X-Git-Url: https://git.ozlabs.org/?p=ccan;a=blobdiff_plain;f=ccan%2Ftdb2%2Fcheck.c;h=54f96bb0ea8283583faf91693decd8d467970423;hp=5500883ff2b04cad383953877f4c8f415b06b0b2;hb=ec88af5dd11a003beb8cd3fbe420f9d5e5dcf5d9;hpb=1a24a8708494668c07e5c02284bfc2ef3b09603b

diff --git a/ccan/tdb2/check.c b/ccan/tdb2/check.c
index 5500883f..54f96bb0 100644
--- a/ccan/tdb2/check.c
+++ b/ccan/tdb2/check.c
@@ -30,66 +30,74 @@ static bool append(tdb_off_t **arr, size_t *num, tdb_off_t off)
 	return true;
 }
 
-static bool check_header(struct tdb_context *tdb)
+static bool check_header(struct tdb_context *tdb, tdb_off_t *recovery)
 {
 	uint64_t hash_test;
+	struct tdb_header hdr;
+
+	if (tdb_read_convert(tdb, 0, &hdr, sizeof(hdr)) == -1)
+		return false;
+	/* magic food should not be converted, so convert back. */
+	tdb_convert(tdb, hdr.magic_food, sizeof(hdr.magic_food));
 
 	hash_test = TDB_HASH_MAGIC;
 	hash_test = tdb_hash(tdb, &hash_test, sizeof(hash_test));
-	if (tdb->header.hash_test != hash_test) {
+	if (hdr.hash_test != hash_test) {
 		tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 			 "check: hash test %llu should be %llu\n",
-			 tdb->header.hash_test, hash_test);
+			 (long long)hdr.hash_test,
+			 (long long)hash_test);
 		return false;
 	}
-	if (strcmp(tdb->header.magic_food, TDB_MAGIC_FOOD) != 0) {
+
+	if (strcmp(hdr.magic_food, TDB_MAGIC_FOOD) != 0) {
 		tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 			 "check: bad magic '%.*s'\n",
-			 sizeof(tdb->header.magic_food),
-			 tdb->header.magic_food);
-		return false;
-	}
-	if (tdb->header.v.hash_bits < INITIAL_HASH_BITS) {
-		tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-			 "check: bad hash bits %llu\n",
-			 (long long)tdb->header.v.hash_bits);
+			 (unsigned)sizeof(hdr.magic_food), hdr.magic_food);
 		return false;
 	}
 
-	/* We check hash_off later. */
+	*recovery = hdr.recovery;
+	if (*recovery) {
+		if (*recovery < sizeof(hdr) || *recovery > tdb->map_size) {
+			tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
+				 "tdb_check: invalid recovery offset %zu\n",
+				 (size_t)*recovery);
+			return false;
+		}
+	}
 
 	/* Don't check reserved: they *can* be used later. */
 	return true;
 }
 
-static int off_cmp(const tdb_off_t *a, const tdb_off_t *b)
-{
-	/* Can overflow an int. */
-	return *a > *b ? 1
-		: *a < *b ? -1
-		: 0;
-}
-
-static bool check_hash_list(struct tdb_context *tdb,
+static bool check_hash_tree(struct tdb_context *tdb,
+			    tdb_off_t off, unsigned int group_bits,
+			    uint64_t hprefix,
+			    unsigned hprefix_bits,
 			    tdb_off_t used[],
-			    size_t num_used)
+			    size_t num_used,
+			    size_t *num_found);
+
+static bool check_hash_record(struct tdb_context *tdb,
+			      tdb_off_t off,
+			      uint64_t hprefix,
+			      unsigned hprefix_bits,
+			      tdb_off_t used[],
+			      size_t num_used,
+			      size_t *num_found)
 {
 	struct tdb_used_record rec;
-	tdb_len_t hashlen, i, num_nonzero;
-	tdb_off_t h;
-	size_t num_found;
-
-	hashlen = sizeof(tdb_off_t) << tdb->header.v.hash_bits;
 
