MDEV-11336: Enable defragmentation on 10.2 when tests pass

Problem was that we could take page latches on different
order than wat is entitled with SX-lock. To follow the
latching order defined in WL#6326, acquire index->lock X-latch.
This entitles us to acquire page latches in any order for the index.

btr0btr.cc
	Document latch rules before and after MariaDB 10.2.2

sync0rw.cc
	Document latch compatibility rules better.

btr_defragment_merge_pages
	Fix parameter value.

btr_defragment_thread
	Acquire X-lock to dict_index_t::lock before restoring
	cursor position and continuing defragmentation.

ha_innobase::optimize
	Restore defragment feature.

Testing
	Add GIS-index and FT-index to table being defragmented.

	Defragmentation is not done to GIS-indexes and FT auxiliary
	tables.

mysql-test/suite/innodb/disabled.def

@@ -10,14 +10,5 @@
 #
 ##############################################################################
 
-innodb_defragment_fill_factor : MDEV-11336 Fix and enable innodb_defragment
-innodb.defrag_mdl-9155 : MDEV-11336 Fix and enable innodb_defragment
-innodb.innodb_defrag_concurrent : MDEV-11336 Fix and enable innodb_defragment
-innodb.innodb_defrag_stats : MDEV-11336 Fix and enable innodb_defragment
-innodb.innodb_defrag_stats_many_tables : MDEV-11336 Fix and enable innodb_defragment
-innodb.innodb_defragment : MDEV-11336 Fix and enable innodb_defragment
-innodb.innodb_defragment_fill_factor : MDEV-11336 Fix and enable innodb_defragment
-innodb.innodb_defragment_small : MDEV-11336 Fix and enable innodb_defragment
-innodb.innodb_defrag_binlog : MDEV-11336 Fix and enable innodb_defragment
 innodb-wl5980-alter : MDEV-9469 / MDEV-13668 extra crash in 10.2
 create-index-debug : MDEV-13680 InnoDB may crash when btr_page_alloc() fails

mysql-test/suite/innodb/r/innodb_defrag_concurrent.result

@@ -3,7 +3,15 @@ select @@global.innodb_stats_persistent;
 @@global.innodb_stats_persistent
 0
 set global innodb_defragment_stats_accuracy = 80;
-CREATE TABLE t1 (a INT NOT NULL PRIMARY KEY AUTO_INCREMENT, b VARCHAR(256), c INT, KEY second(a, b),KEY third(c)) ENGINE=INNODB;
+CREATE TABLE t1 (a INT NOT NULL PRIMARY KEY AUTO_INCREMENT,
+b VARCHAR(256),
+c INT,
+g GEOMETRY NOT NULL,
+t VARCHAR(256),
+KEY second(a, b),
+KEY third(c),
+SPATIAL gk(g),
+FULLTEXT INDEX fti(t)) ENGINE=INNODB;
 connect  con1,localhost,root,,test,$MASTER_MYPORT,$MASTER_MYSOCK;
 connect  con2,localhost,root,,test,$MASTER_MYPORT,$MASTER_MYSOCK;
 connect  con3,localhost,root,,test,$MASTER_MYPORT,$MASTER_MYSOCK;
@@ -40,9 +48,9 @@ count(stat_value) > 0
 connection con1;
 optimize table t1;;
 connection default;
-INSERT INTO t1 VALUES (400000, REPEAT('A', 256),300000);;
+INSERT INTO t1 VALUES (400000, REPEAT('A', 256),300000, Point(1,1),'More like a test but different.');;
 connection con2;
-INSERT INTO t1 VALUES (500000, REPEAT('A', 256),400000);;
+INSERT INTO t1 VALUES (500000, REPEAT('A', 256),400000, Point(1,1),'Totally different text book.');;
 connection con3;
 DELETE FROM t1 where a between 1 and 100;;
 connection con4;
@@ -59,6 +67,9 @@ disconnect con4;
 optimize table t1;
 Table	Op	Msg_type	Msg_text
 test.t1	optimize	status	OK
+check table t1 extended;
+Table	Op	Msg_type	Msg_text
+test.t1	check	status	OK
 select count(*) from t1;
 count(*)
 15723

mysql-test/suite/innodb/t/innodb_defrag_concurrent.test

@@ -16,7 +16,26 @@ select @@global.innodb_stats_persistent;
 set global innodb_defragment_stats_accuracy = 80;
 
