MDEV-29835 InnoDB hang on B-tree split or merge

This is a follow-up to
commit de4030e4 (MDEV-30400),
which fixed some hangs related to B-tree split or merge.

btr_root_block_get(): Use and update the root page guess. This is just
a minor performance optimization, not affecting correctness.

btr_validate_level(): Remove the parameter "lockout", and always
acquire an exclusive dict_index_t::lock in CHECK TABLE without QUICK.
This is needed in order to avoid latching order violation in
btr_page_get_father_node_ptr_for_validate().

btr_cur_need_opposite_intention(): Return true in case
btr_cur_compress_recommendation() would hold later during the
mini-transaction, or if a page underflow or overflow is possible.
If we return true, our caller will escalate to aqcuiring an exclusive
dict_index_t::lock, to prevent a latching order violation and deadlock
during btr_compress() or btr_page_split_and_insert().

btr_cur_t::search_leaf(), btr_cur_t::open_leaf():
Also invoke btr_cur_need_opposite_intention() on the leaf page.

btr_cur_t::open_leaf(): When escalating to exclusive index locking,
acquire exclusive latches on all pages as well.

innobase_instant_try(): Return an error code if the root page cannot
be retrieved.

In addition to the normal stress testing with Random Query Generator (RQG)
this has been tested with
./mtr --mysqld=--loose-innodb-limit-optimistic-insert-debug=2
but with the injection in btr_cur_optimistic_insert() for non-leaf pages
adjusted so that it would use the value 3. (Otherwise, infinite page
splits could occur in some mtr tests.)

Tested by: Matthias Leich

storage/innobase/btr/btr0btr.cc

@@ -254,7 +254,7 @@ Gets the root node of a tree and x- or s-latches it.
 buf_block_t*
 btr_root_block_get(
 /*===============*/
-	const dict_index_t*	index,	/*!< in: index tree */
+	dict_index_t*		index,	/*!< in: index tree */
 	rw_lock_type_t		mode,	/*!< in: either RW_S_LATCH
 					or RW_X_LATCH */
 	mtr_t*			mtr,	/*!< in: mtr */
@@ -266,11 +266,31 @@ btr_root_block_get(
     return nullptr;
   }
 
-  buf_block_t *block = btr_block_get(*index, index->page, mode, false, mtr,
-                                     err);
-  if (block)
+  buf_block_t *block;
+#ifndef BTR_CUR_ADAPT
+  static constexpr buf_block_t *guess= nullptr;
+#else
+  buf_block_t *&guess= btr_search_get_info(index)->root_guess;
+  guess=
+#endif
+  block=
+    buf_page_get_gen(page_id_t{index->table->space->id, index->page},
+                     index->table->space->zip_size(), mode, guess, BUF_GET,
+                     mtr, err, false);
+  ut_ad(!block == (*err != DB_SUCCESS));
+
+  if (UNIV_LIKELY(block != nullptr))
   {
-    if (index->is_ibuf());
+    if (!!page_is_comp(block->page.frame) != index->table->not_redundant() ||
+        btr_page_get_index_id(block->page.frame) != index->id ||
+        !fil_page_index_page_check(block->page.frame) ||
+        index->is_spatial() !=
+        (fil_page_get_type(block->page.frame) == FIL_PAGE_RTREE))
+    {
+      *err= DB_PAGE_CORRUPTED;
+      block= nullptr;
+    }
+    else if (index->is_ibuf());
     else if (!btr_root_fseg_validate(FIL_PAGE_DATA + PAGE_BTR_SEG_LEAF,
                                      *block, *index->table->space) ||
              !btr_root_fseg_validate(FIL_PAGE_DATA + PAGE_BTR_SEG_TOP,
@@ -280,6 +300,9 @@ btr_root_block_get(
       block= nullptr;
     }
   }
+  else if (*err == DB_DECRYPTION_FAILED)
+    btr_decryption_failed(*index);
+
   return block;
 }
 
@@ -290,7 +313,7 @@ static
 page_t*
 btr_root_get(
 /*=========*/
-	const dict_index_t*	index,	/*!< in: index tree */
+	dict_index_t*		index,	/*!< in: index tree */
 	mtr_t*			mtr,	/*!< in: mtr */
 	dberr_t*		err)	/*!< out: error code */
 {
@@ -502,9 +525,7 @@ btr_block_reget(mtr_t *mtr, const dict_index_t &index,
     return block;
   }
 
-#if 0 /* MDEV-29385 FIXME: Acquire the page latch upfront. */
   ut_ad(mtr->memo_contains_flagged(&index.lock, MTR_MEMO_X_LOCK));
-#endif
   return btr_block_get(index, id.page_no(), rw_latch, true, mtr, err);
 }
 
