MDEV-25801: buf_flush_dirty_pages() is very slow

In commit 7cffb5f6 (MDEV-23399)
the implementation of buf_flush_dirty_pages() was replaced with
a slow one, which would perform excessive scans of the
buf_pool.flush_list and make little progress.

buf_flush_list(), buf_flush_LRU(): Split from buf_flush_lists().
Vladislav Vaintroub noticed that we will not need to invoke
log_flush_task.wait() for the LRU eviction flushing.

buf_flush_list_space(): Replaces buf_flush_dirty_pages().
This is like buf_flush_list(), but operating on a single
tablespace at a time. Writes at most innodb_io_capacity
pages. Returns whether some of the tablespace might remain
in the buffer pool.

storage/innobase/buf/buf0buf.cc

@@ -1729,10 +1729,10 @@ inline bool buf_pool_t::withdraw_blocks()
 
 		/* reserve free_list length */
 		if (UT_LIST_GET_LEN(withdraw) < withdraw_target) {
-			ulint n_flushed = buf_flush_lists(
+			ulint n_flushed = buf_flush_LRU(
 				std::max<ulint>(withdraw_target
 						- UT_LIST_GET_LEN(withdraw),
-						srv_LRU_scan_depth), 0);
+						srv_LRU_scan_depth));
 			buf_flush_wait_batch_end_acquiring_mutex(true);
 
 			if (n_flushed) {
@@ -3321,7 +3321,7 @@ buf_page_get_low(
 
 		fix_block->fix();
 		mysql_mutex_unlock(&buf_pool.mutex);
-		buf_flush_lists(ULINT_UNDEFINED, LSN_MAX);
+		buf_flush_list();
 		buf_flush_wait_batch_end_acquiring_mutex(false);
 
 		if (fix_block->page.buf_fix_count() == 1

storage/innobase/buf/buf0lru.cc

@@ -487,7 +487,7 @@ buf_block_t* buf_LRU_get_free_block(bool have_mutex)
 	involved (particularly in case of ROW_FORMAT=COMPRESSED pages). We
 	can do that in a separate patch sometime in future. */
 
-	if (!buf_flush_lists(innodb_lru_flush_size, 0)) {
+	if (!buf_flush_LRU(innodb_lru_flush_size)) {
 		MONITOR_INC(MONITOR_LRU_SINGLE_FLUSH_FAILURE_COUNT);
 		++flush_failures;
 	}

storage/innobase/fil/fil0crypt.cc

@@ -1098,7 +1098,7 @@ static bool fil_crypt_start_encrypting_space(fil_space_t* space)
 		mtr.commit();
 
 		/* 4 - sync tablespace before publishing crypt data */
-		while (buf_flush_dirty_pages(space->id));
+		while (buf_flush_list_space(space));
 
 		/* 5 - publish crypt data */
 		mutex_enter(&fil_crypt_threads_mutex);
@@ -2036,14 +2036,7 @@ fil_crypt_flush_space(
 	if (end_lsn > 0 && !space->is_stopping()) {
 		ulint sum_pages = 0;
 		const ulonglong start = my_interval_timer();
-		do {
-			ulint n_dirty= buf_flush_dirty_pages(state->space->id);
-			if (!n_dirty) {
-				break;
-			}
-			sum_pages += n_dirty;
-		} while (!space->is_stopping());
-
+		while (buf_flush_list_space(space, &sum_pages));
 		if (sum_pages) {
 			const ulonglong end = my_interval_timer();
 

storage/innobase/fil/fil0fil.cc

@@ -1742,7 +1742,7 @@ void fil_close_tablespace(ulint id)
 	can no longer read more pages of this tablespace to buf_pool.
 	Thus we can clean the tablespace out of buf_pool
 	completely and permanently. */
-	while (buf_flush_dirty_pages(id));
+	while (buf_flush_list_space(space));
 	ut_ad(space->is_stopping());
 
 	/* If the free is successful, the X lock will be released before

storage/innobase/include/buf0flu.h

 /*****************************************************************************
 
 Copyright (c) 1995, 2016, Oracle and/or its affiliates. All Rights Reserved.
-Copyright (c) 2014, 2020, MariaDB Corporation.
+Copyright (c) 2014, 2021, MariaDB Corporation.
 
