MDEV-23991 dict_table_stats_lock() has unnecessarily long scope

Patch removes dict_index_t::stats_latch. Table/index statistics now
protected with dict_sys->mutex. That way statistics computation can
happen in parallel in several threads and dict_sys->mutex will be locked
only for a short period of time.

This patch is a joint work with Marko Mäkelä

dict_index_t:🔒 make mutable which allows to pass const pointer
when only lock is touched in an object

btr_height_get()
btr_get_size(): make index argument const for better type safety

btr_estimate_number_of_different_key_vals(): now returns computed values
instead of setting fields in dict_index_t directly

remove everything related to dict_index_t::stats_latch

dict_stats_index_set_n_diff(): now returns computed values instead
of setting fields in dict_index_t directly

dict_stats_analyze_index():  now returns computed values instead
of setting fields in dict_index_t directly

Reviewed by: Marko Mäkelä

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

@@ -0,0 +1,18 @@
+#
+# MDEV-23991 dict_table_stats_lock() has unnecessarily long scope
+#
+CREATE TABLE t1(a INT) ENGINE=INNODB STATS_PERSISTENT=1;
+SET DEBUG_SYNC='dict_stats_update_persistent SIGNAL stop WAIT_FOR go';
+ANALYZE TABLE t1;
+connect con1, localhost, root;
+SET DEBUG_SYNC='now WAIT_FOR stop';
+SELECT ENGINE,SUM(DATA_LENGTH+INDEX_LENGTH),COUNT(ENGINE),SUM(DATA_LENGTH),SUM(INDEX_LENGTH) FROM information_schema.TABLES WHERE ENGINE='InnoDB';
+ENGINE	SUM(DATA_LENGTH+INDEX_LENGTH)	COUNT(ENGINE)	SUM(DATA_LENGTH)	SUM(INDEX_LENGTH)
+InnoDB	49152	3	49152	0
+SET DEBUG_SYNC='now SIGNAL go';
+disconnect con1;
+connection default;
+Table	Op	Msg_type	Msg_text
+test.t1	analyze	status	OK
+SET DEBUG_SYNC= 'RESET';
+DROP TABLE t1;

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

@@ -0,0 +1,27 @@
+--source include/have_innodb.inc
+--source include/have_debug.inc
+--source include/have_debug_sync.inc
+--source include/count_sessions.inc
+
+--echo #
+--echo # MDEV-23991 dict_table_stats_lock() has unnecessarily long scope
+--echo #
+CREATE TABLE t1(a INT) ENGINE=INNODB STATS_PERSISTENT=1;
+
+SET DEBUG_SYNC='dict_stats_update_persistent SIGNAL stop WAIT_FOR go';
+--send ANALYZE TABLE t1
+
+--connect(con1, localhost, root)
+SET DEBUG_SYNC='now WAIT_FOR stop';
+
+SELECT ENGINE,SUM(DATA_LENGTH+INDEX_LENGTH),COUNT(ENGINE),SUM(DATA_LENGTH),SUM(INDEX_LENGTH) FROM information_schema.TABLES WHERE ENGINE='InnoDB';
+
+SET DEBUG_SYNC='now SIGNAL go';
+--disconnect con1
+
+--connection default
+--reap
+SET DEBUG_SYNC= 'RESET';
+DROP TABLE t1;
+
+--source include/wait_until_count_sessions.inc

storage/innobase/btr/btr0btr.cc

@@ -283,7 +283,7 @@ the index.
 ulint
 btr_height_get(
 /*===========*/
-	dict_index_t*	index,	/*!< in: index tree */
+	const dict_index_t*	index,	/*!< in: index tree */
 	mtr_t*		mtr)	/*!< in/out: mini-transaction */
 {
 	ulint		height=0;
@@ -592,7 +592,7 @@ Gets the number of pages in a B-tree.
 ulint
 btr_get_size(
 /*=========*/
-	dict_index_t*	index,	/*!< in: index */
+	const dict_index_t*	index,	/*!< in: index */
 	ulint		flag,	/*!< in: BTR_N_LEAF_PAGES or BTR_TOTAL_SIZE */
 	mtr_t*		mtr)	/*!< in/out: mini-transaction where index
 				is s-latched */

storage/innobase/btr/btr0cur.cc

@@ -6133,21 +6133,19 @@ btr_record_not_null_field_in_rec(
 	}
 }
 
-/*******************************************************************//**
-Estimates the number of different key values in a given index, for
+/** Estimates the number of different key values in a given index, for
 each n-column prefix of the index where 1 <= n <= dict_index_get_n_unique(index).
 The estimates are stored in the array index->stat_n_diff_key_vals[] (indexed
 0..n_uniq-1) and the number of pages that were sampled is saved in
-index->stat_n_sample_sizes[].
+result.n_sample_sizes[].
 If innodb_stats_method is nulls_ignored, we also record the number of
 non-null values for each prefix and stored the estimates in
-array index->stat_n_non_null_key_vals.
-@return true if the index is available and we get the estimated numbers,
-false if the index is unavailable. */
-bool
-btr_estimate_number_of_different_key_vals(
-/*======================================*/
-	dict_index_t*	index)	/*!< in: index */
+array result.n_non_null_key_vals.
+@param[in]	index	index
+@return vector with statistics information
+empty vector if the index is unavailable. */
+std::vector<index_field_stats_t>
+btr_estimate_number_of_different_key_vals(dict_index_t* index)
 {
 	btr_cur_t	cursor;
 	page_t*		page;
@@ -6167,11 +6165,11 @@ btr_estimate_number_of_different_key_vals(
 	rec_offs*	offsets_rec	= NULL;
 	rec_offs*	offsets_next_rec = NULL;
 
+	std::vector<index_field_stats_t> result;
+
 	/* For spatial index, there is no such stats can be
 	fetched. */
-	if (dict_index_is_spatial(index)) {
-		return(false);
-	}
+	ut_ad(!dict_index_is_spatial(index));
 
 	n_cols = dict_index_get_n_unique(index);
 
@@ -6280,7 +6278,7 @@ btr_estimate_number_of_different_key_vals(
 			mtr_commit(&mtr);
 			mem_heap_free(heap);
 