@@ -773,9 +794,7 @@ btr_page_get_father_node_ptr_for_validate(
 	const uint32_t page_no = btr_cur_get_block(cursor)->page.id().page_no();
 	dict_index_t* index = btr_cur_get_index(cursor);
 	ut_ad(!dict_index_is_spatial(index));
-
-	ut_ad(mtr->memo_contains_flagged(&index->lock, MTR_MEMO_X_LOCK
-					 | MTR_MEMO_SX_LOCK));
+	ut_ad(mtr->memo_contains(index->lock, MTR_MEMO_X_LOCK));
 	ut_ad(dict_index_get_page(index) != page_no);
 
 	const auto level = btr_page_get_level(btr_cur_get_page(cursor));
@@ -793,10 +812,6 @@ btr_page_get_father_node_ptr_for_validate(
 	}
 
 	const rec_t* node_ptr = btr_cur_get_rec(cursor);
-#if 0 /* MDEV-29835 FIXME */
-	ut_ad(!btr_cur_get_block(cursor)->page.lock.not_recursive()
-	      || mtr->memo_contains(index->lock, MTR_MEMO_X_LOCK));
-#endif
 
 	offsets = rec_get_offsets(node_ptr, index, offsets, 0,
 				  ULINT_UNDEFINED, &heap);
@@ -2456,11 +2471,10 @@ btr_insert_on_non_leaf_level(
 	}
 
 	ut_ad(cursor.flag == BTR_CUR_BINARY);
-#if 0 /* MDEV-29835 FIXME */
-	ut_ad(!btr_cur_get_block(&cursor)->page.lock.not_recursive()
+	ut_ad(btr_cur_get_block(&cursor)
+	      != mtr->at_savepoint(mtr->get_savepoint() - 1)
 	      || index->is_spatial()
 	      || mtr->memo_contains(index->lock, MTR_MEMO_X_LOCK));
-#endif
 
 	if (UNIV_LIKELY(err == DB_SUCCESS)) {
 		err = btr_cur_optimistic_insert(flags,
@@ -2568,10 +2582,8 @@ btr_attach_half_pages(
 		prev_block = mtr->get_already_latched(id, MTR_MEMO_PAGE_X_FIX);
 #if 1 /* MDEV-29835 FIXME: acquire page latches upfront */
 		if (!prev_block) {
-# if 0 /* MDEV-29835 FIXME */
 			ut_ad(mtr->memo_contains(index->lock,
 						 MTR_MEMO_X_LOCK));
-# endif
 			prev_block = btr_block_get(*index, prev_page_no,
 						   RW_X_LATCH, !level, mtr);
 		}
@@ -2582,10 +2594,8 @@ btr_attach_half_pages(
 		next_block = mtr->get_already_latched(id, MTR_MEMO_PAGE_X_FIX);
 #if 1 /* MDEV-29835 FIXME: acquire page latches upfront */
 		if (!next_block) {
-# if 0 /* MDEV-29835 FIXME */
 			ut_ad(mtr->memo_contains(index->lock,
 						 MTR_MEMO_X_LOCK));
-# endif
 			next_block = btr_block_get(*index, next_page_no,
 						   RW_X_LATCH, !level, mtr);
 		}
@@ -3397,9 +3407,7 @@ dberr_t btr_level_list_remove(const buf_block_t& block,
 #if 1 /* MDEV-29835 FIXME: acquire page latches upfront */
     if (!prev)
     {
-# if 0 /* MDEV-29835 FIXME */
       ut_ad(mtr->memo_contains(index.lock, MTR_MEMO_X_LOCK));
-# endif
       prev= btr_block_get(index, id.page_no(), RW_X_LATCH,
                           page_is_leaf(block.page.frame), mtr, &err);
       if (UNIV_UNLIKELY(!prev))
@@ -3415,9 +3423,7 @@ dberr_t btr_level_list_remove(const buf_block_t& block,
 #if 1 /* MDEV-29835 FIXME: acquire page latches upfront */
     if (!next)
     {
-# if 0 /* MDEV-29835 FIXME */
       ut_ad(mtr->memo_contains(index.lock, MTR_MEMO_X_LOCK));
-# endif
       next= btr_block_get(index, id.page_no(), RW_X_LATCH,
                           page_is_leaf(block.page.frame), mtr, &err);
       if (UNIV_UNLIKELY(!next))
@@ -4291,7 +4297,7 @@ btr_discard_page(
 		if (UNIV_UNLIKELY(!merge_block)) {
 			return err;
 		}
-#if 0 /* MDEV-29385 FIXME: Acquire the page latch upfront. */
+#if 1 /* MDEV-29835 FIXME: Acquire the page latch upfront. */
 		ut_ad(!memcmp_aligned<4>(merge_block->page.frame
 					 + FIL_PAGE_NEXT,
 					 block->page.frame + FIL_PAGE_OFFSET,
@@ -4317,7 +4323,7 @@ btr_discard_page(
 		if (UNIV_UNLIKELY(!merge_block)) {
 			return err;
 		}
-#if 0 /* MDEV-29385 FIXME: Acquire the page latch upfront. */
+#if 1 /* MDEV-29835 FIXME: Acquire the page latch upfront. */
 		ut_ad(!memcmp_aligned<4>(merge_block->page.frame
 					 + FIL_PAGE_PREV,
 					 block->page.frame + FIL_PAGE_OFFSET,
@@ -4898,8 +4904,7 @@ btr_validate_level(
 /*===============*/
 	dict_index_t*	index,	/*!< in: index tree */
 	const trx_t*	trx,	/*!< in: transaction or NULL */
-	ulint		level,	/*!< in: level number */
-	bool		lockout)/*!< in: true if X-latch index is intended */
+	ulint		level)	/*!< in: level number */
 {
 	buf_block_t*	block;
 	page_t*		page;
@@ -4918,18 +4923,10 @@ btr_validate_level(
 #ifdef UNIV_ZIP_DEBUG
 	page_zip_des_t*	page_zip;
 #endif /* UNIV_ZIP_DEBUG */
-	ulint		savepoint = 0;
-	uint32_t	parent_page_no = FIL_NULL;
-	uint32_t	parent_right_page_no = FIL_NULL;
-	bool		rightmost_child = false;
 
