lmdb: major rework of traversal b-tree for mdb_chk.

Change-Id: I9d382516f76092f44fc1a12d7554039582b87656

lmdb.h

@@ -1658,10 +1658,10 @@ typedef void MDB_debug_func(int type, const char *function, int line,
 
 int mdb_setup_debug(int flags, MDB_debug_func* logger, long edge_txn);
 
-typedef int MDB_pgwalk_func(size_t pgno, unsigned pgnumber, void* ctx,
+typedef int MDB_pgvisitor_func(size_t pgno, unsigned pgnumber, void* ctx,
-					const char* dbi, char type,
+					const char* dbi, const char *type,
-					int payload_bytes, int header_bytes);
+					int payload_bytes, int header_bytes, int unused_bytes);
-int mdb_env_pgwalk(MDB_txn *txn, MDB_pgwalk_func* visitor, void* ctx);
+int mdb_env_pgwalk(MDB_txn *txn, MDB_pgvisitor_func* visitor, void* ctx);
 
 char* mdb_dkey(MDB_val *key, char *buf);
 

mdb.c

@@ -1472,8 +1472,7 @@ mdb_page_list(MDB_page *mp)
 		key.mv_data = node->mn_data;
 		nsize = NODESIZE + key.mv_size;
 		if (IS_BRANCH(mp)) {
-			mdb_print("key %d: page %zu, %s\n", i, NODEPGNO(node),
+			mdb_print("key %d: page %zu, %s\n", i, NODEPGNO(node), DKEY(&key));
-				DKEY(&key));
 			total += nsize;
 		} else {
 			if (F_ISSET(node->mn_flags, F_BIGDATA))
@@ -9794,154 +9793,141 @@ mdb_env_get_oomfunc(MDB_env *env)
 struct mdb_walk_ctx {
 	MDB_txn *mw_txn;
 	void *mw_user;
-	MDB_pgwalk_func *mw_visitor;
+	MDB_pgvisitor_func *mw_visitor;
 };
 
 typedef struct mdb_walk_ctx mdb_walk_ctx_t;
 
+
 /** Depth-first tree traversal. */
 static int ESECT
 mdb_env_walk(mdb_walk_ctx_t *ctx, const char* dbi, pgno_t pg, int flags, int deep)
 {
-	MDB_cursor mc;
-	MDB_node *ni;
 	MDB_page *mp;
-	int rc;
+	int rc, i, nkeys;
-	unsigned i;
+	unsigned header_size, unused_size, payload_size, align_bytes;
+	const char* type;
 
-	/* Empty DB, nothing to do */
 	if (pg == P_INVALID)
-		return MDB_SUCCESS;
+		return MDB_CORRUPTED;
 
-	if (deep < 2) {
+	rc = mdb_page_get(ctx->mw_txn, pg, &mp, NULL);
-		if ((rc = mdb_page_get(ctx->mw_txn, pg, &mp, NULL)) != 0)
-			return rc;
-		rc = ctx->mw_visitor(pg, 0, ctx->mw_user, dbi, 'R',
-			ctx->mw_txn->mt_env->me_psize - PAGEHDRSZ - SIZELEFT(mp), PAGEHDRSZ);
-		if (rc)
-			return rc;
-	}
-
-	mc.mc_snum = 1;
-	mc.mc_top = 0;
-	mc.mc_txn = ctx->mw_txn;
-
-	rc = mdb_page_get(ctx->mw_txn, pg, &mc.mc_pg[0], NULL);
 	if (rc)
 		return rc;
+	if (pg != mp->mp_p.p_pgno)
+		return MDB_CORRUPTED;
 
-	for (mp = mc.mc_pg[mc.mc_top]; IS_BRANCH(mp); ) {
+	nkeys = NUMKEYS(mp);
-		MDB_node	*node;
+	header_size = IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower;
+	unused_size = SIZELEFT(mp);
+	payload_size = 0;
 