 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
@@ -54,12 +54,6 @@ the list as they age towards the tail of the LRU.
 @param id    tablespace identifier */
 void buf_flush_remove_pages(ulint id);
 
-/** Try to flush all the dirty pages that belong to a given tablespace.
-@param id    tablespace identifier
-@return number dirty pages that there were for this tablespace */
-ulint buf_flush_dirty_pages(ulint id)
-  MY_ATTRIBUTE((warn_unused_result));
-
 /*******************************************************************//**
 Relocates a buffer control block on the flush_list.
 Note that it is assumed that the contents of bpage has already been
@@ -93,10 +87,23 @@ buf_flush_init_for_writing(
 
 /** Write out dirty blocks from buf_pool.flush_list.
 @param max_n    wished maximum mumber of blocks flushed
-@param lsn      buf_pool.get_oldest_modification(LSN_MAX) target (0=LRU flush)
+@param lsn      buf_pool.get_oldest_modification(LSN_MAX) target
+@return the number of processed pages
+@retval 0 if a buf_pool.flush_list batch is already running */
+ulint buf_flush_list(ulint max_n= ULINT_UNDEFINED, lsn_t lsn= LSN_MAX);
+
+/** Try to flush dirty pages that belong to a given tablespace.
+@param space       tablespace
+@param n_flushed   number of pages written
+@return whether any pages might not have been flushed */
+bool buf_flush_list_space(fil_space_t *space, ulint *n_flushed= nullptr)
+  MY_ATTRIBUTE((warn_unused_result));
+
+/** Write out dirty blocks from buf_pool.LRU.
+@param max_n    wished maximum mumber of blocks flushed
 @return the number of processed pages
-@retval 0 if a batch of the same type (lsn==0 or lsn!=0) is already running */
-ulint buf_flush_lists(ulint max_n, lsn_t lsn);
+@retval 0 if a buf_pool.LRU batch is already running */
+ulint buf_flush_LRU(ulint max_n);
 
 /** Wait until a flush batch ends.
 @param lru    true=buf_pool.LRU; false=buf_pool.flush_list */

storage/innobase/row/row0import.cc

@@ -4232,7 +4232,7 @@ row_import_for_mysql(
 	/* Ensure that all pages dirtied during the IMPORT make it to disk.
 	The only dirty pages generated should be from the pessimistic purge
 	of delete marked records that couldn't be purged in Phase I. */
-	while (buf_flush_dirty_pages(prebuilt->table->space_id));
+	while (buf_flush_list_space(prebuilt->table->space));
 
 	for (ulint count = 0; prebuilt->table->space->referenced(); count++) {
 		/* Issue a warning every 10.24 seconds, starting after

storage/innobase/row/row0quiesce.cc

 /*****************************************************************************
 
 Copyright (c) 2012, 2016, Oracle and/or its affiliates. All Rights Reserved.
-Copyright (c) 2017, 2020, MariaDB Corporation.
+Copyright (c) 2017, 2021, MariaDB Corporation.
 
 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
@@ -536,7 +536,7 @@ row_quiesce_table_start(
 		}
 	}
 
-	while (buf_flush_dirty_pages(table->space_id)) {
+	while (buf_flush_list_space(table->space)) {
 		if (trx_is_interrupted(trx)) {
 			goto aborted;
 		}

storage/innobase/trx/trx0purge.cc

@@ -596,7 +596,7 @@ static void trx_purge_truncate_history()
 			return;
 		}
 
-		const fil_space_t& space = *purge_sys.truncate.current;
+		fil_space_t& space = *purge_sys.truncate.current;
 		/* Undo tablespace always are a single file. */
 		ut_a(UT_LIST_GET_LEN(space.chain) == 1);
 		fil_node_t* file = UT_LIST_GET_FIRST(space.chain);
@@ -672,7 +672,7 @@ static void trx_purge_truncate_history()
 		mini-transaction commit and the server was killed, then
 		discarding the to-be-trimmed pages without flushing would
 		break crash recovery. So, we cannot avoid the write. */
-		while (buf_flush_dirty_pages(space.id));
+		while (buf_flush_list_space(&space));
 
 		log_free_check();