MDEV-30225 RR isolation violation with locking unique search

Before the fix next-key lock was requested only if a record was
delete-marked for locking unique search in RR isolation level.
There can be several delete-marked records for the same unique key,
that's why InnoDB scans the records until eighter non-delete-marked record
is reached or all delete-marked records with the same unique key are
scanned.

For range scan next-key locks are used for RR to protect scanned range from
inserting new records by other transactions. And this is the reason of why
next-key locks are used for delete-marked records for unique searches.

If a record is not delete-marked, the requested lock type was "not-gap".
When a record is not delete-marked during lock request by trx 1, and
some other transaction holds conflicting lock, trx 1 creates waiting
not-gap lock on the record and suspends. During trx 1 suspending the
record can be delete-marked. And when the lock is granted on conflicting
transaction commit or rollback, its type is still "not-gap". So we have
"not-gap" lock on delete-marked record for RR. And this let some other
transaction to insert some record with the same unique key when trx 1 is
not committed, what can cause isolation level violation.

The fix is to set next-key locks for both delete-marked and
non-delete-marked records for unique search in RR.

mysql-test/suite/innodb/r/cursor-restore-locking.result

@@ -11,6 +11,7 @@ SET DEBUG_SYNC = 'lock_wait_suspend_thread_enter SIGNAL first_ins_locked';
 SET DEBUG_SYNC = 'ib_after_row_insert SIGNAL first_ins_row_inserted WAIT_FOR first_ins_cont';
 INSERT INTO t VALUES(10, 20);
 connect con_del_2,localhost,root,,;
+SET TRANSACTION ISOLATION LEVEL READ COMMITTED;
 SET DEBUG_SYNC = 'now WAIT_FOR first_ins_locked';
 SET DEBUG_SYNC = 'lock_wait_suspend_thread_enter SIGNAL second_del_locked';
 DELETE FROM t WHERE b = 20;

mysql-test/suite/innodb/r/insert-before-delete.result

+connect  pause_purge,localhost,root;
+START TRANSACTION WITH CONSISTENT SNAPSHOT;
+connection default;
+CREATE TABLE t (pk int PRIMARY KEY, sk INT UNIQUE) ENGINE=InnoDB;
+INSERT INTO t VALUES (10, 100);
+connect  con1,localhost,root;
+BEGIN;
+SELECT * FROM t WHERE sk = 100 FOR UPDATE;
+pk	sk
+10	100
+connect  con2,localhost,root;
+SET DEBUG_SYNC="lock_wait_suspend_thread_enter SIGNAL insert_wait_started";
+INSERT INTO t VALUES (5, 100) # trx 1;
+connect  con3,localhost,root;
+SET TRANSACTION ISOLATION LEVEL REPEATABLE READ;
+SET DEBUG_SYNC="now WAIT_FOR insert_wait_started";
+SET DEBUG_SYNC="lock_wait_suspend_thread_enter SIGNAL delete_started_waiting";
+DELETE FROM t WHERE sk = 100 # trx 2;
+connection con1;
+SET DEBUG_SYNC="now WAIT_FOR delete_started_waiting";
+DELETE FROM t WHERE sk=100;
+COMMIT;
+disconnect con1;
+connection con2;
+disconnect con2;
+connection con3;
+ERROR 40001: Deadlock found when trying to get lock; try restarting transaction
+disconnect con3;
+connection default;
+SELECT * FROM t;
+pk	sk
+5	100
+disconnect pause_purge;
+SET DEBUG_SYNC="RESET";
+DROP TABLE t;

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

@@ -673,7 +673,7 @@ SET @end = (SELECT COUNT FROM INFORMATION_SCHEMA.INNODB_METRICS WHERE NAME
 = 'lock_rec_lock_created');
 SELECT @end - @start;
 @end - @start
-0
+1
 DROP TABLE t1;
 SET GLOBAL innodb_monitor_enable=default;
 SET GLOBAL innodb_monitor_disable=default;

mysql-test/suite/innodb/t/cursor-restore-locking.test

@@ -27,6 +27,10 @@ SET DEBUG_SYNC = 'ib_after_row_insert SIGNAL first_ins_row_inserted WAIT_FOR fir
 --send INSERT INTO t VALUES(10, 20)
 
