb_context *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)
@@ -455,7 +495,8 @@ static int lock_lists(struct tdb_context *tdb,
 	tdb_off_t i;
 
 	for (i = list; i < list + num; i++) {
-		if (tdb_lock_list(tdb, i, ltype, TDB_LOCK_WAIT) != 0) {
+		if (tdb_lock_list(tdb, i, ltype, TDB_LOCK_WAIT)
+		    == TDB_OFF_ERR) {
 			unlock_lists(tdb, list, i - list, ltype);
 			return -1;
 		}
@@ -469,7 +510,7 @@ static int lock_lists(struct tdb_context *tdb,
 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;
@@ -477,39 +518,45 @@ again:
 	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_lists(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;
-	} else {
-		/* We already have lock on start. */
-		start++;
-		pre_locks = 0;
-	}
-	if (unlikely(lock_lists(tdb, start, len, ltype) == -1)) {
-		if (pre_locks)
+
+		/* Now lock later ones. */
+		if (unlikely(lock_lists(tdb, start, len, ltype) == -1)) {
 			unlock_lists(tdb, 0, pre_locks, ltype);
-		else
+			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;
+			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_lists(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_lists(tdb, start + 1, len - 1, ltype);
 		goto again;
 	}
 
-	return pre_locks + len;
+	return len;
 }
 
 /* FIXME: modify, don't rewrite! */
@@ -522,19 +569,55 @@ static int update_rec_hdr(struct tdb_context *tdb,
 {
 	uint64_t dataroom = rec_data_length(rec) + rec_extra_padding(rec);
 
-	if (set_header(tdb, rec, keylen, datalen, keylen + dataroom, h))
+	if (set_header(tdb, rec, keylen, datalen, keylen + dataroom, h,
+		       rec_zone_bits(rec)))
 		return -1;
 
 	return tdb_write_convert(tdb, off, rec, sizeof(*rec));
 }
 
+static int hash_add(struct tdb_context *tdb,
+		    uint64_t hash, tdb_off_t off)
+{
+	tdb_off_t i, hoff, len, num;
+
+	/* Look for next space. */
+	i = (hash & ((1ULL << tdb->header.v.hash_bits) - 1));
+	len = (1ULL << tdb->header.v.hash_bits) - i;
+	num = tdb_find_zero_off(tdb, hash_off(tdb, i), len);
+
+	if (unlikely(num == len)) {
+		/* We wrapped.  Look through start of hash table. */
+		i = 0;
+		hoff = hash_off(tdb, 0);
+		len = (1ULL << tdb->header.v.hash_bits);
+		num = tdb_find_zero_off(tdb, hoff, len);
+		if (num == len) {
+			tdb->ecode = TDB_ERR_CORRUPT;
+			tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
+				 "hash_add: full hash table!\n");
+			return -1;
+		}
+	}
+	if (tdb_read_off(tdb, hash_off(tdb, i + num)) != 0) {
+		tdb->ecode = TDB_ERR_CORRUPT;
+		tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
+			 "hash_add: overwriting hash table?\n");
+		return -1;
+	}
+
+	/* FIXME: Encode extra hash bits! */
+	return tdb_write_off(tdb, hash_off(tdb, i + num), off);
+}
+
 /* If we fail, others will try after us. */
 static void enlarge_hash(struct tdb_context *tdb)
 {
 	tdb_off_t newoff, oldoff, i;
 	tdb_len_t hlen;
-	uint64_t h, num = 1ULL << tdb->header.v.hash_bits;
+	uint64_t num = 1ULL << tdb->header.v.hash_bits;
 	struct tdb_used_record pad, *r;
+	unsigned int records = 0;
 
 	/* FIXME: We should do this without holding locks throughout. */
 	if (tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_WAIT, false) == -1)
@@ -549,13 +632,6 @@ static void enlarge_hash(struct tdb_context *tdb)
 	newoff = alloc(tdb, 0, hlen, 0, false);
 	if (unlikely(newoff == TDB_OFF_ERR))
 		goto unlock;
-	if (unlikely(newoff == 0)) {
-		if (tdb_expand(tdb, 0, hlen, false) == -1)
-			goto unlock;
-		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);
 