-		rc = ctx->mw_visitor(mp->mp_p.p_pgno, 1, ctx->mw_user, dbi, 'B',
+	/* LY: Don't use mask here, e.g bitwise (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP).
-			ctx->mw_txn->mt_env->me_psize - PAGEHDRSZ - SIZELEFT(mp), PAGEHDRSZ);
+	 * Pages should not me marked dirty/loose or otherwise. */
-		if (rc)
+	switch (mp->mp_flags) {
-			return rc;
+	case P_BRANCH:
-
+		type = "branch";
-		if (NUMKEYS(mp) < 1)
+		if (nkeys < 1)
 			return MDB_CORRUPTED;
-
+		break;
-		mdb_debug("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
+	case P_LEAF:
-		mdb_cassert(&mc, NUMKEYS(mp) > 1);
+		type = "leaf";
-		mdb_debug("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
+		break;
-
+	case P_LEAF|P_SUBP:
-		node = NODEPTR(mp, 0);
+		type = "leaf-dupsort";
-		if ((rc = mdb_page_get(mc.mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
+		break;
-			return rc;
+	case P_LEAF|P_LEAF2:
-
+		/* #MDB_DUPFIXED records */
-		mc.mc_ki[mc.mc_top] = 0;
+		type = "leaf-dupfixed";
-		if ((rc = mdb_cursor_push(&mc, mp)))
+		break;
-			return rc;
+	case P_LEAF|P_LEAF2|P_SUBP:
-	}
+		/* #MDB_DUPSORT sub-pages */
-
+		type = "leaf-dupfixed-dupsort";
-	if (!IS_LEAF(mp)) {
+		break;
-		mdb_debug("internal error, index points to a %02X page!?",
+	case P_META:
-			mp->mp_flags);
+	case P_OVERFLOW:
-		mc.mc_txn->mt_flags |= MDB_TXN_ERROR;
+	default:
 		return MDB_CORRUPTED;
 	}
 
-	mc.mc_flags |= C_INITIALIZED;
+	for (align_bytes = i = 0; i < nkeys;
-	mc.mc_flags &= ~C_EOF;
+		 align_bytes += ((payload_size + align_bytes) & 1), i++) {
+		MDB_node *node;
 
-	rc = ctx->mw_visitor(mp->mp_p.p_pgno, 1, ctx->mw_user, dbi, 'L',
+		if (IS_LEAF2(mp)) {
-		ctx->mw_txn->mt_env->me_psize - PAGEHDRSZ - SIZELEFT(mp), PAGEHDRSZ);
+			/* LEAF2 pages have no mp_ptrs[] or node headers */
-	if (rc)
+			payload_size += mp->mp_ksize;
-		return rc;
+			continue;
-
-	while (mc.mc_snum > 0) {
-		unsigned n;
-		mp = mc.mc_pg[mc.mc_top];
-		n = NUMKEYS(mp);
-
-		if (IS_LEAF(mp)) {
-			if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
-				for (i = 0; i < n; i++) {
-					ni = NODEPTR(mp, i);
-					if (ni->mn_flags & F_BIGDATA) {
-						MDB_page *omp;
-						pgno_t *pg;
-
-						pg = NODEDATA(ni);
-						rc = mdb_page_get(ctx->mw_txn, *pg, &omp, NULL);
-						if (rc)
-							return rc;
-						rc = ctx->mw_visitor(*pg, omp->mp_pages, ctx->mw_user, dbi, 'L',
-							ctx->mw_txn->mt_env->me_psize - PAGEHDRSZ - SIZELEFT(mp), PAGEHDRSZ);
-						if (rc)
-							return rc;
-					} else if (ni->mn_flags & F_SUBDATA) {
-						MDB_db *db = NODEDATA(ni);
-						char* name = NULL;
-						if (! (ni->mn_flags & F_DUPDATA)) {
-							name = NODEKEY(ni);
-							int namelen = (char*) db - name;
-							name = memcpy(alloca(namelen + 1), name, namelen);
-							name[namelen] = 0;
-						}
-						rc = mdb_env_walk(ctx, (name && name[0]) ? name : dbi, db->md_root, ni->mn_flags & F_DUPDATA, deep + 1);
-						if (rc)
-							return rc;
-					}
-				}
-			}
-		} else {
-			mc.mc_ki[mc.mc_top]++;
-			if (mc.mc_ki[mc.mc_top] < n) {
-				pgno_t pg;
-				do {
-					ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
-					pg = NODEPGNO(ni);
-					rc = mdb_page_get(ctx->mw_txn, pg, &mp, NULL);
-					if (rc)
-						return rc;
-					rc = ctx->mw_visitor(pg, 1, ctx->mw_user, dbi, IS_BRANCH(mp) ? 'B' : 'L',
-							ctx->mw_txn->mt_env->me_psize - PAGEHDRSZ - SIZELEFT(mp), PAGEHDRSZ);
-					if (rc)
-						return rc;
-					mc.mc_top++;
-					mc.mc_snum++;
-					mc.mc_ki[mc.mc_top] = 0;
-					mc.mc_pg[mc.mc_top] = mp;
-				}
-				/* Whenever we advance to a sibling branch page,
-				 * we must proceed all the way down to its first leaf.
-				 */
-				while (IS_BRANCH(mp));
-				continue;
-			}
 		}
 
