Bug#32416819 ASSERTION "UT_LIST_GET_LEN.TRX_SYS->MYSQL_TRX_LIST. == 0"

ON SHUTDOWN

Note: This commit in 8.0 is different from 8.4 and trunk due to testcase

Background:
-----------
  The trx_sys->mysql_trx_list maintains a list of trx_t objects created
  for user transactions. When the `USE <database>` statement is run, a
  trx_t object is created for the THD running this query and added to
  this list. When the connection is closed, the trx_t object of the THD
  is freed and removed from this list.

  During shutdown, all the connections are closed before trx_sys is
  closed. When closing trx_sys, it is expected that all trx_t objects
  in trx_sys->mysql_trx_list are freed and the list is empty.

  In replication, when the applier sees an `XA START`, it detaches the
  THD from the handlertons by storing a backup in the engine's ha_data.
  In case of InnoDB, the trx_t created for the THD is stored in ha_data
  and is made nullptr; but not removed from the mysql_trx_list.

  In replication, the detached engine ha_data is reattached after the
  XA transaction is processed. During the reattach, the trx_t is
  restored from the ha_data back to the THD.

  By design, the detach happens when `XA START` is encountered by the
  applier. When the applier sees an `XA PREPARE`, the reattach is done
  via `applier_reset_xa_trans` calling `attach_native_trx`. When the
  applier sees an `XA COMMIT .. ONE PHASE`, the reattach is done in
  `ha_commit_low`. If the server is shutdown, the applier performs the
  reattach in `ha_rollback_low`.

  The binlog applier can be simulated by enabling `pseudo_replica_mode`
  and running a `BINLOG '0'`.

Issue:
------
  When the XA transaction is empty, the ha_list is nullptr. The reattach
  logic is performed only when ha_list is non-empty. Hence, if the
  server is shutdown after the applier detaches the engine (i.e, after
  running an `XA START`), the applier would never reattach the ha_data
  which leads to the assert when InnoDB closes trx_sys. Due to the
  design, the problem is seen only with the reattach in `ha_commit_low`
  and `ha_rollback_low`.

Fix:
----
  Ensured that reattach is done even if ha_list is empty.

Change-Id: Ida617d067d112914c9d55ef4cde9f91956d19a08

mysql-test/include/xa_applier_shutdown.inc

@@ -0,0 +1,143 @@
+# This test checks the behavior of applier thread when shutdown is issued at
+# various stages of the XA transaction. The session thread is marked as an
+# applier thread by setting the pseudo_replica_mode and executing BINLOG''
+# query.
+#
+# References:
+#     Bug#32416819: ASSERTION "UT_LIST_GET_LEN.TRX_SYS->MYSQL_TRX_LIST. == 0" ON
+#     SHUTDOWN
+#
+# The include file lists the scenarios to be checked before running shutdown.
+# Enable the following flags to check different scenarios:
+# $create_table:       Create the table for the test and drop it at the end
+# $insert_table:       Insert 3 rows into the table containing single int col
+# $need_xa_end:        Includes the XA END statement
+# $need_xa_commit_one: Includes the XA COMMIT..ONE PHASE statement followed by
+#                      the previous set of statements
+# $need_xa_prepare:    Includes the XA PREPARE statement
+# $need_xa_commit:     Includes the XA COMMIT statement
+# $need_xa_rollback:   Includes the XA ROLLBACK statement
+# $restart_connection: If set, create new connection and restart the server
+#                      using the new connection
+#                      If unset, restart the server on the connection that
+#                      executes the transaction
+#
+# Note:
+# During shutdown, Innodb closes trx_sys. At this stage, it is expected that all
+# the trx_t's created for mysql is freed (mysql_trx_list - list of trx_t created
+# for mysql is empty). Whenever a `USE <database>` command is issued, Innodb
+# will create a trx_t (associated with the corresponding session THD) and adds
+# it to mysql_trx_list.
+#
+# During replication, if the replication applier sees an 'XA START', it will
+# call detach_native_trx on all handlertons. This function calls
+# replace_native_transaction_in_thd by passing nullptr as the new value and
+# ha_ptr_backup as the buffer to store the handlerton's THD context. In case of
+# InnoDB, the THD context is the trx_t mentioned above.
+#
+# This context is restored at 3 different places depending on the XA transaction
+# 1. In 'XA PREPARE' using attach_native_trx
+# 2. In 'XA COMMIT .. ONE PHASE' using ha_commit_low
+# 3. In 'XA ROLLBACK' or shutdown using ha_rollback_low
+#
+# In the bug scenario, the restore was done only when the ha_list was non-empty.
+# Hence, if shutdown is issued in middle of processing an empty XA transaction
+# (right before XA PREPARE / XA COMMIT .. ONE PHASE), then the reattach step is
+# skipped. Thus the trx_t is never returned, leaving the mysql_trx_list
+# non-empty when closing trx_sys leading to the assert.
+#
+# In case of empty transaction, 'XA COMMIT .. ONE PHASE' does not call reattach.
+# Hence an 'XA START' after this will attempt to detach again, leading to an
+# assertion in the detach function.
+#
+# The current connection can simulate a replication applier by issuing:
+#     SET @@SESSION.pseudo_replica_mode=1;
+#     BINLOG '0';
+#
+# The BINLOG statement will initialize thd->rli_fake, even though the statement
+# fails. The thd->rli_fake will simulate the replication applier. '0' is used
+# for better readability.
+
+if ($create_table) {
+  CREATE TABLE t1 (c1 INT);
+  INSERT INTO t1 VALUES (10);
+}
+
+SET @@SESSION.pseudo_replica_mode=1;
+--error ER_BASE64_DECODE_ERROR
+BINLOG '0';
+
+XA START 'test0';
+
+if ($insert_table) {
+  INSERT INTO t1 VALUES (1);
+  INSERT INTO t1 VALUES (2);
+  INSERT INTO t1 VALUES (3);
+}
+
+if ($need_xa_end) {
+  XA END 'test0';
+}
+
+if ($need_xa_commit_one) {
+  XA COMMIT 'test0' ONE PHASE;
+
+  XA START 'test0';
+  if ($insert_table) {
+    INSERT INTO t1 VALUES (4);
+    INSERT INTO t1 VALUES (5);
+    INSERT INTO t1 VALUES (6);
+  }
+
+  if ($need_xa_end) {
+    XA END 'test0';
+  }
+}
+
+if ($need_xa_prepare) {
+  XA PREPARE 'test0';
+}
+
+XA RECOVER;
+
+if ($restart_connection) {
+  # When restarting from a different connection, default connection must be
+  # handled separately. In this case, the default connection simply waits until
+  # it can reconnect after the server starts.
+
+  --connection default
+  --disable_reconnect
+
+  --connect ($restart_connection, localhost, root,,)
+  --connection $restart_connection
+  --source include/restart_mysqld.inc
+  --disconnect $restart_connection
+
+  --connection default
+  --enable_reconnect
+  --source include/wait_until_connected_again.inc
+}
+
+if (!$restart_connection) {
+  # The `connection default` is used to indicate that restart is issued in the
+  # current connection
+
+  --connection default
+  --source include/restart_mysqld.inc
+  --connection default
+}
+
+XA RECOVER;
+
+if ($need_xa_rollback) {
+  XA ROLLBACK 'test0';
+}
+
+if ($need_xa_commit) {
+  XA COMMIT 'test0';
+}
+
+if ($create_table) {
+  SELECT * FROM t1 ORDER BY c1;
+  DROP TABLE t1;
+}