X-Git-Url: http://git.ozlabs.org/?p=ccan;a=blobdiff_plain;f=ccan%2Ftdb2%2Ffree.c;h=cd9a332abe425407cf27b5e35a49db912e3a92b3;hp=4930400d50f7cabd560d5b31461b6485bea0ab8a;hb=5c4a21ab03a428373e7659b9606facf85dcbe17b;hpb=0aa58f6e6017138a78ce4e86fb758732719acaca

diff --git a/ccan/tdb2/free.c b/ccan/tdb2/free.c
index 4930400d..cd9a332a 100644
--- a/ccan/tdb2/free.c
+++ b/ccan/tdb2/free.c
@@ -59,7 +59,7 @@ tdb_off_t next_ftable(struct tdb_context *tdb, tdb_off_t ftable)
 	return tdb_read_off(tdb, ftable + offsetof(struct tdb_freetable,next));
 }
 
-int tdb_ftable_init(struct tdb_context *tdb)
+enum TDB_ERROR tdb_ftable_init(struct tdb_context *tdb)
 {
 	/* Use reservoir sampling algorithm to select a free list at random. */
 	unsigned int rnd, max = 0, count = 0;
@@ -69,8 +69,9 @@ int tdb_ftable_init(struct tdb_context *tdb)
 	tdb->ftable = 0;
 
 	while (off) {
-		if (off == TDB_OFF_ERR)
-			return -1;
+		if (TDB_OFF_IS_ERR(off)) {
+			return off;
+		}
 
 		rnd = random();
 		if (rnd >= max) {
@@ -82,7 +83,7 @@ int tdb_ftable_init(struct tdb_context *tdb)
 		off = next_ftable(tdb, off);
 		count++;
 	}
-	return 0;
+	return TDB_SUCCESS;
 }
 
 /* Offset of a given bucket. */
@@ -92,7 +93,7 @@ tdb_off_t bucket_off(tdb_off_t ftable_off, unsigned bucket)
 		+ bucket * sizeof(tdb_off_t);
 }
 
-/* Returns free_buckets + 1, or list number to search. */
+/* Returns free_buckets + 1, or list number to search, or -ve error. */
 static tdb_off_t find_free_head(struct tdb_context *tdb,
 				tdb_off_t ftable_off,
 				tdb_off_t bucket)
@@ -102,120 +103,190 @@ static tdb_off_t find_free_head(struct tdb_context *tdb,
 				    bucket, TDB_FREE_BUCKETS);
 }
 
-/* Remove from free bucket. */
-static int remove_from_list(struct tdb_context *tdb,
-			    tdb_off_t b_off, tdb_off_t r_off,
-			    const struct tdb_free_record *r)
+static void check_list(struct tdb_context *tdb, tdb_off_t b_off)
 {
-	tdb_off_t off;
-
-	/* Front of list? */
-	if (frec_prev(r) == 0) {
-		off = b_off;
-	} else {
-		off = frec_prev(r) + offsetof(struct tdb_free_record, next);
+#ifdef CCAN_TDB2_DEBUG
+	tdb_off_t off, prev = 0, first;
+	struct tdb_free_record r;
+
+	first = off = tdb_read_off(tdb, b_off);
+	while (off != 0) {
+		tdb_read_convert(tdb, off, &r, sizeof(r));
+		if (frec_magic(&r) != TDB_FREE_MAGIC)
+			abort();
+		if (prev && frec_prev(&r) != prev)
+			abort();
+		prev = off;
+		off = r.next;
 	}
 
-#ifdef CCAN_TDB2_DEBUG
-	if (tdb_read_off(tdb, off) != r_off) {
-		tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
-			 "remove_from_list: %llu bad prev in list %llu",
-			 (long long)r_off, (long long)b_off);
-		return -1;
+	if (first) {
+		tdb_read_convert(tdb, first, &r, sizeof(r));
+		if (frec_prev(&r) != prev)
+			abort();
 	}
 #endif
+}
 
