MDEV-16515 InnoDB: Failing assertion: ++retries < 10000 in file

dict0dict.cc buf_LRU_drop_page_hash_for_tablespace(): Return whether any adaptive hash index entries existed. If yes, the caller should keep retrying to drop the adaptive hash index. row_import_for_mysql(), row_truncate_table_for_mysql(), row_drop_table_for_mysql(): Ensure that the adaptive hash index was entirely dropped for the table.
MariaDB · Jun 26, 2018 · c4eb4bc · c4eb4bc
1 parent c09a8b5
commit c4eb4bc
Show file tree

Hide file tree

Showing 10 changed files with 118 additions and 73 deletions.
diff --git a/storage/innobase/buf/buf0lru.cc b/storage/innobase/buf/buf0lru.cc
@@ -356,9 +356,10 @@ buf_LRU_drop_page_hash_for_tablespace(
 	ut_free(page_arr);
 }
 
-/** Drop the adaptive hash index for a tablespace.
-@param[in,out]	table	table */
-UNIV_INTERN void buf_LRU_drop_page_hash_for_tablespace(dict_table_t* table)
+/** Try to drop the adaptive hash index for a tablespace.
+@param[in,out]	table	table
+@return	whether anything was dropped */
+UNIV_INTERN bool buf_LRU_drop_page_hash_for_tablespace(dict_table_t* table)
 {
 	for (dict_index_t* index = dict_table_get_first_index(table);
 	     index != NULL;
@@ -369,13 +370,15 @@ UNIV_INTERN void buf_LRU_drop_page_hash_for_tablespace(dict_table_t* table)
 		}
 	}
 
-	return;
+	return false;
 drop_ahi:
 	ulint id = table->space;
 	for (ulint i = 0; i < srv_buf_pool_instances; i++) {
 		buf_LRU_drop_page_hash_for_tablespace(buf_pool_from_array(i),
 						      id);
 	}
+
+	return true;
 }
 
 /******************************************************************//**

diff --git a/storage/innobase/dict/dict0dict.cc b/storage/innobase/dict/dict0dict.cc
@@ -2719,12 +2719,11 @@ dict_index_remove_from_cache_low(
 	zero. See also: dict_table_can_be_evicted() */
 
 	do {
-		if (!btr_search_info_get_ref_count(info)) {
+		if (!btr_search_info_get_ref_count(info)
+		    || !buf_LRU_drop_page_hash_for_tablespace(table)) {
 			break;
 		}
 
-		buf_LRU_drop_page_hash_for_tablespace(table);
-
 		ut_a(++retries < 10000);
 	} while (srv_shutdown_state == SRV_SHUTDOWN_NONE || !lru_evict);
 

diff --git a/storage/innobase/include/buf0lru.h b/storage/innobase/include/buf0lru.h
@@ -53,9 +53,11 @@ These are low-level functions
 /** Minimum LRU list length for which the LRU_old pointer is defined */
 #define BUF_LRU_OLD_MIN_LEN	512	/* 8 megabytes of 16k pages */
 
-/** Drop the adaptive hash index for a tablespace.
-@param[in,out]	table	table */
-UNIV_INTERN void buf_LRU_drop_page_hash_for_tablespace(dict_table_t* table);
+/** Try to drop the adaptive hash index for a tablespace.
+@param[in,out]	table	table
+@return	whether anything was dropped */
+UNIV_INTERN bool buf_LRU_drop_page_hash_for_tablespace(dict_table_t* table)
+	MY_ATTRIBUTE((warn_unused_result,nonnull));
 
 /** Empty the flush list for all pages belonging to a tablespace.
 @param[in]	id		tablespace identifier

diff --git a/storage/innobase/row/row0import.cc b/storage/innobase/row/row0import.cc
@@ -3983,6 +3983,23 @@ row_import_for_mysql(
 	DBUG_EXECUTE_IF("ib_import_reset_space_and_lsn_failure",
 			err = DB_TOO_MANY_CONCURRENT_TRXS;);
 