-	if (tdb_read_convert(tdb, tdb->header.v.hash_off - sizeof(rec),
-			     &rec, sizeof(rec)) == -1)
+	if (tdb_read_convert(tdb, off, &rec, sizeof(rec)) == -1)
 		return false;
 
-	if (rec_data_length(&rec) != hashlen) {
+	if (rec_data_length(&rec)
+	    != sizeof(tdb_off_t) << TDB_SUBLEVEL_HASH_BITS) {
 		tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 			 "tdb_check: Bad hash table length %llu vs %llu\n",
 			 (long long)rec_data_length(&rec),
-			 (long long)hashlen);
+			 (long long)sizeof(tdb_off_t)<<TDB_SUBLEVEL_HASH_BITS);
 		return false;
 	}
 	if (rec_key_length(&rec) != 0) {
@@ -105,69 +113,172 @@ static bool check_hash_list(struct tdb_context *tdb,
 		return false;
 	}
 
-	num_found = 0;
-	num_nonzero = 0;
-	for (i = 0, h = tdb->header.v.hash_off;
-	     i < (1ULL << tdb->header.v.hash_bits);
-	     i++, h += sizeof(tdb_off_t)) {
-		tdb_off_t off, *p, pos;
-		struct tdb_used_record rec;
-		uint64_t hash;
-
-		off = tdb_read_off(tdb, h);
-		if (off == TDB_OFF_ERR)
-			return false;
-		if (!off) {
-			num_nonzero = 0;
-			continue;
-		}
-		/* FIXME: Check hash bits */
-		p = asearch(&off, used, num_used, off_cmp);
-		if (!p) {
-			tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-				 "tdb_check: Invalid offset %llu in hash\n",
-				 (long long)off);
-			return false;
-		}
-		/* Mark it invalid. */
-		*p ^= 1;
-		num_found++;
+	off += sizeof(rec);
+	return check_hash_tree(tdb, off,
+			       TDB_SUBLEVEL_HASH_BITS-TDB_HASH_GROUP_BITS,
+			       hprefix, hprefix_bits,
+			       used, num_used, num_found);
+}
 
-		if (tdb_read_convert(tdb, off, &rec, sizeof(rec)) == -1)
-			return false;
+static int off_cmp(const tdb_off_t *a, const tdb_off_t *b)
+{
+	/* Can overflow an int. */
+	return *a > *b ? 1
+		: *a < *b ? -1
+		: 0;
+}
 
-		/* Check it is hashed correctly. */
-		hash = hash_record(tdb, off);
+static uint64_t get_bits(uint64_t h, unsigned num, unsigned *used)
+{
+	*used += num;
 
