{{ item.name }}
{{ item.name }}

{{ it.name }}

{{ it.text }}

{{ it.name }}

{{ innerIt.name }}

{{ innerIt.text }}

{{news.time}}
{{news.title}}
GDB定位MySQL5.7特定版本hang死的故障分析
2020-05-07发布 1,342浏览

问题背景

某环境上有一组Percona MySQL 5.7.23-23的半同步主从,我们采用Prometheus监控框架,按其接口规范自研了独立的exporter用于监控数据采集,类似于mysqld_exporter,工作方式大致为:

  • 开启一个http端口

  • 采集本机MySQL状态(简单的show语句)

  • 将状态发布到http端口

exporter工作模式比较简单,理论上不会对数据库造成严重影响.

但在数据库节点上部署启动exporter后,发现数据库的连接状态异常,对业务产生了一定影响.

  • 1.检查连接可用性

  • 1.1通过TCP检查连接可用性 -- 新连接无法建立!!

  • 1.2通过Socket检查连接可用性 -- 仍然无法建立连接

  • 2.已连接的业务报错查询超时

  • 3.检查mysql-error.log -- 没有任何内容

  • 4.检查系统日志 -- 没有任何相关信息

  • 5.观察系统资源占用,无异常占用

为何再正常不过的状态采集会导致如此大的问题? 由于日志中已经无法继续判断原因. 所以迅速采集调用栈信息:

pstack $(pgrep -xn mysqld)


问题分析

梳理exporter对数据库发起的SQL语句:

  • show global variables

  • show global status

  • show master status

  • show slave status

通过SQL语句列表和调用栈信息,大致可以观测出MySQL内部出现了Mutex死锁.死锁关系如下:

GDB定位MySQL5.7特定版本hang死的故障分析-爱可生


  • 1.SHOW GLOBAL VARIABLES

  • 持有LOCK_global_system_variables, 等待LOCK_log

  • 2.SLAVE SQL THREAD

  • 持有LOCK_log, 等待LOCK_status

  • 3.SHOW GLOBAL STATUS

  • 持有LOCK_status, 等待LOCK_global_system_variables

我们来看下每部分的调用栈信息:

源码部分基于Percona MySQL 5.7.23-23


SHOW GLOBAL VARIABLES

Thread 18 (Thread 0x7fad69737700 (LWP 3588)):
#0  0x00007fdd3527642d in __lll_lock_wait () from /lib64/libpthread.so.0
#1  0x00007fdd35271dcb in _L_lock_812 () from /lib64/libpthread.so.0
#2  0x00007fdd35271c98 in pthread_mutex_lock () from /lib64/libpthread.so.0
#3  0x0000000000d60a5e in PolyLock_lock_log::rdlock() ()
#4  0x0000000000c3e4d5 in sys_var::value_ptr(THD*, THD*, enum_var_type, st_mysql_lex_string*) ()
#5  0x0000000000d1ce70 in get_one_variable_ext(THD*, THD*, st_mysql_show_var const*, enum_var_type, enum_mysql_show_type, system_status_var*, charset_info_st const**, char*, unsigned long*) ()
#6  0x0000000000d1cee1 in get_one_variable(THD*, st_mysql_show_var const*, enum_var_type, enum_mysql_show_type, system_status_var*, charset_info_st const**, char*, unsigned long*) ()
#7  0x0000000000d2234c in show_status_array(THD*, char const*, st_mysql_show_var*, enum_var_type, system_status_var*, char const*, TABLE_LIST*, bool, Item*) ()
#8  0x0000000000d27ec6 in fill_variables(THD*, TABLE_LIST*, Item*) ()
#9  0x0000000000d1550c in do_fill_table(THD*, TABLE_LIST*, QEP_TAB*) ()
#10 0x0000000000d286ac in get_schema_tables_result(JOIN*, enum_schema_table_state) ()
#11 0x0000000000d0afed in JOIN::prepare_result() ()
#12 0x0000000000c9a2cf in JOIN::exec() ()
#13 0x0000000000d0b93d in handle_query(THD*, LEX*, Query_result*, unsigned long long, unsigned long long) ()
#14 0x00000000007625d8 in execute_sqlcom_select(THD*, TABLE_LIST*) ()
#15 0x0000000000ccd191 in mysql_execute_command(THD*, bool) ()
#16 0x0000000000cd06f5 in mysql_parse(THD*, Parser_state*) ()
#17 0x0000000000cd12bd in dispatch_command(THD*, COM_DATA const*, enum_server_command) ()
#18 0x0000000000cd2cff in do_command(THD*) ()
#19 0x0000000000d9c8f8 in handle_connection ()
#20 0x0000000000f18904 in pfs_spawn_thread ()
#21 0x00007fdd3526fe25 in start_thread () from /lib64/libpthread.so.0
#22 0x00007fdd3344434d in clone () from /lib64/libc.so.6