@@ -563,84 +639,170 @@ static void enlarge_hash(struct tdb_context *tdb)
 	if (zero_out(tdb, newoff, hlen) == -1)
 		goto unlock;
 
+	/* Update header now so we can use normal routines. */
+	oldoff = tdb->header.v.hash_off;
+
+	tdb->header.v.hash_bits++;
+	tdb->header.v.hash_off = newoff;
+
 	/* FIXME: If the space before is empty, we know this is in its ideal
 	 * 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, hash_off(tdb, i), num - i)) {
+	for (i = 0; i < num; i++) {
 		tdb_off_t off;
-		off = tdb_read_off(tdb, hash_off(tdb, i));
+		off = tdb_read_off(tdb, oldoff + i * sizeof(tdb_off_t));
 		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");
-			goto unlock;
-		}
-		h = hash_record(tdb, off);
-		/* FIXME: Encode extra hash bits! */
-		if (tdb_write_off(tdb, newoff
-				  + (h & ((num * 2) - 1)) * sizeof(uint64_t),
-				  off) == -1)
-			goto unlock;
+			goto oldheader;
+		if (off && hash_add(tdb, hash_record(tdb, off), off) == -1)
+			goto oldheader;
+		if (off)
+			records++;
 	}
 
+	tdb->log(tdb, TDB_DEBUG_TRACE, tdb->log_priv,
+		 "enlarge_hash: moved %u records from %llu buckets.\n",
+		 records, (long long)num);
+
 	/* Free up old hash. */
-	oldoff = tdb->header.v.hash_off - sizeof(*r);
-	r = tdb_get(tdb, oldoff, &pad, sizeof(*r));
+	r = tdb_get(tdb, oldoff - sizeof(*r), &pad, sizeof(*r));
 	if (!r)
-		goto unlock;
-	add_free_record(tdb, oldoff,
+		goto oldheader;
+	add_free_record(tdb, rec_zone_bits(r), oldoff - sizeof(*r),
 			sizeof(*r)+rec_data_length(r)+rec_extra_padding(r));
 
 	/* Now we write the modified header. */
-	tdb->header.v.hash_bits++;
-	tdb->header.v.hash_off = newoff;
 	write_header(tdb);
 unlock:
 	tdb_allrecord_unlock(tdb, F_WRLCK);
+	return;
+
+oldheader:
+	tdb->header.v.hash_bits--;
+	tdb->header.v.hash_off = oldoff;
+	goto unlock;
+}
+
+
+/* This is the slow version of the routine which searches the
+ * hashtable for an entry.
+ * We lock every hash bucket up to and including the next zero one.
+ */
+static tdb_off_t find_and_lock_slow(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)
+{
+	/* 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)++) {
+		tdb_off_t 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;
+}
+
+/* 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. */
+	return find_and_lock_slow(tdb, key, h, ltype, start_lock, num_locks,
+				  bucket, rec);
+}
+
+/* Returns -1 on error, 0 on OK" */
+static int replace_data(struct tdb_context *tdb,
+			uint64_t h, struct tdb_data key, struct tdb_data dbuf,
+			tdb_off_t bucket,
+			tdb_off_t old_off, tdb_len_t old_room,
+			unsigned old_zone,
+			bool growing)
+{
+	tdb_off_t new_off;
+
+	/* Allocate a new record. */
+	new_off = alloc(tdb, key.dsize, dbuf.dsize, h, growing);
+	if (unlikely(new_off == TDB_OFF_ERR))
+		return -1;
+
+	/* We didn't like the existing one: remove it. */
+	if (old_off)
+		add_free_record(tdb, old_zone, old_off,
+				sizeof(struct tdb_used_record)
+				+ key.dsize + old_room);
+
+	/* FIXME: Encode extra hash bits! */
+	if (tdb_write_off(tdb, hash_off(tdb, bucket), new_off) == -1)
+		return -1;
+
+	new_off += sizeof(struct tdb_used_record);
+	if (tdb->methods->write(tdb, new_off, key.dptr, key.dsize) == -1)
+		return -1;
+
+	new_off += key.dsize;
+	if (tdb->methods->write(tdb, new_off, dbuf.dptr, dbuf.dsize) == -1)
+		return -1;
+
+	/* FIXME: tdb_increment_seqnum(tdb); */
+	return 0;
 }
 
