MySQL进阶垫脚石:线程长时间处于killed状态怎么破?
作者介绍
猿辅导数据库平台团队,承载猿辅导在线教育全公司的数据库产品研发、运维及服务需求。团队始终致力于新技术的探索实践,结合业务场景不断打磨并提升高可用、可扩展、高可靠的基础设施能力,作为核心基础设施建设者,支持业务快速发展。(猿辅导技术公众号ID:gh_cb5c83bb3ee0)
赵晓杰,猿辅导数据库平台团队成员,主要从事数据库存储、中间件等方向研发工作。
一、背景
MySQL中使用kill命令去杀死连接时,如果使用show processlist会发现线程会处于killed状态一段时间,而不是立即杀掉。一些情况下,killed状态可能会存在很久,甚至可能会一直存在直到发送第二次kill命令才能杀掉连接。下面从MySQL执行kill命令代码流程(基于5.7版本的MySQL)简单分析下出现这种现象的原因。
二、源码分析
MySQL的启动入口函数为mysqld中的main函数,主要流程会启动一个线程去listen端口,accept tcp连接,并创建一个connection并与具体的线程绑定,去处理来自客户端的消息。
执行流程:
日常执行kill流程,一般是通过mysql命令行客户端新起一个连接,通过show processlist找到需要kill掉的连接的conncetion_id,然后发送kill命令。
注:kill + 连接id 默认是kill connection,代表断开连接,如果是kill query + 连接id则只是终止本次执行的语句,连接还会继续监听来自client的命令。(具体执行区别可以参考下面KILL工作流程1中(1)、(2)部分)
概念:
connection: socket连接,默认有一个max_connection,实际上可以接受max_connection + 1个连接,最后一个连接是预留给SUPER用户的。
pthread: mysql的工作线程,每个connection建立时都会分配一个pthread,connection断开后pthread仍旧可以被其他connection复用。
thd: 线程描述类,每个connection对应一个thd,其中包含很多连接状态的描述,其中thd->killed字段标识线程是否需要被kill。
为方便说明,假设有两个连接connection1, connection2, 对应上述流程,则是connection1在do_command或者listen socket event中时,通过connection2发送kill命令,中断connection1的执行流程。
kill connection之后,对应此连接的pthread可能会被新连接复用(具体后面会分析)。
1)执行kill命令的线程发起kill
以connection2的执行流程来分析kill的执行过程,跟踪do_command之后的代码堆栈可以看到:
* frame #0: 0x00000001068a8853 mysqld`THD::awake(this=0x00007fbede88b400, state_to_set=KILL_CONNECTION) at sql_class.cc:2029:27
frame #1: 0x000000010695961f mysqld`kill_one_thread(thd=0x00007fbed6bc9c00, id=2, only_kill_query=false) at sql_parse.cc:6479:14
frame #2: 0x0000000106946529 mysqld`sql_kill(thd=0x00007fbed6bc9c00, id=2, only_kill_query=false) at sql_parse.cc:6507:16
frame #3: 0x000000010694e0fa mysqld`mysql_execute_command(thd=0x00007fbed6bc9c00, first_level=true) at sql_parse.cc:4210:5
frame #4: 0x0000000106945d62 mysqld`mysql_parse(thd=0x00007fbed6bc9c00, parser_state=0x000070000de2f340) at sql_parse.cc:5584:20
frame #5: 0x0000000106942bf0 mysqld`dispatch_command(thd=0x00007fbed6bc9c00, com_data=0x000070000de2fe78, command=COM_QUERY) at sql_parse.cc:1491:5
frame #6: 0x0000000106944e70 mysqld`do_command(thd=0x00007fbed6bc9c00) at sql_parse.cc:1032:17
