MDEV-17158 TRUNCATE is not atomic after MDEV-13564

mysql-test/suite/innodb/disabled.def

@@ -11,4 +11,3 @@
 ##############################################################################
 
 create-index-debug : MDEV-13680 InnoDB may crash when btr_page_alloc() fails
-truncate_crash : MDEV-17158 log_write_up_to() sometimes fails

mysql-test/suite/innodb/r/truncate_crash.result

@@ -1,14 +1,14 @@
 FLUSH TABLES;
 CREATE TABLE t1 (a INT PRIMARY KEY) ENGINE=InnoDB;
-INSERT INTO t1 SET a=1;
+INSERT INTO t1 VALUES (1),(2);
 connect  wait,localhost,root,,test;
-SET DEBUG_SYNC='after_trx_committed_in_memory SIGNAL c WAIT_FOR ever';
+SET DEBUG_SYNC='before_trx_state_committed_in_memory SIGNAL c WAIT_FOR ever';
 TRUNCATE TABLE t1;
 connection default;
 SET DEBUG_SYNC='now WAIT_FOR c';
 disconnect wait;
-SELECT * FROM t1;
-a
-1
+SELECT COUNT(*) FROM t1;
+COUNT(*)
+0
 TRUNCATE TABLE t1;
 DROP TABLE t1;

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

@@ -5,10 +5,10 @@
 
 FLUSH TABLES;
 CREATE TABLE t1 (a INT PRIMARY KEY) ENGINE=InnoDB;
-INSERT INTO t1 SET a=1;
+INSERT INTO t1 VALUES (1),(2);
 
 connect (wait,localhost,root,,test);
-SET DEBUG_SYNC='after_trx_committed_in_memory SIGNAL c WAIT_FOR ever';
+SET DEBUG_SYNC='before_trx_state_committed_in_memory SIGNAL c WAIT_FOR ever';
 send TRUNCATE TABLE t1;
 
 connection default;
@@ -17,6 +17,7 @@ SET DEBUG_SYNC='now WAIT_FOR c';
 --source include/restart_mysqld.inc
 disconnect wait;
 
-SELECT * FROM t1;
+--replace_result 2 0
+SELECT COUNT(*) FROM t1;
 TRUNCATE TABLE t1;
 DROP TABLE t1;

storage/innobase/dict/dict0crea.cc

@@ -38,6 +38,8 @@ Created 1/8/1996 Heikki Tuuri
 #include "row0mysql.h"
 #include "pars0pars.h"
 #include "trx0roll.h"
+#include "trx0rseg.h"
+#include "trx0undo.h"
 #include "ut0vec.h"
 #include "dict0priv.h"
 #include "fts0priv.h"
@@ -370,7 +372,33 @@ dict_build_table_def_step(
 
 		ut_ad(DICT_TF_GET_ZIP_SSIZE(table->flags) == 0
 		      || dict_table_has_atomic_blobs(table));
-
+		trx_t* trx = thr_get_trx(thr);
+		ut_ad(trx->table_id);
+		mtr_t mtr;
+		trx_undo_t* undo = trx->rsegs.m_redo.undo;
+		if (undo && !undo->table_id
+		    && trx_get_dict_operation(trx) == TRX_DICT_OP_TABLE) {
+			/* This must be a TRUNCATE operation where
+			the empty table is created after the old table
+			was renamed. Be sure to mark the transaction
+			associated with the new empty table, so that
+			we can remove it on recovery. */
+			mtr.start();
+			undo->table_id = trx->table_id;
+			undo->dict_operation = TRUE;
+			page_t* page = trx_undo_page_get(
+				page_id_t(trx->rsegs.m_redo.rseg->space->id,
+					  undo->hdr_page_no),
+				&mtr);
+			mlog_write_ulint(page + undo->hdr_offset
+					 + TRX_UNDO_DICT_TRANS,
+					 TRUE, MLOG_1BYTE, &mtr);
+			mlog_write_ull(page + undo->hdr_offset
+				       + TRX_UNDO_TABLE_ID,
+				       trx->table_id, &mtr);
+			mtr.commit();
+			log_write_up_to(mtr.commit_lsn(), true);
+		}
 		/* Get a new tablespace ID */
 		ulint space_id;
 		dict_hdr_get_new_id(NULL, NULL, &space_id, table, false);
@@ -381,7 +409,7 @@ dict_build_table_def_step(
 		);
 
 		if (space_id == ULINT_UNDEFINED) {
-			return(DB_ERROR);
+			return DB_ERROR;
 		}
 