 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 off, bucket, start, num;
+	tdb_len_t old_room = 0;
 	struct tdb_used_record rec;
 	uint64_t h;
-	bool growing = false;
+	int 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_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) {
@@ -649,18 +811,22 @@ int tdb_store(struct tdb_context *tdb,
 		}
 	} else {
 		if (off) {
-			if (rec_data_length(&rec) + rec_extra_padding(&rec)
-			    >= dbuf.dsize) {
-				new_off = off;
+			old_room = rec_data_length(&rec)
+				+ rec_extra_padding(&rec);
+			if (old_room >= dbuf.dsize) {
+				/* Can modify in-place.  Easy! */
 				if (update_rec_hdr(tdb, off,
 						   key.dsize, dbuf.dsize,
 						   &rec, h))
 					goto fail;
-				goto write;
+				if (tdb->methods->write(tdb, off + sizeof(rec)
+							+ key.dsize,
+							dbuf.dptr, dbuf.dsize))
+					goto fail;
+				unlock_lists(tdb, start, num, F_WRLCK);
+				return 0;
 			}
 			/* FIXME: See if right record is free? */
-			/* Hint to allocator that we've realloced. */
-			growing = true;
 		} else {
 			if (flag == TDB_MODIFY) {
 				/* if the record doesn't exist and we
@@ -672,202 +838,179 @@ 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, num_locks, F_WRLCK);
-		/* Expand, then try again... */
-		if (tdb_expand(tdb, key.dsize, dbuf.dsize, growing) == -1)
-			return -1;
-		return tdb_store(tdb, key, dbuf, flag);
-	}
+	/* If we didn't use the old record, this implies we're growing. */
+	ret = replace_data(tdb, h, key, dbuf, bucket, off, old_room,
+			   rec_zone_bits(&rec), off != 0);
+	unlock_lists(tdb, start, num, F_WRLCK);
+
+	/* FIXME: by simple simulation, this approximated 60% full.
+	 * Check in real case! */
+	if (unlikely(num > 4 * tdb->header.v.hash_bits - 30))
+		enlarge_hash(tdb);
+
+	return ret;
+
+fail:
+	unlock_lists(tdb, start, num, F_WRLCK);
+	return -1;
+}
+
+int tdb_append(struct tdb_context *tdb,
+	       struct tdb_data key, struct tdb_data dbuf)
+{
+	tdb_off_t off, bucket, start, num;
+	struct tdb_used_record rec;
+	tdb_len_t old_room = 0, old_dlen;
+	uint64_t h;
+	unsigned char *newdata;
+	struct tdb_data new_dbuf;
+	int ret;
+
+	h = tdb_hash(tdb, key.dptr, key.dsize);
+	off = find_and_lock(tdb, key, h, F_WRLCK, &start, &num, &bucket, &rec);
+	if (unlikely(off == TDB_OFF_ERR))
+		return -1;
 
-	/* We didn't like the existing one: remove it. */
 	if (off) {
-		add_free_record(tdb, off, sizeof(struct tdb_used_record)
-				+ rec_key_length(&rec)
-				+ rec_data_length(&rec)
-				+ rec_extra_padding(&rec));
-	}
+		old_dlen = rec_data_length(&rec);
+		old_room = old_dlen + rec_extra_padding(&rec);
 
-	/* FIXME: Encode extra hash bits! */
-	if (tdb_write_off(tdb, hash_off(tdb, old_bucket), new_off) == -1)
-		goto fail;
+		/* Fast path: can append in place. */
+		if (rec_extra_padding(&rec) >= dbuf.dsize) {
+			if (update_rec_hdr(tdb, off, key.dsize,
+					   old_dlen + dbuf.dsize, &rec, h))
+				goto fail;
 
-write:
-	off = new_off + sizeof(struct tdb_used_record);
-	if (tdb->methods->write(tdb, off, key.dptr, key.dsize) == -1)
-		goto fail;
-	off += key.dsize;
-	if (tdb->methods->write(tdb, off, dbuf.dptr, dbuf.dsize) == -1)
-		goto fail;
+			off += sizeof(rec) + key.dsize + old_dlen;
+			if (tdb->methods->write(tdb, off, dbuf.dptr,
+						dbuf.dsize) == -1)
+				goto fail;
 