-	/* r->prev->next = r->next */
-	if (tdb_write_off(tdb, off, r->next)) {
-		return -1;
-	}
+/* Remove from free bucket. */
+static enum TDB_ERROR remove_from_list(struct tdb_context *tdb,
+				       tdb_off_t b_off, tdb_off_t r_off,
+				       const struct tdb_free_record *r)
+{
+	tdb_off_t off, prev_next, head;
+	enum TDB_ERROR ecode;
+
+	/* Is this only element in list?  Zero out bucket, and we're done. */
+	if (frec_prev(r) == r_off)
+		return tdb_write_off(tdb, b_off, 0);
+
+	/* off = &r->prev->next */
+	off = frec_prev(r) + offsetof(struct tdb_free_record, next);
 
-	if (r->next != 0) {
-		off = r->next + offsetof(struct tdb_free_record,magic_and_prev);
-		/* r->next->prev = r->prev */
+	/* Get prev->next */
+	prev_next = tdb_read_off(tdb, off);
+	if (TDB_OFF_IS_ERR(prev_next))
+		return 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, off) & TDB_OFF_MASK != r_off) {
-			tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
-				   "remove_from_list: %llu bad list %llu",
-				   (long long)r_off, (long long)b_off);
-			return -1;
+		if (tdb_read_off(tdb, b_off) != 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)b_off);
 		}
 #endif
+		ecode = tdb_write_off(tdb, b_off, r->next);
+		if (ecode != TDB_SUCCESS)
+			return ecode;
+	} else {
+		/* r->prev->next = r->next */
+		ecode = tdb_write_off(tdb, off, r->next);
+		if (ecode != TDB_SUCCESS)
+			return ecode;
+	}
 
-		if (tdb_write_off(tdb, off, r->magic_and_prev)) {
-			return -1;
-		}
+	/* If we were the tail, off = &head->prev. */
+	if (r->next == 0) {
+		head = tdb_read_off(tdb, b_off);
+		if (TDB_OFF_IS_ERR(head))
+			return head;
+		off = head + offsetof(struct tdb_free_record, magic_and_prev);
+	} else {
+		/* off = &r->next->prev */
+		off = r->next + offsetof(struct tdb_free_record,
+					 magic_and_prev);
 	}
-	return 0;
+
+#ifdef CCAN_TDB2_DEBUG
+	/* *off == r */
+	if ((tdb_read_off(tdb, off) & TDB_OFF_MASK) != r_off) {
+		return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
+				  "remove_from_list:"
+				  " %llu bad prev in list %llu",
+				  (long long)r_off, (long long)b_off);
+	}
+#endif
+	/* r->next->prev = r->prev */
+	return tdb_write_off(tdb, off, r->magic_and_prev);
 }
 
 /* Enqueue in this free bucket. */
-static int enqueue_in_free(struct tdb_context *tdb,
-			   tdb_off_t b_off,
-			   tdb_off_t off,
-			   tdb_len_t len)
+static enum TDB_ERROR enqueue_in_free(struct tdb_context *tdb,
+				      tdb_off_t b_off,
+				      tdb_off_t off,
+				      tdb_len_t len)
 {
 	struct tdb_free_record new;
+	enum TDB_ERROR ecode;
+	tdb_off_t prev;
 	uint64_t magic = (TDB_FREE_MAGIC << (64 - TDB_OFF_UPPER_STEAL));
 
 	/* 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))
 		| len;
-	/* prev = 0. */
-	new.magic_and_prev = magic;
 