-		/* Top bits must match header. */
-		if (hash >> (64 - 5) != rec_hash(&rec)) {
-			tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-				 "tdb_check: Bad hash magic at offset %llu"
-				 " (0x%llx vs 0x%llx)\n",
-				 (long long)off,
-				 (long long)hash, (long long)rec_hash(&rec));
-			return false;
-		}
+	return (h >> (64 - *used)) & ((1U << num) - 1);
+}
+
+static bool check_hash_tree(struct tdb_context *tdb,
+			    tdb_off_t off, unsigned int group_bits,
+			    uint64_t hprefix,
+			    unsigned hprefix_bits,
+			    tdb_off_t used[],
+			    size_t num_used,
+			    size_t *num_found)
+{
+	unsigned int g, b;
+	const tdb_off_t *hash;
+	struct tdb_used_record rec;
+
+	hash = tdb_access_read(tdb, off,
+			       sizeof(tdb_off_t)
+			       << (group_bits + TDB_HASH_GROUP_BITS),
+			       true);
+	if (!hash)
+		return false;
+
+	for (g = 0; g < (1 << group_bits); g++) {
+		const tdb_off_t *group = hash + (g << TDB_HASH_GROUP_BITS);
+		for (b = 0; b < (1 << TDB_HASH_GROUP_BITS); b++) {
+			unsigned int bucket, i, used_bits;
+			uint64_t h;
+			tdb_off_t *p;
+			if (group[b] == 0)
+				continue;
+
+			off = group[b] & TDB_OFF_MASK;
+			p = asearch(&off, used, num_used, off_cmp);
+			if (!p) {
+				tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
+					 "tdb_check: Invalid offset %llu "
+					 "in hash\n",
+					 (long long)off);
+				goto fail;
+			}
+			/* Mark it invalid. */
+			*p ^= 1;
+			(*num_found)++;
+
+			if (is_subhash(group[b])) {
+				uint64_t subprefix;
+				subprefix = (hprefix 
+				     << (group_bits + TDB_HASH_GROUP_BITS))
+					+ g * (1 << TDB_HASH_GROUP_BITS) + b;
+
+				if (!check_hash_record(tdb,
+					       group[b] & TDB_OFF_MASK,
+					       subprefix,
+					       hprefix_bits
+						       + group_bits
+						       + TDB_HASH_GROUP_BITS,
+					       used, num_used, num_found))
+					goto fail;
+				continue;
+			}
+			/* A normal entry */
+
+			/* Does it belong here at all? */
+			h = hash_record(tdb, off);
+			used_bits = 0;
+			if (get_bits(h, hprefix_bits, &used_bits) != hprefix
+			    && hprefix_bits) {
+				tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
+					 "check: bad hash placement"
+					 " 0x%llx vs 0x%llx\n",
+					 (long long)h, (long long)hprefix);
+				goto fail;
+			}
+
+			/* Does it belong in this group? */
+			if (get_bits(h, group_bits, &used_bits) != g) {
+				tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
+					 "check: bad group %llu vs %u\n",
+					 (long long)h, g);
+				goto fail;
+			}
 
-		/* It must be in the right place in hash array. */
-		pos = hash & ((1ULL << tdb->header.v.hash_bits)-1);
-		if (pos < i - num_nonzero || pos > i) {
-			/* Could be wrap from end of array?  FIXME: check? */
-			if (i != num_nonzero) {
+			/* Are bucket bits correct? */
+			bucket = group[b] & TDB_OFF_HASH_GROUP_MASK;
+			if (get_bits(h, TDB_HASH_GROUP_BITS, &used_bits)
+			    != bucket) {
+				used_bits -= TDB_HASH_GROUP_BITS;
 				tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-					 "tdb_check: Bad hash position %llu at"
-					 " offset %llu hash 0x%llx\n",
-					 (long long)i,
+					 "check: bad bucket %u vs %u\n",
+					 (unsigned)get_bits(h,
+							    TDB_HASH_GROUP_BITS,
+							    &used_bits),
+					 bucket);
+				goto fail;
+			}
+
+			/* There must not be any zero entries between
+			 * the bucket it belongs in and this one! */
+			for (i = bucket;
+			     i != b;
+			     i = (i + 1) % (1 << TDB_HASH_GROUP_BITS)) {
+				if (group[i] == 0) {
+					tdb->log(tdb, TDB_DEBUG_ERROR,
+						 tdb->log_priv,
+						 "check: bad group placement"
+						 " %u vs %u\n",
+						 b, bucket);
+					goto fail;
+				}
+			}
+
+			if (tdb_read_convert(tdb, off, &rec, sizeof(rec)) == -1)
+				goto fail;
+
+			/* Bottom bits must match header. */
+			if ((h & ((1 << 11)-1)) != rec_hash(&rec)) {
+				tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
+					 "tdb_check: Bad hash magic at"
+					 " offset %llu (0x%llx vs 0x%llx)\n",
 					 (long long)off,
-					 (long long)hash);
-				return false;
+					 (long long)h,
+					 (long long)rec_hash(&rec));
+				goto fail;
 			}
 		}
-		num_nonzero++;
 	}
+	tdb_access_release(tdb, hash);
+	return true;
 
