X-Git-Url: http://git.ozlabs.org/?p=ccan;a=blobdiff_plain;f=ccan%2Ftdb2%2Ftdb.c;h=79c76b1991fe5eca77edfdad99ab8d07e3a79d15;hp=5d69eedcdc4ecf58079609628d59b902c4623b56;hb=d1383862ad9a74e713dc915d351b74da4db35078;hpb=06e0037d97f5e1d83667ec40627cef862f3b7b85

diff --git a/ccan/tdb2/tdb.c b/ccan/tdb2/tdb.c
index 5d69eedc..79c76b19 100644
--- a/ccan/tdb2/tdb.c
+++ b/ccan/tdb2/tdb.c
@@ -1,6 +1,5 @@
 #include "private.h"
 #include <ccan/tdb2/tdb2.h>
-#include <ccan/hash/hash.h>
 #include <ccan/build_assert/build_assert.h>
 #include <ccan/likely/likely.h>
 #include <assert.h>
@@ -11,52 +10,13 @@ struct tdb_data tdb_null = { .dptr = NULL, .dsize = 0 };
 /* all contexts, to ensure no double-opens (fcntl locks don't nest!) */
 static struct tdb_context *tdbs = NULL;
 
-PRINTF_ATTRIBUTE(4, 5) static void
+PRINTF_FMT(4, 5) static void
 null_log_fn(struct tdb_context *tdb,
 	    enum tdb_debug_level level, void *priv,
 	    const char *fmt, ...)
 {
 }
 
-/* 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)
-{
-	struct tdb_header_volatile pad, *v;
-
-	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->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;
-	}
-	tdb->header.v = *v;
-	return true;
-}
-
-static uint64_t jenkins_hash(const void *key, size_t length, uint64_t seed,
-			     void *arg)
-{
-	return hash64_stable((const unsigned char *)key, length, seed);
-}
-
-uint64_t tdb_hash(struct tdb_context *tdb, const void *ptr, size_t len)
-{
-	return tdb->khash(ptr, len, tdb->header.hash_seed, tdb->hash_priv);
-}
-
 static bool tdb_already_open(dev_t device, ino_t ino)
 {
 	struct tdb_context *i;
@@ -120,59 +80,49 @@ static uint64_t random_number(struct tdb_context *tdb)
 
 struct new_database {
 	struct tdb_header hdr;
-	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 */
+	/* Initial free zone. */
+	struct free_zone_header zhdr;
+	tdb_off_t free[BUCKETS_FOR_ZONE(INITIAL_ZONE_BITS) + 1];
 };
 
 /* initialise a new database */
-static int tdb_new_database(struct tdb_context *tdb)
+static int tdb_new_database(struct tdb_context *tdb,
+			    struct tdb_attribute_seed *seed,
+			    struct tdb_header *hdr)
 {
 	/* 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 magic_len;
 
 	/* Fill in the header */
 	newdb.hdr.version = TDB_VERSION;
-	newdb.hdr.hash_seed = random_number(tdb);
+	if (seed)
+		newdb.hdr.hash_seed = seed->seed;
+	else
+		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);
+	memset(newdb.hdr.reserved, 0, sizeof(newdb.hdr.reserved));
+	/* Initial hashes are empty. */
+	memset(newdb.hdr.hashtable, 0, sizeof(newdb.hdr.hashtable));
 
-	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);
+	/* Free is empty. */
+	newdb.zhdr.zone_bits = INITIAL_ZONE_BITS;
 	memset(newdb.free, 0, sizeof(newdb.free));
 
-	/* 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));
-
 	/* Magic food */
 	memset(newdb.hdr.magic_food, 0, sizeof(newdb.hdr.magic_food));
 	strcpy(newdb.hdr.magic_food, TDB_MAGIC_FOOD);
 
