MDEV-23693 Failing assertion: my_atomic_load32_explicit(&lock->lock_word, MY_MEMORY_ORDER_RELAXED) == X_LOCK_DECR

InnoDB frees the block lock during buffer pool shrinking when other
thread is yet to release the block lock.  While shrinking the
buffer pool, InnoDB allows the page to be freed unless it is buffer
fixed. In some cases, InnoDB releases the latch after unfixing the
block.

Fix:
====
- InnoDB should unfix the block after releases the latch.

- Add more assertion to check buffer fix while accessing the page.

- Introduced block_hint structure to store buf_block_t pointer
and allow accessing the buf_block_t pointer only by passing a
functor. It returns original buf_block_t* pointer if it is valid
or nullptr if the pointer become stale.

- Replace buf_block_is_uncompressed() with
buf_pool_t::is_block_pointer()

This change is motivated by a change in mysql-5.7.32:
mysql/mysql-server@46e60de
Bug #31036301 ASSERTION FAILURE: SYNC0RW.IC:429:LOCK->LOCK_WORD

Author: Thirunarayanan

storage/innobase/CMakeLists.txt

@@ -28,6 +28,7 @@ SET(INNOBASE_SOURCES
 	btr/btr0scrub.cc
 	btr/btr0sea.cc
 	btr/btr0defragment.cc
+	buf/buf0block_hint.cc
 	buf/buf0buddy.cc
 	buf/buf0buf.cc
 	buf/buf0dblwr.cc

storage/innobase/btr/btr0bulk.cc

@@ -695,6 +695,8 @@ PageBulk::latch()
 		m_mtr.set_named_space(m_index->space);
 	}
 
+	ut_ad(m_block->page.buf_fix_count);
+
 	/* In case the block is S-latched by page_cleaner. */
 	if (!buf_page_optimistic_get(RW_X_LATCH, m_block, m_modify_clock,
 				     __FILE__, __LINE__, &m_mtr)) {
@@ -713,6 +715,8 @@ PageBulk::latch()
 
 	buf_block_buf_fix_dec(m_block);
 
+	ut_ad(m_block->page.buf_fix_count);
+
 	ut_ad(m_cur_rec > m_page && m_cur_rec < m_heap_top);
 
 	return (m_err);

storage/innobase/btr/btr0cur.cc

@@ -416,6 +416,8 @@ btr_cur_optimistic_latch_leaves(
 	ulint		mode;
 	ulint		left_page_no;
 	ulint		curr_page_no;
+	ut_ad(block->page.buf_fix_count);
+	ut_ad(buf_block_get_state(block) == BUF_BLOCK_FILE_PAGE);
 
 	switch (*latch_mode) {
 	case BTR_SEARCH_LEAF:
@@ -427,20 +429,10 @@ btr_cur_optimistic_latch_leaves(
 		mode = *latch_mode == BTR_SEARCH_PREV
 			? RW_S_LATCH : RW_X_LATCH;
 
-		buf_page_mutex_enter(block);
-		if (buf_block_get_state(block) != BUF_BLOCK_FILE_PAGE) {
-			buf_page_mutex_exit(block);
-			return(false);
-		}
-		/* pin the block not to be relocated */
-		buf_block_buf_fix_inc(block, file, line);
-		buf_page_mutex_exit(block);
-
 		rw_lock_s_lock(&block->lock);
 		if (block->modify_clock != modify_clock) {
 			rw_lock_s_unlock(&block->lock);
-
-			goto unpin_failed;
+			return false;
 		}
 
 		curr_page_no = block->page.id.page_no();
@@ -470,7 +462,7 @@ btr_cur_optimistic_latch_leaves(
 				/* release the left block */
 				btr_leaf_page_release(
 					cursor->left_block, mode, mtr);
-				goto unpin_failed;
+				return false;
 			}
 		} else {
 			cursor->left_block = NULL;
@@ -480,23 +472,28 @@ btr_cur_optimistic_latch_leaves(
 					    file, line, mtr)) {
 			if (btr_page_get_prev(buf_block_get_frame(block))
 			    == left_page_no) {
-				buf_block_buf_fix_dec(block);
+				/* block was already buffer-fixed while
+				entering the function and
+				buf_page_optimistic_get() buffer-fixes
+				it again. */
+				ut_ad(2 <= block->page.buf_fix_count);
 				*latch_mode = mode;
 				return(true);
 			} else {
-				/* release the block */
+				/* release the block and decrement of
+				buf_fix_count which was incremented
+				in buf_page_optimistic_get() */
 				btr_leaf_page_release(block, mode, mtr);
 			}
 		}
 