-	/* hash table is one of the used blocks. */
-	if (num_found != num_used - 1) {
+fail:
+	tdb_access_release(tdb, hash);
+	return false;
+}
+
+static bool check_hash(struct tdb_context *tdb,
+		       tdb_off_t used[],
+		       size_t num_used, size_t num_flists)
+{
+	/* Free lists also show up as used. */
+	size_t num_found = num_flists;
+
+	if (!check_hash_tree(tdb, offsetof(struct tdb_header, hashtable),
+			     TDB_TOPLEVEL_HASH_BITS-TDB_HASH_GROUP_BITS,
+			     0, 0, used, num_used, &num_found))
+		return false;
+
+	if (num_found != num_used) {
 		tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 			 "tdb_check: Not all entries are in hash\n");
 		return false;
@@ -178,8 +289,7 @@ static bool check_hash_list(struct tdb_context *tdb,
 static bool check_free(struct tdb_context *tdb,
 		       tdb_off_t off,
 		       const struct tdb_free_record *frec,
-		       tdb_off_t prev,
-		       tdb_off_t zone_off, unsigned int bucket)
+		       tdb_off_t prev, tdb_off_t flist_off, unsigned int bucket)
 {
 	if (frec_magic(frec) != TDB_FREE_MAGIC) {
 		tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
@@ -187,24 +297,22 @@ static bool check_free(struct tdb_context *tdb,
 			 (long long)off, (long long)frec->magic_and_meta);
 		return false;
 	}
-	if (tdb->methods->oob(tdb, off
-			      + frec->data_len-sizeof(struct tdb_used_record),
-			      true))
-		return false;
-	if (off < zone_off || off >= zone_off + (1ULL<<frec_zone_bits(frec))) {
+	if (frec_flist(frec) != flist_off) {
 		tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-			 "tdb_check: offset %llu outside zone %llu-%llu\n",
-			 (long long)off,
-			 (long long)zone_off,
-			 (long long)zone_off + (1ULL<<frec_zone_bits(frec)));
+			 "tdb_check: offset %llu bad freelist 0x%llx\n",
+			 (long long)off, (long long)frec_flist(frec));
 		return false;
 	}
-	if (size_to_bucket(frec_zone_bits(frec), frec->data_len) != bucket) {
+
+	if (tdb->methods->oob(tdb, off
+			      + frec->data_len+sizeof(struct tdb_used_record),
+			      false))
+		return false;
+	if (size_to_bucket(frec->data_len) != bucket) {
 		tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 			 "tdb_check: offset %llu in wrong bucket %u vs %u\n",
 			 (long long)off,
-			 bucket,
-			 size_to_bucket(frec_zone_bits(frec), frec->data_len));
+			 bucket, size_to_bucket(frec->data_len));
 		return false;
 	}
 	if (prev != frec->prev) {
@@ -217,30 +325,40 @@ static bool check_free(struct tdb_context *tdb,
 	return true;
 }
 		       
-static tdb_len_t check_free_list(struct tdb_context *tdb,
-				 tdb_off_t zone_off,
-				 tdb_off_t free[],
-				 size_t num_free,
-				 size_t *num_found)
+static bool check_free_list(struct tdb_context *tdb,
+			    tdb_off_t flist_off,
+			    tdb_off_t free[],
+			    size_t num_free,
+			    size_t *num_found)
 {
-	struct free_zone_header zhdr;
+	struct tdb_freelist flist;
 	tdb_off_t h;
 	unsigned int i;
 
-	if (tdb_read_convert(tdb, zone_off, &zhdr, sizeof(zhdr)) == -1)
-		return TDB_OFF_ERR;
+	if (tdb_read_convert(tdb, flist_off, &flist, sizeof(flist)) == -1)
+		return false;
 
-	for (i = 0; i <= BUCKETS_FOR_ZONE(zhdr.zone_bits); i++) {
+	if (rec_magic(&flist.hdr) != TDB_MAGIC
+	    || rec_key_length(&flist.hdr) != 0
+	    || rec_data_length(&flist.hdr) != sizeof(flist) - sizeof(flist.hdr)
+	    || rec_hash(&flist.hdr) != 1) {
+		tdb->log(tdb, TDB_DEBUG_ERROR,
+			 tdb->log_priv,
+			 "tdb_check: Invalid header on free list\n");
+		return false;
+	}
+
+	for (i = 0; i < TDB_FREE_BUCKETS; i++) {
 		tdb_off_t off, prev = 0, *p;
 		struct tdb_free_record f;
 
