X-Git-Url: http://git.ozlabs.org/?p=ccan;a=blobdiff_plain;f=ccan%2Ftdb2%2Ffree.c;h=e693fe828ae3bebf801c653184256ace07e9d23e;hp=eaaeb3cf2bca594d8294ac9a25b533d45d671eea;hb=c438ec17d7b2efe76e56e5fc5ab88bd4a02735e8;hpb=6b999f4511e8458bd79eddc2951ab91511d23fa8

diff --git a/ccan/tdb2/free.c b/ccan/tdb2/free.c
index eaaeb3cf..e693fe82 100644
--- a/ccan/tdb2/free.c
+++ b/ccan/tdb2/free.c
@@ -65,18 +65,18 @@ enum TDB_ERROR tdb_ftable_init(struct tdb_context *tdb)
 	unsigned int rnd, max = 0, count = 0;
 	tdb_off_t off;
 
-	tdb->ftable_off = off = first_ftable(tdb);
-	tdb->ftable = 0;
+	tdb->tdb2.ftable_off = off = first_ftable(tdb);
+	tdb->tdb2.ftable = 0;
 
 	while (off) {
 		if (TDB_OFF_IS_ERR(off)) {
-			return off;
+			return TDB_OFF_TO_ERR(off);
 		}
 
 		rnd = random();
 		if (rnd >= max) {
-			tdb->ftable_off = off;
-			tdb->ftable = count;
+			tdb->tdb2.ftable_off = off;
+			tdb->tdb2.ftable = count;
 			max = rnd;
 		}
 
@@ -109,7 +109,7 @@ static void check_list(struct tdb_context *tdb, tdb_off_t b_off)
 	tdb_off_t off, prev = 0, first;
 	struct tdb_free_record r;
 
-	first = off = tdb_read_off(tdb, b_off);
+	first = off = (tdb_read_off(tdb, b_off) & TDB_OFF_MASK);
 	while (off != 0) {
 		tdb_read_convert(tdb, off, &r, sizeof(r));
 		if (frec_magic(&r) != TDB_FREE_MAGIC)
@@ -146,21 +146,25 @@ static enum TDB_ERROR remove_from_list(struct tdb_context *tdb,
 	/* Get prev->next */
 	prev_next = tdb_read_off(tdb, off);
 	if (TDB_OFF_IS_ERR(prev_next))
-		return prev_next;
+		return TDB_OFF_TO_ERR(prev_next);
 
 	/* If prev->next == 0, we were head: update bucket to point to next. */
 	if (prev_next == 0) {
-#ifdef CCAN_TDB2_DEBUG
-		if (tdb_read_off(tdb, b_off) != r_off) {
+		/* We must preserve upper bits. */
+		head = tdb_read_off(tdb, b_off);
+		if (TDB_OFF_IS_ERR(head))
+			return TDB_OFF_TO_ERR(head);
+
+		if ((head & TDB_OFF_MASK) != r_off) {
 			return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
 					  "remove_from_list:"
 					  " %llu head %llu on list %llu",
 					  (long long)r_off,
-					  (long long)tdb_read_off(tdb, b_off),
+					  (long long)head,
 					  (long long)b_off);
 		}
-#endif
-		ecode = tdb_write_off(tdb, b_off, r->next);
+		head = ((head & ~TDB_OFF_MASK) | r->next);
+		ecode = tdb_write_off(tdb, b_off, head);
 		if (ecode != TDB_SUCCESS)
 			return ecode;
 	} else {
@@ -174,7 +178,8 @@ static enum TDB_ERROR remove_from_list(struct tdb_context *tdb,
 	if (r->next == 0) {
 		head = tdb_read_off(tdb, b_off);
 		if (TDB_OFF_IS_ERR(head))
-			return head;
+			return TDB_OFF_TO_ERR(head);
+		head &= TDB_OFF_MASK;
 		off = head + offsetof(struct tdb_free_record, magic_and_prev);
 	} else {
 		/* off = &r->next->prev */
@@ -195,26 +200,29 @@ static enum TDB_ERROR remove_from_list(struct tdb_context *tdb,
 	return tdb_write_off(tdb, off, r->magic_and_prev);
 }
 