-			return(false);
+			return result;
 		}
 
 		/* Count the number of different key values for each prefix of
@@ -6386,8 +6384,12 @@ exit_loop:
 	also the pages used for external storage of fields (those pages are
 	included in index->stat_n_leaf_pages) */
 
+	result.reserve(n_cols);
+
 	for (j = 0; j < n_cols; j++) {
-		index->stat_n_diff_key_vals[j]
+		index_field_stats_t stat;
+
+		stat.n_diff_key_vals
 			= BTR_TABLE_STATS_FROM_SAMPLE(
 				n_diff[j], index, n_sample_pages,
 				total_external_size, not_empty_flag);
@@ -6408,25 +6410,23 @@ exit_loop:
 			add_on = n_sample_pages;
 		}
 
-		index->stat_n_diff_key_vals[j] += add_on;
+		stat.n_diff_key_vals += add_on;
 
-		index->stat_n_sample_sizes[j] = n_sample_pages;
+		stat.n_sample_sizes = n_sample_pages;
 
-		/* Update the stat_n_non_null_key_vals[] with our
-		sampled result. stat_n_non_null_key_vals[] is created
-		and initialized to zero in dict_index_add_to_cache(),
-		along with stat_n_diff_key_vals[] array */
 		if (n_not_null != NULL) {
-			index->stat_n_non_null_key_vals[j] =
+			stat.n_non_null_key_vals =
 				 BTR_TABLE_STATS_FROM_SAMPLE(
 					n_not_null[j], index, n_sample_pages,
 					total_external_size, not_empty_flag);
 		}
+
+		result.push_back(stat);
 	}
 
 	mem_heap_free(heap);
 
-	return(true);
+	return result;
 }
 
 /*================== EXTERNAL STORAGE OF BIG FIELDS ===================*/

storage/innobase/dict/dict0dict.cc

@@ -267,56 +267,6 @@ dict_mutex_exit_for_mysql(void)
 	mutex_exit(&dict_sys->mutex);
 }
 
-/** Lock the appropriate latch to protect a given table's statistics.
-@param[in]	table		table whose stats to lock
-@param[in]	latch_mode	RW_S_LATCH or RW_X_LATCH */
-void
-dict_table_stats_lock(
-	dict_table_t*	table,
-	ulint		latch_mode)
-{
-	ut_ad(table != NULL);
-	ut_ad(table->magic_n == DICT_TABLE_MAGIC_N);
-
-	switch (latch_mode) {
-	case RW_S_LATCH:
-		rw_lock_s_lock(&table->stats_latch);
-		break;
-	case RW_X_LATCH:
-		rw_lock_x_lock(&table->stats_latch);
-		break;
-	case RW_NO_LATCH:
-		/* fall through */
-	default:
-		ut_error;
-	}
-}
-
-/** Unlock the latch that has been locked by dict_table_stats_lock().
-@param[in]	table		table whose stats to unlock
-@param[in]	latch_mode	RW_S_LATCH or RW_X_LATCH */
-void
-dict_table_stats_unlock(
-	dict_table_t*	table,
-	ulint		latch_mode)
-{
-	ut_ad(table != NULL);
-	ut_ad(table->magic_n == DICT_TABLE_MAGIC_N);
-
-	switch (latch_mode) {
-	case RW_S_LATCH:
-		rw_lock_s_unlock(&table->stats_latch);
-		break;
-	case RW_X_LATCH:
-		rw_lock_x_unlock(&table->stats_latch);
-		break;
-	case RW_NO_LATCH:
-		/* fall through */
-	default:
-		ut_error;
-	}
-}
-
 /**********************************************************************//**
 Try to drop any indexes after an aborted index creation.
 This can also be after a server kill during DROP INDEX. */

storage/innobase/dict/dict0mem.cc

@@ -125,8 +125,7 @@ dict_mem_table_create(
 	ulint		n_cols,
 	ulint		n_v_cols,
 	ulint		flags,
-	ulint		flags2,
-	bool		init_stats_latch)
+	ulint		flags2)
 {
 	dict_table_t*	table;
 	mem_heap_t*	heap;
@@ -182,12 +181,6 @@ dict_mem_table_create(
 	new(&table->foreign_set) dict_foreign_set();
 	new(&table->referenced_set) dict_foreign_set();
 
-	if (init_stats_latch) {
-		rw_lock_create(dict_table_stats_key, &table->stats_latch,
-			       SYNC_INDEX_TREE);
-		table->stats_latch_inited = true;
-	}
-
 	return(table);
 }
 
@@ -237,10 +230,6 @@ dict_mem_table_free(
 		UT_DELETE(table->s_cols);
 	}
 
-	if (table->stats_latch_inited) {
-		rw_lock_free(&table->stats_latch);
-	}
-
 	mem_heap_free(table->heap);
 }
 

storage/innobase/dict/dict0stats.cc

@@ -486,8 +486,6 @@ dict_stats_table_clone_create(
 
 	ut_d(t->magic_n = DICT_TABLE_MAGIC_N);
 
-	rw_lock_create(dict_table_stats_key, &t->stats_latch, SYNC_INDEX_TREE);
-
 	return(t);
 }
 
@@ -500,15 +498,12 @@ dict_stats_table_clone_free(
 /*========================*/
 	dict_table_t*	t)	/*!< in: dummy table object to free */
 {
-	rw_lock_free(&t->stats_latch);
 	mem_heap_free(t->heap);
 }
 
 /*********************************************************************//**
 Write all zeros (or 1 where it makes sense) into an index
-statistics members. The resulting stats correspond to an empty index.
-The caller must own index's table stats latch in X mode
-(dict_table_stats_lock(table, RW_X_LATCH)) */
+statistics members. The resulting stats correspond to an empty index. */
 static
 void
 dict_stats_empty_index(
@@ -519,6 +514,7 @@ dict_stats_empty_index(
 {
 	ut_ad(!(index->type & DICT_FTS));
 	ut_ad(!dict_index_is_ibuf(index));
+	ut_ad(mutex_own(&dict_sys->mutex));
 
 	ulint	n_uniq = index->n_uniq;
 