-		h = bucket_off(zone_off, i);
+		h = bucket_off(flist_off, i);
 		for (off = tdb_read_off(tdb, h); off; off = f.next) {
 			if (off == TDB_OFF_ERR)
 				return false;
 			if (tdb_read_convert(tdb, off, &f, sizeof(f)))
 				return false;
-			if (!check_free(tdb, off, &f, prev, zone_off, i))
+			if (!check_free(tdb, off, &f, prev, flist_off, i))
 				return false;
 
 			/* FIXME: Check hash bits */
@@ -259,74 +377,88 @@ static tdb_len_t check_free_list(struct tdb_context *tdb,
 			prev = off;
 		}
 	}
-	return 1ULL << zhdr.zone_bits;
+	return true;
 }
 
-static tdb_off_t check_zone(struct tdb_context *tdb, tdb_off_t zone_off,
-			    tdb_off_t **used, size_t *num_used,
-			    tdb_off_t **free, size_t *num_free,
-			    bool *hash_found, unsigned int *max_zone_bits)
+/* Slow, but should be very rare. */
+size_t dead_space(struct tdb_context *tdb, tdb_off_t off)
 {
-	struct free_zone_header zhdr;
-	tdb_off_t off, hdrlen;
-	tdb_len_t len;
-
-	if (tdb_read_convert(tdb, zone_off, &zhdr, sizeof(zhdr)) == -1)
-		return TDB_OFF_ERR;
-
-	if (zhdr.zone_bits < INITIAL_ZONE_BITS) {
-		tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-			 "check: bad zone_bits %llu at zone %llu\n",
-			 (long long)zhdr.zone_bits, (long long)zone_off);
-		return TDB_OFF_ERR;
-	}
-
-	/* Zone bits can only increase... */
-	if (zhdr.zone_bits > *max_zone_bits)
-		*max_zone_bits = zhdr.zone_bits;
-	else if (zhdr.zone_bits < *max_zone_bits) {
-		tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-			 "check: small zone_bits %llu at zone %llu\n",
-			 (long long)zhdr.zone_bits, (long long)zone_off);
-		return TDB_OFF_ERR;
+	size_t len;
+
+	for (len = 0; off + len < tdb->map_size; len++) {
+		char c;
+		if (tdb->methods->read(tdb, off, &c, 1))
+			return 0;
+		if (c != 0 && c != 0x43)
+			break;
 	}
+	return len;
+}
 
-	/* Zone must be within file! */
-	if (tdb->methods->oob(tdb, zone_off + (1ULL << zhdr.zone_bits), false))
-		return TDB_OFF_ERR;
+static bool check_linear(struct tdb_context *tdb,
+			 tdb_off_t **used, size_t *num_used,
+			 tdb_off_t **free, size_t *num_free,
+			 tdb_off_t recovery)
+{
+	tdb_off_t off;
+	tdb_len_t len;
+	bool found_recovery = false;
 