-		if (! mc.mc_top)
+		node = NODEPTR(mp, i);
-			break;
+		payload_size += NODESIZE + node->mn_ksize;
 
-		mdb_cursor_pop(&mc);
+		if (IS_BRANCH(mp)) {
+			rc = mdb_env_walk(ctx, dbi, NODEPGNO(node), flags, deep);
+			if (rc)
+				return rc;
+			continue;
+		}
+
+		assert(IS_LEAF(mp));
+		if (node->mn_ksize < 1)
+			return MDB_CORRUPTED;
+		if (node->mn_flags & F_BIGDATA) {
+			MDB_page *omp;
+			pgno_t *opg;
+			size_t over_header, over_payload, over_unused;
+
+			payload_size += sizeof(pgno_t);
+			opg = NODEDATA(node);
+			rc = mdb_page_get(ctx->mw_txn, *opg, &omp, NULL);
+			if (rc)
+				return rc;
+			if (*opg != omp->mp_p.p_pgno)
+				return MDB_CORRUPTED;
+			/* LY: Don't use mask here, e.g bitwise (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP).
+			 * Pages should not me marked dirty/loose or otherwise. */
+			if (P_OVERFLOW != omp->mp_flags)
+				return MDB_CORRUPTED;
+
+			over_header = PAGEHDRSZ;
+			over_payload = NODEDSZ(node);
+			over_unused = omp->mp_pages * ctx->mw_txn->mt_env->me_psize
+					- over_payload - over_header;
+
+			rc = ctx->mw_visitor(*opg, omp->mp_pages, ctx->mw_user, dbi, "overflow-data",
+					over_payload, over_header, over_unused);
+			if (rc)
+				return rc;
+			continue;
+		}
+
+		payload_size += NODEDSZ(node);
+		if (node->mn_flags & F_SUBDATA) {
+			MDB_db *db = NODEDATA(node);
+			char* name = NULL;
+
+			if (NODEDSZ(node) < 1)
+				return MDB_CORRUPTED;
+			if (! (node->mn_flags & F_DUPDATA)) {
+				name = NODEKEY(node);
+				int namelen = (char*) db - name;
+				name = memcpy(alloca(namelen + 1), name, namelen);
+				name[namelen] = 0;
+			}
+			rc = mdb_env_walk(ctx, (name && name[0]) ? name : dbi,
+					db->md_root, node->mn_flags & F_DUPDATA, deep + 1);
+			if (rc)
+				return rc;
+		}
 	}
-	return rc;
+
+	return ctx->mw_visitor(mp->mp_p.p_pgno, 1, ctx->mw_user, dbi,
+		type, payload_size, header_size, unused_size + align_bytes);
 }
 
 int ESECT
-mdb_env_pgwalk(MDB_txn *txn, MDB_pgwalk_func* visitor, void* user)
+mdb_env_pgwalk(MDB_txn *txn, MDB_pgvisitor_func* visitor, void* user)
 {
 	mdb_walk_ctx_t ctx;
 	int rc;
@@ -9950,13 +9936,14 @@ mdb_env_pgwalk(MDB_txn *txn, MDB_pgwalk_func* visitor, void* user)
 	ctx.mw_user = user;
 	ctx.mw_visitor = visitor;
 