frame #7: 0x0000000106ad3976 mysqld`::handle_connection(arg=0x00007fbee220b8d0) at
connection_handler_per_thread.cc:313:13
frame #8: 0x000000010749e74c mysqld`::pfs_spawn_thread(arg=0x00007fbee15dcf90) at pfs.cc:2197:3
frame #9: 0x00007fff734b6109 libsystem_pthread.dylib`_pthread_start + 148
frame #10: 0x00007fff734b1b8b libsystem_pthread.dylib`thread_start + 15
核心代码为awake函数(为方便,分为3段分析):
① 设置线程killed flag状态
if (this->m_server_idle && state_to_set == KILL_QUERY)
{ /* nothing */ }
else
{
killed= state_to_set;
}
如果当前线程处于idle状态(代表命令已执行完),而且kill级别只是终止查询,而不是kill整个连接,那么不会去设置thd->killed状态,防止影响下一次正常的请求。
(认为需要被kill的查询已经执行结束了,不需要再做操作了)
② 关闭socket连接&中断引擎等待
if (state_to_set != THD::KILL_QUERY && state_to_set != THD::KILL_TIMEOUT)
{
if (this != current_thd)
{
shutdown_active_vio();
}
/* Send an event to the scheduler that a thread should be killed. */
if (!slave_thread)
MySQL_CALLBACK(Connection_handler_manager::event_functions, post_kill_notification, (this)); //post_kill
}
if (state_to_set != THD::NOT_KILLED)
ha_kill_connection(this);
之后会首先关闭socket连接(注如果是kill query,则不会关闭连接)不再接收新的命令。客户端报下面这个错就是在这步之后:
另外会执行ha_close_connection,这里实际是将处于innodb层等待状态的线程唤醒,具体代码在ha_innodb.cc中innobase_kill_connection里会调用lock_trx_handle_wait方法。
trx: 一个mysql线程对应的innodb的事务描述类。
③ 通过信号量通知处于wait状态的线程
if (is_killable)
{
mysql_mutex_lock(&LOCK_current_cond);
if (current_cond && current_mutex)
{
DBUG_EXECUTE_IF("before_dump_thread_acquires_current_mutex",
{
const char act[]=
"now signal dump_thread_signal wait_for go_dump_thread";
DBUG_ASSERT(!debug_sync_set_action(current_thd,
STRING_WITH_LEN(act)));
};);
mysql_mutex_lock(current_mutex); mysql_cond_broadcast(current_cond); mysql_mutex_unlock(current_mutex);
}
mysql_mutex_unlock(&LOCK_current_cond);
这里看到除了设置connection1的thd->killed状态外,还会获取current_mutex锁,唤醒wait条件变量current_cond的线程(connection2)。注意上述②和③中唤醒的对象不同:
ha_close_connection唤醒的是本次对应的innodb事务中的锁(trx->lock.wait_lock),对应的一般是在innodb层事务中等待的某个行锁。mysql_cond_broadcast(current_cond)则是唤醒thd中的锁,等待锁是通过THD::enter_cond()进入(如open table时获取表锁,或者sleep等)
具体可参考下面本地debug复现部分的分析。
为什么在发送信号量之前先关闭socket连接?