 # Create table.
-CREATE TABLE t1 (a INT NOT NULL PRIMARY KEY AUTO_INCREMENT, b VARCHAR(256), c INT, KEY second(a, b),KEY third(c)) ENGINE=INNODB;
+#
+# TODO: Currently we do not defragment  spatial indexes,
+# because doing it properly would require
+# appropriate logic around the SSN (split
+# sequence number).
+#
+# Also do not defragment auxiliary tables related to FULLTEXT INDEX.
+#
+# Both types added to this test to make sure they do not cause
+# problems.
+#
+CREATE TABLE t1 (a INT NOT NULL PRIMARY KEY AUTO_INCREMENT,
+b VARCHAR(256),
+c INT,
+g GEOMETRY NOT NULL,
+t VARCHAR(256),
+KEY second(a, b),
+KEY third(c),
+SPATIAL gk(g),
+FULLTEXT INDEX fti(t)) ENGINE=INNODB;
 
 connect (con1,localhost,root,,test,$MASTER_MYPORT,$MASTER_MYSOCK);
 connect (con2,localhost,root,,test,$MASTER_MYPORT,$MASTER_MYSOCK);
@@ -36,7 +55,7 @@ let $i = $data_size;
 while ($i)
 {
   eval
-    INSERT INTO t1 VALUES ($data_size + 1 - $i, REPEAT('A', 256), $i);
+    INSERT INTO t1 VALUES ($data_size + 1 - $i, REPEAT('A', 256), $i, Point($i,$i), 'This is a test message.');
   dec $i;
 }
 --enable_query_log
@@ -69,10 +88,10 @@ connection con1;
 --send optimize table t1;
 
 connection default;
---send INSERT INTO t1 VALUES (400000, REPEAT('A', 256),300000);
+--send INSERT INTO t1 VALUES (400000, REPEAT('A', 256),300000, Point(1,1),'More like a test but different.');
 
 connection con2;
---send INSERT INTO t1 VALUES (500000, REPEAT('A', 256),400000);
+--send INSERT INTO t1 VALUES (500000, REPEAT('A', 256),400000, Point(1,1),'Totally different text book.');
 
 connection con3;
 --send DELETE FROM t1 where a between 1 and 100;
@@ -103,6 +122,7 @@ disconnect con3;
 disconnect con4;
 
 optimize table t1;
+check table t1 extended;
 
 select count(*) from t1;
 select count(*) from t1 force index (second);

storage/innobase/btr/btr0btr.cc

@@ -77,22 +77,85 @@ btr_corruption_report(
 /*
 Latching strategy of the InnoDB B-tree
 --------------------------------------
-A tree latch protects all non-leaf nodes of the tree. Each node of a tree
-also has a latch of its own.
 
-A B-tree operation normally first acquires an S-latch on the tree. It
+Node pointer page latches acquisition is protected by index->lock latch.
-searches down the tree and releases the tree latch when it has the
-leaf node latch. To save CPU time we do not acquire any latch on
-non-leaf nodes of the tree during a search, those pages are only bufferfixed.
 