@@ -548,10 +544,9 @@ dict_stats_empty_table(
 	bool		empty_defrag_stats)
 				/*!< in: whether to empty defrag stats */
 {
+	mutex_enter(&dict_sys->mutex);
+
 	/* Zero the stats members */
-
-	dict_table_stats_lock(table, RW_X_LATCH);
-
 	table->stat_n_rows = 0;
 	table->stat_clustered_index_size = 1;
 	/* 1 page for each index, not counting the clustered */
@@ -575,8 +570,7 @@ dict_stats_empty_table(
 	}
 
 	table->stat_initialized = TRUE;
-
-	dict_table_stats_unlock(table, RW_X_LATCH);
+	mutex_exit(&dict_sys->mutex);
 }
 
 /*********************************************************************//**
@@ -675,6 +669,8 @@ dict_stats_copy(
                                              to have the same statistics as if
                                              the table was empty */
 {
+	ut_ad(mutex_own(&dict_sys->mutex));
+
 	dst->stats_last_recalc = src->stats_last_recalc;
 	dst->stat_n_rows = src->stat_n_rows;
 	dst->stat_clustered_index_size = src->stat_clustered_index_size;
@@ -792,8 +788,6 @@ dict_stats_snapshot_create(
 {
 	mutex_enter(&dict_sys->mutex);
 
-	dict_table_stats_lock(table, RW_S_LATCH);
-
 	dict_stats_assert_initialized(table);
 
 	dict_table_t*	t;
@@ -807,8 +801,6 @@ dict_stats_snapshot_create(
 	t->stats_sample_pages = table->stats_sample_pages;
 	t->stats_bg_flag = table->stats_bg_flag;
 
-	dict_table_stats_unlock(table, RW_S_LATCH);
-
 	mutex_exit(&dict_sys->mutex);
 
 	return(t);
@@ -848,10 +840,14 @@ dict_stats_update_transient_for_index(
 		Initialize some bogus index cardinality
 		statistics, so that the data can be queried in
 		various means, also via secondary indexes. */
+		mutex_enter(&dict_sys->mutex);
 		dict_stats_empty_index(index, false);
+		mutex_exit(&dict_sys->mutex);
 #if defined UNIV_DEBUG || defined UNIV_IBUF_DEBUG
 	} else if (ibuf_debug && !dict_index_is_clust(index)) {
+		mutex_enter(&dict_sys->mutex);
 		dict_stats_empty_index(index, false);
+		mutex_exit(&dict_sys->mutex);
 #endif /* UNIV_DEBUG || UNIV_IBUF_DEBUG */
 	} else {
 		mtr_t	mtr;
@@ -874,7 +870,9 @@ dict_stats_update_transient_for_index(
 
 		switch (size) {
 		case ULINT_UNDEFINED:
+			mutex_enter(&dict_sys->mutex);
 			dict_stats_empty_index(index, false);
+			mutex_exit(&dict_sys->mutex);
 			return;
 		case 0:
 			/* The root node of the tree is a leaf */
@@ -886,11 +884,23 @@ dict_stats_update_transient_for_index(
 		/* Do not continue if table decryption has failed or
 		table is already marked as corrupted. */
 		if (index->is_readable()) {
-			/* We don't handle the return value since it
-			will be false only when some thread is
-			dropping the table and we don't have to empty
-			the statistics of the to be dropped index */
-			btr_estimate_number_of_different_key_vals(index);
+			std::vector<index_field_stats_t> stats
+				= btr_estimate_number_of_different_key_vals(
+					index);
+
+			if (!stats.empty()) {
+				ut_ad(!mutex_own(&dict_sys->mutex));
+				mutex_enter(&dict_sys->mutex);
+				for (size_t i = 0; i < stats.size(); ++i) {
+					index->stat_n_diff_key_vals[i]
+						= stats[i].n_diff_key_vals;
+					index->stat_n_sample_sizes[i]
+						= stats[i].n_sample_sizes;
+					index->stat_n_non_null_key_vals[i]
+						= stats[i].n_non_null_key_vals;
+				}
+				mutex_exit(&dict_sys->mutex);
+			}
 		}
 	}
 }
@@ -907,6 +917,8 @@ dict_stats_update_transient(
 /*========================*/
 	dict_table_t*	table)	/*!< in/out: table */
 {
+	ut_ad(!mutex_own(&dict_sys->mutex));
+
 	dict_index_t*	index;
 	ulint		sum_of_index_sizes	= 0;
 
@@ -932,27 +944,25 @@ dict_stats_update_transient(
 
 		ut_ad(!dict_index_is_ibuf(index));
 
-		if (index->type & DICT_FTS || dict_index_is_spatial(index)) {
+		if (index->type & (DICT_FTS | DICT_SPATIAL)) {
 			continue;
 		}
 
+		if (dict_stats_should_ignore_index(index)
+		    || !index->is_readable()) {
+			mutex_enter(&dict_sys->mutex);
 			dict_stats_empty_index(index, false);
-
-		if (dict_stats_should_ignore_index(index)) {
+			mutex_exit(&dict_sys->mutex);
 			continue;
 		}
 
-		/* Do not continue if table decryption has failed or
-		table is already marked as corrupted. */
-		if (!index->is_readable()) {
-			break;
-		}
-
 		dict_stats_update_transient_for_index(index);
 
 		sum_of_index_sizes += index->stat_index_size;
 	}
 
+	mutex_enter(&dict_sys->mutex);
+
 	index = dict_table_get_first_index(table);
 
 	table->stat_n_rows = index->stat_n_diff_key_vals[
@@ -968,6 +978,8 @@ dict_stats_update_transient(
 	table->stat_modified_counter = 0;
 
 	table->stat_initialized = TRUE;
+
+	mutex_exit(&dict_sys->mutex);
 }
 
 /* @{ Pseudo code about the relation between the following functions
@@ -1808,16 +1820,31 @@ dict_stats_analyze_index_for_n_prefix(
 	btr_pcur_close(&pcur);
 }
 