不关闭socket连接,并发情况下有什么问题?代码中提及了一种case,假设connection1运行已经过了主动检查flag的点,之后connection2调用awake设置flag及发送信号量唤醒,然后connection1进入到socket read中,那么相当于这次信号量会丢失,connection1就会一直阻塞在read中,所以需要关闭socket 连接中断read。BUG#37780
相当于是通过io中断解决信号量丢失的情况。
所以如果connection1在其他阶段发生信号量丢失(如connection2先broadcast,connection1再wait),就需要发送第二次kill命令才能唤醒。
(sql_class.cc 2090,但是注意KILL_CONNECTION是不会重复进入awake的)
注: 一般出现这种情况是,connection2修改了killed状态,但是由于cpu缓存一致性等问题,connection1看不到killed状态,然后通过了主动检查点,进入了wait状态。
2)被kill线程响应kill命令
被kill线程感知(响应)kill命令主要有两种方式:
主动检查:connection1在一些代码处会去主动检查killed状态;
被通过信号量唤醒:connection1在执行某些命令时(如引擎层去做一些操作),会主动去await一个condition,释放掉相应的锁,connection2执行kill命令时,会通过锁和condition唤醒connection1,执行终止操作;
killed状态:
enum killed_state
{
NOT_KILLED=0, KILL_BAD_DATA=1,
KILL_CONNECTION=ER_SERVER_SHUTDOWN, KILL_QUERY=ER_QUERY_INTERRUPTED, KILL_TIMEOUT=ER_QUERY_TIMEOUT,
KILLED_NO_VALUE /* means neither of the states */
};
3)connection真正被kill掉
真正被kill指的是show processlist看不到这个线程的时机。mysql在新建一个connection之后,会不断的去监听连接(do_command),前面提到kill时会主动把连接的socket关闭(shutdown_active_vio)。所以真正连接终止的逻辑是在此处,判断thd_connection_alive的状态是待杀死之后,会关闭connection,并且release_resources,此时再去show processlist,则killed的线程才会消失。相应的pthread也会等待其他连接复用。
killed状态:
if (thd_prepare_connection(thd))
handler_manager->inc_aborted_connects();
else
{
while (thd_connection_alive(thd))
{
if (do_command(thd)) break;
}
end_connection(thd);
}
close_connection(thd, 0, false, false);
thd->get_stmt_da()->reset_diagnostics_area();
thd->release_resources();
// Clean up errors now, before possibly waiting for a new connection.
#if OPENSSL_VERSION_NUMBER < 0x10100000L
ERR_remove_thread_state(0);
#endif /* OPENSSL_VERSION_NUMBER < 0x10100000L */
thd_manager->remove_thd(thd);
Connection_handler_manager::dec_connection_count();
........
channel_info= Per_thread_connection_handler::block_until_new_connection();
//等待新连接复用线程
3、被KILL线程主动检查点和唤醒机制分析
由前面kill过程分析,大致可以分为两种情况,一种是connection1代码一直在执行中(占据cpu),那么总会执行到某些地方可以检查thd->killed状态,另外一种是connection1线程wait状态,需要其他线程通过信号量唤醒connection1的线程,实现kill中断目的。具体地,这两类又可以分为下面4种情况:
1)第一类:主动检查断点
① connection2发送kill命令时,connection1已执行完命令 (主动检查)
此时connection1阻塞在socket_read上,由于前面提到connection2会去shutdown_active_vio,connection1很容易感知到,执行后续操作,如回滚等。
if (thd_prepare_connection(thd))
handler_manager->inc_aborted_connects();
else
{
while (thd_connection_alive(thd))
{
if (do_command(thd))
break;
}
end_connection(thd);
}
close_connection(thd, 0, false, false);
thd->get_stmt_da()->reset_diagnostics_area();
thd->release_resources(); //clean_up
②内部innodb引擎在获取记录时,也会去判断thd->killed状态,决定是否中断操作,进行返回。
这一类检查点很多。如下面两处:
select每行读取都会检查;
int rr_sequential(READ_RECORD *info)
{
int tmp;
while ((tmp=info->table->file->ha_rnd_next(info->record)))
{
/*
ha_rnd_next can return RECORD_DELETED for MyISAM when one thread is reading and another deleting without locks.
