MDEV-26445 innodb_undo_log_truncate is unnecessarily slow

trx_purge_truncate_history(): Do not force a write of the undo tablespace that is being truncated. Instead, prevent page writes by acquiring an exclusive latch on all dirty pages of the tablespace. fseg_create(): Relax an assertion that could fail if a dirty undo page is being initialized during undo tablespace truncation (and trx_purge_truncate_history() already acquired an exclusive latch on it). fsp_page_create(): If we are truncating a tablespace, try to reuse a page that we may have already latched exclusively (because it was in buf_pool.flush_list). To some extent, this helps the test innodb.undo_truncate,16k to avoid running out of buffer pool. mtr_t::commit_shrink(): Mark as clean all pages that are outside the new bounds of the tablespace, and only add the newly reinitialized pages to the buf_pool.flush_list. buf_page_create(): Do not unnecessarily invoke change buffer merge on undo tablespaces. buf_page_t::clear_oldest_modification(bool temporary): Move some assertions to the caller buf_page_write_complete(). innodb.undo_truncate: Use a bigger innodb_buffer_pool_size=24M. On my system, it would otherwise hang 1 out of 1547 attempts (on the 40th repeat of innodb.undo_truncate,16k). Other page sizes were not affected.

Author: dr-m

mysql-test/suite/innodb/t/undo_truncate.opt

@@ -0,0 +1 @@
+--innodb-buffer-pool-size=24M

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

@@ -3,6 +3,10 @@
 --source include/have_undo_tablespaces.inc
 --source include/have_sequence.inc
 
+--disable_query_log
+call mtr.add_suppression("InnoDB: Difficult to find free blocks in the buffer pool");
+--enable_query_log
+
 SET @save_frequency = @@GLOBAL.innodb_purge_rseg_truncate_frequency;
 SET @save_truncate = @@GLOBAL.innodb_undo_log_truncate;
 SET GLOBAL innodb_undo_log_truncate = 0;
@@ -46,17 +50,5 @@ drop table t1, t2;
 
 --source include/wait_all_purged.inc
 
-# Truncation will normally not occur with innodb_page_size=64k,
-# and occasionally not with innodb_page_size=32k,
-# because the undo log will not grow enough.
-# TODO: For some reason this does not occur on 4k either!
-if (`select @@innodb_page_size IN (8192,16384)`)
-{
- let $wait_condition = (SELECT variable_value!=@trunc_start
- FROM information_schema.global_status
- WHERE variable_name = 'innodb_undo_truncations');
- source include/wait_condition.inc;
-}
-
 SET GLOBAL innodb_purge_rseg_truncate_frequency = @save_frequency;
 SET GLOBAL innodb_undo_log_truncate = @save_truncate;

storage/innobase/buf/buf0buf.cc

@@ -3796,6 +3796,7 @@ buf_page_create(fil_space_t *space, uint32_t offset,
  /* Delete possible entries for the page from the insert buffer:
  such can exist if the page belonged to an index which was dropped */
  if (page_id < page_id_t{SRV_SPACE_ID_UPPER_BOUND, 0} &&
+ !srv_is_undo_tablespace(page_id.space()) &&
  !recv_recovery_is_on())
  ibuf_merge_or_delete_for_page(nullptr, page_id, zip_size);
 

storage/innobase/buf/buf0flu.cc

@@ -363,10 +363,12 @@ void buf_page_write_complete(const IORequest &request)
  const bool temp= fsp_is_system_temporary(bpage->id().space());
 
  mysql_mutex_lock(&buf_pool.mutex);
+ mysql_mutex_assert_not_owner(&buf_pool.flush_list_mutex);
  buf_pool.stat.n_pages_written++;
  /* While we do not need any mutex for clearing oldest_modification
  here, we hope that it will be in the same cache line with io_fix,
  whose changes must be protected by buf_pool.mutex. */
+ ut_ad(temp || bpage->oldest_modification() > 2);
  bpage->clear_oldest_modification(temp);
  ut_ad(bpage->io_fix() == BUF_IO_WRITE);
  bpage->set_io_fix(BUF_IO_NONE);

storage/innobase/fsp/fsp0fsp.cc

@@ -1054,11 +1054,36 @@ static
 buf_block_t*
 fsp_page_create(fil_space_t *space, page_no_t offset, mtr_t *mtr)
 {
- buf_block_t *free_block= buf_LRU_get_free_block(false);
- buf_block_t *block= buf_page_create(space, static_cast<uint32_t>(offset),
- space->zip_size(), mtr, free_block);
+ buf_block_t *block, *free_block;
+
+ if (UNIV_UNLIKELY(space->is_being_truncated))
+ {
+ const page_id_t page_id{space->id, offset};
+ const ulint fold= page_id.fold();
+ mysql_mutex_lock(&buf_pool.mutex);
+ block= reinterpret_cast<buf_block_t*>
+ (buf_pool.page_hash_get_low(page_id, fold));
+ if (block && block->page.oldest_modification() <= 1)
+ block= nullptr;
+ mysql_mutex_unlock(&buf_pool.mutex);
+
+ if (block)
+ {
+ ut_ad(block->page.buf_fix_count() >= 1);
+ ut_ad(rw_lock_get_x_lock_count(&block->lock) == 1);
+ ut_ad(mtr->have_x_latch(*block));
+ free_block= block;
+ goto got_free_block;
+ }
+ }
+
+ free_block= buf_LRU_get_free_block(false);
+got_free_block:
+ block= buf_page_create(space, static_cast<uint32_t>(offset),
+ space->zip_size(), mtr, free_block);
  if (UNIV_UNLIKELY(block != free_block))
  buf_pool.free_block(free_block);
+
  fsp_init_file_page(space, block, mtr);
  return block;
 }
@@ -1728,7 +1753,10 @@ fseg_create(fil_space_t *space, ulint byte_offset, mtr_t *mtr,
 goto funct_exit;
 }
 