+		ut_ad(block->page.buf_fix_count);
 		/* release the left block */
 		if (cursor->left_block != NULL) {
 			btr_leaf_page_release(cursor->left_block,
 					      mode, mtr);
 		}
-unpin_failed:
-		/* unpin the block */
-		buf_block_buf_fix_dec(block);
+
 		return(false);
 
 	default:
@@ -1066,12 +1063,7 @@ btr_cur_search_to_nth_level_func(
 	guess = NULL;
 #else
 	info = btr_search_get_info(index);
-
-	if (!buf_pool_is_obsolete(info->withdraw_clock)) {
-		guess = info->root_guess;
-	} else {
-		guess = NULL;
-	}
+	guess = info->root_guess;
 
 #ifdef BTR_CUR_HASH_ADAPT
 	rw_lock_t* const search_latch = btr_get_search_latch(index);
@@ -1509,10 +1501,7 @@ btr_cur_search_to_nth_level_func(
 		}
 
 #ifdef BTR_CUR_ADAPT
-		if (block != guess) {
-			info->root_guess = block;
-			info->withdraw_clock = buf_withdraw_clock;
-		}
+		info->root_guess = block;
 #endif
 	}
 

storage/innobase/btr/btr0pcur.cc

@@ -165,11 +165,10 @@ btr_pcur_store_position(
 		index, rec, &cursor->old_n_fields,
 		&cursor->old_rec_buf, &cursor->buf_size);
 
-	cursor->block_when_stored = block;
+	cursor->block_when_stored.store(block);
 
 	/* Function try to check if block is S/X latch. */
 	cursor->modify_clock = buf_block_get_modify_clock(block);
-	cursor->withdraw_clock = buf_withdraw_clock;
 }
 
 /**************************************************************//**
@@ -199,6 +198,26 @@ btr_pcur_copy_stored_position(
 	pcur_receive->old_n_fields = pcur_donate->old_n_fields;
 }
 
+/** Structure acts as functor to do the latching of leaf pages.
+It returns true if latching of leaf pages succeeded and false
+otherwise. */
+struct optimistic_latch_leaves
+{
+  btr_pcur_t *const cursor;
+  ulint *latch_mode;
+  mtr_t *const mtr;
+
+  optimistic_latch_leaves(btr_pcur_t *cursor, ulint *latch_mode, mtr_t *mtr)
+  :cursor(cursor), latch_mode(latch_mode), mtr(mtr) {}
+
+  bool operator() (buf_block_t *hint) const
+  {
+    return hint && btr_cur_optimistic_latch_leaves(
+             hint, cursor->modify_clock, latch_mode,
+             btr_pcur_get_btr_cur(cursor), __FILE__, __LINE__, mtr);
+  }
+};
+
 /**************************************************************//**
 Restores the stored position of a persistent cursor bufferfixing the page and
 obtaining the specified latches. If the cursor position was saved when the
@@ -261,7 +280,7 @@ btr_pcur_restore_position_func(
 		cursor->latch_mode =
 			BTR_LATCH_MODE_WITHOUT_INTENTION(latch_mode);
 		cursor->pos_state = BTR_PCUR_IS_POSITIONED;
-		cursor->block_when_stored = btr_pcur_get_block(cursor);
+		cursor->block_when_stored.clear();
 
 		return(FALSE);
 	}
@@ -276,12 +295,9 @@ btr_pcur_restore_position_func(
 	case BTR_MODIFY_PREV:
 		/* Try optimistic restoration. */
 
-		if (!buf_pool_is_obsolete(cursor->withdraw_clock)
-		    && btr_cur_optimistic_latch_leaves(
-			cursor->block_when_stored, cursor->modify_clock,
-			&latch_mode, btr_pcur_get_btr_cur(cursor),
-			file, line, mtr)) {
-
+		if (cursor->block_when_stored.run_with_hint(
+			optimistic_latch_leaves(cursor, &latch_mode,
+						mtr))) {
 			cursor->pos_state = BTR_PCUR_IS_POSITIONED;
 			cursor->latch_mode = latch_mode;
 
@@ -378,11 +394,10 @@ btr_pcur_restore_position_func(
 		since the cursor can now be on a different page!
 		But we can retain the value of old_rec */
 
