文 | 杨一 on 运维

转 | 来源：公众号yangyidba

一、前言

死锁，其实是一个很有意思也很有挑战的技术问题，大概每个 DBA 和部分开发同学都会在工作过程中遇见 。关于死锁我会持续写一个系列的案例分析，希望能够对想了解死锁的朋友有所帮助。本文介绍一例三个并发 insert 导致的死锁，根本原因还是在于 insert 唯一键申请插入意向锁这个特殊的 GAP 锁。其实称呼插入意向锁 为 Insert Intention Gap Lock 更为合理。

二、案例分析

2.1 环境准备

Percona server 5.6 RR模式

1. CREATE TABLE `t6` (

2.  `id` int(11) NOT NULL AUTO_INCREMENT,

3.  `a` int(11) DEFAULT NULL,

4.  PRIMARY KEY (`id`),

5.  unique KEY `idx_a` (`a`)

6. ) ENGINE=InnoDB AUTO_INCREMENT=1 DEFAULT CHARSET=utf8mb4;

7. 
   

8. insert into t6 values(1,2),(2,8),(3,9),(4,11),(5,19);

2.2 死锁日志

1. ------------------------

2. LATEST DETECTED DEADLOCK

3. ------------------------

4. 2017-09-18 10:03:50 7f78eae30700

5. *** (1) TRANSACTION:

6. TRANSACTION 462308725, ACTIVE 18 sec inserting, thread declared inside InnoDB 1

7. mysql tables in use 1, locked 1

8. LOCK WAIT 4 lock struct(s), heap size 1184, 2 row lock(s), undo log entries 1

9. MySQL thread id 3825465, OS thread handle 0x7f78eaef4700, query id 781148519 localhost root update

10. insert into t6(id,a) values(7,15)

11. *** (1) WAITING FOR THIS LOCK TO BE GRANTED:

12. RECORD LOCKS space id 227 page no 4 n bits 80 index `idx_a` of table `test`.`t6` trx id 462308725 lock_mode X insert intention waiting

13. *** (2) TRANSACTION:

14. TRANSACTION 462308726, ACTIVE 10 sec inserting, thread declared inside InnoDB 1

15. mysql tables in use 1, locked 1

16. 4 lock struct(s), heap size 1184, 2 row lock(s), undo log entries 1

17. MySQL thread id 3825581, OS thread handle 0x7f78eae30700, query id 781148528 localhost root update

18. insert into t6(id,a) values(8,15)

19. *** (2) HOLDS THE LOCK(S):

20. RECORD LOCKS space id 227 page no 4 n bits 80 index `idx_a` of table `test`.`t6` trx id 462308726 lock mode S

21. *** (2) WAITING FOR THIS LOCK TO BE GRANTED:

22. RECORD LOCKS space id 227 page no 4 n bits 80 index `idx_a` of table `test`.`t6` trx id 462308726 lock_mode X insert intention waiting

23. *** WE ROLL BACK TRANSACTION (2)

2.3 死锁分析

首先依然要再次强调 insert 插入操作的加锁逻辑。

第一阶段：唯一性约束检查，先申请 LOCK_INSERT_INTENTION

第二阶段：获取阶段一的锁并且 insert 成功之后,插入的位置有 Gap 锁：LOCK_S + LOCK_ORDINARY，为了防止其他 insert 唯一键冲突。

新数据插入：LOCK_X + LOCK_REC_NOT_GAP

对于 insert 操作来说，若发生唯一约束冲突，则需要对冲突的唯一索引加上 S Next-key Lock。从这里会发现，即使是 RC 事务隔离级别，也同样会存在 Next-Key Lock锁，从而阻塞并发。然而，文档没有说明的是，对于检测到冲突的唯一索引，等待线程在获得 S Lock 之后，还需要对下一个记录进行加锁，在源码中由函数 row_ins_scan_sec_index_for_duplicate 进行判断。

其次我们需要了解锁的兼容性矩阵。

从兼容性矩阵我们可以得到如下结论：

1. INSERT操作之间不会有冲突。

2. GAP,Next-Key会阻止Insert。

3. GAP和Record,Next-Key不会冲突

4. Record和Record、Next-Key之间相互冲突。

已有的 Insert 锁不阻止任何准备加的锁。

这个案例是三个会话并发执行的，我打算一步一步来分析每个步骤执行完之后的事务日志。

第一步 sess1 执行插入操作 insert into t6(id,a) values(6,15);