-/* Enqueue in this free bucket. */
+/* Enqueue in this free bucket: sets coalesce if we've added 128
+ * entries to it. */
 static enum TDB_ERROR enqueue_in_free(struct tdb_context *tdb,
 				      tdb_off_t b_off,
 				      tdb_off_t off,
-				      tdb_len_t len)
+				      tdb_len_t len,
+				      bool *coalesce)
 {
 	struct tdb_free_record new;
 	enum TDB_ERROR ecode;
-	tdb_off_t prev;
+	tdb_off_t prev, head;
 	uint64_t magic = (TDB_FREE_MAGIC << (64 - TDB_OFF_UPPER_STEAL));
 
+	head = tdb_read_off(tdb, b_off);
+	if (TDB_OFF_IS_ERR(head))
+		return TDB_OFF_TO_ERR(head);
+
 	/* We only need to set ftable_and_len; rest is set in enqueue_in_free */
-	new.ftable_and_len = ((uint64_t)tdb->ftable << (64 - TDB_OFF_UPPER_STEAL))
+	new.ftable_and_len = ((uint64_t)tdb->tdb2.ftable << (64 - TDB_OFF_UPPER_STEAL))
 		| len;
 
 	/* new->next = head. */
-	new.next = tdb_read_off(tdb, b_off);
-	if (TDB_OFF_IS_ERR(new.next)) {
-		return new.next;
-	}
+	new.next = (head & TDB_OFF_MASK);
 
 	/* First element?  Prev points to ourselves. */
 	if (!new.next) {
@@ -255,64 +263,23 @@ static enum TDB_ERROR enqueue_in_free(struct tdb_context *tdb,
 		}
 #endif
 	}
-	/* head = new */
-	ecode = tdb_write_off(tdb, b_off, off);
-	if (ecode != TDB_SUCCESS) {
-		return ecode;
-	}
 
-	return tdb_write_convert(tdb, off, &new, sizeof(new));
-}
+	/* Update enqueue count, but don't set high bit: see TDB_OFF_IS_ERR */
+	if (*coalesce)
+		head += (1ULL << (64 - TDB_OFF_UPPER_STEAL));
+	head &= ~(TDB_OFF_MASK | (1ULL << 63));
+	head |= off;
 
-/* List need not be locked. */
-enum TDB_ERROR add_free_record(struct tdb_context *tdb,
-			       tdb_off_t off, tdb_len_t len_with_header)
-{
-	tdb_off_t b_off;
-	tdb_len_t len;
-	enum TDB_ERROR ecode;
-
-	assert(len_with_header >= sizeof(struct tdb_free_record));
-
-	len = len_with_header - sizeof(struct tdb_used_record);
-
-	b_off = bucket_off(tdb->ftable_off, size_to_bucket(len));
-	ecode = tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT);
+	ecode = tdb_write_off(tdb, b_off, head);
 	if (ecode != TDB_SUCCESS) {
 		return ecode;
 	}
 
-	ecode = enqueue_in_free(tdb, b_off, off, len);
-	check_list(tdb, b_off);
-	tdb_unlock_free_bucket(tdb, b_off);
-	return ecode;
-}
-
-static size_t adjust_size(size_t keylen, size_t datalen)
-{
-	size_t size = keylen + datalen;
-
-	if (size < TDB_MIN_DATA_LEN)
-		size = TDB_MIN_DATA_LEN;
-
-	/* Round to next uint64_t boundary. */
-	return (size + (sizeof(uint64_t) - 1ULL)) & ~(sizeof(uint64_t) - 1ULL);
-}
+	/* It's time to coalesce if counter wrapped. */
+	if (*coalesce)
+		*coalesce = ((head & ~TDB_OFF_MASK) == 0);
 
-/* If we have enough left over to be useful, split that off. */
-static size_t record_leftover(size_t keylen, size_t datalen,
-			      bool want_extra, size_t total_len)
-{
-	ssize_t leftover;
-
-	if (want_extra)
-		datalen += datalen / 2;
-	leftover = total_len - adjust_size(keylen, datalen);
-
-	if (leftover < (ssize_t)sizeof(struct tdb_free_record))
-		return 0;
-
-	return leftover;
+	return tdb_write_convert(tdb, off, &new, sizeof(new));
 }
 
 static tdb_off_t ftable_offset(struct tdb_context *tdb, unsigned int ftable)
@@ -320,8 +287,8 @@ static tdb_off_t ftable_offset(struct tdb_context *tdb, unsigned int ftable)
 	tdb_off_t off;
 	unsigned int i;
 