+/** statistics for an index */
+struct index_stats_t
+{
+  std::vector<index_field_stats_t> stats;
+  ulint index_size;
+  ulint n_leaf_pages;
+
+  index_stats_t(ulint n_uniq) : index_size(1), n_leaf_pages(1)
+  {
+    stats.reserve(n_uniq);
+    for (ulint i= 0; i < n_uniq; ++i)
+      stats.push_back(index_field_stats_t(0, 1, 0));
+  }
+};
+
 /** Set dict_index_t::stat_n_diff_key_vals[] and stat_n_sample_sizes[].
 @param[in]	n_diff_data	input data to use to derive the results
-@param[in,out]	index		index whose stat_n_diff_key_vals[] to set */
+@param[in,out]	index_stats	index stats to set */
 UNIV_INLINE
 void
 dict_stats_index_set_n_diff(
 	const n_diff_data_t*	n_diff_data,
-	dict_index_t*		index)
+	index_stats_t&		index_stats)
 {
-	for (ulint n_prefix = dict_index_get_n_unique(index);
+	for (ulint n_prefix = index_stats.stats.size();
 	     n_prefix >= 1;
 	     n_prefix--) {
 		/* n_diff_all_analyzed_pages can be 0 here if
@@ -1848,14 +1875,14 @@ dict_stats_index_set_n_diff(
 			that the total number of ordinary leaf pages is
 			T * D / (D + E). */
 			n_ordinary_leaf_pages
-				= index->stat_n_leaf_pages
+				= index_stats.n_leaf_pages
 				* data->n_leaf_pages_to_analyze
 				/ (data->n_leaf_pages_to_analyze
 				   + data->n_external_pages_sum);
 		}
 
 		/* See REF01 for an explanation of the algorithm */
-		index->stat_n_diff_key_vals[n_prefix - 1]
+		index_stats.stats[n_prefix - 1].n_diff_key_vals
 			= n_ordinary_leaf_pages
 
 			* data->n_diff_on_level
@@ -1864,7 +1891,7 @@ dict_stats_index_set_n_diff(
 			* data->n_diff_all_analyzed_pages
 			/ data->n_leaf_pages_to_analyze;
 
-		index->stat_n_sample_sizes[n_prefix - 1]
+		index_stats.stats[n_prefix - 1].n_sample_sizes
 			= data->n_leaf_pages_to_analyze;
 
 		DEBUG_PRINTF("    %s(): n_diff=" UINT64PF
@@ -1873,9 +1900,9 @@ dict_stats_index_set_n_diff(
 			     " * " UINT64PF " / " UINT64PF
 			     " * " UINT64PF " / " UINT64PF ")\n",
 			     __func__,
-			     index->stat_n_diff_key_vals[n_prefix - 1],
+			     index_stats.stats[n_prefix - 1].n_diff_key_vals,
 			     n_prefix,
-			     index->stat_n_leaf_pages,
+			     index_stats.n_leaf_pages,
 			     data->n_diff_on_level,
 			     data->n_recs_on_level,
 			     data->n_diff_all_analyzed_pages,
@@ -1883,15 +1910,12 @@ dict_stats_index_set_n_diff(
 	}
 }
 
-/*********************************************************************//**
-Calculates new statistics for a given index and saves them to the index
+/** Calculates new statistics for a given index and saves them to the index
 members stat_n_diff_key_vals[], stat_n_sample_sizes[], stat_index_size and
-stat_n_leaf_pages. This function could be slow. */
-static
-void
-dict_stats_analyze_index(
-/*=====================*/
-	dict_index_t*	index)	/*!< in/out: index to analyze */
+stat_n_leaf_pages. This function can be slow.
+@param[in]	index	index to analyze
+@return index stats */
+static index_stats_t dict_stats_analyze_index(dict_index_t* index)
 {
 	ulint		root_level;
 	ulint		level;
@@ -1902,20 +1926,22 @@ dict_stats_analyze_index(
 	ib_uint64_t	total_pages;
 	mtr_t		mtr;
 	ulint		size;
+	index_stats_t	result(index->n_uniq);
 	DBUG_ENTER("dict_stats_analyze_index");
 
 	DBUG_PRINT("info", ("index: %s, online status: %d", index->name(),
 			    dict_index_get_online_status(index)));
 
+	ut_ad(!mutex_own(&dict_sys->mutex)); // because this function is slow
+	ut_ad(index->table->get_ref_count());
+
 	/* Disable update statistic for Rtree */
 	if (dict_index_is_spatial(index)) {
-		DBUG_VOID_RETURN;
+		DBUG_RETURN(result);
 	}
 
 	DEBUG_PRINTF("  %s(index=%s)\n", __func__, index->name());
 
-	dict_stats_empty_index(index, false);
-
 	mtr_start(&mtr);
 
 	mtr_s_lock(dict_index_get_lock(index), &mtr);
@@ -1923,7 +1949,7 @@ dict_stats_analyze_index(
 	size = btr_get_size(index, BTR_TOTAL_SIZE, &mtr);
 
 	if (size != ULINT_UNDEFINED) {
-		index->stat_index_size = size;
+		result.index_size = size;
 		size = btr_get_size(index, BTR_N_LEAF_PAGES, &mtr);
 	}
 
@@ -1933,13 +1959,13 @@ dict_stats_analyze_index(
 	switch (size) {
 	case ULINT_UNDEFINED:
 		dict_stats_assert_initialized_index(index);
-		DBUG_VOID_RETURN;
+		DBUG_RETURN(result);
 	case 0:
 		/* The root node of the tree is a leaf */
 		size = 1;
 	}
 
-	index->stat_n_leaf_pages = size;
+	result.n_leaf_pages = size;
 
 	mtr_start(&mtr);
 
@@ -1980,14 +2006,18 @@ dict_stats_analyze_index(
 					       NULL /* boundaries not needed */,
 					       &mtr);
 
-		for (ulint i = 0; i < n_uniq; i++) {
-			index->stat_n_sample_sizes[i] = total_pages;
-		}
-
 		mtr_commit(&mtr);
 
-		dict_stats_assert_initialized_index(index);
-		DBUG_VOID_RETURN;
+		mutex_enter(&dict_sys->mutex);
+		for (ulint i = 0; i < n_uniq; i++) {
+			result.stats[i].n_diff_key_vals = index->stat_n_diff_key_vals[i];
+			result.stats[i].n_sample_sizes = total_pages;
+			result.stats[i].n_non_null_key_vals = index->stat_n_non_null_key_vals[i];
+		}
+		result.n_leaf_pages = index->stat_n_leaf_pages;
+		mutex_exit(&dict_sys->mutex);
+
+		DBUG_RETURN(result);
 	}
 