 	/* new->next = head. */
 	new.next = tdb_read_off(tdb, b_off);
-	if (new.next == TDB_OFF_ERR)
-		return -1;
+	if (TDB_OFF_IS_ERR(new.next)) {
+		return new.next;
+	}
+
+	/* First element?  Prev points to ourselves. */
+	if (!new.next) {
+		new.magic_and_prev = (magic | off);
+	} else {
+		/* new->prev = next->prev */
+		prev = tdb_read_off(tdb,
+				    new.next + offsetof(struct tdb_free_record,
+							magic_and_prev));
+		new.magic_and_prev = prev;
+		if (frec_magic(&new) != TDB_FREE_MAGIC) {
+			return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
+					  "enqueue_in_free: %llu bad head"
+					  " prev %llu",
+					  (long long)new.next,
+					  (long long)prev);
+		}
+		/* next->prev = new. */
+		ecode = tdb_write_off(tdb, new.next
+				      + offsetof(struct tdb_free_record,
+						 magic_and_prev),
+				      off | magic);
+		if (ecode != TDB_SUCCESS) {
+			return ecode;
+		}
 
-	if (new.next) {
 #ifdef CCAN_TDB2_DEBUG
-		if (tdb_read_off(tdb,
-				 new.next + offsetof(struct tdb_free_record,
-						     magic_and_prev))
-		    != magic) {
-			tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
-				   "enqueue_in_free: %llu bad head"
-				   " prev %llu",
-				   (long long)new.next, (long long)b_off);
-			return -1;
+		prev = tdb_read_off(tdb, frec_prev(&new)
+				    + offsetof(struct tdb_free_record, next));
+		if (prev != 0) {
+			return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
+					  "enqueue_in_free:"
+					  " %llu bad tail next ptr %llu",
+					  (long long)frec_prev(&new)
+					  + offsetof(struct tdb_free_record,
+						     next),
+					  (long long)prev);
 		}
 #endif
-		/* next->prev = new. */
-		if (tdb_write_off(tdb, new.next
-				  + offsetof(struct tdb_free_record,
-					     magic_and_prev),
-				  off | magic) != 0)
-			return -1;
 	}
 	/* head = new */
-	if (tdb_write_off(tdb, b_off, off) != 0)
-		return -1;
+	ecode = tdb_write_off(tdb, b_off, off);
+	if (ecode != TDB_SUCCESS) {
+		return ecode;
+	}
 
 	return tdb_write_convert(tdb, off, &new, sizeof(new));
 }
 
 /* List need not be locked. */
-int add_free_record(struct tdb_context *tdb,
-		    tdb_off_t off, tdb_len_t len_with_header)
+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)
 {
 	tdb_off_t b_off;
 	tdb_len_t len;
-	int ret;
+	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));
-	if (tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT) != 0)
-		return -1;
+	ecode = tdb_lock_free_bucket(tdb, b_off, waitflag);
+	if (ecode != TDB_SUCCESS) {
+		return ecode;
+	}
 
-	ret = enqueue_in_free(tdb, b_off, off, len);
+	ecode = enqueue_in_free(tdb, b_off, off, len);
+	check_list(tdb, b_off);
 	tdb_unlock_free_bucket(tdb, b_off);
-	return ret;
+	return ecode;
 }
 
 static size_t adjust_size(size_t keylen, size_t datalen)
@@ -254,29 +325,40 @@ static tdb_off_t ftable_offset(struct tdb_context *tdb, unsigned int ftable)
 		return tdb->ftable_off;
 
 	off = first_ftable(tdb);
-	for (i = 0; i < ftable; i++)
+	for (i = 0; i < ftable; i++) {
+		if (TDB_OFF_IS_ERR(off)) {
+			break;
+		}
 		off = next_ftable(tdb, off);
+	}
 	return off;
 }
 
-/* Note: we unlock the current bucket if we coalesce or fail. */
-static int coalesce(struct tdb_context *tdb,
-		    tdb_off_t off, tdb_off_t b_off, tdb_len_t data_len)
+/* Note: we unlock the current bucket if fail (-ve), or coalesce (-ve) and
+ * need to blatt either of the *protect records (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_off_t *protect1,
+			  tdb_off_t *protect2)
 {
 	tdb_off_t end;
 	struct tdb_free_record rec;
+	enum TDB_ERROR ecode;
 