-	if (likely(tdb->ftable == ftable))
-		return tdb->ftable_off;
+	if (likely(tdb->tdb2.ftable == ftable))
+		return tdb->tdb2.ftable_off;
 
 	off = first_ftable(tdb);
 	for (i = 0; i < ftable; i++) {
@@ -333,10 +300,12 @@ static tdb_off_t ftable_offset(struct tdb_context *tdb, unsigned int ftable)
 	return off;
 }
 
-/* Note: we unlock the current bucket if we coalesce (> 0) or fail (-ve). */
+/* Note: we unlock the current bucket if fail (-ve), or coalesce (+ve) and
+ * need to blatt the *protect record (which is set to an error). */
 static tdb_len_t coalesce(struct tdb_context *tdb,
 			  tdb_off_t off, tdb_off_t b_off,
-			  tdb_len_t data_len)
+			  tdb_len_t data_len,
+			  tdb_off_t *protect)
 {
 	tdb_off_t end;
 	struct tdb_free_record rec;
@@ -367,7 +336,7 @@ static tdb_len_t coalesce(struct tdb_context *tdb,
 		nb_off = ftable_offset(tdb, ftable);
 		if (TDB_OFF_IS_ERR(nb_off)) {
 			tdb_access_release(tdb, r);
-			ecode = nb_off;
+			ecode = TDB_OFF_TO_ERR(nb_off);
 			goto err;
 		}
 		nb_off = bucket_off(nb_off, bucket);
@@ -400,6 +369,12 @@ static tdb_len_t coalesce(struct tdb_context *tdb,
 			break;
 		}
 
+		/* Did we just mess up a record you were hoping to use? */
+		if (end == *protect) {
+			tdb->stats.alloc_coalesce_iterate_clash++;
+			*protect = TDB_ERR_TO_OFF(TDB_ERR_NOEXIST);
+		}
+
 		ecode = remove_from_list(tdb, nb_off, end, &rec);
 		check_list(tdb, nb_off);
 		if (ecode != TDB_SUCCESS) {
@@ -416,6 +391,12 @@ static tdb_len_t coalesce(struct tdb_context *tdb,
 	if (end == off + sizeof(struct tdb_used_record) + data_len)
 		return 0;
 
+	/* Before we expand, check this isn't one you wanted protected? */
+	if (off == *protect) {
+		*protect = TDB_ERR_TO_OFF(TDB_ERR_EXISTS);
+		tdb->stats.alloc_coalesce_iterate_clash++;
+	}
+
 	/* OK, expand initial record */
 	ecode = tdb_read_convert(tdb, off, &rec, sizeof(rec));
 	if (ecode != TDB_SUCCESS) {
@@ -435,33 +416,229 @@ static tdb_len_t coalesce(struct tdb_context *tdb,
 		goto err;
 	}
 
-	/* We have to drop this to avoid deadlocks, so make sure record
-	 * doesn't get coalesced by someone else! */
-	rec.ftable_and_len = (TDB_FTABLE_NONE << (64 - TDB_OFF_UPPER_STEAL))
-		| (end - off - sizeof(struct tdb_used_record));
-	ecode = tdb_write_off(tdb, off + offsetof(struct tdb_free_record,
-						  ftable_and_len),
-			      rec.ftable_and_len);
+	/* Try locking violation first.  We don't allow coalesce recursion! */
+	ecode = add_free_record(tdb, off, end - off, TDB_LOCK_NOWAIT, false);
 	if (ecode != TDB_SUCCESS) {
-		goto err;
-	}
+		/* Need to drop lock.  Can't rely on anything stable. */
+		tdb->stats.alloc_coalesce_lockfail++;
+		*protect = TDB_ERR_TO_OFF(TDB_ERR_CORRUPT);
+
+		/* We have to drop this to avoid deadlocks, so make sure record
+		 * doesn't get coalesced by someone else! */
+		rec.ftable_and_len = (TDB_FTABLE_NONE
+				      << (64 - TDB_OFF_UPPER_STEAL))
+			| (end - off - sizeof(struct tdb_used_record));
+		ecode = tdb_write_off(tdb,
+				      off + offsetof(struct tdb_free_record,
+						     ftable_and_len),
+				      rec.ftable_and_len);
+		if (ecode != TDB_SUCCESS) {
+			goto err;
+		}
 