 --connect(con_del_2,localhost,root,,)
+# After MDEV-30225 is fixed, the following DELETE creates next-key lock for
+# unqique search for RR, and the above INSERT kills it as deadlock victim.
+# But it still requests not-gap lock for RC.
+SET TRANSACTION ISOLATION LEVEL READ COMMITTED;
 SET DEBUG_SYNC = 'now WAIT_FOR first_ins_locked';
 SET DEBUG_SYNC = 'lock_wait_suspend_thread_enter SIGNAL second_del_locked';
 ###############################################################################

mysql-test/suite/innodb/t/insert-before-delete.test

+--source include/have_innodb.inc
+--source include/have_debug.inc
+--source include/have_debug_sync.inc
+--source include/count_sessions.inc
+
+--connect (pause_purge,localhost,root)
+START TRANSACTION WITH CONSISTENT SNAPSHOT;
+
+--connection default
+CREATE TABLE t (pk int PRIMARY KEY, sk INT UNIQUE) ENGINE=InnoDB;
+INSERT INTO t VALUES (10, 100);
+
+--connect (con1,localhost,root)
+BEGIN; # trx 0
+SELECT * FROM t WHERE sk = 100 FOR UPDATE;
+
+--connect (con2,localhost,root)
+SET DEBUG_SYNC="lock_wait_suspend_thread_enter SIGNAL insert_wait_started";
+# trx 1 is locked on try to read the record in secondary index during duplicates
+# check. It's the first in waiting queue, that's why it will be woken up firstly
+# when trx 0 commits.
+--send INSERT INTO t VALUES (5, 100) # trx 1
+
+--connect (con3,localhost,root)
+# MDEV-30225 is fixed only for RR
+SET TRANSACTION ISOLATION LEVEL REPEATABLE READ;
+SET DEBUG_SYNC="now WAIT_FOR insert_wait_started";
+SET DEBUG_SYNC="lock_wait_suspend_thread_enter SIGNAL delete_started_waiting";
+# trx 2 can delete (5, 100) on master, but not on slave, as on slave trx 1
+# can insert (5, 100) after trx 2 positioned it's cursor. Trx 2 lock is placed
+# in waiting queue after trx 1 lock, but its persistent cursor position was
+# stored on (100, 10) record in secondary index before suspending. After trx 1
+# is committed, trx 2 will restore persistent cursor position on (100, 10). As
+# (100, 5) secondary index record was inserted before (100, 10) in logical
+# order, and (100, 10) record is delete-marked, trx 2 just continues scanning.
+#
+# Note. There can be several records with the same key in unique secondary
+# index, but only one of them must be non-delete-marked. That's why when we do
+# point query, cursor position is set in the first record in logical order, and
+# then records are iterated until either non-delete-marked record is found or
+# all records with the same unique fields are iterated.
+--send DELETE FROM t WHERE sk = 100 # trx 2
+
+--connection con1
+SET DEBUG_SYNC="now WAIT_FOR delete_started_waiting";
+DELETE FROM t WHERE sk=100; # trx 0
+COMMIT;
+--disconnect con1
+
+--connection con2
+--reap
+--disconnect con2
+
+--connection con3
+# If the bug is fixed, deadlock error will be there, as trx 2 owns
+# next-key lock waiting for trx 1, and trx 1 requests
+# insert-intention lock, conflicting with trx 2 next-key lock.
+--error ER_LOCK_DEADLOCK
+--reap
+--disconnect con3
+
+--connection default
+# If the bug is not fixed, we will see the row inserted by trx 1 here. This can
+# cause duplicate key error on slave, when some other trx tries in insert row
+# with the same secondary key, as was inserted by trx 1, and not deleted by trx
+# 2.
+SELECT * FROM t;
+
+--disconnect pause_purge
+SET DEBUG_SYNC="RESET";
+DROP TABLE t;
+--source include/wait_until_count_sessions.inc

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

@@ -435,6 +435,9 @@ INSERT INTO t1 VALUES(1,1,'a'),(2,9999,'b'),(3,10000,'c'),(4,4,'d');
 DELETE FROM t1 WHERE a = 9999 AND b='b';
 COMMIT;
 
+# After MDEV-30225 is fixed, the above DELETE creates next-key lock during
+# secondary index unique search. That's why the result of the following must
+# be 1.
 SET @end = (SELECT COUNT FROM INFORMATION_SCHEMA.INNODB_METRICS WHERE NAME
 = 'lock_rec_lock_created');
 SELECT @end - @start;

storage/innobase/row/row0sel.cc

@@ -5114,8 +5114,10 @@ row_search_mvcc(
 			goto no_gap_lock;
 		}
 
+		/* Set next-key lock both for delete- and non-delete-marked
+		records for unique search, because non-delete-marked record can
+		be marked as deleted while transaction suspends. */
 		if (!set_also_gap_locks
-		    || (unique_search && !rec_get_deleted_flag(rec, comp))
 		    || dict_index_is_spatial(index)) {
 
 			goto no_gap_lock;