A fix and test case for Bug#5987 "subselect in bool function
crashes server (prepared statements)": the bug was that all boolean items always recovered its original arguments at statement cleanup stage. This collided with Item_subselect::select_transformer, which tries to permanently change the item tree to use a transformed subselect instead of original one. So we had this call sequence for prepare: mysql_stmt_prepare -> JOIN::prepare -> Item_subselect::fix_fields -> the item tree gets transformed -> Item_bool_rowready_func2::cleanup, item tree is recovered to original state, while it shouldn't have been; mysql_stmt_execute -> attempts to execute a broken tree -> crash. Now instead of bluntly recovering all arguments of bool functions in Item_bool_rowready_func2::cleanup, we recover only those which were changed, and do it in one place. There still would exist a possibility for a collision with subselect tranformation, if permanent and temporary changes were performed at the same stage. But fortunately subselect transformation is always done first, so it doesn't conflict with the optimization done by propogate_cond_constants. Now we have: mysql_stmt_prepare -> JOIN::prepare -> subselect transformation permanently changes the tree -> cleanup doesn't recover anything, because nothing was registered for recovery. mysql_stmt_execute -> JOIN::prepare (the tree is already transformed, so it doesn't change), JOIN::optimize -> propogate_cond_constants -> temporary changes the item tree with constants -> JOIN::execute -> cleanup -> the changes done by propogate_cond_constants are recovered, as they were registered for recovery. mysql-test/r/ps.result: Bug#5987: test results fixed. mysql-test/t/ps.test: A test for bug#5987 "subselect in bool function crashes server (prepared statements)" sql/item.cc: resolve_const_item is now responsible to register all changes of the item tree for recovery sql/item.h: resolve_const_item signagture changed sql/item_cmpfunc.h: Arguments of boolean functions are now recovered using the centralized registry of THD. sql/sql_class.cc: It's crucial to add new items to the beginning of the recovery list, so that the recovery is performed in LIFO mode: otherwise if we change one node of a tree twice, it will be recovered to some intermediate state. sql/sql_select.cc: change_cond_ref_to_const and propogate_cond_constants are now responsible to register all changes of the item tree for recovery. The recovery is done using the centralized THD registry of changed tree items.
Showing
Please register or sign in to comment