-	tdb->stats.alloc_coalesce_succeeded++;
-	tdb_unlock_free_bucket(tdb, b_off);
+		tdb_unlock_free_bucket(tdb, b_off);
 
-	ecode = add_free_record(tdb, off, end - off);
-	if (ecode != TDB_SUCCESS) {
-		return ecode;
+		ecode = add_free_record(tdb, off, end - off, TDB_LOCK_WAIT,
+					false);
+		if (ecode != TDB_SUCCESS) {
+			return TDB_ERR_TO_OFF(ecode);
+		}
+	} else if (TDB_OFF_IS_ERR(*protect)) {
+		/* For simplicity, we always drop lock if they can't continue */
+		tdb_unlock_free_bucket(tdb, b_off);
 	}
+	tdb->stats.alloc_coalesce_succeeded++;
+
 	/* Return usable length. */
 	return end - off - sizeof(struct tdb_used_record);
 
 err:
 	/* To unify error paths, we *always* unlock bucket on error. */
 	tdb_unlock_free_bucket(tdb, b_off);
+	return TDB_ERR_TO_OFF(ecode);
+}
+
+/* List is locked: we unlock it. */
+static enum TDB_ERROR coalesce_list(struct tdb_context *tdb,
+				    tdb_off_t ftable_off,
+				    tdb_off_t b_off,
+				    unsigned int limit)
+{
+	enum TDB_ERROR ecode;
+	tdb_off_t off;
+
+	off = tdb_read_off(tdb, b_off);
+	if (TDB_OFF_IS_ERR(off)) {
+		ecode = TDB_OFF_TO_ERR(off);
+		goto unlock_err;
+	}
+	/* A little bit of paranoia: counter should be 0. */
+	off &= TDB_OFF_MASK;
+
+	while (off && limit--) {
+		struct tdb_free_record rec;
+		tdb_len_t coal;
+		tdb_off_t next;
+
+		ecode = tdb_read_convert(tdb, off, &rec, sizeof(rec));
+		if (ecode != TDB_SUCCESS)
+			goto unlock_err;
+
+		next = rec.next;
+		coal = coalesce(tdb, off, b_off, frec_len(&rec), &next);
+		if (TDB_OFF_IS_ERR(coal)) {
+			/* This has already unlocked on error. */
+			return TDB_OFF_TO_ERR(coal);
+		}
+		if (TDB_OFF_IS_ERR(next)) {
+			/* Coalescing had to unlock, so stop. */
+			return TDB_SUCCESS;
+		}
+		/* Keep going if we're doing well... */
+		limit += size_to_bucket(coal / 16 + TDB_MIN_DATA_LEN);
+		off = next;
+	}
+
+	/* Now, move those elements to the tail of the list so we get something
+	 * else next time. */
+	if (off) {
+		struct tdb_free_record oldhrec, newhrec, oldtrec, newtrec;
+		tdb_off_t oldhoff, oldtoff, newtoff;
+
+		/* The record we were up to is the new head. */
+		ecode = tdb_read_convert(tdb, off, &newhrec, sizeof(newhrec));
+		if (ecode != TDB_SUCCESS)
+			goto unlock_err;
+
+		/* Get the new tail. */
+		newtoff = frec_prev(&newhrec);
+		ecode = tdb_read_convert(tdb, newtoff, &newtrec,
+					 sizeof(newtrec));
+		if (ecode != TDB_SUCCESS)
+			goto unlock_err;
+
+		/* Get the old head. */
+		oldhoff = tdb_read_off(tdb, b_off);
+		if (TDB_OFF_IS_ERR(oldhoff)) {
+			ecode = TDB_OFF_TO_ERR(oldhoff);
+			goto unlock_err;
+		}
+
+		/* This could happen if they all coalesced away. */
+		if (oldhoff == off)
+			goto out;
+
+		ecode = tdb_read_convert(tdb, oldhoff, &oldhrec,
+					 sizeof(oldhrec));
+		if (ecode != TDB_SUCCESS)
+			goto unlock_err;
+
+		/* Get the old tail. */
+		oldtoff = frec_prev(&oldhrec);
+		ecode = tdb_read_convert(tdb, oldtoff, &oldtrec,
+					 sizeof(oldtrec));
+		if (ecode != TDB_SUCCESS)
+			goto unlock_err;
+
+		/* Old tail's next points to old head. */
+		oldtrec.next = oldhoff;
+
+		/* Old head's prev points to old tail. */
+		oldhrec.magic_and_prev
+			= (TDB_FREE_MAGIC << (64 - TDB_OFF_UPPER_STEAL))
+			| oldtoff;
+
+		/* New tail's next is 0. */
+		newtrec.next = 0;
+
+		/* Write out the modified versions. */
+		ecode = tdb_write_convert(tdb, oldtoff, &oldtrec,
+					  sizeof(oldtrec));
+		if (ecode != TDB_SUCCESS)
+			goto unlock_err;
+
+		ecode = tdb_write_convert(tdb, oldhoff, &oldhrec,
+					  sizeof(oldhrec));
+		if (ecode != TDB_SUCCESS)
+			goto unlock_err;
+
+		ecode = tdb_write_convert(tdb, newtoff, &newtrec,
+					  sizeof(newtrec));
+		if (ecode != TDB_SUCCESS)
+			goto unlock_err;
+		
+		/* And finally link in new head. */
+		ecode = tdb_write_off(tdb, b_off, off);
+		if (ecode != TDB_SUCCESS)
+			goto unlock_err;
+	}
+out:
+	tdb_unlock_free_bucket(tdb, b_off);
+	return TDB_SUCCESS;
+
+unlock_err:
+	tdb_unlock_free_bucket(tdb, b_off);
+	return ecode;
+}
+
+/* List must not be locked if coalesce_ok is set. */
+enum TDB_ERROR add_free_record(struct tdb_context *tdb,
+			       tdb_off_t off, tdb_len_t len_with_header,
+			       enum tdb_lock_flags waitflag,
+			       bool coalesce)
+{
+	tdb_off_t b_off;
+	tdb_len_t len;
+	enum TDB_ERROR ecode;
+
+	assert(len_with_header >= sizeof(struct tdb_free_record));
+
+	len = len_with_header - sizeof(struct tdb_used_record);
+
+	b_off = bucket_off(tdb->tdb2.ftable_off, size_to_bucket(len));
+	ecode = tdb_lock_free_bucket(tdb, b_off, waitflag);
+	if (ecode != TDB_SUCCESS) {
+		return ecode;
+	}
+
+	ecode = enqueue_in_free(tdb, b_off, off, len, &coalesce);
+	check_list(tdb, b_off);
+
+	/* Coalescing unlocks free list. */
+	if (!ecode && coalesce)
+		ecode = coalesce_list(tdb, tdb->tdb2.ftable_off, b_off, 2);
+	else
+		tdb_unlock_free_bucket(tdb, b_off);
 	return ecode;
 }
 