-If an operation needs to restructure the tree, it acquires an X-latch on
+Before MariaDB 10.2.2, all node pointer pages were protected by index->lock
-the tree before searching to a leaf node. If it needs, for example, to
+either in S (shared) or X (exclusive) mode and block->lock was not acquired on
-split a leaf,
+node pointer pages.
-(1) InnoDB decides the split point in the leaf,
+
-(2) allocates a new page,
+After MariaDB 10.2.2, block->lock S-latch or X-latch is used to protect
-(3) inserts the appropriate node pointer to the first non-leaf level,
+node pointer pages and obtaiment of node pointer page latches is protected by
-(4) releases the tree X-latch,
+index->lock.
-(5) and then moves records from the leaf to the new allocated page.
+
+(0) Definition: B-tree level.
+
+(0.1) The leaf pages of the B-tree are at level 0.
+
+(0.2) The parent of a page at level L has level L+1. (The level of the
+root page is equal to the tree height.)
+
+(0.3) The B-tree lock (index->lock) is the parent of the root page and
+has a level = tree height + 1.
+
+Index->lock has 3 possible locking modes:
+
+(1) S-latch:
+
+(1.1) All latches for pages must be obtained in descending order of tree level.
+
+(1.2) Before obtaining the first node pointer page latch at a given B-tree
+level, parent latch must be held (at level +1 ).
+
+(1.3) If a node pointer page is already latched at the same level
+we can only obtain latch to its right sibling page latch at the same level.
+
+(1.4) Release of the node pointer page latches must be done in
+child-to-parent order. (Prevents deadlocks when obtained index->lock
+in SX mode).
+
+(1.4.1) Level L node pointer page latch can be released only when
+no latches at children level i.e. level < L are hold.
+
+(1.4.2) All latches from node pointer pages must be released so
+that no latches are obtained between.
+
+(1.5) [implied by (1.1), (1.2)] Root page latch must be first node pointer
+latch obtained.
+
+(2) SX-latch:
+
+In this case rules (1.2) and (1.3) from S-latch case are relaxed and
+merged into (2.2) and rule (1.4) is removed. Thus, latch acquisition
+can be skipped at some tree levels and latches can be obtained in
+a less restricted order.
+
+(2.1) [identical to (1.1)]: All latches for pages must be obtained in descending
+order of tree level.
+
+(2.2) When a node pointer latch at level L is obtained,
+the left sibling page latch in the same level or some ancestor
+page latch (at level > L) must be hold.
+
+(2.3) [implied by (2.1), (2.2)] The first node pointer page latch obtained can
+be any node pointer page.
+
+(3) X-latch:
+
+Node pointer latches can be obtained in any order.
+
+NOTE: New rules after MariaDB 10.2.2 does not affect the latching rules of leaf pages:
+
+index->lock S-latch is needed in read for the node pointer traversal. When the leaf
+level is reached, index-lock can be released (and with the MariaDB 10.2.2 changes, all
+node pointer latches). Left to right index travelsal in leaf page level can be safely done
+by obtaining right sibling leaf page latch and then releasing the old page latch.
+
+Single leaf page modifications (BTR_MODIFY_LEAF) are protected by index->lock
+S-latch.
+
+B-tree operations involving page splits or merges (BTR_MODIFY_TREE) and page
+allocations are protected by index->lock X-latch.
 
 Node pointers
 -------------

storage/innobase/btr/btr0defragment.cc

@@ -564,7 +564,7 @@ btr_defragment_merge_pages(
 				page_get_infimum_rec(from_page));
 			node_ptr = dict_index_build_node_ptr(
 				index, rec, page_get_page_no(from_page),
-				heap, level + 1);
+				heap, level);
 			btr_insert_on_non_leaf_level(0, index, level+1,
 						     node_ptr, mtr);
 		}
@@ -797,11 +797,16 @@ DECLARE_THREAD(btr_defragment_thread)(void*)
 
 		now = ut_timer_now();
 		mtr_start(&mtr);
-		btr_pcur_restore_position(BTR_MODIFY_TREE, pcur, &mtr);
 		cursor = btr_pcur_get_btr_cur(pcur);
 		index = btr_cur_get_index(cursor);
-		first_block = btr_cur_get_block(cursor);
 		mtr.set_named_space(index->space);
+		/* To follow the latching order defined in WL#6326, acquire index->lock X-latch.
+		This entitles us to acquire page latches in any order for the index. */
+		mtr_x_lock(&index->lock, &mtr);
+		/* This will acquire index->lock SX-latch, which per WL#6363 is allowed
+		when we are already holding the X-latch. */
+		btr_pcur_restore_position(BTR_MODIFY_TREE, pcur, &mtr);
+		first_block = btr_cur_get_block(cursor);
 
 		last_block = btr_defragment_n_pages(first_block, index,
 						    srv_defragment_n_pages,

storage/innobase/handler/ha_innodb.cc

@@ -15066,7 +15066,7 @@ ha_innobase::optimize(
 	calls to OPTIMIZE, which is undesirable. */
 
 	/* TODO: Defragment is disabled for now */
-	if (0) {
+	if (srv_defragment) {
 		int err;
 
 		err = defragment_table(m_prebuilt->table->name.m_name, NULL, false);

storage/innobase/sync/sync0rw.cc

@@ -84,10 +84,15 @@ lock_word < -(X_LOCK_DECR + X_LOCK_HALF_DECR):
 				2 - (lock_word + X_LOCK_DECR + X_LOCK_HALF_DECR)
 
  LOCK COMPATIBILITY MATRIX
-    S SX  X
+
- S  +  +  -
+      | S|SX| X|
- SX +  -  -
+    --+--+--+--+
- X  -  -  -
+     S| +| +| -|
+    --+--+--+--+
+    SX| +| -| -|
+    --+--+--+--+
+     X| -| -| -|
+    --+--+--+--+
 
 The lock_word is always read and updated atomically and consistently, so that
 it always represents the state of the lock, and the state of the lock changes