 	/* This creates an endian-converted database, as if read from disk */
+	magic_len = sizeof(newdb.hdr.magic_food);
 	tdb_convert(tdb,
-		    (char *)&newdb.hdr + magic_off,
-		    sizeof(newdb) - magic_off);
+		    (char *)&newdb.hdr + magic_len, sizeof(newdb) - magic_len);
 
-	tdb->header = newdb.hdr;
+	*hdr = newdb.hdr;
 
 	if (tdb->flags & TDB_INTERNAL) {
 		tdb->map_size = sizeof(newdb);
@@ -206,6 +156,8 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
 	int save_errno;
 	uint64_t hash_test;
 	unsigned v;
+	struct tdb_header hdr;
+	struct tdb_attribute_seed *seed = NULL;
 
 	tdb = malloc(sizeof(*tdb));
 	if (!tdb) {
@@ -215,27 +167,40 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
 	}
 	tdb->name = NULL;
 	tdb->map_ptr = NULL;
+	tdb->direct_access = 0;
 	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;
 	tdb->flags = tdb_flags;
 	tdb->log = null_log_fn;
 	tdb->log_priv = NULL;
-	tdb->khash = jenkins_hash;
-	tdb->hash_priv = NULL;
 	tdb->transaction = NULL;
+	tdb_hash_init(tdb);
 	/* 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;
+		case TDB_ATTRIBUTE_SEED:
+			seed = &attr->seed;
+			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) {
@@ -249,19 +214,23 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
 		tdb->read_only = true;
 		/* read only databases don't do locking */
 		tdb->flags |= TDB_NOLOCK;
-	} else
+		tdb->mmap_flags = PROT_READ;
+	} else {
 		tdb->read_only = false;
+		tdb->mmap_flags = PROT_READ | PROT_WRITE;
+	}
 
 	/* 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) {
+		if (tdb_new_database(tdb, seed, &hdr) != 0) {
 			tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 				 "tdb_open: tdb_new_database failed!");
 			goto fail;
 		}
-		TEST_IT(tdb->flags & TDB_CONVERT);
-		tdb_convert(tdb, &tdb->header, sizeof(tdb->header));
+		tdb_convert(tdb, &hdr.hash_seed, sizeof(hdr.hash_seed));
+		tdb->hash_seed = hdr.hash_seed;
+		tdb_zone_init(tdb);
 		return tdb;
 	}
 
@@ -284,32 +253,33 @@ struct tdb_context *tdb_open(const char *name, int tdb_flags,
 		goto fail;	/* errno set by tdb_brlock */
 	}
 
-	if (!tdb_pread_all(tdb->fd, &tdb->header, sizeof(tdb->header), 0)
-	    || strcmp(tdb->header.magic_food, TDB_MAGIC_FOOD) != 0) {
-		if (!(open_flags & O_CREAT) || tdb_new_database(tdb) == -1) {
+	if (!tdb_pread_all(tdb->fd, &hdr, sizeof(hdr), 0)
+	    || strcmp(hdr.magic_food, TDB_MAGIC_FOOD) != 0) {
+		if (!(open_flags & O_CREAT)
+		    || tdb_new_database(tdb, seed, &hdr) == -1) {
 			if (errno == 0) {
 				errno = EIO; /* ie bad format or something */
 			}
 			goto fail;
 		}
-	} else if (tdb->header.version != TDB_VERSION) {
-		if (tdb->header.version == bswap_64(TDB_VERSION))
+	} else if (hdr.version != TDB_VERSION) {
+		if (hdr.version == bswap_64(TDB_VERSION))
 			tdb->flags |= TDB_CONVERT;
 		else {
 			/* wrong version */
 			tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 				 "tdb_open: %s is unknown version 0x%llx\n",
-				 name, (long long)tdb->header.version);
+				 name, (long long)hdr.version);
 			errno = EIO;
 			goto fail;
 		}
 	}
 