+static size_t adjust_size(size_t keylen, size_t datalen)
+{
+	size_t size = keylen + datalen;
+
+	if (size < TDB_MIN_DATA_LEN)
+		size = TDB_MIN_DATA_LEN;
+
+	/* Round to next uint64_t boundary. */
+	return (size + (sizeof(uint64_t) - 1ULL)) & ~(sizeof(uint64_t) - 1ULL);
+}
+
+/* If we have enough left over to be useful, split that off. */
+static size_t record_leftover(size_t keylen, size_t datalen,
+			      bool want_extra, size_t total_len)
+{
+	ssize_t leftover;
+
+	if (want_extra)
+		datalen += datalen / 2;
+	leftover = total_len - adjust_size(keylen, datalen);
+
+	if (leftover < (ssize_t)sizeof(struct tdb_free_record))
+		return 0;
+
+	return leftover;
+}
+
 /* We need size bytes to put our key and data in. */
 static tdb_off_t lock_and_alloc(struct tdb_context *tdb,
 				tdb_off_t ftable_off,
@@ -478,14 +655,13 @@ static tdb_off_t lock_and_alloc(struct tdb_context *tdb,
 	enum TDB_ERROR ecode;
 
 	tdb->stats.allocs++;
-again:
 	b_off = bucket_off(ftable_off, bucket);
 