*/
if (info->thd->killed || (tmp != HA_ERR_RECORD_DELETED))
{
tmp= rr_handle_error(info, tmp); break;
}
}
return tmp;
}
内部innodb引擎在获取记录时,也会去判断thd->killed状态,决定是否中断操作,进行返回。
* frame #0: 0x0000000106629a7e mysqld`trx_is_interrupted(trx=0x00007f8266dbd8a0) at ha_innodb.cc:3234:9
frame #1: 0x000000010686d901 mysqld`row_search_mvcc(buf=", mode=PAGE_CUR_G, prebuilt=0x00007f826b81faa0,
match_mode=0, direction=0) at row0sel.cc:5245:6
frame #2: 0x0000000106636cda mysqld`ha_innobase::index_read(this=0x00007f825c8c1430, buf=", key_ptr=0x0000000000000000, key_len=0, find_flag=HA_READ_AFTER_KEY) at ha_innodb.cc:8768:10
frame #3: 0x000000010663798c mysqld`ha_innobase::index_first(this=0x00007f825c8c1430, buf=") at ha_innodb. cc:9186:14
frame #4: 0x0000000106637c2a mysqld`ha_innobase::rnd_next(this=0x00007f825c8c1430, buf=") at ha_innodb.cc: 9284:11
frame #5: 0x000000010563f7a5 mysqld`handler::ha_rnd_next(this=0x00007f825c8c1430, buf=") at handler.cc:2963:
3
frame #6: 0x0000000105d954d4 mysqld`rr_sequential(info=0x00007f826780d208) at records.cc:517:34
frame #7: 0x0000000105e85c78 mysqld`join_init_read_record(tab=0x00007f826780d1b8) at sql_executor.cc:2504:10
frame #8: 0x0000000105e82a1c mysqld`sub_select(join=0x00007f825e37a4d0, qep_tab=0x00007f826780d1b8,
end_of_records=false) at sql_executor.cc:1284:14
frame #9: 0x0000000105e7f299 mysqld`do_select(join=0x00007f825e37a4d0) at sql_executor.cc:957:12
frame #10: 0x0000000105e7ec26 mysqld`JOIN::exec(this=0x00007f825e37a4d0) at sql_executor.cc:206:10
frame #11: 0x0000000105f5fe90 mysqld`handle_query(thd=0x00007f825e35ec00, lex=0x00007f825e361058,
result=0x00007f825e379cf8, added_options=0, removed_options=0) at sql_select.cc:191:21
frame #12: 0x0000000105f006f7 mysqld`execute_sqlcom_select(thd=0x00007f825e35ec00,
all_tables=0x00007f825e3796b0) at sql_parse.cc:5155:12
frame #13: 0x0000000105ef6527 mysqld`mysql_execute_command(thd=0x00007f825e35ec00, first_level=true) at sql_parse.cc:2826:12
frame #14: 0x0000000105ef3d62 mysqld`mysql_parse(thd=0x00007f825e35ec00, parser_state=0x00007000097d5340) at sql_parse.cc:5584:20
frame #15: 0x0000000105ef0bf0 mysqld`dispatch_command(thd=0x00007f825e35ec00, com_data=0x00007000097d5e78, command=COM_QUERY) at sql_parse.cc:1491:5
frame #16: 0x0000000105ef2e70 mysqld`do_command(thd=0x00007f825e35ec00) at sql_parse.cc:1032:17
frame #17: 0x0000000106081976 mysqld`::handle_connection(arg=0x00007f82681d84f0) at
connection_handler_per_thread.cc:313:13
frame #18: 0x0000000106a4c74c mysqld`::pfs_spawn_thread(arg=0x00007f8268012fc0) at pfs.cc:2197:3
frame #19: 0x00007fff734b6109 libsystem_pthread.dylib`_pthread_start + 148
frame #20: 0x00007fff734b1b8b libsystem_pthread.dylib`thread_start + 15
2)第二类:需要其他线程通过信号量唤醒
① connection2发送kill命令时,connection1处于innodb层wait行锁状态