-	tdb_convert(tdb, &tdb->header, sizeof(tdb->header));
+	tdb_convert(tdb, &hdr, sizeof(hdr));
+	tdb->hash_seed = hdr.hash_seed;
 	hash_test = TDB_HASH_MAGIC;
-	hash_test = tdb->khash(&hash_test, sizeof(hash_test),
-			       tdb->header.hash_seed, tdb->hash_priv);
-	if (tdb->header.hash_test != hash_test) {
+	hash_test = tdb_hash(tdb, &hash_test, sizeof(hash_test));
+	if (hdr.hash_test != hash_test) {
 		/* wrong hash variant */
 		tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 			 "tdb_open: %s uses a different hash function\n",
@@ -337,12 +307,16 @@ 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;
@@ -374,215 +348,73 @@ 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)
-{
-	do {
-		tdb_unlock_list(tdb, start, ltype);
-		start = (start + 1) & ((1ULL << tdb->header.v.hash_bits) - 1);
-	} while (start != end);
-}
-
-/* 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)
-{
-	uint64_t hextra;
-	tdb_off_t off;
-
-	/* 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;
-
-	/* FIXME: can we avoid locks for some fast paths? */
-	if (tdb_lock_list(tdb, *end, ltype, TDB_LOCK_WAIT) == -1)
-		return TDB_OFF_ERR;
-
-	/* We only need to check this for first lock. */
-	if (unlikely(update_header(tdb))) {
-		tdb_unlock_list(tdb, *end, ltype);
-		goto again;
-	}
-
-	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;
-
-		/* Didn't find it? */
-		if (!off)
-			return 0;
-
-#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
-
-		r = tdb_get(tdb, off, &pad, sizeof(*r));
-		if (!r)
-			goto unlock_err;
-
-		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;
-		}
-
-		/* 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;
-		}
-
-	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;
-		}
-		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;
-}
-
+/* 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,
+		       rec_zone_bits(rec)))
 		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)
+/* Returns -1 on error, 0 on OK */
+static int replace_data(struct tdb_context *tdb,
+			struct hash_info *h,
+			struct tdb_data key, struct tdb_data dbuf,
+			tdb_off_t old_off, tdb_len_t old_room,
+			unsigned old_zone,
+			bool growing)
 {
-	tdb_off_t newoff, i;
-	uint64_t h, num = 1ULL << tdb->header.v.hash_bits;
-	struct tdb_used_record pad, *r;
-
-	/* FIXME: We should do this without holding locks throughout. */
-	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;
-	}
+	tdb_off_t new_off;
 
-	newoff = alloc(tdb, 0, num * 2, 0, false);
-	if (unlikely(newoff == TDB_OFF_ERR))
-		goto unlock;
-	if (unlikely(newoff == 0)) {
-		if (tdb_expand(tdb, 0, num * 2, false) == -1)
-			goto unlock;
-		newoff = alloc(tdb, 0, num * 2, 0, false);
-		if (newoff == TDB_OFF_ERR || newoff == 0)
-			goto unlock;
-	}
+	/* Allocate a new record. */
+	new_off = alloc(tdb, key.dsize, dbuf.dsize, h->h, growing);
+	if (unlikely(new_off == TDB_OFF_ERR))
+		return -1;
 
-	/* 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);
-	     i < num;
-	     i += tdb_find_nonzero_off(tdb, tdb->header.v.hash_off
-				       + i*sizeof(tdb_off_t), num - i)) {
-		tdb_off_t off;
-		off = tdb_read_off(tdb, tdb->header.v.hash_off
-				   + 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;
+	/* 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);
+		if (replace_in_hash(tdb, h, new_off) == -1)
+			return -1;
+	} else {
+		if (add_to_hash(tdb, h, new_off) == -1)
+			return -1;
 	}
 