-	rc = visitor(0, 2, user, "meta", 'M', sizeof(MDB_meta), PAGEHDRSZ);
+	rc = visitor(0, 2, user, "lmdb", "meta", sizeof(MDB_meta)*2, PAGEHDRSZ*2,
-	if (! rc)
+				 (txn->mt_env->me_psize - sizeof(MDB_meta) - PAGEHDRSZ) *2);
+	if (! rc && txn->mt_dbs[FREE_DBI].md_root != P_INVALID)
 		rc = mdb_env_walk(&ctx, "free", txn->mt_dbs[FREE_DBI].md_root, 0, 0);
-	if (! rc)
+	if (! rc && txn->mt_dbs[MAIN_DBI].md_root != P_INVALID)
 		rc = mdb_env_walk(&ctx, "main", txn->mt_dbs[MAIN_DBI].md_root, 0, 0);
 	if (! rc)
-		rc = visitor(P_INVALID, 0, user, NULL, 0, -1, 0);
+		rc = visitor(P_INVALID, 0, user, NULL, NULL, -1, 0, 0);
 	return rc;
 }
 

mdb_chk.c

@@ -134,6 +134,10 @@ static int pagemap_lookup_dbi(const char* dbi) {
 		return last = -1;
 
 	walk.dbi_names[last] = strdup(dbi);
+
+	if (verbose > 2)
+		print(" - found '%s' area\n", dbi);
+
 	return last;
 }
 
@@ -201,39 +205,55 @@ static size_t problems_pop(struct problem* list) {
 }
 
 static int pgvisitor(size_t pgno, unsigned pgnumber, void* ctx, const char* dbi,
-					 char type, int payload_bytes, int header_bytes)
+	const char* type, int payload_bytes, int header_bytes, int unused_bytes)
 {
-	if (pgnumber) {
+	if (type) {
+		size_t page_bytes = payload_bytes + header_bytes + unused_bytes;
+		size_t page_size = pgnumber * stat.ms_psize;
 		int index = pagemap_lookup_dbi(dbi);
 		if (index < 0)
 			return ENOMEM;
 
+		if (verbose > 3) {
+			print((pgnumber < 2) ? "     %s-page %zu" : "     %s-span %zu..%zu (%u pages)",
+				type, pgno, pgno + pgnumber - 1, pgnumber);
+			print(" of %s: header %i, payload %i, unused %i\n",
+				dbi, header_bytes, payload_bytes, unused_bytes);
+		}
+
 		walk.pgcount += pgnumber;
 
+		if (unused_bytes < 0 || (size_t) unused_bytes > page_size)
+			problem_add(pgno, "illegal unused-bytes", "(%zu < %i < %zu)",
+				0, unused_bytes, stat.ms_psize);
+
 		if (header_bytes < sizeof(long) || header_bytes >= stat.ms_psize - sizeof(long))
-			problem_add(pgno, "wrong header-length", "(%zu < %i < %zu)",
+			problem_add(pgno, "illegal header-length", "(%zu < %i < %zu)",
-				sizeof(long), header_bytes, header_bytes >= stat.ms_psize - sizeof(long));
+				sizeof(long), header_bytes, stat.ms_psize - sizeof(long));
 		else if (payload_bytes < 1)
-			problem_add(pgno, "empty page",  "(payload %zu bytes)", payload_bytes);
+			problem_add(pgno, "empty page", "(payload %i bytes)", payload_bytes);
-		else if (payload_bytes + header_bytes > pgnumber * stat.ms_psize)
+
-			problem_add(pgno, "overflowed page",  "(%zu + %zu > %zu)",
+		if (page_bytes != page_size)
-						payload_bytes, header_bytes, pgnumber * stat.ms_psize);
+			problem_add(pgno, "misused page",  "(%zu != %zu (%ih + %ip + %iu))",
+				page_size, page_bytes, header_bytes, payload_bytes, unused_bytes);
 		else {
 			walk.dbi_payload_bytes[index] += payload_bytes + header_bytes;
 			walk.total_payload_bytes += payload_bytes + header_bytes;
 		}
 