 	/* FIXME: Try non-blocking wait first, to measure contention. */
 	/* Lock this bucket. */
 	ecode = tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT);
 	if (ecode != TDB_SUCCESS) {
-		return ecode;
+		return TDB_ERR_TO_OFF(ecode);
 	}
 
 	best.ftable_and_len = -1ULL;
@@ -501,13 +677,14 @@ again:
 	 * as we go. */
 	off = tdb_read_off(tdb, b_off);
 	if (TDB_OFF_IS_ERR(off)) {
-		ecode = off;
+		ecode = TDB_OFF_TO_ERR(off);
 		goto unlock_err;
 	}
+	off &= TDB_OFF_MASK;
 
 	while (off) {
 		const struct tdb_free_record *r;
-		tdb_len_t len, coal;
+		tdb_len_t len;
 		tdb_off_t next;
 
 		r = tdb_access_read(tdb, off, sizeof(*r), true);
@@ -541,17 +718,6 @@ again:
 		next = r->next;
 		len = frec_len(r);
 		tdb_access_release(tdb, r);
-
-		/* Since we're going slow anyway, try coalescing here. */
-		coal = coalesce(tdb, off, b_off, len);
-		if (TDB_OFF_IS_ERR(coal)) {
-			/* This has already unlocked on error. */
-			return coal;
-		}
-		if (coal > 0) {
-			/* This has unlocked list, restart. */
-			goto again;
-		}
 		off = next;
 	}
 
@@ -586,7 +752,7 @@ again:
 
 		/* For futureproofing, we put a 0 in any unused space. */
 		if (rec_extra_padding(&rec)) {
-			ecode = tdb->methods->twrite(tdb, best_off + sizeof(rec)
+			ecode = tdb->tdb2.io->twrite(tdb, best_off + sizeof(rec)
 						     + keylen + datalen, "", 1);
 			if (ecode != TDB_SUCCESS) {
 				goto unlock_err;
@@ -600,9 +766,9 @@ again:
 			ecode = add_free_record(tdb,
 						best_off + sizeof(rec)
 						+ frec_len(&best) - leftover,
-						leftover);
+						leftover, TDB_LOCK_WAIT, false);
 			if (ecode != TDB_SUCCESS) {
-				best_off = ecode;
+				best_off = TDB_ERR_TO_OFF(ecode);
 			}
 		}
 		tdb_unlock_free_bucket(tdb, b_off);
@@ -615,7 +781,7 @@ again:
 
 unlock_err:
 	tdb_unlock_free_bucket(tdb, b_off);
-	return ecode;
+	return TDB_ERR_TO_OFF(ecode);
 }
 
 /* Get a free block from current free list, or 0 if none, -ve on error. */
@@ -634,9 +800,9 @@ static tdb_off_t get_free(struct tdb_context *tdb,
 	else
 		start_b = size_to_bucket(adjust_size(keylen, datalen));
 
-	ftable_off = tdb->ftable_off;
-	ftable = tdb->ftable;
-	while (!wrapped || ftable_off != tdb->ftable_off) {
+	ftable_off = tdb->tdb2.ftable_off;
+	ftable = tdb->tdb2.ftable;
+	while (!wrapped || ftable_off != tdb->tdb2.ftable_off) {
 		/* Start at exact size bucket, and search up... */
 		for (b = find_free_head(tdb, ftable_off, start_b);
 		     b < TDB_FREE_BUCKETS;
@@ -653,8 +819,8 @@ static tdb_off_t get_free(struct tdb_context *tdb,
 				if (b == TDB_FREE_BUCKETS - 1)
 					tdb->stats.alloc_bucket_max++;
 				/* Worked?  Stay using this list. */
-				tdb->ftable_off = ftable_off;
-				tdb->ftable = ftable;
+				tdb->tdb2.ftable_off = ftable_off;
+				tdb->tdb2.ftable = ftable;
 				return off;
 			}
 			/* Didn't work.  Try next bucket. */
@@ -710,10 +876,38 @@ enum TDB_ERROR set_header(struct tdb_context *tdb,
 	return TDB_SUCCESS;
 }
 