-	add_stat(tdb, alloc_coalesce_tried, 1);
+	tdb->stats.alloc_coalesce_tried++;
 	end = off + sizeof(struct tdb_used_record) + data_len;
 
-	while (end < tdb->map_size) {
+	while (end < tdb->file->map_size) {
 		const struct tdb_free_record *r;
 		tdb_off_t nb_off;
 		unsigned ftable, bucket;
 
 		r = tdb_access_read(tdb, end, sizeof(*r), true);
-		if (!r)
+		if (TDB_PTR_IS_ERR(r)) {
+			ecode = TDB_PTR_ERR(r);
 			goto err;
+		}
 
 		if (frec_magic(r) != TDB_FREE_MAGIC
 		    || frec_ftable(r) == TDB_FTABLE_NONE) {
@@ -286,82 +368,123 @@ static int coalesce(struct tdb_context *tdb,
 
 		ftable = frec_ftable(r);
 		bucket = size_to_bucket(frec_len(r));
-		nb_off = bucket_off(ftable_offset(tdb, ftable), bucket);
+		nb_off = ftable_offset(tdb, ftable);
+		if (TDB_OFF_IS_ERR(nb_off)) {
+			tdb_access_release(tdb, r);
+			ecode = nb_off;
+			goto err;
+		}
+		nb_off = bucket_off(nb_off, bucket);
 		tdb_access_release(tdb, r);
 
 		/* We may be violating lock order here, so best effort. */
-		if (tdb_lock_free_bucket(tdb, nb_off, TDB_LOCK_NOWAIT) == -1) {
-			add_stat(tdb, alloc_coalesce_lockfail, 1);
+		if (tdb_lock_free_bucket(tdb, nb_off, TDB_LOCK_NOWAIT)
+		    != TDB_SUCCESS) {
+			tdb->stats.alloc_coalesce_lockfail++;
 			break;
 		}
 
 		/* Now we have lock, re-check. */
-		if (tdb_read_convert(tdb, end, &rec, sizeof(rec))) {
+		ecode = tdb_read_convert(tdb, end, &rec, sizeof(rec));
+		if (ecode != TDB_SUCCESS) {
 			tdb_unlock_free_bucket(tdb, nb_off);
 			goto err;
 		}
 
 		if (unlikely(frec_magic(&rec) != TDB_FREE_MAGIC)) {
-			add_stat(tdb, alloc_coalesce_race, 1);
+			tdb->stats.alloc_coalesce_race++;
 			tdb_unlock_free_bucket(tdb, nb_off);
 			break;
 		}
 
 		if (unlikely(frec_ftable(&rec) != ftable)
 		    || unlikely(size_to_bucket(frec_len(&rec)) != bucket)) {
-			add_stat(tdb, alloc_coalesce_race, 1);
+			tdb->stats.alloc_coalesce_race++;
 			tdb_unlock_free_bucket(tdb, nb_off);
 			break;
 		}
 
-		if (remove_from_list(tdb, nb_off, end, &rec) == -1) {
+		/* Did we just mess up a record you were hoping to use? */
+		if (end == *protect1 || end == *protect2)
+			*protect1 = TDB_ERR_NOEXIST;
+
+		ecode = remove_from_list(tdb, nb_off, end, &rec);
+		check_list(tdb, nb_off);
+		if (ecode != TDB_SUCCESS) {
 			tdb_unlock_free_bucket(tdb, nb_off);
 			goto err;
 		}
 
 		end += sizeof(struct tdb_used_record) + frec_len(&rec);
 		tdb_unlock_free_bucket(tdb, nb_off);
-		add_stat(tdb, alloc_coalesce_num_merged, 1);
+		tdb->stats.alloc_coalesce_num_merged++;
 	}
 
