Bug#22179133 - INNODB: TOO SMALL BUFFER POOL FOR INNODB_PAGE_SIZE=64K

Problem:
--------
When initializing rollback segment headers(max TRX_SYS_N_RSEGS), all
rollback segment headers are accessed using single mtr.

With buf_pool_size of 5M and innodb_page_size=64k,
the maximum number of pages available is 80.
When a single mtr accesses more than 80 pages, all the blocks
are buffer fixed and cannot be replaced from buf pool.

So we cannot find a free block.

Fix:
----
Access each rollback segment header in a separate mini-transaction.

Reviewed-By: Marko Mäkelä <[email protected]>
RB: 11367

Author: satya461

mysql-test/suite/innodb/t/innodb_64k-master.opt

@@ -0,0 +1 @@
+--innodb_buffer_pool_size=5M

mysql-test/suite/innodb/t/innodb_64k.test

@@ -6,6 +6,13 @@
 # Test COMPRESSED table
 -- source suite/innodb/include/create_options_invalid.inc
 
+# This test is run with 5M buffer pool size
+let $buf_pool_size = `SELECT @@innodb_buffer_pool_size`;
+
+if ($buf_pool_size != 5242880) {
+  --skip "Test requires 5M buffer pool size"
+}
+
 # Test max inline field length should < 16384.
 CREATE TABLE t1 (id int, a TEXT) ROW_FORMAT=COMPACT;
 

storage/innobase/include/trx0rseg.h

@@ -1,6 +1,6 @@
 /*****************************************************************************
 
-Copyright (c) 1996, 2014, Oracle and/or its affiliates. All Rights Reserved.
+Copyright (c) 1996, 2015, Oracle and/or its affiliates. All Rights Reserved.
 
 This program is free software; you can redistribute it and/or modify it under
 the terms of the GNU General Public License as published by the Free Software
@@ -122,9 +122,8 @@ rseg array in trx_sys at a database startup. */
 void
 trx_rseg_array_init(
 /*================*/
-	trx_sysf_t*	sys_header,	/*!< in/out: trx system header */
-	purge_pq_t*	purge_queue,	/*!< in: rseg queue */
-	mtr_t*		mtr);		/*!< in/out: mtr */
+	purge_pq_t*	purge_queue);	/*!< in: rseg queue */
+
 /***************************************************************************
 Free's an instance of the rollback segment in memory. */
 void

storage/innobase/trx/trx0rseg.cc

@@ -319,15 +319,19 @@ static
 void
 trx_rseg_create_instance(
 /*=====================*/
-	trx_sysf_t*	sys_header,	/*!< in: trx system header */
-	purge_pq_t*	purge_queue,	/*!< in/out: rseg queue */
-	mtr_t*		mtr)		/*!< in: mtr */
+	purge_pq_t*	purge_queue)	/*!< in/out: rseg queue */
 {
 	ulint		i;
 
 	for (i = 0; i < TRX_SYS_N_RSEGS; i++) {
 		ulint	page_no;
 
+		mtr_t	mtr;
+		mtr.start();
+		trx_sysf_t* sys_header = trx_sysf_get(&mtr);
+
+		page_no = trx_sysf_rseg_get_page_no(sys_header, i, &mtr);
+
 		/* Slot-1....Slot-n are reserved for non-redo rsegs.
 		Non-redo rsegs are recreated on server re-start so
 		avoid initializing the existing non-redo rsegs. */
@@ -337,20 +341,15 @@ trx_rseg_create_instance(
 			in range from slot-1....slot-n needs to be scheduled
 			for purge if there are pending purge operation. */
 			trx_rseg_schedule_pending_purge(
-				sys_header, purge_queue, i, mtr);
+				sys_header, purge_queue, i, &mtr);
 
-			continue;
-		}
-
-		page_no = trx_sysf_rseg_get_page_no(sys_header, i, mtr);
-
-		if (page_no != FIL_NULL) {
+		} else if (page_no != FIL_NULL) {
 			ulint		space;
 			trx_rseg_t*	rseg = NULL;
 
 			ut_a(!trx_rseg_get_on_id(i, true));
 
-			space = trx_sysf_rseg_get_space(sys_header, i, mtr);
+			space = trx_sysf_rseg_get_space(sys_header, i, &mtr);
 
 			bool			found = true;
 			const page_size_t&	page_size
@@ -365,12 +364,13 @@ trx_rseg_create_instance(
 
 			rseg = trx_rseg_mem_create(
 				i, space, page_no, page_size,
-				purge_queue, rseg_array, mtr);
+				purge_queue, rseg_array, &mtr);
 
 			ut_a(rseg->id == i);
 		} else {
 			ut_a(trx_sys->rseg_array[i] == NULL);
 		}
+		mtr.commit();
 	}
 }
 
@@ -447,13 +447,11 @@ rseg array in trx_sys at a database startup. */
 void
 trx_rseg_array_init(
 /*================*/
-	trx_sysf_t*	sys_header,	/* in/out: trx system header */
-	purge_pq_t*	purge_queue,	/*!< in: rseg queue */
-	mtr_t*		mtr)		/*!< in: mtr */
+	purge_pq_t*	purge_queue)	/*!< in: rseg queue */
 {
 	trx_sys->rseg_history_len = 0;
 
-	trx_rseg_create_instance(sys_header, purge_queue, mtr);
+	trx_rseg_create_instance(purge_queue);
 }
 
 /********************************************************************

storage/innobase/trx/trx0sys.cc

@@ -453,7 +453,6 @@ purge_pq_t*
 trx_sys_init_at_db_start(void)
 /*==========================*/
 {
-	mtr_t		mtr;
 	purge_pq_t*	purge_queue;
 	trx_sysf_t*	sys_header;
 	ib_uint64_t	rows_to_undo	= 0;
@@ -465,12 +464,8 @@ trx_sys_init_at_db_start(void)
 	purge_queue = UT_NEW_NOKEY(purge_pq_t());
 	ut_a(purge_queue != NULL);
 
-	mtr_start(&mtr);
-
-	sys_header = trx_sysf_get(&mtr);
-
 	if (srv_force_recovery < SRV_FORCE_NO_UNDO_LOG_SCAN) {
-		trx_rseg_array_init(sys_header, purge_queue, &mtr);
+		trx_rseg_array_init(purge_queue);
 	}
 
 	/* VERY important: after the database is started, max_trx_id value is
@@ -480,11 +475,17 @@ trx_sys_init_at_db_start(void)
 	to the disk-based header! Thus trx id values will not overlap when
 	the database is repeatedly started! */
 
+	mtr_t	mtr;
+	mtr.start();
+
+	sys_header = trx_sysf_get(&mtr);
+
 	trx_sys->max_trx_id = 2 * TRX_SYS_TRX_ID_WRITE_MARGIN
 		+ ut_uint64_align_up(mach_read_from_8(sys_header
 						   + TRX_SYS_TRX_ID_STORE),
 				     TRX_SYS_TRX_ID_WRITE_MARGIN);
 
+	mtr.commit();
 	ut_d(trx_sys->rw_max_trx_id = trx_sys->max_trx_id);
 
 	trx_dummy_sess = sess_open();
@@ -527,8 +528,6 @@ trx_sys_init_at_db_start(void)
 
 	trx_sys_mutex_exit();
 
-	mtr_commit(&mtr);
-
 	return(purge_queue);
 }