  • 关键代码

/**sql\set_var.cc:267**/
uchar *sys_var::value_ptr(THD *running_thd, ...)
{
 mysql_mutex_assert_owner(&LOCK_global_system_variables);
}

/**sql\sys_vars.cc:3788**/
void PolyLock_lock_log::rdlock()
{
 mysql_mutex_lock(mysql_bin_log.get_log_lock());
}


SLAVE SQL THREAD

Thread 25 (Thread 0x7fad7c0f5700 (LWP 47842)):
#0  0x00007fdd3527642d in __lll_lock_wait () from /lib64/libpthread.so.0
#1  0x00007fdd35271e01 in _L_lock_1022 () from /lib64/libpthread.so.0
#2  0x00007fdd35271da2 in pthread_mutex_lock () from /lib64/libpthread.so.0
#3  0x0000000000e92d37 in MYSQL_BIN_LOG::publish_coordinates_for_global_status() const ()
#4  0x0000000000e99791 in MYSQL_BIN_LOG::ordered_commit(THD*, bool, bool) ()
#5  0x0000000000e9ccbf in MYSQL_BIN_LOG::commit(THD*, bool) ()
#6  0x00000000008084d1 in ha_commit_trans(THD*, bool, bool) ()
#7  0x0000000000d7c9c9 in trans_commit(THD*) ()
#8  0x0000000000e65efb in Xid_log_event::do_commit(THD*) ()
#9  0x0000000000e69b45 in Xid_apply_log_event::do_apply_event(Relay_log_info const*) ()
#10 0x0000000000e6f5dd in Log_event::apply_event(Relay_log_info*) ()
#11 0x0000000000eb87f6 in apply_event_and_update_pos(Log_event**, THD*, Relay_log_info*) ()
#12 0x0000000000ec4e90 in handle_slave_sql ()
#13 0x0000000000f18904 in pfs_spawn_thread ()
#14 0x00007fdd3526fe25 in start_thread () from /lib64/libpthread.so.0
#15 0x00007fdd3344434d in clone () from /lib64/libc.so.6


  • 关键代码

/**sql\binlog.cc:10424*//
void MYSQL_BIN_LOG::publish_coordinates_for_global_status(void) const
{
 mysql_mutex_assert_owner(&LOCK_log);

 mysql_mutex_lock(&LOCK_status);
}


SHOW GLOBAL STATUS

Thread 14 (Thread 0x7fad69633700 (LWP 3613)):
#0  0x00007fdd3527642d in __lll_lock_wait () from /lib64/libpthread.so.0
#1  0x00007fdd35271e01 in _L_lock_1022 () from /lib64/libpthread.so.0
#2  0x00007fdd35271da2 in pthread_mutex_lock () from /lib64/libpthread.so.0
#3  0x0000000000d222eb in show_status_array(THD*, char const*, st_mysql_show_var*, enum_var_type, system_status_var*, char const*, TABLE_LIST*, bool, Item*) ()
#4  0x0000000000d29860 in fill_status(THD*, TABLE_LIST*, Item*) ()
#5  0x0000000000d1550c in do_fill_table(THD*, TABLE_LIST*, QEP_TAB*) ()
#6  0x0000000000d286ac in get_schema_tables_result(JOIN*, enum_schema_table_state) ()
#7  0x0000000000d0afed in JOIN::prepare_result() ()
#8  0x0000000000c9a2cf in JOIN::exec() ()
#9  0x0000000000d0b93d in handle_query(THD*, LEX*, Query_result*, unsigned long long, unsigned long long) ()
#10 0x00000000007625d8 in execute_sqlcom_select(THD*, TABLE_LIST*) ()
#11 0x0000000000cca8a6 in mysql_execute_command(THD*, bool) ()
#12 0x0000000000cd06f5 in mysql_parse(THD*, Parser_state*) ()
#13 0x0000000000cd12bd in dispatch_command(THD*, COM_DATA const*, enum_server_command) ()
#14 0x0000000000cd2cff in do_command(THD*) ()
#15 0x0000000000d9c8f8 in handle_connection ()
#16 0x0000000000f18904 in pfs_spawn_thread ()
#17 0x00007fdd3526fe25 in start_thread () from /lib64/libpthread.so.0
#18 0x00007fdd3344434d in clone () from /lib64/libc.so.6


  • 关键代码

/**sql\sql_show.cc:8020*//
int fill_status(THD *thd, TABLE_LIST *tables, Item *cond){
   ...
   mysql_mutex_lock(&LOCK_status);
}


/**sql\sql_show.cc:3061**/
static bool show_status_array(THD *thd, const char *wild,SHOW_VAR *variables, ...){
   ...
   mysql_mutex_lock(&LOCK_global_system_variables);


问题1: 新连接为什么无法建立?