-		cursor->block_when_stored = btr_pcur_get_block(cursor);
+		cursor->block_when_stored.store(btr_pcur_get_block(cursor));
 		cursor->modify_clock = buf_block_get_modify_clock(
-						cursor->block_when_stored);
+					cursor->block_when_stored.block());
 		cursor->old_stored = true;
-		cursor->withdraw_clock = buf_withdraw_clock;
 
 		mem_heap_free(heap);
 

storage/innobase/buf/buf0block_hint.cc

@@ -0,0 +1,78 @@
+/*****************************************************************************
+
+Copyright (c) 2020, Oracle and/or its affiliates. All Rights Reserved.
+Copyright (c) 2020, MariaDB Corporation.
+
+This program is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License, version 2.0, as published by the
+Free Software Foundation.
+
+This program is also distributed with certain software (including but not
+limited to OpenSSL) that is licensed under separate terms, as designated in a
+particular file or component or in included license documentation. The authors
+of MySQL hereby grant you an additional permission to link the program and
+your derivative works with the separately licensed software that they have
+included with MySQL.
+
+This program is distributed in the hope that it will be useful, but WITHOUT
+ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+FOR A PARTICULAR PURPOSE. See the GNU General Public License, version 2.0,
+for more details.
+
+You should have received a copy of the GNU General Public License along with
+this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin St, Fifth Floor, Boston, MA 02110-1301  USA
+
+*****************************************************************************/
+
+#include "buf0block_hint.h"
+namespace buf {
+
+void Block_hint::buffer_fix_block_if_still_valid()
+{
+  /* We need to check if m_block points to one of chunks. For this to be
+  meaningful we need to prevent freeing memory while we check, and until we
+  buffer-fix the block. For this purpose it is enough to latch any of the many
+  latches taken by buf_resize().
+  However, for buffer-fixing to be meaningful, the block has to contain a page
+  (as opposed to being already empty, which might mean that buf_pool_resize()
+  can proceed and free it once we free the s-latch), so we confirm that the
+  block contains a page. However, it is not sufficient to check that this is
+  just any page, because just after we check it could get freed, unless we
+  have a latch which prevents this. This is tricky because page_hash latches
+  are sharded by page_id and we don't know the page_id until we look into the
+  block. To solve this chicken-and-egg problem somewhat, we latch the shard
+  for the m_page_id and compare block->page.id to it - so if is equal then we
+  can be reasonably sure that we have the correct latch.
+  There is still a theoretical problem here, where other threads might try
+  to modify the m_block->page.id while we are comparing it, but the chance of
+  accidentally causing the old space_id == m_page_id.m_space and the new
+  page_no == m_page_id.m_page_no is minimal as compilers emit a single 8-byte
+  comparison instruction to compare both at the same time atomically, and f()
+    will probably double-check the block->page.id again, anyway.
+  Finally, assuming that we have correct hash bucket latched, we should check if
+  the state of the block is BUF_BLOCK_FILE_PAGE before buffer-fixing the block,
+  as otherwise we risk buffer-fixing and operating on a block, which is already
+  meant to be freed. In particular, buf_LRU_free_page() first calls
+  buf_LRU_block_remove_hashed() under hash bucket latch protection to change the
+  state to BUF_BLOCK_REMOVE_HASH and then releases the latch. Later it calls
+  buf_LRU_block_free_hashed_page() without any latch to change the state to
+  BUF_BLOCK_MEMORY and reset the page's id, which means buf_resize() can free it
+  regardless of our buffer-fixing. */
+  if (m_block)
+  {
+    const buf_pool_t *const buf_pool= buf_pool_get(m_page_id);
+    rw_lock_t *latch= buf_page_hash_lock_get(buf_pool, m_page_id);
+    rw_lock_s_lock(latch);
+    /* If not own buf_pool_mutex, page_hash can be changed. */
+    latch= buf_page_hash_lock_s_confirm(latch, buf_pool, m_page_id);
+    if (buf_pool->is_block_field(m_block) &&
+        m_page_id == m_block->page.id &&
+        buf_block_get_state(m_block) == BUF_BLOCK_FILE_PAGE)
+      buf_block_buf_fix_inc(m_block, __FILE__, __LINE__);
+    else
+      clear();
+    rw_lock_s_unlock(latch);
+  }
+}
+}  // namespace buf