 	mtr.start();
 
-	if (lockout) {
-		mtr_x_lock_index(index, &mtr);
-	} else {
-		mtr_sx_lock_index(index, &mtr);
-	}
+	mtr_x_lock_index(index, &mtr);
 
 	dberr_t err;
 	block = btr_root_block_get(index, RW_SX_LATCH, &mtr, &err);
@@ -5025,11 +5022,7 @@ btr_validate_level(
 	mem_heap_empty(heap);
 	offsets = offsets2 = NULL;
 
-	if (lockout) {
-		mtr_x_lock_index(index, &mtr);
-	} else {
-		mtr_sx_lock_index(index, &mtr);
-	}
+	mtr_x_lock_index(index, &mtr);
 
 	page = block->page.frame;
 
@@ -5073,7 +5066,6 @@ btr_validate_level(
 
 	if (right_page_no != FIL_NULL) {
 		const rec_t*	right_rec;
-		savepoint = mtr.get_savepoint();
 
 		right_block = btr_block_get(*index, right_page_no, RW_SX_LATCH,
 					    !level, &mtr, &err);
@@ -5177,11 +5169,6 @@ btr_validate_level(
 		father_page = btr_cur_get_page(&node_cur);
 		node_ptr = btr_cur_get_rec(&node_cur);
 
-		parent_page_no = page_get_page_no(father_page);
-		parent_right_page_no = btr_page_get_next(father_page);
-		rightmost_child = page_rec_is_supremum(
-					page_rec_get_next(node_ptr));
-
 		rec = page_rec_get_prev(page_get_supremum_rec(page));
 		if (rec) {
 			btr_cur_position(index, rec, block, &node_cur);
@@ -5263,37 +5250,6 @@ btr_validate_level(
 			}
 		} else if (const rec_t* right_node_ptr
 			   = page_rec_get_next(node_ptr)) {
-			if (!lockout && rightmost_child) {
-
-				/* To obey latch order of tree blocks,
-				we should release the right_block once to
-				obtain lock of the uncle block. */
-				ut_ad(right_block
-				      == mtr.at_savepoint(savepoint));
-				mtr.rollback_to_savepoint(savepoint,
-							  savepoint + 1);
-
-				if (parent_right_page_no != FIL_NULL) {
-					btr_block_get(*index,
-						      parent_right_page_no,
-						      RW_SX_LATCH, false,
-						      &mtr);
-				}
-
-				right_block = btr_block_get(*index,
-							    right_page_no,
-							    RW_SX_LATCH,
-							    !level, &mtr,
-							    &err);
-				if (!right_block) {
-					btr_validate_report1(index, level,
-							     block);
-					fputs("InnoDB: broken FIL_PAGE_NEXT"
-					      " link\n", stderr);
-					goto invalid_page;
-				}
-			}
-
 			btr_cur_position(
 				index,
 				page_get_infimum_rec(right_block->page.frame),
@@ -5365,20 +5321,6 @@ btr_validate_level(
 
 		mtr.start();
 