+	/* On DISCARD TABLESPACE, we did not drop any adaptive hash
+	index entries. If we replaced the discarded tablespace with a
+	smaller one here, there could still be some adaptive hash
+	index entries that point to cached garbage pages in the buffer
+	pool, because PageConverter::operator() only evicted those
+	pages that were replaced by the imported pages. We must
+	discard all remaining adaptive hash index entries, because the
+	adaptive hash index must be a subset of the table contents;
+	false positives are not tolerated. */
+	while (buf_LRU_drop_page_hash_for_tablespace(table)) {
+		if (trx_is_interrupted(trx)
+		    || srv_shutdown_state != SRV_SHUTDOWN_NONE) {
+			err = DB_INTERRUPTED;
+			break;
+		}
+	}
+
 	if (err != DB_SUCCESS) {
 		char	table_name[MAX_FULL_NAME_LEN + 1];
 
@@ -4000,17 +4017,6 @@ row_import_for_mysql(
 		return(row_import_cleanup(prebuilt, trx, err));
 	}
 
-	/* On DISCARD TABLESPACE, we did not drop any adaptive hash
-	index entries. If we replaced the discarded tablespace with a
-	smaller one here, there could still be some adaptive hash
-	index entries that point to cached garbage pages in the buffer
-	pool, because PageConverter::operator() only evicted those
-	pages that were replaced by the imported pages. We must
-	discard all remaining adaptive hash index entries, because the
-	adaptive hash index must be a subset of the table contents;
-	false positives are not tolerated. */
-	buf_LRU_drop_page_hash_for_tablespace(table);
-
 	row_mysql_lock_data_dictionary(trx);
 
 	/* If the table is stored in a remote tablespace, we need to

diff --git a/storage/innobase/row/row0mysql.cc b/storage/innobase/row/row0mysql.cc
@@ -3516,7 +3516,13 @@ row_truncate_table_for_mysql(
 			fil_space_release(space);
 		}
 
-		buf_LRU_drop_page_hash_for_tablespace(table);
+		while (buf_LRU_drop_page_hash_for_tablespace(table)) {
+			if (trx_is_interrupted(trx)
+			    || srv_shutdown_state != SRV_SHUTDOWN_NONE) {
+				err = DB_INTERRUPTED;
+				goto funct_exit;
+			}
+		}
 
 		if (flags != ULINT_UNDEFINED
 		    && fil_discard_tablespace(space_id) == DB_SUCCESS) {
@@ -4172,6 +4178,27 @@ row_drop_table_for_mysql(
 
 	ut_a(!lock_table_has_locks(table));
 
+	if (table->space != TRX_SYS_SPACE) {
+		/* On DISCARD TABLESPACE, we would not drop the
+		adaptive hash index entries. If the tablespace is
+		missing here, delete-marking the record in SYS_INDEXES
+		would not free any pages in the buffer pool. Thus,
+		dict_index_remove_from_cache() would hang due to
+		adaptive hash index entries existing in the buffer
+		pool.  To prevent this hang, and also to guarantee
+		that btr_search_drop_page_hash_when_freed() will avoid
+		calling btr_search_drop_page_hash_index() while we
+		hold the InnoDB dictionary lock, we will drop any
+		adaptive hash index entries upfront. */
+		while (buf_LRU_drop_page_hash_for_tablespace(table)) {
+			if (trx_is_interrupted(trx)
+			    || srv_shutdown_state != SRV_SHUTDOWN_NONE) {
+				err = DB_INTERRUPTED;
+				goto funct_exit;
+			}
+		}
+	}
+
 	switch (trx_get_dict_operation(trx)) {
 	case TRX_DICT_OP_NONE:
 		trx_set_dict_operation(trx, TRX_DICT_OP_TABLE);
@@ -4211,21 +4238,6 @@ row_drop_table_for_mysql(
 		rw_lock_x_unlock(dict_index_get_lock(index));
 	}
 
-	if (table->space != TRX_SYS_SPACE) {
-		/* On DISCARD TABLESPACE, we would not drop the
-		adaptive hash index entries. If the tablespace is
-		missing here, delete-marking the record in SYS_INDEXES
-		would not free any pages in the buffer pool. Thus,
-		dict_index_remove_from_cache() would hang due to
-		adaptive hash index entries existing in the buffer
-		pool.  To prevent this hang, and also to guarantee
-		that btr_search_drop_page_hash_when_freed() will avoid
-		calling btr_search_drop_page_hash_index() while we
-		hold the InnoDB dictionary lock, we will drop any
-		adaptive hash index entries upfront. */
-		buf_LRU_drop_page_hash_for_tablespace(table);
-	}
-
 	/* We use the private SQL parser of Innobase to generate the
 	query graphs needed in deleting the dictionary data from system
 	tables in Innobase. Deleting a row from SYS_INDEXES table also

diff --git a/storage/xtradb/buf/buf0lru.cc b/storage/xtradb/buf/buf0lru.cc
@@ -354,9 +354,10 @@ buf_LRU_drop_page_hash_for_tablespace(
 	ut_free(page_arr);
 }
 