 	/* For each level that is being scanned in the btree, this contains the
@@ -2179,13 +2209,12 @@ found_level:
 	/* n_prefix == 0 means that the above loop did not end up prematurely
 	due to tree being changed and so n_diff_data[] is set up. */
 	if (n_prefix == 0) {
-		dict_stats_index_set_n_diff(n_diff_data, index);
+		dict_stats_index_set_n_diff(n_diff_data, result);
 	}
 
 	UT_DELETE_ARRAY(n_diff_data);
 
-	dict_stats_assert_initialized_index(index);
-	DBUG_VOID_RETURN;
+	DBUG_RETURN(result);
 }
 
 /*********************************************************************//**
@@ -2203,7 +2232,7 @@ dict_stats_update_persistent(
 
 	DEBUG_PRINTF("%s(table=%s)\n", __func__, table->name);
 
-	dict_table_stats_lock(table, RW_X_LATCH);
+	DEBUG_SYNC_C("dict_stats_update_persistent");
 
 	/* analyze the clustered index first */
 
@@ -2214,7 +2243,6 @@ dict_stats_update_persistent(
 	    || (index->type | DICT_UNIQUE) != (DICT_CLUSTERED | DICT_UNIQUE)) {
 
 		/* Table definition is corrupt */
-		dict_table_stats_unlock(table, RW_X_LATCH);
 		dict_stats_empty_table(table, true);
 
 		return(DB_CORRUPTION);
@@ -2222,7 +2250,16 @@ dict_stats_update_persistent(
 
 	ut_ad(!dict_index_is_ibuf(index));
 
-	dict_stats_analyze_index(index);
+	index_stats_t stats = dict_stats_analyze_index(index);
+
+	mutex_enter(&dict_sys->mutex);
+	index->stat_index_size = stats.index_size;
+	index->stat_n_leaf_pages = stats.n_leaf_pages;
+	for (size_t i = 0; i < stats.stats.size(); ++i) {
+		index->stat_n_diff_key_vals[i] = stats.stats[i].n_diff_key_vals;
+		index->stat_n_sample_sizes[i] = stats.stats[i].n_sample_sizes;
+		index->stat_n_non_null_key_vals[i] = stats.stats[i].n_non_null_key_vals;
+	}
 
 	ulint	n_unique = dict_index_get_n_unique(index);
 
@@ -2240,7 +2277,7 @@ dict_stats_update_persistent(
 
 		ut_ad(!dict_index_is_ibuf(index));
 
-		if (index->type & DICT_FTS || dict_index_is_spatial(index)) {
+		if (index->type & (DICT_FTS | DICT_SPATIAL)) {
 			continue;
 		}
 
@@ -2251,7 +2288,20 @@ dict_stats_update_persistent(
 		}
 
 		if (!(table->stats_bg_flag & BG_STAT_SHOULD_QUIT)) {
-			dict_stats_analyze_index(index);
+			mutex_exit(&dict_sys->mutex);
+			stats = dict_stats_analyze_index(index);
+			mutex_enter(&dict_sys->mutex);
+
+			index->stat_index_size = stats.index_size;
+			index->stat_n_leaf_pages = stats.n_leaf_pages;
+			for (size_t i = 0; i < stats.stats.size(); ++i) {
+				index->stat_n_diff_key_vals[i]
+					= stats.stats[i].n_diff_key_vals;
+				index->stat_n_sample_sizes[i]
+					= stats.stats[i].n_sample_sizes;
+				index->stat_n_non_null_key_vals[i]
+					= stats.stats[i].n_non_null_key_vals;
+			}
 		}
 
 		table->stat_sum_of_other_index_sizes
@@ -2266,7 +2316,7 @@ dict_stats_update_persistent(
 
 	dict_stats_assert_initialized(table);
 
-	dict_table_stats_unlock(table, RW_X_LATCH);
+	mutex_exit(&dict_sys->mutex);
 
 	return(DB_SUCCESS);
 }
@@ -3094,11 +3144,22 @@ dict_stats_update_for_index(
 	if (dict_stats_is_persistent_enabled(index->table)) {
 
 		if (dict_stats_persistent_storage_check(false)) {
-			dict_table_stats_lock(index->table, RW_X_LATCH);
-			dict_stats_analyze_index(index);
+			index_stats_t stats = dict_stats_analyze_index(index);
+			mutex_enter(&dict_sys->mutex);
+			index->stat_index_size = stats.index_size;
+			index->stat_n_leaf_pages = stats.n_leaf_pages;
+			for (size_t i = 0; i < stats.stats.size(); ++i) {
+				index->stat_n_diff_key_vals[i]
+					= stats.stats[i].n_diff_key_vals;
+				index->stat_n_sample_sizes[i]
+					= stats.stats[i].n_sample_sizes;
+				index->stat_n_non_null_key_vals[i]
+					= stats.stats[i].n_non_null_key_vals;
+			}
 			index->table->stat_sum_of_other_index_sizes
 				+= index->stat_index_size;
-			dict_table_stats_unlock(index->table, RW_X_LATCH);
+			mutex_exit(&dict_sys->mutex);
+
 			dict_stats_save(index->table, &index->id);
 			DBUG_VOID_RETURN;
 		}
@@ -3119,9 +3180,7 @@ dict_stats_update_for_index(
 		}
 	}
 
-	dict_table_stats_lock(index->table, RW_X_LATCH);
 	dict_stats_update_transient_for_index(index);
-	dict_table_stats_unlock(index->table, RW_X_LATCH);
 
 	DBUG_VOID_RETURN;
 }
@@ -3275,7 +3334,7 @@ dict_stats_update(
 		switch (err) {
 		case DB_SUCCESS:
 
-			dict_table_stats_lock(table, RW_X_LATCH);
+			mutex_enter(&dict_sys->mutex);
 
 			/* Pass reset_ignored_indexes=true as parameter
 			to dict_stats_copy. This will cause statictics
@@ -3284,7 +3343,7 @@ dict_stats_update(
 
 			dict_stats_assert_initialized(table);
 