-		if (!lockout) {
-			if (rightmost_child) {
-				if (parent_right_page_no != FIL_NULL) {
-					btr_block_get(*index,
-						      parent_right_page_no,
-						      RW_SX_LATCH, false,
-						      &mtr);
-				}
-			} else if (parent_page_no != FIL_NULL) {
-				btr_block_get(*index, parent_page_no,
-					      RW_SX_LATCH, false, &mtr);
-			}
-		}
-
 		block = btr_block_get(*index, right_page_no, RW_SX_LATCH,
 				      !level, &mtr, &err);
 		goto loop;
@@ -5396,21 +5338,16 @@ btr_validate_index(
 	dict_index_t*	index,	/*!< in: index */
 	const trx_t*	trx)	/*!< in: transaction or NULL */
 {
-  const bool lockout= index->is_spatial();
-
   mtr_t mtr;
   mtr.start();
 
-  if (lockout)
-    mtr_x_lock_index(index, &mtr);
-  else
-    mtr_sx_lock_index(index, &mtr);
+  mtr_x_lock_index(index, &mtr);
 
   dberr_t err;
   if (page_t *root= btr_root_get(index, &mtr, &err))
     for (auto level= btr_page_get_level(root);; level--)
     {
-      if (dberr_t err_level= btr_validate_level(index, trx, level, lockout))
+      if (dberr_t err_level= btr_validate_level(index, trx, level))
         err= err_level;
       if (!level)
         break;

storage/innobase/handler/handler0alter.cc

@@ -6104,6 +6104,7 @@ static bool innobase_instant_try(
 			id, MTR_MEMO_PAGE_SX_FIX);
 
 		if (UNIV_UNLIKELY(!root)) {
+			err = DB_CORRUPTION;
 			goto func_exit;
 		}
 

storage/innobase/include/btr0btr.h

@@ -445,7 +445,7 @@ Gets the root node of a tree and x- or s-latches it.
 buf_block_t*
 btr_root_block_get(
 /*===============*/
-	const dict_index_t*	index,	/*!< in: index tree */
+	dict_index_t*		index,	/*!< in: index tree */
 	rw_lock_type_t		mode,	/*!< in: either RW_S_LATCH
 					or RW_X_LATCH */
 	mtr_t*			mtr,	/*!< in: mtr */

storage/innobase/include/btr0types.h

@@ -103,6 +103,9 @@ enum btr_latch_mode {
 	dict_index_t::lock is being held in non-exclusive mode. */
 	BTR_MODIFY_LEAF_ALREADY_LATCHED = BTR_MODIFY_LEAF
 	| BTR_ALREADY_S_LATCHED,
+	/** Attempt to modify records in an x-latched tree. */
+	BTR_MODIFY_TREE_ALREADY_LATCHED = BTR_MODIFY_TREE
+	| BTR_ALREADY_S_LATCHED,
 	/** U-latch root and X-latch a leaf page, assuming that
 	dict_index_t::lock is being held in U mode. */
 	BTR_MODIFY_ROOT_AND_LEAF_ALREADY_LATCHED = BTR_MODIFY_ROOT_AND_LEAF

storage/innobase/include/mtr0mtr.h

@@ -344,6 +344,9 @@ struct mtr_t {
   /** Upgrade U locks on a block to X */
   void page_lock_upgrade(const buf_block_t &block);
 
+  /** Upgrade index U lock to X */
+  ATTRIBUTE_COLD void index_lock_upgrade();
+
   /** Check if we are holding tablespace latch
   @param space  tablespace to search for
   @return whether space.latch is being held */

storage/innobase/row/row0log.cc

@@ -3078,6 +3078,9 @@ row_log_apply_op_low(
 	mtr_start(&mtr);
 	index->set_modified(mtr);
 	cursor.page_cur.index = index;
+	if (has_index_lock) {
+		mtr_x_lock_index(index, &mtr);
+	}
 
 	/* We perform the pessimistic variant of the operations if we
 	already hold index->lock exclusively. First, search the
@@ -3085,7 +3088,8 @@ row_log_apply_op_low(
 	depending on when the row in the clustered index was
 	scanned. */
 	*error = cursor.search_leaf(entry, PAGE_CUR_LE, has_index_lock
-				    ? BTR_MODIFY_TREE : BTR_MODIFY_LEAF, &mtr);
+				    ? BTR_MODIFY_TREE_ALREADY_LATCHED
+				    : BTR_MODIFY_LEAF, &mtr);
 	if (UNIV_UNLIKELY(*error != DB_SUCCESS)) {
 		goto func_exit;
 	}