主要通过awake中的下面这行触发唤醒 (也可能由系统的后台线程lock_wait_timeout_thread唤醒)
/* Interrupt target waiting inside a storage engine. */
if (state_to_set != THD::NOT_KILLED)
ha_kill_connection(this);
参 考 下 面 debug 分 析 的 case3 。
② connection2发送kill命令时,connection1处于msyql层wait状态(由connection2唤醒)
主要通过下面的方法实现唤醒:
/* Broadcast a condition to kick the target if it is waiting on it. */
if (is_killable)
{
mysql_mutex_lock(&LOCK_current_cond);
if (current_cond && current_mutex)
{
DBUG_EXECUTE_IF("before_dump_thread_acquires_current_mutex",
{
const char act[]=
"now signal dump_thread_signal wait_for go_dump_thread";
DBUG_ASSERT(!debug_sync_set_action(current_thd,
STRING_WITH_LEN(act)));
};);
mysql_mutex_lock(current_mutex);
mysql_cond_broadcast(current_cond); mysql_mutex_unlock(current_mutex);
}
mysql_mutex_unlock(&LOCK_current_cond);
参考下面debug分析的case4。
三、原因总结
通过上面代码分析可以得知,kill之后会进行回滚操作(大事务)或清理临时表(比如较慢的ddl),都有可能导致长时间处于killed状态。
具体回滚是上面提到的handle_connection中的thd->release_resources()中执行clean_up进行回滚或者在sql_parse中trans_rollback_stmt中。
除开回滚操作的影响,如果本身mysql机器负载较高,一样会导致主动检查thd->killed会有延迟或者影响线程的唤醒调度。
四、案例复现
1)case 1
connection1已执行完命令,connection2去kill连接1
(较为简单,略)
2)case2
connection1正在执行(如parse阶段,还没有真正到innodb层),connection2去kill连接1
(较为简单,略)
3)case3(innodb层唤醒)
新起一个session作为connection0开启事务,update某一行row1
再起一个session作为connection1 update同一行row1
* thread #29, stop reason = breakpoint 127.1
* frame #0: 0x00000001060f87cb mysqld`os_event::wait(this=0x00007fcce2b16598) at os0event.cc:180:4
frame #1: 0x00000001060f8773 mysqld`os_event::wait_low(this=0x00007fcce2b16598, reset_sig_count=1) at
os0event.cc:366:3
frame #2: 0x00000001060f91ed mysqld`os_event_wait_low(event=0x00007fcce2b16598, reset_sig_count=0) at os0event.cc:611:9
frame #3: 0x00000001060a9332 mysqld`lock_wait_suspend_thread(thr=0x00007fccd90936f0) at lock0wait.cc:315:2
frame #4: 0x00000001061b4f77 mysqld`row_mysql_handle_errors(new_err=0x0000700009f3af6c,
trx=0x00007fcce1cc0ce0, thr=0x00007fccd90936f0, savept=0x0000000000000000) at row0mysql.cc:783:3
frame #5: 0x0000000106212fa6 mysqld`row_search_mvcc(buf=", mode=PAGE_CUR_GE, prebuilt=0x00007fccd9092ea0, match_mode=1, direction=0) at row0sel.cc:6292:6
frame #6: 0x0000000105fd8cda mysqld`ha_innobase::index_read(this=0x00007fccd9091430, buf=", key_ptr="\x1e", key_len=514, find_flag=HA_READ_KEY_EXACT) at ha_innodb.cc:8768:10
frame #7: 0x0000000104ff2c67 mysqld`handler::index_read_map(this=0x00007fccd9091430, buf=", key="\x1e", keypart_map=1, find_flag=HA_READ_KEY_EXACT) at handler.h:2824:13
frame #8: 0x0000000104fe1f14 mysqld`handler::ha_index_read_map(this=0x00007fccd9091430, buf=", key="\x1e", keypart_map=1, find_flag=HA_READ_KEY_EXACT) at handler.cc:3047:3