-	/* FIXME: tdb_increment_seqnum(tdb); */
-	unlock_lists(tdb, start, num_locks, F_WRLCK);
+			/* FIXME: tdb_increment_seqnum(tdb); */
+			unlock_lists(tdb, start, num, F_WRLCK);
+			return 0;
+		}
+		/* FIXME: Check right record free? */
+
+		/* Slow path. */
+		newdata = malloc(key.dsize + old_dlen + dbuf.dsize);
+		if (!newdata) {
+			tdb->ecode = TDB_ERR_OOM;
+			tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
+				 "tdb_append: cannot allocate %llu bytes!\n",
+				 (long long)key.dsize + old_dlen + dbuf.dsize);
+			goto fail;
+		}
+		if (tdb->methods->read(tdb, off + sizeof(rec) + key.dsize,
+				       newdata, old_dlen) != 0) {
+			free(newdata);
+			goto fail;
+		}
+		memcpy(newdata + old_dlen, dbuf.dptr, dbuf.dsize);
+		new_dbuf.dptr = newdata;
+		new_dbuf.dsize = old_dlen + dbuf.dsize;
+	} else {
+		newdata = NULL;
+		new_dbuf = dbuf;
+	}
+
+	/* If they're using tdb_append(), it implies they're growing record. */
+	ret = replace_data(tdb, h, key, new_dbuf, bucket, off, old_room,
+			   rec_zone_bits(&rec), true);
+	unlock_lists(tdb, start, num, F_WRLCK);
+	free(newdata);
 
 	/* 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;
+	return ret;
 
 fail:
-	unlock_lists(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_lists(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_lists(tdb, start, num_locks, F_RDLCK);
+	unlock_lists(tdb, start, num, F_RDLCK);
 	return ret;
 }
 
-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));
-	len = (1ULL << tdb->header.v.hash_bits) - i;
-	num = tdb_find_zero_off(tdb, hash_off(tdb, i), len);
-
-	if (unlikely(num == len)) {
-		/* 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) {
-			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, hash_off(tdb, i + num), off);
-}
-
 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)
+	if (unlikely(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;
-
-		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;
-	}
+	/* FIXME: Fastpath: if next is zero, we can delete without lock,
+	 * since this lock protects us. */
+	off = find_and_lock_slow(tdb, key, h, F_WRLCK,
+				 &start, &num, &bucket, &rec);
+	if (unlikely(off == TDB_OFF_ERR))
+		return -1;
 
 	if (!off) {
-		unlock_lists(tdb, start, num_locks, F_WRLCK);
+		/* FIXME: We could optimize not found case if it mattered, by
+		 * reading offset after first lock: if it's zero, goto here. */
+		unlock_lists(tdb, start, num, F_WRLCK);
 		tdb->ecode = TDB_ERR_NOEXIST;
 		return -1;
 	}
+	/* Since we found the entry, we must have locked it and a zero. */
+	assert(num >= 2);
 
-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);
-	     i += sizeof(tdb_off_t)) {
-		tdb_off_t off2;
+	for (i = bucket+1; i != bucket + num - 1; i++) {
+		tdb_off_t hoff, off2;
 		uint64_t h2;
 
-		off2 = tdb_read_off(tdb, i);
+		hoff = hash_off(tdb, i);
+		off2 = tdb_read_off(tdb, hoff);
 		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? */
-		if (hash_off(tdb, h2) == i)
+		if (hash_off(tdb, h2) == hoff)
 			continue;
 
 		/* Remove it. */
-		if (tdb_write_off(tdb, i, 0) == -1)
+		if (tdb_write_off(tdb, hoff, 0) == -1)
 			goto unlock_err;
 
 		/* Rehash it. */
@@ -876,18 +1019,18 @@ delete:
 	}
 
 	/* Free the deleted entry. */
-	if (add_free_record(tdb, off,
+	if (add_free_record(tdb, rec_zone_bits(&rec), 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);
+	unlock_lists(tdb, start, num, F_WRLCK);
 	return 0;
 
 unlock_err:
-	unlock_lists(tdb, start, num_locks, F_WRLCK);
+	unlock_lists(tdb, start, num, F_WRLCK);
 	return -1;
 }