| 1 |
# |
| 2 |
# MDEV-33133: MDL conflict handling code should skip transactions |
| 3 |
# BF-aborted before. |
| 4 |
# |
| 5 |
# It's possible that MDL conflict handling code is called more |
| 6 |
# than once for a transaction when: |
| 7 |
# - it holds more than one conflicting MDL lock |
| 8 |
# - reschedule_waiters() is executed, |
| 9 |
# which results in repeated attempts to BF-abort already aborted |
| 10 |
# transaction. |
| 11 |
# In such situations, it might be that BF-aborting logic sees |
| 12 |
# a partially rolled back transaction and erroneously decides |
| 13 |
# on future actions for such a transaction. |
| 14 |
# |
| 15 |
# The specific situation tested and fixed is when a SR transaction |
| 16 |
# applied in the node gets BF-aborted by a started TOI operation. |
| 17 |
# It's then caught with the server transaction already rolled back, |
| 18 |
# but with no MDL locks yet released. This caused wrong state |
| 19 |
# detection for such a transaction during repeated MDL conflict |
| 20 |
# handling code execution. |
| 21 |
# |
| 22 |
|
| 23 |
--source include/galera_cluster.inc |
| 24 |
--source include/have_debug_sync.inc |
| 25 |
--source include/have_debug.inc |
| 26 |
|
| 27 |
--connect node_1a,127.0.0.1,root,,test,$NODE_MYPORT_1 |
| 28 |
|
| 29 |
--connection node_1 |
| 30 |
CREATE TABLE t1 (f1 INTEGER PRIMARY KEY) ENGINE=InnoDB; |
| 31 |
SET GLOBAL DEBUG_DBUG = 'd,sync.wsrep_rollback_mdl_release'; |
| 32 |
|
| 33 |
--connection node_2 |
| 34 |
SET SESSION wsrep_trx_fragment_size = 1; |
| 35 |
START TRANSACTION; |
| 36 |
INSERT INTO t1 VALUES (1); |
| 37 |
|
| 38 |
--connection node_1a |
| 39 |
# Sync wait for SR transaction to replicate and apply fragment. |
| 40 |
SELECT COUNT(*) FROM t1; |
| 41 |
SET SESSION wsrep_retry_autocommit = 0; |
| 42 |
SET DEBUG_SYNC = 'ha_write_row_start SIGNAL may_toi WAIT_FOR bf_abort'; |
| 43 |
--send |
| 44 |
INSERT INTO t1 VALUES (2); |
| 45 |
|
| 46 |
--connection node_1 |
| 47 |
SET DEBUG_SYNC = 'now WAIT_FOR may_toi'; |
| 48 |
# BF-abort SR transaction and wait until it reaches the point |
| 49 |
# prior to release MDL locks. |
| 50 |
# Then abort local INSERT, which will go through rescedule_waiters() |
| 51 |
# and see SR transaction holding MDL locks but already rolled back. |
| 52 |
# In this case SR transaction should be skipped in MDL conflict |
| 53 |
# handling code. |
| 54 |
SET DEBUG_SYNC = 'after_wsrep_thd_abort WAIT_FOR sync.wsrep_rollback_mdl_release_reached'; |
| 55 |
--send |
| 56 |
TRUNCATE TABLE t1; |
| 57 |
|
| 58 |
--connection node_1a |
| 59 |
# Local INSERT gets aborted. |
| 60 |
--error ER_LOCK_DEADLOCK |
| 61 |
--reap |
| 62 |
# Let the aborted SR transaction continue and finally release MDL locks, |
| 63 |
# which in turn allows TRUNCATE to complete. |
| 64 |
SET DEBUG_SYNC = 'now SIGNAL signal.wsrep_rollback_mdl_release'; |
| 65 |
|
| 66 |
--connection node_2 |
| 67 |
# SR transaction has been BF-aborted. |
| 68 |
--error ER_LOCK_DEADLOCK |
| 69 |
INSERT INTO t1 VALUES (3); |
| 70 |
|
| 71 |
--connection node_1 |
| 72 |
# TRUNCATE completes. |
| 73 |
--reap |
| 74 |
|
| 75 |
# Cleanup |
| 76 |
SET GLOBAL DEBUG_DBUG = ''; |
| 77 |
SET DEBUG_SYNC = 'RESET'; |
| 78 |
DROP TABLE t1; |
| 79 |
--disconnect node_1a |
| 80 |
--source include/galera_end.inc |