-	hdrlen = sizeof(zhdr)
-		+ (BUCKETS_FOR_ZONE(zhdr.zone_bits) + 1) * sizeof(tdb_off_t);
-	for (off = zone_off + hdrlen;
-	     off < zone_off + (1ULL << zhdr.zone_bits);
-	     off += len) {
+	for (off = sizeof(struct tdb_header); off < tdb->map_size; off += len) {
 		union {
 			struct tdb_used_record u;
 			struct tdb_free_record f;
+			struct tdb_recovery_record r;
 		} pad, *p;
 		p = tdb_get(tdb, off, &pad, sizeof(pad));
 		if (!p)
-			return TDB_OFF_ERR;
-		if (frec_magic(&p->f) == TDB_FREE_MAGIC) {
-			if (frec_zone_bits(&p->f) != zhdr.zone_bits) {
+			return false;
+
+		/* If we crash after ftruncate, we can get zeroes or fill. */
+		if (p->r.magic == TDB_RECOVERY_INVALID_MAGIC
+		    || p->r.magic ==  0x4343434343434343ULL) {
+			if (recovery == off) {
+				found_recovery = true;
+				len = sizeof(p->r) + p->r.max_len;
+			} else {
+				len = dead_space(tdb, off);
+				if (len < sizeof(p->r)) {
+					tdb->log(tdb, TDB_DEBUG_ERROR,
+						 tdb->log_priv,
+						 "tdb_check: invalid dead space"
+						 " at %zu\n", (size_t)off);
+					return false;
+				}
+
+				tdb->log(tdb, TDB_DEBUG_WARNING, tdb->log_priv,
+					 "Dead space at %zu-%zu (of %zu)\n",
+					 (size_t)off, (size_t)(off + len),
+					 (size_t)tdb->map_size);
+			}
+		} else if (p->r.magic == TDB_RECOVERY_MAGIC) {
+			if (recovery != off) {
 				tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-					 "tdb_check: Bad free zone bits %u"
-					 " at offset %llu\n",
-					 frec_zone_bits(&p->f),
-					 (long long)off);
-				return TDB_OFF_ERR;
+					 "tdb_check: unexpected recovery"
+					 " record at offset %zu\n",
+					 (size_t)off);
+				return false;
 			}
-			/* This record is free! */
-			if (!append(free, num_free, off))
-				return TDB_OFF_ERR;
+			found_recovery = true;
+			len = sizeof(p->r) + p->r.max_len;
+		} else if (frec_magic(&p->f) == TDB_FREE_MAGIC
+			   || frec_magic(&p->f) == TDB_COALESCING_MAGIC) {
 			len = sizeof(p->u) + p->f.data_len;
-			if (off + len > zone_off + (1ULL << zhdr.zone_bits)) {
+			if (off + len > tdb->map_size) {
 				tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 					 "tdb_check: free overlength %llu"
 					 " at offset %llu\n",
 					 (long long)len, (long long)off);
-				return TDB_OFF_ERR;
+				return false;
 			}
+			/* This record is free! */
+			if (frec_magic(&p->f) == TDB_FREE_MAGIC
+			    && !append(free, num_free, off))
+				return false;
 		} else {
 			uint64_t klen, dlen, extra;
 
@@ -337,32 +469,23 @@ static tdb_off_t check_zone(struct tdb_context *tdb, tdb_off_t zone_off,
 					 " at offset %llu\n",
 					 (long long)rec_magic(&p->u),
 					 (long long)off);
-				return TDB_OFF_ERR;
+				return false;
 			}
 
-			if (rec_zone_bits(&p->u) != zhdr.zone_bits) {
-				tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-					 "tdb_check: Bad zone bits %u"
-					 " at offset %llu\n",
-					 rec_zone_bits(&p->u),
-					 (long long)off);
-				return TDB_OFF_ERR;
-			}
-			
 			if (!append(used, num_used, off))
-				return TDB_OFF_ERR;
+				return false;
 
 			klen = rec_key_length(&p->u);
 			dlen = rec_data_length(&p->u);
 			extra = rec_extra_padding(&p->u);
 
 			len = sizeof(p->u) + klen + dlen + extra;
-			if (off + len > zone_off + (1ULL << zhdr.zone_bits)) {
+			if (off + len > tdb->map_size) {
 				tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 					 "tdb_check: used overlength %llu"
 					 " at offset %llu\n",
 					 (long long)len, (long long)off);
-				return TDB_OFF_ERR;
+				return false;
 			}
 
 			if (len < sizeof(p->f)) {
@@ -370,14 +493,20 @@ static tdb_off_t check_zone(struct tdb_context *tdb, tdb_off_t zone_off,
 					 "tdb_check: too short record %llu at"
 					 " %llu\n",
 					 (long long)len, (long long)off);
-				return TDB_OFF_ERR;
+				return false;
 			}
-
-			if (off + sizeof(p->u) == tdb->header.v.hash_off)
-				*hash_found = true;
 		}
 	}
-	return 1ULL << zhdr.zone_bits;
+
+	/* We must have found recovery area if there was one. */
+	if (recovery != 0 && !found_recovery) {
+		tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
+			 "tdb_check: expected a recovery area at %zu\n",
+			 (size_t)recovery);
+		return false;
+	}
+
+	return true;
 }
 