frame #9: 0x0000000104feeb62 mysqld`handler::read_range_first(this=0x00007fccd9091430, start_key=0x00007fccd9091518, end_key=0x00007fccd9091538, eq_range_arg=true, sorted=true) at handler.cc:7412:13
frame #10: 0x0000000104fec01b mysqld`handler::multi_range_read_next(this=0x00007fccd9091430, range_info=0x0000700009f3bc10) at handler.cc:6477:15
frame #11: 0x0000000104fed4ba mysqld`DsMrr_impl::dsmrr_next(this=0x00007fccd9091698, range_info=0x0000700009f3bc10) at handler.cc:6869:24
frame #12: 0x0000000105fee7f6 mysqld`ha_innobase::multi_range_read_next(this=0x00007fccd9091430, range_info=0x0000700009f3bc10) at ha_innodb.cc:20585:18
frame #13: 0x00000001056ec7c8 mysqld`QUICK_RANGE_SELECT::get_next(this=0x00007fcce2b15fa0) at opt_range.cc: 11247:21
frame #14: 0x00000001057371ad mysqld`rr_quick(info=0x0000700009f3c320) at records.cc:405:29
frame #15: 0x0000000105984e35 mysqld`mysql_update(thd=0x00007fcce585f400, fields=0x00007fcce5863618, values=0x00007fcce58643d8, limit=18446744073709551615, handle_duplicates=DUP_ERROR, found_return=0x0000700009f3cb48, updated_return=0x0000700009f3cb40) at sql_update.cc:819:14
frame #16: 0x000000010598cc87 mysqld`Sql_cmd_update::try_single_table_update(this=0x00007fcce58643c8, thd=0x00007fcce585f400, switch_to_multitable=0x0000700009f3cc07) at sql_update.cc:2927:21
frame #17: 0x000000010598d457 mysqld`Sql_cmd_update::execute(this=0x00007fcce58643c8, thd=0x00007fcce585f400) at sql_update.cc:3058:7
frame #18: 0x000000010589b475 mysqld`mysql_execute_command(thd=0x00007fcce585f400, first_level=true) at sql_parse.cc:3616:26
frame #19: 0x0000000105895d62 mysqld`mysql_parse(thd=0x00007fcce585f400, parser_state=0x0000700009f40340) at sql_parse.cc:5584:20
frame #20: 0x0000000105892bf0 mysqld`dispatch_command(thd=0x00007fcce585f400, com_data=0x0000700009f40e78, command=COM_QUERY) at sql_parse.cc:1491:5
frame #21: 0x0000000105894e70 mysqld`do_command(thd=0x00007fcce585f400) at sql_parse.cc:1032:17
frame #22: 0x0000000105a23976 mysqld`::handle_connection(arg=0x00007fcce2555370) at
connection_handler_per_thread.cc:313:13
frame #23: 0x00000001063ee74c mysqld`::pfs_spawn_thread(arg=0x00007fcce2555f80) at pfs.cc:2197:3
frame #24: 0x00007fff71032109 libsystem_pthread.dylib`_pthread_start + 148
frame #25: 0x00007fff7102db8b libsystem_pthread.dylib`thread_start + 15
可以看到connection1 在等待row1的行锁
开启新的session作为connection2执行kill connection1的命令,走到前面代码中分析到的ha_close_connection时,会去中断innodb行锁的等待。堆栈为
thread #30
frame #0: 0x00000001060f910e mysqld`os_event::set(this=0x00007fcce2b16598) at os0event.cc:93:2
frame #1: 0x00000001060f90b5 mysqld`os_event_set(event=0x00007fcce2b16598) at os0event.cc:560:9
frame #2: 0x00000001060aa06e mysqld`lock_wait_release_thread_if_suspended(thr=0x00007fccd90936f0) at lock0wait.cc:411:3
frame #3: 0x0000000106089a99 mysqld`lock_cancel_waiting_and_release(lock=0x00007fccd8866c18) at lock0lock. cc:6789:3