-			dict_table_stats_unlock(table, RW_X_LATCH);
+			mutex_exit(&dict_sys->mutex);
 
 			dict_stats_table_clone_free(t);
 
@@ -3338,13 +3397,8 @@ dict_stats_update(
 	}
 
 transient:
-
-	dict_table_stats_lock(table, RW_X_LATCH);
-
 	dict_stats_update_transient(table);
 
-	dict_table_stats_unlock(table, RW_X_LATCH);
-
 	return(DB_SUCCESS);
 }
 

storage/innobase/handler/ha_innodb.cc

@@ -672,7 +672,6 @@ static PSI_rwlock_info all_innodb_rwlocks[] = {
 	PSI_RWLOCK_KEY(trx_purge_latch),
 	PSI_RWLOCK_KEY(index_tree_rw_lock),
 	PSI_RWLOCK_KEY(index_online_log),
-	PSI_RWLOCK_KEY(dict_table_stats),
 	PSI_RWLOCK_KEY(hash_table_locks)
 };
 # endif /* UNIV_PFS_RWLOCK */
@@ -14341,6 +14340,8 @@ ha_innobase::info_low(
 
 	DEBUG_SYNC_C("ha_innobase_info_low");
 
+	ut_ad(!mutex_own(&dict_sys->mutex));
+
 	/* If we are forcing recovery at a high level, we will suppress
 	statistics calculation on tables, because that may crash the
 	server if an index is badly corrupted. */
@@ -14377,7 +14378,6 @@ ha_innobase::info_low(
 				opt = DICT_STATS_RECALC_TRANSIENT;
 			}
 
-			ut_ad(!mutex_own(&dict_sys->mutex));
 			ret = dict_stats_update(ib_table, opt);
 
 			if (ret != DB_SUCCESS) {
@@ -14398,9 +14398,7 @@ ha_innobase::info_low(
 		ulint	stat_clustered_index_size;
 		ulint	stat_sum_of_other_index_sizes;
 
-		if (!(flag & HA_STATUS_NO_LOCK)) {
-			dict_table_stats_lock(ib_table, RW_S_LATCH);
-		}
+		mutex_enter(&dict_sys->mutex);
 
 		ut_a(ib_table->stat_initialized);
 
@@ -14412,9 +14410,7 @@ ha_innobase::info_low(
 		stat_sum_of_other_index_sizes
 			= ib_table->stat_sum_of_other_index_sizes;
 
-		if (!(flag & HA_STATUS_NO_LOCK)) {
-			dict_table_stats_unlock(ib_table, RW_S_LATCH);
-		}
+		mutex_exit(&dict_sys->mutex);
 
 		/*
 		The MySQL optimizer seems to assume in a left join that n_rows
@@ -14563,10 +14559,27 @@ ha_innobase::info_low(
 			ib_push_frm_error(m_user_thd, ib_table, table, num_innodb_index, true);
 		}
 
-		if (!(flag & HA_STATUS_NO_LOCK)) {
-			dict_table_stats_lock(ib_table, RW_S_LATCH);
+
+		snprintf(path, sizeof(path), "%s/%s%s",
+			 mysql_data_home, table->s->normalized_path.str,
+			 reg_ext);
+
+		unpack_filename(path,path);
+
+		/* Note that we do not know the access time of the table,
+		nor the CHECK TABLE time, nor the UPDATE or INSERT time. */
+
+		if (os_file_get_status(
+			    path, &stat_info, false,
+			    srv_read_only_mode) == DB_SUCCESS) {
+			stats.create_time = (ulong) stat_info.ctime;
 		}
 
+		struct Locking {
+			Locking() { mutex_enter(&dict_sys->mutex); }
+			~Locking() { mutex_exit(&dict_sys->mutex); }
+		} locking;
+
 		ut_a(ib_table->stat_initialized);
 
 		for (i = 0; i < table->s->keys; i++) {
@@ -14644,25 +14657,6 @@ ha_innobase::info_low(
 				key->rec_per_key[j] = rec_per_key_int;
 			}
 		}
-
-		if (!(flag & HA_STATUS_NO_LOCK)) {
-			dict_table_stats_unlock(ib_table, RW_S_LATCH);
-		}
-
-		snprintf(path, sizeof(path), "%s/%s%s",
-			 mysql_data_home, table->s->normalized_path.str,
-			 reg_ext);
-
-		unpack_filename(path,path);
-
-		/* Note that we do not know the access time of the table,
-		nor the CHECK TABLE time, nor the UPDATE or INSERT time. */
-
-		if (os_file_get_status(
-			    path, &stat_info, false,
-			    srv_read_only_mode) == DB_SUCCESS) {
-			stats.create_time = (ulong) stat_info.ctime;
-		}
 	}
 
 	if (srv_force_recovery >= SRV_FORCE_NO_IBUF_MERGE) {

storage/innobase/handler/i_s.cc

@@ -6317,20 +6317,26 @@ i_s_dict_fill_sys_tablestats(
 	OK(field_store_string(fields[SYS_TABLESTATS_NAME],
 			      table->name.m_name));
 
-	dict_table_stats_lock(table, RW_S_LATCH);
+	{
+		struct Locking
+		{
+			Locking() { mutex_enter(&dict_sys->mutex); }
+			~Locking() { mutex_exit(&dict_sys->mutex); }
+		} locking;
 
 		if (table->stat_initialized) {
 			OK(field_store_string(fields[SYS_TABLESTATS_INIT],
 					      "Initialized"));
 
-		OK(fields[SYS_TABLESTATS_NROW]->store(table->stat_n_rows,
-						      true));
+			OK(fields[SYS_TABLESTATS_NROW]->store(
+				   table->stat_n_rows, true));
 
 			OK(fields[SYS_TABLESTATS_CLUST_SIZE]->store(
 				   table->stat_clustered_index_size, true));
 
 			OK(fields[SYS_TABLESTATS_INDEX_SIZE]->store(
-			   table->stat_sum_of_other_index_sizes, true));
+				   table->stat_sum_of_other_index_sizes,
+				   true));
 
 			OK(fields[SYS_TABLESTATS_MODIFIED]->store(
 				   table->stat_modified_counter, true));
@@ -6346,8 +6352,7 @@ i_s_dict_fill_sys_tablestats(
 