 	/* Didn't find any adjacent free? */
 	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 == *protect1 || off == *protect2)
+		*protect1 = TDB_ERR_EXISTS;
+
 	/* OK, expand initial record */
-	if (tdb_read_convert(tdb, off, &rec, sizeof(rec)))
+	ecode = tdb_read_convert(tdb, off, &rec, sizeof(rec));
+	if (ecode != TDB_SUCCESS) {
 		goto err;
+	}
 
 	if (frec_len(&rec) != data_len) {
-		tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
-			   "coalesce: expected data len %zu not %zu",
-			   (size_t)data_len, (size_t)frec_len(&rec));
+		ecode = tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
+				   "coalesce: expected data len %zu not %zu",
+				   (size_t)data_len, (size_t)frec_len(&rec));
 		goto err;
 	}
 
-	if (remove_from_list(tdb, b_off, off, &rec) == -1)
+	ecode = remove_from_list(tdb, b_off, off, &rec);
+	check_list(tdb, b_off);
+	if (ecode != TDB_SUCCESS) {
 		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));
-	if (tdb_write_off(tdb, off + offsetof(struct tdb_free_record,
-					      ftable_and_len),
-			  rec.ftable_and_len) != 0)
-		goto err;
+	/* Try locking violation first... */
+	ecode = add_free_record(tdb, off, end - off, TDB_LOCK_NOWAIT);
+	if (ecode != TDB_SUCCESS) {
+		/* Need to drop lock.  Can't rely on anything stable. */
+		*protect1 = 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;
+		}
 
-	add_stat(tdb, alloc_coalesce_succeeded, 1);
-	tdb_unlock_free_bucket(tdb, b_off);
+		tdb->stats.alloc_coalesce_succeeded++;
+		tdb_unlock_free_bucket(tdb, b_off);
 
-	if (add_free_record(tdb, off, end - off) == -1)
-		return -1;
-	return 1;
+		ecode = add_free_record(tdb, off, end - off, TDB_LOCK_WAIT);
+		if (ecode != TDB_SUCCESS) {
+			return ecode;
+		}
+	} else if (TDB_OFF_IS_ERR(*protect1)) {
+		/* For simplicity, we always drop lock if they can't continue */
+		tdb_unlock_free_bucket(tdb, b_off);
+	}
+
+	/* 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 -1;
+	return ecode;
 }
 
 /* We need size bytes to put our key and data in. */
@@ -376,16 +499,19 @@ static tdb_off_t lock_and_alloc(struct tdb_context *tdb,
 	tdb_off_t off, b_off,best_off;
 	struct tdb_free_record best = { 0 };
 	double multiplier;
+	bool coalesce_after_best = false; /* Damn GCC warning! */
 	size_t size = adjust_size(keylen, datalen);
+	enum TDB_ERROR ecode;
 
-	add_stat(tdb, allocs, 1);
+	tdb->stats.allocs++;
 again:
 	b_off = bucket_off(ftable_off, bucket);
 
 	/* FIXME: Try non-blocking wait first, to measure contention. */
 	/* Lock this bucket. */
-	if (tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT) == -1) {
-		return TDB_OFF_ERR;
+	ecode = tdb_lock_free_bucket(tdb, b_off, TDB_LOCK_WAIT);
+	if (ecode != TDB_SUCCESS) {
+		return ecode;
 	}
 
 	best.ftable_and_len = -1ULL;
@@ -400,32 +526,39 @@ again:
 	/* Walk the list to see if any are large enough, getting less fussy
 	 * as we go. */
 	off = tdb_read_off(tdb, b_off);
-	if (unlikely(off == TDB_OFF_ERR))
+	if (TDB_OFF_IS_ERR(off)) {
+		ecode = off;
 		goto unlock_err;
+	}
 
 	while (off) {
 		const struct tdb_free_record *r;
-		tdb_len_t len;
+		tdb_len_t len, coal;
 		tdb_off_t next;
 
 		r = tdb_access_read(tdb, off, sizeof(*r), true);
-		if (!r)
+		if (TDB_PTR_IS_ERR(r)) {
+			ecode = TDB_PTR_ERR(r);
 			goto unlock_err;
+		}
 