-/** Drop the adaptive hash index for a tablespace.
-@param[in,out]	table	table */
-UNIV_INTERN void buf_LRU_drop_page_hash_for_tablespace(dict_table_t* table)
+/** Try to drop the adaptive hash index for a tablespace.
+@param[in,out]	table	table
+@return	whether anything was dropped */
+UNIV_INTERN bool buf_LRU_drop_page_hash_for_tablespace(dict_table_t* table)
 {
 	for (dict_index_t* index = dict_table_get_first_index(table);
 	     index != NULL;
@@ -367,13 +368,15 @@ UNIV_INTERN void buf_LRU_drop_page_hash_for_tablespace(dict_table_t* table)
 		}
 	}
 
-	return;
+	return false;
 drop_ahi:
 	ulint id = table->space;
 	for (ulint i = 0; i < srv_buf_pool_instances; i++) {
 		buf_LRU_drop_page_hash_for_tablespace(buf_pool_from_array(i),
 						      id);
 	}
+
+	return true;
 }
 
 /******************************************************************//**

diff --git a/storage/xtradb/dict/dict0dict.cc b/storage/xtradb/dict/dict0dict.cc
@@ -2729,11 +2729,11 @@ dict_index_remove_from_cache_low(
 	zero. See also: dict_table_can_be_evicted() */
 
 	do {
-		if (!btr_search_info_get_ref_count(info, index)) {
+		if (!btr_search_info_get_ref_count(info, index)
+		    || !buf_LRU_drop_page_hash_for_tablespace(table)) {
 			break;
 		}
 
-		buf_LRU_drop_page_hash_for_tablespace(table);
 		ut_a(++retries < 10000);
 	} while (srv_shutdown_state == SRV_SHUTDOWN_NONE || !lru_evict);
 

diff --git a/storage/xtradb/include/buf0lru.h b/storage/xtradb/include/buf0lru.h
@@ -55,9 +55,11 @@ These are low-level functions
 /** Minimum LRU list length for which the LRU_old pointer is defined */
 #define BUF_LRU_OLD_MIN_LEN	512	/* 8 megabytes of 16k pages */
 
-/** Drop the adaptive hash index for a tablespace.
-@param[in,out]	table	table */
-UNIV_INTERN void buf_LRU_drop_page_hash_for_tablespace(dict_table_t* table);
+/** Try to drop the adaptive hash index for a tablespace.
+@param[in,out]	table	table
+@return	whether anything was dropped */
+UNIV_INTERN bool buf_LRU_drop_page_hash_for_tablespace(dict_table_t* table)
+	MY_ATTRIBUTE((warn_unused_result,nonnull));
 
 /** Empty the flush list for all pages belonging to a tablespace.
 @param[in]	id		tablespace identifier

diff --git a/storage/xtradb/row/row0import.cc b/storage/xtradb/row/row0import.cc
@@ -3982,6 +3982,23 @@ row_import_for_mysql(
 	DBUG_EXECUTE_IF("ib_import_reset_space_and_lsn_failure",
 			err = DB_TOO_MANY_CONCURRENT_TRXS;);
 
+	/* On DISCARD TABLESPACE, we did not drop any adaptive hash
+	index entries. If we replaced the discarded tablespace with a
+	smaller one here, there could still be some adaptive hash
+	index entries that point to cached garbage pages in the buffer
+	pool, because PageConverter::operator() only evicted those
+	pages that were replaced by the imported pages. We must
+	discard all remaining adaptive hash index entries, because the
+	adaptive hash index must be a subset of the table contents;
+	false positives are not tolerated. */
+	while (buf_LRU_drop_page_hash_for_tablespace(table)) {
+		if (trx_is_interrupted(trx)
+		    || srv_shutdown_state != SRV_SHUTDOWN_NONE) {
+			err = DB_INTERRUPTED;
+			break;
+		}
+	}
+
 	if (err != DB_SUCCESS) {
 		char	table_name[MAX_FULL_NAME_LEN + 1];
 