 			OK(fields[SYS_TABLESTATS_MODIFIED]->store(0, true));
 		}
-
-	dict_table_stats_unlock(table, RW_S_LATCH);
+	}
 
 	OK(fields[SYS_TABLESTATS_AUTONINC]->store(table->autoinc, true));
 

storage/innobase/ibuf/ibuf0ibuf.cc

@@ -1447,7 +1447,7 @@ ibuf_dummy_index_create(
 
 	table = dict_mem_table_create("IBUF_DUMMY",
 				      DICT_HDR_SPACE, n, 0,
-				      comp ? DICT_TF_COMPACT : 0, 0, false);
+				      comp ? DICT_TF_COMPACT : 0, 0);
 
 	index = dict_mem_index_create("IBUF_DUMMY", "IBUF_DUMMY",
 				      DICT_HDR_SPACE, 0, n);

storage/innobase/include/btr0btr.h

@@ -213,7 +213,7 @@ the index.
 ulint
 btr_height_get(
 /*===========*/
-	dict_index_t*	index,	/*!< in: index tree */
+	const dict_index_t*	index,	/*!< in: index tree */
 	mtr_t*		mtr)	/*!< in/out: mini-transaction */
 	MY_ATTRIBUTE((warn_unused_result));
 
@@ -585,7 +585,7 @@ Gets the number of pages in a B-tree.
 ulint
 btr_get_size(
 /*=========*/
-	dict_index_t*	index,	/*!< in: index */
+	const dict_index_t*	index,	/*!< in: index */
 	ulint		flag,	/*!< in: BTR_N_LEAF_PAGES or BTR_TOTAL_SIZE */
 	mtr_t*		mtr)	/*!< in/out: mini-transaction where index
 				is s-latched */

storage/innobase/include/btr0cur.h

@@ -592,8 +592,24 @@ btr_estimate_n_rows_in_range(
 	const dtuple_t*	tuple2,
 	page_cur_mode_t	mode2);
 
-/*******************************************************************//**
-Estimates the number of different key values in a given index, for
+
+/** Statistics for one field of an index. */
+struct index_field_stats_t
+{
+  ib_uint64_t n_diff_key_vals;
+  ib_uint64_t n_sample_sizes;
+  ib_uint64_t n_non_null_key_vals;
+
+  index_field_stats_t(ib_uint64_t n_diff_key_vals= 0,
+                      ib_uint64_t n_sample_sizes= 0,
+                      ib_uint64_t n_non_null_key_vals= 0)
+      : n_diff_key_vals(n_diff_key_vals), n_sample_sizes(n_sample_sizes),
+        n_non_null_key_vals(n_non_null_key_vals)
+  {
+  }
+};
+
+/** Estimates the number of different key values in a given index, for
 each n-column prefix of the index where 1 <= n <= dict_index_get_n_unique(index).
 The estimates are stored in the array index->stat_n_diff_key_vals[] (indexed
 0..n_uniq-1) and the number of pages that were sampled is saved in
@@ -601,12 +617,11 @@ index->stat_n_sample_sizes[].
 If innodb_stats_method is nulls_ignored, we also record the number of
 non-null values for each prefix and stored the estimates in
 array index->stat_n_non_null_key_vals.
-@return true if the index is available and we get the estimated numbers,
-false if the index is unavailable. */
-bool
-btr_estimate_number_of_different_key_vals(
-/*======================================*/
-	dict_index_t*	index);	/*!< in: index */
+@param[in]	index	index
+@return stat vector if the index is available and we get the estimated numbers,
+empty vector if the index is unavailable. */
+std::vector<index_field_stats_t>
+btr_estimate_number_of_different_key_vals(dict_index_t* index);
 
 /** Gets the externally stored size of a record, in units of a database page.
 @param[in]	rec	record

storage/innobase/include/dict0dict.h

@@ -1459,7 +1459,7 @@ UNIV_INLINE
 rw_lock_t*
 dict_index_get_lock(
 /*================*/
-	dict_index_t*	index)	/*!< in: index */
+	const dict_index_t*	index)	/*!< in: index */
 	MY_ATTRIBUTE((nonnull, warn_unused_result));
 /********************************************************************//**
 Returns free space reserved for future updates of records. This is
@@ -1524,22 +1524,6 @@ void
 dict_mutex_exit_for_mysql(void);
 /*===========================*/
 
-/** Lock the appropriate latch to protect a given table's statistics.
-@param[in]	table		table whose stats to lock
-@param[in]	latch_mode	RW_S_LATCH or RW_X_LATCH */
-void
-dict_table_stats_lock(
-	dict_table_t*	table,
-	ulint		latch_mode);
-
-/** Unlock the latch that has been locked by dict_table_stats_lock().
-@param[in]	table		table whose stats to unlock
-@param[in]	latch_mode	RW_S_LATCH or RW_X_LATCH */
-void
-dict_table_stats_unlock(
-	dict_table_t*	table,
-	ulint		latch_mode);
-
 /********************************************************************//**
 Checks if the database name in two table names is the same.
 @return TRUE if same db name */

storage/innobase/include/dict0dict.ic

@@ -1181,7 +1181,7 @@ UNIV_INLINE
 rw_lock_t*
 dict_index_get_lock(
 /*================*/
-	dict_index_t*	index)	/*!< in: index */
+	const dict_index_t*	index)	/*!< in: index */
 {
 	ut_ad(index->magic_n == DICT_INDEX_MAGIC_N);
 

storage/innobase/include/dict0mem.h

@@ -307,7 +307,6 @@ before proceeds. */
 @param[in]      n_v_cols        number of virtual columns
 @param[in]      flags           table flags
 @param[in]      flags2          table flags2
-@param[in]      init_stats_latch whether to init the stats latch
 @return own: table object */
 dict_table_t*
 dict_mem_table_create(
@@ -316,8 +315,7 @@ dict_mem_table_create(
 	ulint           n_cols,
         ulint           n_v_cols,
         ulint           flags,
-        ulint           flags2,
-        bool            init_stats_latch=true);
+        ulint           flags2);
 
