X-Git-Url: https://git.ozlabs.org/?p=ccan;a=blobdiff_plain;f=ccan%2Ftdb2%2Ffree.c;h=be9b18455342a76ee316f98eff4363ade141849e;hp=631eeb3051a2babc71c72c3ed9081c35e9b82417;hb=6b3c079f3135b8276659e788da80b17ebd82ad59;hpb=1444b09cd357c145b6e3fdb1bce3b6ace6862f56

diff --git a/ccan/tdb2/free.c b/ccan/tdb2/free.c
index 631eeb30..be9b1845 100644
--- a/ccan/tdb2/free.c
+++ b/ccan/tdb2/free.c
@@ -103,92 +103,143 @@ static tdb_off_t find_free_head(struct tdb_context *tdb,
 				    bucket, TDB_FREE_BUCKETS);
 }
 
+static void check_list(struct tdb_context *tdb, tdb_off_t b_off)
+{
+#ifdef CCAN_TDB2_DEBUG
+	tdb_off_t off, prev = 0, first;
+	struct tdb_free_record r;
+
+	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)
+			abort();
+		if (prev && frec_prev(&r) != prev)
+			abort();
+		prev = off;
+		off = r.next;
+	}
+
+	if (first) {
+		tdb_read_convert(tdb, first, &r, sizeof(r));
+		if (frec_prev(&r) != prev)
+			abort();
+	}
+#endif
+}
+
 /* 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;
+	tdb_off_t off, prev_next, head;
 	enum TDB_ERROR ecode;
 
-	/* Front of list? */
-	if (frec_prev(r) == 0) {
-		off = b_off;
+	/* 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);
+
+	/* 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) {
+		/* We must preserve upper bits. */
+		head = tdb_read_off(tdb, b_off);
+		if (TDB_OFF_IS_ERR(head))
+			return 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)head,
+					  (long long)b_off);
+		}
+		head = ((head & ~TDB_OFF_MASK) | r->next);
+		ecode = tdb_write_off(tdb, b_off, head);
+		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 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;
+		head &= TDB_OFF_MASK;
+		off = head + offsetof(struct tdb_free_record, magic_and_prev);
 	} else {
-		off = frec_prev(r) + offsetof(struct tdb_free_record, next);
+		/* off = &r->next->prev */
+		off = r->next + offsetof(struct tdb_free_record,
+					 magic_and_prev);
 	}
 
 #ifdef CCAN_TDB2_DEBUG
-	if (tdb_read_off(tdb, off) != r_off) {
+	/* *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->prev->next = r->next */
-	ecode = tdb_write_off(tdb, off, r->next);
-	if (ecode != TDB_SUCCESS) {
-		return ecode;
-	}
-
-	if (r->next != 0) {
-		off = r->next + offsetof(struct tdb_free_record,magic_and_prev);
-		/* r->next->prev = r->prev */
-
-#ifdef CCAN_TDB2_DEBUG
-		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 list %llu",
-					  (long long)r_off, (long long)b_off);
-		}
-#endif
-
-		ecode = tdb_write_off(tdb, off, r->magic_and_prev);
-		if (ecode != TDB_SUCCESS) {
-			return ecode;
-		}
-	}
-	return TDB_SUCCESS;
+	/* r->next->prev = r->prev */
+	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, 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 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))
 		| len;
-	/* prev = 0. */
-	new.magic_and_prev = magic;
 
 	/* 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);
 
-	if (new.next) {
-#ifdef CCAN_TDB2_DEBUG
-		if (tdb_read_off(tdb,
-				 new.next + offsetof(struct tdb_free_record,
-						     magic_and_prev))
-		    != magic) {
+	/* 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)b_off);
+					  (long long)prev);
 		}
-#endif
 		/* next->prev = new. */
 		ecode = tdb_write_off(tdb, new.next
 				      + offsetof(struct tdb_free_record,
@@ -197,64 +248,38 @@ static enum TDB_ERROR enqueue_in_free(struct tdb_context *tdb,
 		if (ecode != TDB_SUCCESS) {
 			return ecode;
 		}
-	}
-	/* 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));
-}
-
-/* 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));
+#ifdef CCAN_TDB2_DEBUG
+		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
+	}
 
-	len = len_with_header - sizeof(struct tdb_used_record);
+	/* 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;
 
-	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);
-	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;
+	/* It's time to coalesce if counter wrapped. */
+	if (*coalesce)
+		*coalesce = ((head & ~TDB_OFF_MASK) == 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)
@@ -275,16 +300,18 @@ static tdb_off_t ftable_offset(struct tdb_context *tdb, unsigned int ftable)
 	return off;
 }
 
-/* Note: we unlock the current bucket if we coalesce or fail. */
-static tdb_bool_err 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 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_off_t *protect)
 {
 	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->file->map_size) {
@@ -318,7 +345,7 @@ static tdb_bool_err coalesce(struct tdb_context *tdb,
 		/* We may be violating lock order here, so best effort. */
 		if (tdb_lock_free_bucket(tdb, nb_off, TDB_LOCK_NOWAIT)
 		    != TDB_SUCCESS) {
-			add_stat(tdb, alloc_coalesce_lockfail, 1);
+			tdb->stats.alloc_coalesce_lockfail++;
 			break;
 		}
 