1. ---TRANSACTION 462308737, ACTIVE 5 sec

2. 1 lock struct(s), heap size 360, 0 row lock(s), undo log entries 1

3. MySQL thread id 3825779, OS thread handle 0x7f78eacd9700, query id 781149440 localhost root init

4. show engine innodb status

5. TABLE LOCK table `test`.`t6` trx id 462308737 lock mode IX

因为第一个插入的语句，所以唯一性冲突检查通过，成功插入(6,15)。此时 sess1 会话持有(6,15)的 LOCK_X|LOCK_REC_NOT_GAP 锁。参考:

"INSERT sets an exclusive lock on the inserted row. This lock is an index-record lock, not a next-key lock (that is, there is no gap lock) and does not prevent other sessions from inserting into the gap before the inserted row."

第二步 sess2 执行插入操作

insert into t6(id,a) values(7,15);

1. ---TRANSACTION 462308738, ACTIVE 4 sec inserting

2. mysql tables in use 1, locked 1

3. LOCK WAIT 2 lock struct(s), heap size 360, 1 row lock(s), undo log entries 1

4. MySQL thread id 3825768, OS thread handle 0x7f78ea9c9700, query id 781149521 localhost root update

5. insert into t6(id,a) values(7,15)

6. ------- TRX HAS BEEN WAITING 4 SEC FOR THIS LOCK TO BE GRANTED:

7. RECORD LOCKS space id 227 page no 4 n bits 80 index `idx_a` of table `test`.`t6` trx id 462308738 lock more S waiting

8. ------------------

9. TABLE LOCK table `test`.`t6` trx id 462308738 lock mode IX

10. RECORD LOCKS space id 227 page no 4 n bits 80 index `idx_a` of table `test`.`t6` trx id 462308738 lock mode S waiting

11. ---TRANSACTION 462308737, ACTIVE 66 sec

12. 2 lock struct(s), heap size 360, 1 row lock(s), undo log entries 1

13. MySQL thread id 3825779, OS thread handle 0x7f78eacd9700, query id 781149526 localhost root init

14. show engine innodb status

15. TABLE LOCK table `test`.`t6` trx id 462308737 lock mode IX

16. RECORD LOCKS space id 227 page no 4 n bits 80 index `idx_a` of table `test`.`t6` trx id 462308737 lock_mode X locks rec but not gap

首先 sess2 的 insert 申请了 IX 锁，因为 sess1 会话已经插入成功并且持有唯一键 a=15 的 X 行锁 ，故而 sess2 insert 进行唯一性检查，先申请 LOCK_S + LOCK_ORDINARY ，事务日志列表中提示 lock mode S waiting。

第三步 sess3 执行插入操作 insert into t6(id,a) values(8,15);

1. ---TRANSACTION 462308739, ACTIVE 3 sec inserting

2. mysql tables in use 1, locked 1

3. LOCK WAIT 2 lock struct(s), heap size 360, 1 row lock(s), undo log entries 1

4. MySQL thread id 3825764, OS thread handle 0x7f78ea593700, query id 781149555 localhost root update

5. insert into t6(id,a) values(8,15)

6. ------- TRX HAS BEEN WAITING 3 SEC FOR THIS LOCK TO BE GRANTED:

7. RECORD LOCKS space id 227 page no 4 n bits 80 index `idx_a` of table `test`.`t6` trx id 462308739 lock mode S waiting

8. ------------------

9. TABLE LOCK table `test`.`t6` trx id 462308739 lock mode IX

10. RECORD LOCKS space id 227 page no 4 n bits 80 index `idx_a` of table `test`.`t6` trx id 462308739 lock mode S waiting

11. ---TRANSACTION 462308738, ACTIVE 35 sec inserting

12. mysql tables in use 1, locked 1

13. LOCK WAIT 2 lock struct(s), heap size 360, 1 row lock(s), undo log entries 1

14. MySQL thread id 3825768, OS thread handle 0x7f78ea9c9700, query id 781149521 localhost root update

15. insert into t6(id,a) values(7,15)

16. ------- TRX HAS BEEN WAITING 35 SEC FOR THIS LOCK TO BE GRANTED:

17. RECORD LOCKS space id 227 page no 4 n bits 80 index `idx_a` of table `test`.`t6` trx id 462308738 lock mode S waiting