-	/* Free up old hash. */
-	r = tdb_get(tdb, tdb->header.v.hash_off, &pad, sizeof(*r));
-	if (!r)
-		goto unlock;
-	add_free_record(tdb, tdb->header.v.hash_off,
-			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));
-unlock:
-	tdb_allrecord_unlock(tdb, F_WRLCK);
+	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, start, end, room;
-	uint64_t h;
-	bool growing = false;
+	struct hash_info h;
+	tdb_off_t off;
+	tdb_len_t old_room = 0;
+	struct tdb_used_record rec;
+	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)
+	off = find_and_lock(tdb, key, F_WRLCK, &h, &rec, NULL);
+	if (unlikely(off == TDB_OFF_ERR))
 		return -1;
 
 	/* Now we have lock on this hash bucket. */
@@ -593,17 +425,23 @@ int tdb_store(struct tdb_context *tdb,
 		}
 	} else {
 		if (off) {
-			if (room >= key.dsize + 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,
-						   room, h))
+						   &rec, h.h))
+					goto fail;
+				if (tdb->methods->write(tdb, off + sizeof(rec)
+							+ key.dsize,
+							dbuf.dptr, dbuf.dsize))
 					goto fail;
-				goto write;
+				tdb_unlock_hashes(tdb, h.hlock_start,
+						  h.hlock_range, 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
@@ -615,233 +453,148 @@ 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);
-		/* 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, off, old_room,
+			   rec_zone_bits(&rec), off != 0);
+	tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range, F_WRLCK);
+	return ret;
+
+fail:
+	tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range, F_WRLCK);
+	return -1;
+}
+
+int tdb_append(struct tdb_context *tdb,
+	       struct tdb_data key, struct tdb_data dbuf)
+{
+	struct hash_info h;
+	tdb_off_t off;
+	struct tdb_used_record rec;
+	tdb_len_t old_room = 0, old_dlen;
+	unsigned char *newdata;
+	struct tdb_data new_dbuf;
+	int ret;
+
+	off = find_and_lock(tdb, key, F_WRLCK, &h, &rec, NULL);
+	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)
-				+ key.dsize + room);
-	}
+		old_dlen = rec_data_length(&rec);
+		old_room = old_dlen + 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)
-		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.h))
+				goto fail;
 
-	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, end, F_WRLCK);
+			/* FIXME: tdb_increment_seqnum(tdb); */
+			tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range,
+					  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;
+	}
 
-	/* By simple trial and error, this roughly approximates a 60%
-	 * full measure. */
-	if (unlikely(end - start > 4 * tdb->header.v.hash_bits - 32))
-		enlarge_hash(tdb);
+	/* If they're using tdb_append(), it implies they're growing record. */
+	ret = replace_data(tdb, &h, key, new_dbuf, off,
+			   old_room, rec_zone_bits(&rec), true);
+	tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range, F_WRLCK);
+	free(newdata);
 
-	return 0;
+	return ret;
 
 fail:
-	unlock_lists(tdb, start, end, F_WRLCK);
+	tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range, F_WRLCK);
 	return -1;
 }
 
 struct tdb_data tdb_fetch(struct tdb_context *tdb, struct tdb_data key)
 {
-	tdb_off_t off, start, end, room;
-	uint64_t h;
-	struct tdb_used_record pad, *r;
+	tdb_off_t off;
+	struct tdb_used_record rec;
+	struct hash_info h;
 	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)
+	off = find_and_lock(tdb, key, F_RDLCK, &h, &rec, NULL);
+	if (unlikely(off == TDB_OFF_ERR))
 		return tdb_null;
 
 	if (!off) {
-		unlock_lists(tdb, start, end, F_RDLCK);
-		tdb->ecode = TDB_SUCCESS;
-		return tdb_null;
-	}
-
-	r = tdb_get(tdb, off, &pad, sizeof(*r));
-	if (!r) {
-		unlock_lists(tdb, start, end, F_RDLCK);
-		return tdb_null;
+		tdb->ecode = TDB_ERR_NOEXIST;
+		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(r);
-	ret.dptr = tdb_alloc_read(tdb, off + sizeof(*r) + key.dsize,
-				  ret.dsize);
-	unlock_lists(tdb, start, end, F_RDLCK);
+	tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range, F_RDLCK);
 	return ret;
 }
 