 		if (frec_magic(r) != TDB_FREE_MAGIC) {
+			ecode = tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
+					   "lock_and_alloc:"
+					   " %llu non-free 0x%llx",
+					   (long long)off,
+					   (long long)r->magic_and_prev);
 			tdb_access_release(tdb, r);
-			tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_DEBUG_FATAL,
-				 "lock_and_alloc: %llu non-free 0x%llx",
-				 (long long)off, (long long)r->magic_and_prev);
 			goto unlock_err;
 		}
 
 		if (frec_len(r) >= size && frec_len(r) < frec_len(&best)) {
 			best_off = off;
 			best = *r;
+			coalesce_after_best = false;
 		}
 
-		if (frec_len(&best) < size * multiplier && best_off) {
+		if (frec_len(&best) <= size * multiplier && best_off) {
 			tdb_access_release(tdb, r);
 			break;
 		}
@@ -437,14 +570,17 @@ again:
 		tdb_access_release(tdb, r);
 
 		/* Since we're going slow anyway, try coalescing here. */
-		switch (coalesce(tdb, off, b_off, len)) {
-		case -1:
+		coal = coalesce(tdb, off, b_off, len, &best_off, &next);
+		if (TDB_OFF_IS_ERR(coal)) {
 			/* This has already unlocked on error. */
-			return -1;
-		case 1:
+			return coal;
+		}
+		if (TDB_OFF_IS_ERR(best_off)) {
 			/* This has unlocked list, restart. */
 			goto again;
 		}
+		if (coal > 0)
+			coalesce_after_best = true;
 		off = next;
 	}
 
@@ -453,9 +589,20 @@ again:
 		struct tdb_used_record rec;
 		size_t leftover;
 
+		/* If we coalesced, we might have change prev/next ptrs. */
+		if (coalesce_after_best) {
+			ecode = tdb_read_convert(tdb, best_off, &best,
+						 sizeof(best));
+			if (ecode != TDB_SUCCESS)
+				goto unlock_err;
+		}
+
 		/* We're happy with this size: take it. */
-		if (remove_from_list(tdb, b_off, best_off, &best) != 0)
+		ecode = remove_from_list(tdb, b_off, best_off, &best);
+		check_list(tdb, b_off);
+		if (ecode != TDB_SUCCESS) {
 			goto unlock_err;
+		}
 
 		leftover = record_leftover(keylen, datalen, want_extra,
 					   frec_len(&best));
@@ -463,22 +610,37 @@ again:
 		assert(keylen + datalen + leftover <= frec_len(&best));
 		/* We need to mark non-free before we drop lock, otherwise
 		 * coalesce() could try to merge it! */
-		if (set_header(tdb, &rec, magic, keylen, datalen,
-			       frec_len(&best) - leftover, hashlow) != 0)
+		ecode = set_header(tdb, &rec, magic, keylen, datalen,
+				   frec_len(&best) - leftover, hashlow);
+		if (ecode != TDB_SUCCESS) {
 			goto unlock_err;
+		}
 
-		if (tdb_write_convert(tdb, best_off, &rec, sizeof(rec)) != 0)
+		ecode = tdb_write_convert(tdb, best_off, &rec, sizeof(rec));
+		if (ecode != TDB_SUCCESS) {
 			goto unlock_err;
+		}
+
+		/* For futureproofing, we put a 0 in any unused space. */
+		if (rec_extra_padding(&rec)) {
+			ecode = tdb->methods->twrite(tdb, best_off + sizeof(rec)
+						     + keylen + datalen, "", 1);
+			if (ecode != TDB_SUCCESS) {
+				goto unlock_err;
+			}
+		}
 
 		/* Bucket of leftover will be <= current bucket, so nested
 		 * locking is allowed. */
 		if (leftover) {
-			add_stat(tdb, alloc_leftover, 1);
-			if (add_free_record(tdb,
-					    best_off + sizeof(rec)
-					    + frec_len(&best) - leftover,
-					    leftover))
-				best_off = TDB_OFF_ERR;
+			tdb->stats.alloc_leftover++;
+			ecode = add_free_record(tdb,
+						best_off + sizeof(rec)
+						+ frec_len(&best) - leftover,
+						leftover, TDB_LOCK_WAIT);
+			if (ecode != TDB_SUCCESS) {
+				best_off = ecode;
+			}
 		}
 		tdb_unlock_free_bucket(tdb, b_off);
 