@@ -330,19 +357,24 @@ static tdb_bool_err coalesce(struct tdb_context *tdb,
 		}
 
 		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;
 		}
 
+		/* Did we just mess up a record you were hoping to use? */
+		if (end == *protect)
+			*protect = 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;
@@ -350,12 +382,16 @@ static tdb_bool_err coalesce(struct tdb_context *tdb,
 
 		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 false;
+		return 0;
+
+	/* Before we expand, check this isn't one you wanted protected? */
+	if (off == *protect)
+		*protect = TDB_ERR_EXISTS;
 
 	/* OK, expand initial record */
 	ecode = tdb_read_convert(tdb, off, &rec, sizeof(rec));
@@ -371,36 +407,233 @@ static tdb_bool_err coalesce(struct tdb_context *tdb,
 	}
 
 	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));
-	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. */
+		*protect = 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_unlock_free_bucket(tdb, b_off);
+
+		ecode = add_free_record(tdb, off, end - off, TDB_LOCK_WAIT,
+					false);
+		if (ecode != TDB_SUCCESS) {
+			return 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);
 
-	add_stat(tdb, alloc_coalesce_succeeded, 1);
+err:
+	/* To unify error paths, we *always* unlock bucket on error. */
 	tdb_unlock_free_bucket(tdb, b_off);
+	return ecode;
+}
 
-	ecode = add_free_record(tdb, off, end - off);
+/* 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 = 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 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 = 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->ftable_off, size_to_bucket(len));
+	ecode = tdb_lock_free_bucket(tdb, b_off, waitflag);
 	if (ecode != TDB_SUCCESS) {
 		return ecode;
 	}
-	return true;
 
-err:
-	/* To unify error paths, we *always* unlock bucket on error. */
-	tdb_unlock_free_bucket(tdb, b_off);
+	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->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,
@@ -416,8 +649,7 @@ static tdb_off_t lock_and_alloc(struct tdb_context *tdb,
 	size_t size = adjust_size(keylen, datalen);
 	enum TDB_ERROR ecode;
 
-	add_stat(tdb, allocs, 1);
-again:
+	tdb->stats.allocs++;
 	b_off = bucket_off(ftable_off, bucket);
 
 	/* FIXME: Try non-blocking wait first, to measure contention. */
@@ -443,12 +675,12 @@ again:
 		ecode = off;
 		goto unlock_err;
 	}
+	off &= TDB_OFF_MASK;
 
 	while (off) {
 		const struct tdb_free_record *r;
 		tdb_len_t len;
 		tdb_off_t next;
-		int coal;
 
 		r = tdb_access_read(tdb, off, sizeof(*r), true);
 		if (TDB_PTR_IS_ERR(r)) {
@@ -481,17 +713,6 @@ 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 (coal == 1) {
-			/* This has unlocked list, restart. */
-			goto again;
-		}
-		if (coal < 0) {
-			/* This has already unlocked on error. */
-			return coal;
-		}
 		off = next;
 	}
 
@@ -502,6 +723,7 @@ again:
 
 		/* We're happy with this size: take it. */
 		ecode = remove_from_list(tdb, b_off, best_off, &best);
+		check_list(tdb, b_off);
 		if (ecode != TDB_SUCCESS) {
 			goto unlock_err;
 		}
@@ -535,11 +757,11 @@ again:
 		/* Bucket of leftover will be <= current bucket, so nested
 		 * locking is allowed. */
 		if (leftover) {
-			add_stat(tdb, alloc_leftover, 1);
+			tdb->stats.alloc_leftover++;
 			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;
 			}
@@ -588,9 +810,9 @@ static tdb_off_t get_free(struct tdb_context *tdb,
 				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;
@@ -652,13 +874,10 @@ enum TDB_ERROR set_header(struct tdb_context *tdb,
 /* Expand the database. */
 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;
 	enum TDB_ERROR ecode;
 
-	/* We need room for the record header too. */
-	wanted = sizeof(struct tdb_used_record) + size;
-
 	/* Need to hold a hash lock to expand DB: transactions rely on it. */
 	if (!(tdb->flags & TDB_NOLOCK)
 	    && !tdb->file->allrecord_lock.count && !tdb_has_hash_locks(tdb)) {
@@ -666,14 +885,6 @@ static enum TDB_ERROR tdb_expand(struct tdb_context *tdb, tdb_len_t size)
 				  "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->file->map_size / 4)
-		wanted = size * TDB_EXTENSION_FACTOR;
-	else
-		wanted = tdb->file->map_size / 4;
-	wanted = adjust_size(0, wanted);
-
 	/* Only one person can expand file at a time. */
 	ecode = tdb_lock_expand(tdb, F_WRLCK);
 	if (ecode != TDB_SUCCESS) {
@@ -688,6 +899,32 @@ static enum TDB_ERROR tdb_expand(struct tdb_context *tdb, tdb_len_t size)
 		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;
+	}
+
+	/* 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);
@@ -697,8 +934,8 @@ static enum TDB_ERROR tdb_expand(struct tdb_context *tdb, tdb_len_t size)
 	/* 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, true);
 }
 
 /* This won't fail: it will expand the database if it has to. */