+/* You need 'size', this tells you how much you should expand by. */
+tdb_off_t tdb_expand_adjust(tdb_off_t map_size, tdb_off_t size)
+{
+	tdb_off_t new_size, top_size;
+
+	/* limit size in order to avoid using up huge amounts of memory for
+	 * in memory tdbs if an oddball huge record creeps in */
+	if (size > 100 * 1024) {
+		top_size = map_size + size * 2;
+	} else {
+		top_size = map_size + size * 100;
+	}
+
+	/* always make room for at least top_size more records, and at
+	   least 25% more space. if the DB is smaller than 100MiB,
+	   otherwise grow it by 10% only. */
+	if (map_size > 100 * 1024 * 1024) {
+		new_size = map_size * 1.10;
+	} else {
+		new_size = map_size * 1.25;
+	}
+
+	/* Round the database up to a multiple of the page size */
+	if (new_size < top_size)
+		new_size = top_size;
+	return new_size - map_size;
+}
+
 /* Expand the database. */
 static enum TDB_ERROR tdb_expand(struct tdb_context *tdb, tdb_len_t size)
 {
-	uint64_t old_size, rec_size, map_size;
+	uint64_t old_size;
 	tdb_len_t wanted;
 	enum TDB_ERROR ecode;
 
@@ -732,39 +926,18 @@ static enum TDB_ERROR tdb_expand(struct tdb_context *tdb, tdb_len_t size)
 
 	/* Someone else may have expanded the file, so retry. */
 	old_size = tdb->file->map_size;
-	tdb->methods->oob(tdb, tdb->file->map_size + 1, true);
+	tdb->tdb2.io->oob(tdb, tdb->file->map_size, 1, true);
 	if (tdb->file->map_size != old_size) {
 		tdb_unlock_expand(tdb, F_WRLCK);
 		return TDB_SUCCESS;
 	}
 
-	/* limit size in order to avoid using up huge amounts of memory for
-	 * in memory tdbs if an oddball huge record creeps in */
-	if (size > 100 * 1024) {
-		rec_size = size * 2;
-	} else {
-		rec_size = size * 100;
-	}
-
-	/* always make room for at least rec_size more records, and at
-	   least 25% more space. if the DB is smaller than 100MiB,
-	   otherwise grow it by 10% only. */
-	if (old_size > 100 * 1024 * 1024) {
-		map_size = old_size / 10;
-	} else {
-		map_size = old_size / 4;
-	}
-
-	if (map_size > rec_size) {
-		wanted = map_size;
-	} else {
-		wanted = rec_size;
-	}
-
+	/* Overallocate. */
+	wanted = tdb_expand_adjust(old_size, size);
 	/* We need room for the record header too. */
 	wanted = adjust_size(0, sizeof(struct tdb_used_record) + wanted);
 
-	ecode = tdb->methods->expand_file(tdb, wanted);
+	ecode = tdb->tdb2.io->expand_file(tdb, wanted);
 	if (ecode != TDB_SUCCESS) {
 		tdb_unlock_expand(tdb, F_WRLCK);
 		return ecode;
@@ -774,7 +947,7 @@ static enum TDB_ERROR tdb_expand(struct tdb_context *tdb, tdb_len_t size)
 	tdb_unlock_expand(tdb, F_WRLCK);
 
 	tdb->stats.expands++;
-	return add_free_record(tdb, old_size, wanted);
+	return add_free_record(tdb, old_size, wanted, TDB_LOCK_WAIT, true);
 }
 
 /* This won't fail: it will expand the database if it has to. */
@@ -784,7 +957,7 @@ tdb_off_t alloc(struct tdb_context *tdb, size_t keylen, size_t datalen,
 	tdb_off_t off;
 
 	/* We can't hold pointers during this: we could unmap! */
-	assert(!tdb->direct_access);
+	assert(!tdb->tdb2.direct_access);
 
 	for (;;) {
 		enum TDB_ERROR ecode;
@@ -794,7 +967,7 @@ tdb_off_t alloc(struct tdb_context *tdb, size_t keylen, size_t datalen,
 
 		ecode = tdb_expand(tdb, adjust_size(keylen, datalen));
 		if (ecode != TDB_SUCCESS) {
-			return ecode;
+			return TDB_ERR_TO_OFF(ecode);
 		}
 	}