frame #4: 0x0000000106096679 mysqld`lock_trx_handle_wait(trx=0x00007fcce1cc0ce0) at lock0lock.cc:6972:3
frame #5: 0x0000000105ff93a6 mysqld`innobase_kill_connection(hton=0x00007fccd7e094d0,
thd=0x00007fcce585f400) at ha_innodb.cc:4868:3
frame #6: 0x0000000104fdcb96 mysqld`kill_handlerton(thd=0x00007fcce585f400, plugin=0x0000700009f809d8, (null)=0x0000000000000000) at handler.cc:1052:7
frame #7: 0x00000001058d659c mysqld`plugin_foreach_with_mask(thd=0x00007fcce585f400, funcs=0x0000700009f80a60, type=1, state_mask=4294967287, arg=0x0000000000000000)(THD*, st_plugin_int**, void*), int, unsigned int, void*) at sql_plugin.cc:2524:21
frame #8: 0x00000001058d66a2 mysqld`plugin_foreach_with_mask(thd=0x00007fcce585f400, func= (mysqld`kill_handlerton(THD*, st_plugin_int**, void*) at handler.cc:1046), type=1, state_mask=8, arg=0x0000000000000000)(THD*, st_plugin_int**, void*), int, unsigned int, void*) at sql_plugin.cc:2539:10
frame #9: 0x0000000104fdcb1b mysqld`ha_kill_connection(thd=0x00007fcce585f400) at handler.cc:1060:3
frame #10: 0x00000001057f8923 mysqld`THD::awake(this=0x00007fcce585f400, state_to_set=KILL_CONNECTION) at sql_class.cc:2077:5
frame #11: 0x00000001058a961f mysqld`kill_one_thread(thd=0x00007fccd8b9ea00, id=3, only_kill_query=false) at sql_parse.cc:6479:14
frame #12: 0x0000000105896529 mysqld`sql_kill(thd=0x00007fccd8b9ea00, id=3, only_kill_query=false) at sql_parse.cc:6507:16
frame #13: 0x000000010589e0fa mysqld`mysql_execute_command(thd=0x00007fccd8b9ea00, first_level=true) at sql_parse.cc:4210:5
frame #14: 0x0000000105895d62 mysqld`mysql_parse(thd=0x00007fccd8b9ea00, parser_state=0x0000700009f84340) at sql_parse.cc:5584:20
frame #15: 0x0000000105892bf0 mysqld`dispatch_command(thd=0x00007fccd8b9ea00, com_data=0x0000700009f84e78, command=COM_QUERY) at sql_parse.cc:1491:5
frame #16: 0x0000000105894e70 mysqld`do_command(thd=0x00007fccd8b9ea00) at sql_parse.cc:1032:17
frame #17: 0x0000000105a23976 mysqld`::handle_connection(arg=0x00007fcce5307d10) at
connection_handler_per_thread.cc:313:13
frame #18: 0x00000001063ee74c mysqld`::pfs_spawn_thread(arg=0x00007fcce2405030) at pfs.cc:2197:3
frame #19: 0x00007fff71032109 libsystem_pthread.dylib`_pthread_start + 148
frame #20: 0x00007fff7102db8b libsystem_pthread.dylib`thread_start + 15
注 :mysql启动时也会启动一个线程检测锁是否超时(间隔1s),也会去调用lock_cancel_waiting_and_release中断等待行锁的线程。
这里的超时机制也可以防止信号量丢失无法唤醒的问题。
4)case4 (mysql层唤醒)
开启connection0,lock一张表。
然后开启connection1, 去update这张表,可以看到线程会阻塞在wait表锁状态。
thread #29
frame #0: 0x00007fff733f5882 libsystem_kernel.dylib` psynch_cvwait + 10
frame #1: 0x00007fff734b6425 libsystem_pthread.dylib`_pthread_cond_wait + 698
frame #2: 0x000000010ef6d495 mysqld`native_cond_timedwait(cond=0x00007f96cb1a6538,