@@ -490,16 +652,16 @@ again:
 
 unlock_err:
 	tdb_unlock_free_bucket(tdb, b_off);
-	return TDB_OFF_ERR;
+	return ecode;
 }
 
-/* Get a free block from current free list, or 0 if none. */
+/* Get a free block from current free list, or 0 if none, -ve on error. */
 static tdb_off_t get_free(struct tdb_context *tdb,
 			  size_t keylen, size_t datalen, bool want_extra,
 			  unsigned magic, unsigned hashlow)
 {
 	tdb_off_t off, ftable_off;
-	unsigned start_b, b, ftable;
+	tdb_off_t start_b, b, ftable;
 	bool wrapped = false;
 
 	/* If they are growing, add 50% to get to higher bucket. */
@@ -520,13 +682,13 @@ static tdb_off_t get_free(struct tdb_context *tdb,
 			off = lock_and_alloc(tdb, ftable_off,
 					     b, keylen, datalen, want_extra,
 					     magic, hashlow);
-			if (off == TDB_OFF_ERR)
-				return TDB_OFF_ERR;
+			if (TDB_OFF_IS_ERR(off))
+				return off;
 			if (off != 0) {
 				if (b == start_b)
-					add_stat(tdb, alloc_bucket_exact, 1);
+					tdb->stats.alloc_bucket_exact++;
 				if (b == TDB_FREE_BUCKETS - 1)
-					add_stat(tdb, alloc_bucket_max, 1);
+					tdb->stats.alloc_bucket_max++;
 				/* Worked?  Stay using this list. */
 				tdb->ftable_off = ftable_off;
 				tdb->ftable = ftable;
@@ -535,13 +697,23 @@ static tdb_off_t get_free(struct tdb_context *tdb,
 			/* Didn't work.  Try next bucket. */
 		}
 
+		if (TDB_OFF_IS_ERR(b)) {
+			return b;
+		}
+
 		/* Hmm, try next table. */
 		ftable_off = next_ftable(tdb, ftable_off);
+		if (TDB_OFF_IS_ERR(ftable_off)) {
+			return ftable_off;
+		}
 		ftable++;
 
 		if (ftable_off == 0) {
 			wrapped = true;
 			ftable_off = first_ftable(tdb);
+			if (TDB_OFF_IS_ERR(ftable_off)) {
+				return ftable_off;
+			}
 			ftable = 0;
 		}
 	}
@@ -549,10 +721,10 @@ static tdb_off_t get_free(struct tdb_context *tdb,
 	return 0;
 }
 
-int set_header(struct tdb_context *tdb,
-	       struct tdb_used_record *rec,
-	       unsigned magic, uint64_t keylen, uint64_t datalen,
-	       uint64_t actuallen, unsigned hashlow)
+enum TDB_ERROR set_header(struct tdb_context *tdb,
+			  struct tdb_used_record *rec,
+			  unsigned magic, uint64_t keylen, uint64_t datalen,
+			  uint64_t actuallen, unsigned hashlow)
 {
 	uint64_t keybits = (fls64(keylen) + 1) / 2;
 
@@ -567,62 +739,79 @@ int set_header(struct tdb_context *tdb,
 	if (rec_key_length(rec) != keylen
 	    || rec_data_length(rec) != datalen
 	    || rec_extra_padding(rec) != actuallen - (keylen + datalen)) {
-		tdb_logerr(tdb, TDB_ERR_IO, TDB_DEBUG_ERROR,
-			 "Could not encode k=%llu,d=%llu,a=%llu",
-			 (long long)keylen, (long long)datalen,
-			 (long long)actuallen);
-		return -1;
+		return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
+				  "Could not encode k=%llu,d=%llu,a=%llu",
+				  (long long)keylen, (long long)datalen,
+				  (long long)actuallen);
 	}
-	return 0;
+	return TDB_SUCCESS;
 }
 