-ut_ad(rw_lock_get_x_lock_count(&block->lock) == 1);
+ut_d(const auto x = rw_lock_get_x_lock_count(&block->lock));
+ut_ad(x > 0);
+ut_ad(x == 1 || space->is_being_truncated);
+ut_ad(x <= 2);
 ut_ad(!fil_page_get_type(block->frame));
 mtr->write<1>(*block, FIL_PAGE_TYPE + 1 + block->frame,
  FIL_PAGE_TYPE_SYS);

storage/innobase/include/buf0buf.h

@@ -2232,9 +2232,7 @@ inline void buf_page_t::clear_oldest_modification()
 it from buf_pool.flush_list */
 inline void buf_page_t::clear_oldest_modification(bool temporary)
 {
- mysql_mutex_assert_not_owner(&buf_pool.flush_list_mutex);
  ut_ad(temporary == fsp_is_system_temporary(id().space()));
- ut_ad(io_fix_ == BUF_IO_WRITE);
  if (temporary)
  {
  ut_ad(oldest_modification() == 2);

storage/innobase/mtr/mtr0mtr.cc

@@ -459,15 +459,15 @@ void mtr_t::commit()
  release_resources();
 }
 
-#ifdef UNIV_DEBUG
-/** Check that all pages belong to a shrunk tablespace. */
+/** Shrink a tablespace. */
 struct Shrink
 {
- const page_id_t low, high;
- Shrink(const fil_space_t &space) :
- low({space.id, 0}), high({space.id, space.size}) {}
+ /** the first non-existing page in the tablespace */
+ const page_id_t high;
 
- bool operator()(const mtr_memo_slot_t *slot) const
+ Shrink(const fil_space_t &space) : high({space.id, space.size}) {}
+
+ bool operator()(mtr_memo_slot_t *slot) const
  {
  if (!slot->object)
  return true;
@@ -476,29 +476,31 @@ struct Shrink
  ut_ad("invalid type" == 0);
  return false;
  case MTR_MEMO_SPACE_X_LOCK:
- ut_ad(low.space() == static_cast<fil_space_t*>(slot->object)->id);
+ ut_ad(high.space() == static_cast<fil_space_t*>(slot->object)->id);
  return true;
  case MTR_MEMO_PAGE_X_MODIFY:
  case MTR_MEMO_PAGE_SX_MODIFY:
  case MTR_MEMO_PAGE_X_FIX:
  case MTR_MEMO_PAGE_SX_FIX:
- const auto &bpage= static_cast<buf_block_t*>(slot->object)->page;
+ auto &bpage= static_cast<buf_block_t*>(slot->object)->page;
+ ut_ad(bpage.io_fix() == BUF_IO_NONE);
  const auto id= bpage.id();
- if (id == page_id_t{0, TRX_SYS_PAGE_NO})
+ if (id < high)
  {
- ut_ad(srv_is_undo_tablespace(low.space()));
+ ut_ad(id.space() == high.space() ||
+ (id == page_id_t{0, TRX_SYS_PAGE_NO} &&
+ srv_is_undo_tablespace(high.space())));
  break;
  }
- ut_ad(id >= low);
- ut_ad(id < high);
+ ut_ad(id.space() == high.space());
  ut_ad(bpage.state() == BUF_BLOCK_FILE_PAGE);
- ut_ad(bpage.oldest_modification() <= 1);
- break;
+ if (bpage.oldest_modification() > 1)
+ bpage.clear_oldest_modification(false);
+ slot->type= static_cast<mtr_memo_type_t>(slot->type & ~MTR_MEMO_MODIFY);
  }
  return true;
  }
 };
-#endif
 
 /** Commit a mini-transaction that is shrinking a tablespace.
 @param space tablespace that is being shrunk */
@@ -542,7 +544,7 @@ void mtr_t::commit_shrink(fil_space_t &space)
  else
  ut_ad(!m_freed_space);
 
- ut_d(m_memo.for_each_block_in_reverse(CIterate<Shrink>{space}));
+ m_memo.for_each_block_in_reverse(CIterate<Shrink>{space});
 
  m_memo.for_each_block_in_reverse(CIterate<const ReleaseBlocks>
  (ReleaseBlocks(start_lsn, m_commit_lsn,