@@ -3999,17 +4016,6 @@ row_import_for_mysql(
 		return(row_import_cleanup(prebuilt, trx, err));
 	}
 
-	/* On DISCARD TABLESPACE, we did not drop any adaptive hash
-	index entries. If we replaced the discarded tablespace with a
-	smaller one here, there could still be some adaptive hash
-	index entries that point to cached garbage pages in the buffer
-	pool, because PageConverter::operator() only evicted those
-	pages that were replaced by the imported pages. We must
-	discard all remaining adaptive hash index entries, because the
-	adaptive hash index must be a subset of the table contents;
-	false positives are not tolerated. */
-	buf_LRU_drop_page_hash_for_tablespace(table);
-
 	row_mysql_lock_data_dictionary(trx);
 
 	/* If the table is stored in a remote tablespace, we need to

diff --git a/storage/xtradb/row/row0mysql.cc b/storage/xtradb/row/row0mysql.cc
@@ -3540,7 +3540,13 @@ row_truncate_table_for_mysql(
 			fil_space_release(space);
 		}
 
-		buf_LRU_drop_page_hash_for_tablespace(table);
+		while (buf_LRU_drop_page_hash_for_tablespace(table)) {
+			if (trx_is_interrupted(trx)
+			    || srv_shutdown_state != SRV_SHUTDOWN_NONE) {
+				err = DB_INTERRUPTED;
+				goto funct_exit;
+			}
+		}
 
 		if (flags != ULINT_UNDEFINED
 		    && fil_discard_tablespace(space_id) == DB_SUCCESS) {
@@ -4202,6 +4208,27 @@ row_drop_table_for_mysql(
 
 	ut_a(!lock_table_has_locks(table));
 
+	if (table->space != TRX_SYS_SPACE) {
+		/* On DISCARD TABLESPACE, we would not drop the
+		adaptive hash index entries. If the tablespace is
+		missing here, delete-marking the record in SYS_INDEXES
+		would not free any pages in the buffer pool. Thus,
+		dict_index_remove_from_cache() would hang due to
+		adaptive hash index entries existing in the buffer
+		pool.  To prevent this hang, and also to guarantee
+		that btr_search_drop_page_hash_when_freed() will avoid
+		calling btr_search_drop_page_hash_index() while we
+		hold the InnoDB dictionary lock, we will drop any
+		adaptive hash index entries upfront. */
+		while (buf_LRU_drop_page_hash_for_tablespace(table)) {
+			if (trx_is_interrupted(trx)
+			    || srv_shutdown_state != SRV_SHUTDOWN_NONE) {
+				err = DB_INTERRUPTED;
+				goto funct_exit;
+			}
+		}
+	}
+
 	switch (trx_get_dict_operation(trx)) {
 	case TRX_DICT_OP_NONE:
 		trx_set_dict_operation(trx, TRX_DICT_OP_TABLE);
@@ -4241,21 +4268,6 @@ row_drop_table_for_mysql(
 		rw_lock_x_unlock(dict_index_get_lock(index));
 	}
 
-	if (table->space != TRX_SYS_SPACE) {
-		/* On DISCARD TABLESPACE, we would not drop the
-		adaptive hash index entries. If the tablespace is
-		missing here, delete-marking the record in SYS_INDEXES
-		would not free any pages in the buffer pool. Thus,
-		dict_index_remove_from_cache() would hang due to
-		adaptive hash index entries existing in the buffer
-		pool.  To prevent this hang, and also to guarantee
-		that btr_search_drop_page_hash_when_freed() will avoid
-		calling btr_search_drop_page_hash_index() while we
-		hold the InnoDB dictionary lock, we will drop any
-		adaptive hash index entries upfront. */
-		buf_LRU_drop_page_hash_for_tablespace(table);
-	}
-
 	/* We use the private SQL parser of Innobase to generate the
 	query graphs needed in deleting the dictionary data from system
 	tables in Innobase. Deleting a row from SYS_INDEXES table also