 /* FIXME: call check() function. */
@@ -385,60 +514,36 @@ int tdb_check(struct tdb_context *tdb,
 	      int (*check)(TDB_DATA key, TDB_DATA data, void *private_data),
 	      void *private_data)
 {
-	tdb_off_t *free = NULL, *used = NULL, off;
-	tdb_len_t len;
-	size_t num_free = 0, num_used = 0, num_found = 0;
-	bool hash_found = false;
-	unsigned max_zone_bits = INITIAL_ZONE_BITS;
-	uint8_t tailer;
+	tdb_off_t *free = NULL, *used = NULL, flist, recovery;
+	size_t num_free = 0, num_used = 0, num_found = 0, num_flists = 0;
 
-	/* FIXME: need more locking? against expansion? */
-	/* This always ensures the header is uptodate. */
 	if (tdb_allrecord_lock(tdb, F_RDLCK, TDB_LOCK_WAIT, false) != 0)
 		return -1;
 
-	if (!check_header(tdb))
-		goto fail;
-
-	/* First we do a linear scan, checking all records. */
-	for (off = sizeof(struct tdb_header);
-	     off < tdb->map_size - 1;
-	     off += len) {
-		len = check_zone(tdb, off, &used, &num_used, &free, &num_free,
-				 &hash_found, &max_zone_bits);
-		if (len == TDB_OFF_ERR)
-			goto fail;
+	if (tdb_lock_expand(tdb, F_RDLCK) != 0) {
+		tdb_allrecord_unlock(tdb, F_RDLCK);
+		return -1;
 	}
 
-	/* Check tailer. */
-	if (tdb->methods->read(tdb, tdb->map_size - 1, &tailer, 1) == -1)
-		goto fail;
-	if (tailer != max_zone_bits) {
-		tdb->ecode = TDB_ERR_CORRUPT;
-		tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-			 "tdb_check: Bad tailer value %u vs %u\n", tailer,
-			 max_zone_bits);
+	if (!check_header(tdb, &recovery))
 		goto fail;
-	}
 
-	if (!hash_found) {
-		tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
-			 "tdb_check: hash table not found at %llu\n",
-			 (long long)tdb->header.v.hash_off);
+	/* First we do a linear scan, checking all records. */
+	if (!check_linear(tdb, &used, &num_used, &free, &num_free, recovery))
 		goto fail;
+
+	for (flist = first_flist(tdb); flist; flist = next_flist(tdb, flist)) {
+		if (flist == TDB_OFF_ERR)
+			goto fail;
+		if (!check_free_list(tdb, flist, free, num_free, &num_found))
+			goto fail;
+		num_flists++;
 	}
 
 	/* FIXME: Check key uniqueness? */
-	if (!check_hash_list(tdb, used, num_used))
+	if (!check_hash(tdb, used, num_used, num_flists))
 		goto fail;
 
-	for (off = sizeof(struct tdb_header);
-	     off < tdb->map_size - 1;
-	     off += len) {
-		len = check_free_list(tdb, off, free, num_free, &num_found);
-		if (len == TDB_OFF_ERR)
-			goto fail;
-	}
 	if (num_found != num_free) {
 		tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 			 "tdb_check: Not all entries are in free table\n");
@@ -446,9 +551,11 @@ int tdb_check(struct tdb_context *tdb,
 	}
 
 	tdb_allrecord_unlock(tdb, F_RDLCK);
+	tdb_unlock_expand(tdb, F_RDLCK);
 	return 0;
 
 fail:
 	tdb_allrecord_unlock(tdb, F_RDLCK);
+	tdb_unlock_expand(tdb, F_RDLCK);
 	return -1;
 }