  • 连接请求时的调用栈

Thread 6 (Thread 0x7fad6942b700 (LWP 4343)):
#0  0x00007fdd3527642d in __lll_lock_wait () from /lib64/libpthread.so.0
#1  0x00007fdd35271e01 in _L_lock_1022 () from /lib64/libpthread.so.0
#2  0x00007fdd35271da2 in pthread_mutex_lock () from /lib64/libpthread.so.0
#3  0x0000000000c88ff7 in THD::init() ()
#4  0x0000000000c8be9f in THD::THD(bool) ()
#5  0x0000000000e54a1a in Channel_info::create_thd() ()
#6  0x0000000000d9fede in Channel_info_tcpip_socket::create_thd() ()
#7  0x0000000000d9c7c2 in handle_connection ()
#8  0x0000000000f18904 in pfs_spawn_thread ()
#9  0x00007fdd3526fe25 in start_thread () from /lib64/libpthread.so.0
#10 0x00007fdd3344434d in clone () from /lib64/libc.so.6


连接的初始化需要LOCK_global_system_variables, 此mutex被占用,导致新连接无法初始化

[root@localhost /]# mysql -uroot -p -S /mysql/data/mysqld.sock
Enter password:

^C^C^C^C^C


新连接请求发起时一直被挂起,没有任何响应,现象如下:

[root@localhost /]# mysql -uroot -p -S /mysql/data/mysqld.sock
Enter password:

^C^C^C^C^C


问题2:查询操作为什么超时?

  • 查询操作的调用栈

/**sql\item_func.cc:7275**/
void Item_func_get_system_var::fix_length_and_dec()
{
 ...
 mysql_mutex_lock(&LOCK_global_system_variables);
}


SQL查询时也需要LOCK_global_system_variables, 此mutex被占用,导致查询被阻塞.

/**sql\item_func.cc:7275**/
void Item_func_get_system_var::fix_length_and_dec()
{
 ...
 mysql_mutex_lock(&LOCK_global_system_variables);
}


问题结论

经过一轮分析,我们基本找到问题原因. 但心中还存在疑问.

  • LOCK_global_system_variables是热点的mutex,大概率会出现争用,但不至于死锁.

  • show global variables操作为何需要LOCK_log?

  • 这样常用的查询真的会稳定导致数据库死锁么?

进一步搜索,发现社区已有类似的缺陷反馈:

该缺陷可理解为:

MySQL 5.7.22引入了如下两个参数:

  • binlog_transaction_dependency_trackinghttps://dev.mysql.com/doc/refman/5.7/en/replication-options-binary-log.html#sysvar_binlog_transaction_dependency_tracking

  • binlog_transaction_dependency_history_sizehttps://dev.mysql.com/doc/refman/5.7/en/replication-options-binary-log.html#sysvar_binlog_transaction_dependency_history_size

在5.7.23版本中,读取这两个参数需要LOCK_log.从而会导致各种各样的死锁(bug正文中提到了另外一种场景).

在5.7.25版本以后,这两个参数的读取使用了另外一种影响更低的lock,从而避免了本文以及bug中提到的死锁.


问题处理

对于环境中出现死锁的这一组Percona MySQL 5.7.23-23半同步主从,我们使用5.7.25-28版本进行了升级后此问题消失,至此可以基本确定导致问题的根本原因.

上一篇
MySQL之MGR相同GTID产生不同transaction故障分析
400-820-6580 13916131869
marketing@actionsky.com
上海市闵行区万源路2138号泓茂中心2号楼
产品详情
关系型数据库
AI数据库
数据库智能管理平台
数据库生态产品
行业案例
金融行业
新零售行业
制造业
通信行业
更多
公司动态
最新新闻
国产化信息
技术分享
关于我们
公司简介
公司分布
国家专利
资质认证
扫码关注公众号
© Copyright 2017, All rights reserved by: 上海爱可生信息技术股份有限公司 沪ICP备12003970号-1 | 法律声明 | 网站地图
沪公网安备 31010402003331号