-		do {
+		if (pgnumber) {
-			if (pgno >= lastpgno)
+			do {
-				problem_add(pgno, "wrong page-no", "(> %zi)", lastpgno);
+				if (pgno >= lastpgno)
-			else if (walk.pagemap[pgno])
+					problem_add(pgno, "wrong page-no", "(> %zi)", lastpgno);
-				problem_add(pgno, "page already used", "(in %s)", walk.dbi_names[walk.pagemap[pgno]]);
+				else if (walk.pagemap[pgno])
-			else {
+					problem_add(pgno, "page already used", "(in %s)", walk.dbi_names[walk.pagemap[pgno]]);
-				walk.pagemap[pgno] = index;
+				else {
-				walk.dbi_pages[index] += 1;
+					walk.pagemap[pgno] = index;
-			}
+					walk.dbi_pages[index] += 1;
-			++pgno;
+				}
-		} while(--pgnumber);
+				++pgno;
+			} while(--pgnumber);
+		}
 	}
 
 	return gotsignal ? EINTR : MDB_SUCCESS;
@@ -285,7 +305,7 @@ static int handle_freedb(size_t record_number, MDB_val *key, MDB_val* data) {
 				for (; i >= span && iptr[i - span] == pg; span++, pg++) ;
 			}
 			if (verbose > 2)
-				print(" - transaction %zu, %zd pages, maxspan %zd%s\n",
+				print("     transaction %zu, %zd pages, maxspan %zd%s\n",
 					*(size_t *)key->mv_data, number, span, bad);
 			if (verbose > 3) {
 				int j = number - 1;
@@ -655,8 +675,11 @@ int main(int argc, char *argv[])
 			  info.me_mapsize / k, sf[i]);
 		if (info.me_mapaddr)
 			print(" - mapaddr %p\n", info.me_mapaddr);
-		print(" - pagesize %u, max keysize %zu, max readers %u\n",
+		print(" - pagesize %u, max keysize %zu (%s), max readers %u\n",
-				  stat.ms_psize, maxkeysize, info.me_maxreaders);
+				stat.ms_psize, maxkeysize,
+				(maxkeysize == 511) ? "default" :
+				(maxkeysize == 0) ? "devel" : "custom",
+				info.me_maxreaders);
 		print(" - transactions: last %zu, bottom %zu, lag reading %zi\n", info.me_last_txnid,
 			  info.me_tail_txnid, info.me_last_txnid - info.me_tail_txnid);
 
@@ -712,6 +735,9 @@ int main(int argc, char *argv[])
 	}
 
 	if (!dont_traversal) {
+		struct problem* saved_list;
+		size_t traversal_problems;
+
 		print("Traversal b-tree...\n");
 		fflush(NULL);
 		walk.pagemap = calloc(lastpgno, sizeof(*walk.pagemap));
@@ -721,7 +747,10 @@ int main(int argc, char *argv[])
 			goto bailout;
 		}
 
+		saved_list = problems_push();
 		rc = mdb_env_pgwalk(txn, pgvisitor, NULL);
+		traversal_problems = problems_pop(saved_list);
+
 		if (rc) {
 			if (rc == EINTR && gotsignal) {
 				print(" - interrupted by signal\n");
@@ -759,7 +788,7 @@ int main(int argc, char *argv[])
 				}
 			}
 			print(" - summary: average fill %.1f%%, %zu problems\n",
-				walk.total_payload_bytes * 100.0 / total_page_bytes, total_problems);
+				walk.total_payload_bytes * 100.0 / total_page_bytes, traversal_problems);
 		}
 	} else if (verbose) {
 		print("Skipping b-tree walk...\n");