 /* Expand the database. */
-static int tdb_expand(struct tdb_context *tdb, tdb_len_t size)
+static enum TDB_ERROR tdb_expand(struct tdb_context *tdb, tdb_len_t size)
 {
-	uint64_t old_size;
+	uint64_t old_size, rec_size, map_size;
 	tdb_len_t wanted;
-
-	/* We need room for the record header too. */
-	wanted = sizeof(struct tdb_used_record) + size;
+	enum TDB_ERROR ecode;
 
 	/* Need to hold a hash lock to expand DB: transactions rely on it. */
 	if (!(tdb->flags & TDB_NOLOCK)
-	    && !tdb->allrecord_lock.count && !tdb_has_hash_locks(tdb)) {
-		tdb_logerr(tdb, TDB_ERR_LOCK, TDB_DEBUG_ERROR,
-			   "tdb_expand: must hold lock during expand");
-		return -1;
+	    && !tdb->file->allrecord_lock.count && !tdb_has_hash_locks(tdb)) {
+		return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_ERROR,
+				  "tdb_expand: must hold lock during expand");
 	}
 
-	/* always make room for at least 100 more records, and at
-           least 25% more space. */
-	if (size * TDB_EXTENSION_FACTOR > tdb->map_size / 4)
-		wanted = size * TDB_EXTENSION_FACTOR;
-	else
-		wanted = tdb->map_size / 4;
-	wanted = adjust_size(0, wanted);
-
 	/* Only one person can expand file at a time. */
-	if (tdb_lock_expand(tdb, F_WRLCK) != 0)
-		return -1;
+	ecode = tdb_lock_expand(tdb, F_WRLCK);
+	if (ecode != TDB_SUCCESS) {
+		return ecode;
+	}
 
 	/* Someone else may have expanded the file, so retry. */
-	old_size = tdb->map_size;
-	tdb->methods->oob(tdb, tdb->map_size + 1, true);
-	if (tdb->map_size != old_size) {
+	old_size = tdb->file->map_size;
+	tdb->methods->oob(tdb, tdb->file->map_size + 1, true);
+	if (tdb->file->map_size != old_size) {
 		tdb_unlock_expand(tdb, F_WRLCK);
-		return 0;
+		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 (tdb->methods->expand_file(tdb, wanted) == -1) {
+	if (map_size > rec_size) {
+		wanted = map_size;
+	} else {
+		wanted = rec_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);
+	if (ecode != TDB_SUCCESS) {
 		tdb_unlock_expand(tdb, F_WRLCK);
-		return -1;
+		return ecode;
 	}
 
 	/* We need to drop this lock before adding free record. */
 	tdb_unlock_expand(tdb, F_WRLCK);
 
-	add_stat(tdb, expands, 1);
-	return add_free_record(tdb, old_size, wanted);
+	tdb->stats.expands++;
+	return add_free_record(tdb, old_size, wanted, TDB_LOCK_WAIT);
 }
 
 /* This won't fail: it will expand the database if it has to. */
@@ -635,12 +824,15 @@ tdb_off_t alloc(struct tdb_context *tdb, size_t keylen, size_t datalen,
 	assert(!tdb->direct_access);
 
 	for (;;) {
+		enum TDB_ERROR ecode;
 		off = get_free(tdb, keylen, datalen, growing, magic, hash);
 		if (likely(off != 0))
 			break;
 
-		if (tdb_expand(tdb, adjust_size(keylen, datalen)))
-			return TDB_OFF_ERR;
+		ecode = tdb_expand(tdb, adjust_size(keylen, datalen));
+		if (ecode != TDB_SUCCESS) {
+			return ecode;
+		}
 	}
 
 	return off;