 /****************************************************************//**
 Free a table memory object. */
@@ -970,7 +968,7 @@ struct dict_index_t{
 				when InnoDB was started up */
 	zip_pad_info_t	zip_pad;/*!< Information about state of
 				compression failures and successes */
-	rw_lock_t	lock;	/*!< read-write lock protecting the
+	mutable rw_lock_t	lock;	/*!< read-write lock protecting the
 				upper levels of the index tree */
 
 	/** Determine if the index has been committed to the
@@ -1691,23 +1689,8 @@ struct dict_table_t {
 	/*!< set of foreign key constraints which refer to this table */
 	dict_foreign_set			referenced_set;
 
-	/** Statistics for query optimization. @{ */
-
-	/** Creation state of 'stats_latch'. */
-	bool					stats_latch_inited;
-
-	/** This latch protects:
-	dict_table_t::stat_initialized,
-	dict_table_t::stat_n_rows (*),
-	dict_table_t::stat_clustered_index_size,
-	dict_table_t::stat_sum_of_other_index_sizes,
-	dict_table_t::stat_modified_counter (*),
-	dict_table_t::indexes*::stat_n_diff_key_vals[],
-	dict_table_t::indexes*::stat_index_size,
-	dict_table_t::indexes*::stat_n_leaf_pages.
-	(*) Those are not always protected for
-	performance reasons. */
-	rw_lock_t				stats_latch;
+	/** Statistics for query optimization. Mostly protected by
+	dict_sys->mutex. @{ */
 
 	/** TRUE if statistics have been calculated the first time after
 	database startup or table creation. */

storage/innobase/include/dict0stats.ic

@@ -75,7 +75,7 @@ dict_stats_is_persistent_enabled(const dict_table_t* table)
 	+ dict_stats_update(DICT_STATS_RECALC_TRANSIENT) on a table that has
 	  just been PS-enabled.
 	This is acceptable. Avoiding this would mean that we would have to
-	protect the ::stat_persistent with dict_table_stats_lock() like the
+	protect the ::stat_persistent with dict_sys->mutex like the
 	other ::stat_ members which would be too big performance penalty,
 	especially when this function is called from
 	dict_stats_update_if_needed(). */
@@ -178,10 +178,7 @@ dict_stats_deinit(
 
 	ut_a(table->get_ref_count() == 0);
 
-	dict_table_stats_lock(table, RW_X_LATCH);
-
 	if (!table->stat_initialized) {
-		dict_table_stats_unlock(table, RW_X_LATCH);
 		return;
 	}
 
@@ -221,6 +218,4 @@ dict_stats_deinit(
 			sizeof(index->stat_n_leaf_pages));
 	}
 #endif /* HAVE_valgrind_or_MSAN */
-
-	dict_table_stats_unlock(table, RW_X_LATCH);
 }

storage/innobase/include/sync0sync.h

@@ -126,7 +126,6 @@ extern	mysql_pfs_key_t	trx_i_s_cache_lock_key;
 extern	mysql_pfs_key_t	trx_purge_latch_key;
 extern	mysql_pfs_key_t	index_tree_rw_lock_key;
 extern	mysql_pfs_key_t	index_online_log_key;
-extern	mysql_pfs_key_t	dict_table_stats_key;
 extern  mysql_pfs_key_t trx_sys_rw_lock_key;
 extern  mysql_pfs_key_t hash_table_locks_key;
 #endif /* UNIV_PFS_RWLOCK */

storage/innobase/row/row0mysql.cc

@@ -1512,7 +1512,7 @@ error_exit:
 		srv_stats.n_rows_inserted.inc(size_t(trx->id));
 	}
 
-	/* Not protected by dict_table_stats_lock() for performance
+	/* Not protected by dict_sys->mutex for performance
 	reasons, we would rather get garbage in stat_n_rows (which is
 	just an estimate anyway) than protecting the following code
 	with a latch. */
@@ -1870,7 +1870,7 @@ row_update_for_mysql(row_prebuilt_t* prebuilt)
 	ut_ad(node->is_delete == is_delete);
 
 	if (/*node->*/is_delete) {
-		/* Not protected by dict_table_stats_lock() for performance
+		/* Not protected by dict_sys->mutex for performance
 		reasons, we would rather get garbage in stat_n_rows (which is
 		just an estimate anyway) than protecting the following code
 		with a latch. */
@@ -2131,7 +2131,7 @@ row_update_cascade_for_mysql(
 
 			if (node->is_delete) {
 				/* Not protected by
-				dict_table_stats_lock() for
+				dict_sys->mutex for
 				performance reasons, we would rather
 				get garbage in stat_n_rows (which is
 				just an estimate anyway) than

storage/innobase/row/row0uins.cc

@@ -524,7 +524,7 @@ row_undo_ins(
 		}
 
 		if (err == DB_SUCCESS && node->table->stat_initialized) {
-			/* Not protected by dict_table_stats_lock() for
+			/* Not protected by dict_sys->mutex for
 			performance reasons, we would rather get garbage
 			in stat_n_rows (which is just an estimate anyway)
 			than protecting the following code with a latch. */

storage/innobase/sync/sync0debug.cc

@@ -1473,9 +1473,6 @@ sync_latch_meta_init()
 
 	LATCH_ADD_RWLOCK(INDEX_TREE, SYNC_INDEX_TREE, index_tree_rw_lock_key);
 
-	LATCH_ADD_RWLOCK(DICT_TABLE_STATS, SYNC_INDEX_TREE,
-			 dict_table_stats_key);
-
 	LATCH_ADD_RWLOCK(HASH_TABLE_RW_LOCK, SYNC_BUF_PAGE_HASH,
 		  hash_table_locks_key);
 

storage/innobase/sync/sync0sync.cc

@@ -102,7 +102,6 @@ mysql_pfs_key_t	buf_block_debug_latch_key;
 # endif /* UNIV_DEBUG */
 mysql_pfs_key_t	checkpoint_lock_key;
 mysql_pfs_key_t	dict_operation_lock_key;
-mysql_pfs_key_t	dict_table_stats_key;
 mysql_pfs_key_t	hash_table_locks_key;
 mysql_pfs_key_t	index_tree_rw_lock_key;
 mysql_pfs_key_t	index_online_log_key;