 		/* Determine the tablespace flags. */
@@ -416,7 +444,6 @@ dict_build_table_def_step(
 		}
 
 		table->space_id = space_id;
-		mtr_t mtr;
 		mtr.start();
 		mtr.set_named_space(table->space);
 		fsp_header_init(table->space, FIL_IBD_FILE_INITIAL_SIZE, &mtr);
@@ -426,8 +453,6 @@ dict_build_table_def_step(
 		      != REC_FORMAT_COMPRESSED);
 		table->space = fil_system.sys_space;
 		table->space_id = TRX_SYS_SPACE;
-		DBUG_EXECUTE_IF("ib_ddl_crash_during_tablespace_alloc",
-				DBUG_SUICIDE(););
 	}
 
 	ins_node_set_new_row(node->tab_def,

storage/innobase/dict/dict0dict.cc

@@ -1556,9 +1556,14 @@ dict_table_rename_in_cache(
 /*=======================*/
 	dict_table_t*	table,		/*!< in/out: table */
 	const char*	new_name,	/*!< in: new name */
-	ibool		rename_also_foreigns)/*!< in: in ALTER TABLE we want
+	bool		rename_also_foreigns,
+					/*!< in: in ALTER TABLE we want
 					to preserve the original table name
 					in constraints which reference it */
+	bool		replace_new_file)
+					/*!< in: whether to replace the
+					file with the new name
+					(as part of rolling back TRUNCATE) */
 {
 	dberr_t		err;
 	dict_foreign_t*	foreign;
@@ -1658,7 +1663,8 @@ dict_table_rename_in_cache(
 		}
 
 		/* New filepath must not exist. */
-		err = table->space->rename(new_name, new_path, true);
+		err = table->space->rename(new_name, new_path, true,
+					   replace_new_file);
 		ut_free(new_path);
 
 		/* If the tablespace is remote, a new .isl file was created

storage/innobase/fil/fil0fil.cc

@@ -74,12 +74,14 @@ if that the old filepath exists and the new filepath does not exist.
 @param[in]	old_path	old filepath
 @param[in]	new_path	new filepath
 @param[in]	is_discarded	whether the tablespace is discarded
+@param[in]	replace_new	whether to ignore the existence of new_path
 @return innodb error code */
 static dberr_t
 fil_rename_tablespace_check(
 	const char*	old_path,
 	const char*	new_path,
-	bool		is_discarded);
+	bool		is_discarded,
+	bool		replace_new = false);
 /** Rename a single-table tablespace.
 The tablespace must exist in the memory cache.
 @param[in]	id		tablespace identifier
@@ -2866,12 +2868,14 @@ if that the old filepath exists and the new filepath does not exist.
 @param[in]	old_path	old filepath
 @param[in]	new_path	new filepath
 @param[in]	is_discarded	whether the tablespace is discarded
+@param[in]	replace_new	whether to ignore the existence of new_path
 @return innodb error code */
 static dberr_t
 fil_rename_tablespace_check(
 	const char*	old_path,
 	const char*	new_path,
-	bool		is_discarded)
+	bool		is_discarded,
+	bool		replace_new)
 {
 	bool	exists = false;
 	os_file_type_t	ftype;
@@ -2887,25 +2891,58 @@ fil_rename_tablespace_check(
 	}
 
 	exists = false;
-	if (!os_file_status(new_path, &exists, &ftype) || exists) {
+	if (os_file_status(new_path, &exists, &ftype) && !exists) {
+		return DB_SUCCESS;
+	}
+
+	if (!replace_new) {
 		ib::error() << "Cannot rename '" << old_path
 			<< "' to '" << new_path
 			<< "' because the target file exists."
 			" Remove the target file and try again.";
 		return(DB_TABLESPACE_EXISTS);
 	}
 
+	/* This must be during the ROLLBACK of TRUNCATE TABLE.
+	Because InnoDB only allows at most one data dictionary
+	transaction at a time, and because this incomplete TRUNCATE
+	would have created a new tablespace file, we must remove
+	a possibly existing tablespace that is associated with the
+	new tablespace file. */
+retry:
+	mutex_enter(&fil_system.mutex);
+	for (fil_space_t* space = UT_LIST_GET_FIRST(fil_system.space_list);
+	     space; space = UT_LIST_GET_NEXT(space_list, space)) {
+		ulint id = space->id;
+		if (id && id < SRV_LOG_SPACE_FIRST_ID
+		    && space->purpose == FIL_TYPE_TABLESPACE
+		    && !strcmp(new_path,
+			       UT_LIST_GET_FIRST(space->chain)->name)) {
+			ib::info() << "TRUNCATE rollback: " << id
+				<< "," << new_path;
+			mutex_exit(&fil_system.mutex);
+			dberr_t err = fil_delete_tablespace(id);
+			if (err != DB_SUCCESS) {
+				return err;
+			}
+			goto retry;
+		}
+	}
+	mutex_exit(&fil_system.mutex);
+	fil_delete_file(new_path);
+
 	return(DB_SUCCESS);
 }
 
-dberr_t fil_space_t::rename(const char* name, const char* path, bool log)
+dberr_t fil_space_t::rename(const char* name, const char* path, bool log,
+			    bool replace)
 {
 	ut_ad(UT_LIST_GET_LEN(chain) == 1);
 	ut_ad(!is_system_tablespace(id));
 
 	if (log) {
 		dberr_t err = fil_rename_tablespace_check(
-			chain.start->name, path, false);
+			chain.start->name, path, false, replace);
 		if (err != DB_SUCCESS) {
 			return(err);
 		}