-static int hash_add(struct tdb_context *tdb, uint64_t h, tdb_off_t off)
+int tdb_delete(struct tdb_context *tdb, struct tdb_data key)
 {
-	tdb_off_t i, hoff, len, num;
-
-	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;
-
-	/* Look for next space. */
-	num = tdb_find_zero_off(tdb, hoff, len);
-	if (unlikely(num == len)) {
-		hoff = tdb->header.v.hash_off;
-		len = (1ULL << tdb->header.v.hash_bits);
-		num = tdb_find_zero_off(tdb, hoff, len);
-		if (i == len)
-			return -1;
-	}
-	/* FIXME: Encode extra hash bits! */
-	return tdb_write_off(tdb, hoff + num * sizeof(tdb_off_t), off);
-}
+	tdb_off_t off;
+	struct tdb_used_record rec;
+	struct hash_info h;
 
-static int unlink_used_record(struct tdb_context *tdb, tdb_off_t chain,
-			      uint64_t *extra_locks)
-{
-	tdb_off_t num, len, i, hoff;
-
-	/* 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);
-			return -1;
-		}
-	}
+	off = find_and_lock(tdb, key, F_WRLCK, &h, &rec, NULL);
+	if (unlikely(off == 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;
-			}
-		}
-		num += *extra_locks;
+	if (!off) {
+		tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range, F_WRLCK);
+		tdb->ecode = TDB_ERR_NOEXIST;
+		return -1;
 	}
 
-	/* 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 (delete_from_hash(tdb, &h) == -1)
+		goto unlock_err;
 
-	/* 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)
+	/* Free the deleted entry. */
+	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;
 
-	/* Rehash the rest. */
-	for (i = 1; i < num; i++) {
-		tdb_off_t off;
-		uint64_t h;
-
-		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))
-			goto unlock_err;
-
-		/* Maybe use a bit to indicate it is in ideal place? */
-		h = hash_record(tdb, off);
-		/* Is it happy where it is? */
-		if ((h & ((1ULL << tdb->header.v.hash_bits)-1)) == (chain + i))
-			continue;
-
-		/* Remove it. */
-		if (tdb_write_off(tdb, hoff, 0) == -1)
-			goto unlock_err;
-
-		/* Rehash it. */
-		if (hash_add(tdb, h, off) == -1)
-			goto unlock_err;
-	}
-	unlock_lists(tdb, chain + 1, chain + num, F_WRLCK);
+	tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range, F_WRLCK);
 	return 0;
 
 unlock_err:
-	unlock_lists(tdb, chain + 1, chain + num, F_WRLCK);
+	tdb_unlock_hashes(tdb, h.hlock_start, h.hlock_range, 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;
@@ -887,3 +640,21 @@ enum TDB_ERROR tdb_error(struct tdb_context *tdb)
 {
 	return tdb->ecode;
 }
+
+const char *tdb_errorstr(struct tdb_context *tdb)
+{
+	/* Gcc warns if you miss a case in the switch, so use that. */
+	switch (tdb->ecode) {
+	case TDB_SUCCESS: return "Success";
+	case TDB_ERR_CORRUPT: return "Corrupt database";
+	case TDB_ERR_IO: return "IO Error";
+	case TDB_ERR_LOCK: return "Locking error";
+	case TDB_ERR_OOM: return "Out of memory";
+	case TDB_ERR_EXISTS: return "Record exists";
+	case TDB_ERR_NESTING: return "Transaction already started";
+	case TDB_ERR_EINVAL: return "Invalid parameter";
+	case TDB_ERR_NOEXIST: return "Record does not exist";
+	case TDB_ERR_RDONLY: return "write not permitted";
+	}
+	return "Invalid error code";
+}