18. ------------------

19. TABLE LOCK table `test`.`t6` trx id 462308738 lock mode IX

20. RECORD LOCKS space id 227 page no 4 n bits 80 index `idx_a` of table `test`.`t6` trx id 462308738 lock mode S waiting

21. ---TRANSACTION 462308737, ACTIVE 97 sec

22. 2 lock struct(s), heap size 360, 1 row lock(s), undo log entries 1

23. MySQL thread id 3825779, OS thread handle 0x7f78eacd9700, query id 781149560 localhost root init

24. show engine innodb status

25. TABLE LOCK table `test`.`t6` trx id 462308737 lock mode IX

26. RECORD LOCKS space id 227 page no 4 n bits 80 index `idx_a` of table `test`.`t6` trx id 462308737 lock_mode X locks rec but not gap

与会话 sess2 的加锁申请流程一致，都在等待 sess1 释放锁资源。

第四步 sess1 执行回滚操作，sess2 不提交sess1 rollback;

此时 sess2 插入成功，sess3 出现死锁，此时 sess2 insert 插入成功，还未提交，事务列表如下：

1. ------------

2. TRANSACTIONS

3. ------------

4. Trx id counter 462308744

5. Purge done for trx s n:o < 462308744 undo n:o < 0 state: running but idle

6. History list length 1866

7. LIST OF TRANSACTIONS FOR EACH SESSION:

8. ---TRANSACTION 462308737, not started

9. MySQL thread id 3825779, OS thread handle 0x7f78eacd9700, query id 781149626 localhost root init

10. show engine innodb status

11. ---TRANSACTION 462308739, not started

12. MySQL thread id 3825764, OS thread handle 0x7f78ea593700, query id 781149555 localhost root cleaning up

13. ---TRANSACTION 462308738, ACTIVE 75 sec

14. 5 lock struct(s), heap size 1184, 3 row lock(s), undo log entries 1

15. MySQL thread id 3825768, OS thread handle 0x7f78eadce700, query id 781149608 localhost root cleaning up

16. TABLE LOCK table `test`.`t6` trx id 462308738 lock mode IX

17. RECORD LOCKS space id 227 page no 4 n bits 80 index `idx_a` of table `test`.`t6` trx id 462308738 lock mode S

18. RECORD LOCKS space id 227 page no 4 n bits 80 index `idx_a` of table `test`.`t6` trx id 462308738 lock mode S

19. RECORD LOCKS space id 227 page no 4 n bits 80 index `idx_a` of table `test`.`t6` trx id 462308738 lock_mode X insert intention

20. RECORD LOCKS space id 227 page no 4 n bits 80 index `idx_a` of table `test`.`t6` trx id 462308738 lock mode S locks gap before rec

死锁的原因

sess1 insert成功并针对a=15的唯一键加上X锁。

sess2 执行insert 插入(6,15), 在插入之前进行唯一性检查发现和sess1的已经插入的记录重复键需要申请LOCK_S|LOCK_ORDINARY, 但与sess1 的(LOCK_X | LOCK_REC_NOT_GAP)冲突,加入等待队列,等待sess1 释放锁。

sess3 执行insert 插入(7,15), 在插入之前进行唯一性检查发现和sess1的已经插入的记录重复键需要申请LOCK_S|LOCK_ORDINARY, 但与sess1 的(LOCK_X | LOCK_REC_NOT_GAP)冲突,加入等待队列,等待sess1 释放锁。

sess1 执行rollback, sess1 释放索引a=15 上的排他记录锁(LOCK_X | LOCK_REC_NOT_GAP),此后 sess2和sess3 获得S锁(LOCK_S|LOCK_ORDINARY)成功,sess2和sess3都要请求索引a=15上的排他记录锁(LOCK_X | LOCK_REC_NOT_GAP),日志中提示 lock_mode X insert intention。由于X锁与S锁互斥，

sess2和sess3都等待对方释放S锁，于是出现死锁，MySQL 选择回滚其中之一。

四、总结

死锁分析是一件很有挑战的事情，尤其对于 insert 唯一键冲突,要分多个阶段去申请，也要理解锁的兼容矩阵。对于这块我还有需要在学习了解的知识点，本文算是抛砖引玉，如有分析理解不正确的地方，望大家指正。

扩展阅读

1. 漫谈死锁

2. 如何阅读死锁日志

3. 死锁案例一

4. 死锁案例二

5. 死锁案例三

-The End-

Vol.155