mutex=0x00007f96cb1a64d8, abstime=0x000070000bc67e70) at thr_cond.h:136:10
frame #3: 0x000000010ef6d37a mysqld`safe_cond_timedwait(cond=0x00007f96cb1a6538, mp=0x00007f96cb1a6498, abstime=0x000070000bc67e70, file="/Users/zhaoxiaojie/CLionProjects/mysql-5.7.33/sql/mdl.cc", line=1868) at thr_cond.c:100:10
frame #4: 0x000000010ea69955 mysqld`my_cond_timedwait(cond=0x00007f96cb1a6538, mp=0x00007f96cb1a6498, abstime=0x000070000bc67e70, file="/Users/zhaoxiaojie/CLionProjects/mysql-5.7.33/sql/mdl.cc", line=1868) at thr_cond.h:187:10
frame #5: 0x000000010ea63341 mysqld`inline_mysql_cond_timedwait(that=0x00007f96cb1a6538, mutex=0x00007f96cb1a6498, abstime=0x000070000bc67e70, src_file="/Users/zhaoxiaojie/CLionProjects/mysql-5.7.33
/sql/mdl.cc", src_line=1868) at mysql_thread.h:1232:13
frame #6: 0x000000010ea631f8 mysqld`MDL_wait::timed_wait(this=0x00007f96cb1a6498, owner=0x00007f96cb1a6400, abs_timeout=0x000070000bc67e70, set_status_on_timeout=true, wait_state_name=0x00000001100b4b58) at mdl.cc:1867: 18
frame #7: 0x000000010ea660e5 mysqld`MDL_context::acquire_lock(this=0x00007f96cb1a6498, mdl_request=0x00007f96cb0e71a0, lock_wait_timeout=31536000) at mdl.cc:3699:25
frame #8: 0x000000010eb8a090 mysqld`open_table_get_mdl_lock(thd=0x00007f96cb1a6400, ot_ctx=0x000070000bc689a0, table_list=0x00007f96cb0e6e00, flags=0, mdl_ticket=0x000070000bc68318) at sql_base. cc:2914:35
frame #9: 0x000000010eb885ec mysqld`open_table(thd=0x00007f96cb1a6400, table_list=0x00007f96cb0e6e00, ot_ctx=0x000070000bc689a0) at sql_base.cc:3296:9
frame #10: 0x000000010eb8e77b mysqld`open_and_process_table(thd=0x00007f96cb1a6400, lex=0x00007f96cb1a8858, tables=0x00007f96cb0e6e00, counter=0x00007f96cb1a8918, flags=0, prelocking_strategy=0x000070000bc68ab8, has_prelocking_list=false, ot_ctx=0x000070000bc689a0) at sql_base.cc:5260:14
frame #11: 0x000000010eb8d8dc mysqld`open_tables(thd=0x00007f96cb1a6400, start=0x000070000bc68ac8, counter=0x00007f96cb1a8918, flags=0, prelocking_strategy=0x000070000bc68ab8) at sql_base.cc:5883:14
frame #12: 0x000000010eb90c6d mysqld`open_tables_for_query(thd=0x00007f96cb1a6400, tables=0x00007f96cb0e6e00, flags=0) at sql_base.cc:6660:7
frame #13: 0x000000010ed48b0c mysqld`Sql_cmd_update::try_single_table_update(this=0x00007f96cb0e6dc8, thd=0x00007f96cb1a6400, switch_to_multitable=0x000070000bc68c07) at sql_update.cc:2911:7
frame #14: 0x000000010ed49457 mysqld`Sql_cmd_update::execute(this=0x00007f96cb0e6dc8, thd=0x00007f96cb1a6400) at sql_update.cc:3058:7
frame #15: 0x000000010ec57475 mysqld`mysql_execute_command(thd=0x00007f96cb1a6400, first_level=true) at sql_parse.cc:3616:26
frame #16: 0x000000010ec51d62 mysqld`mysql_parse(thd=0x00007f96cb1a6400, parser_state=0x000070000bc6c340) at sql_parse.cc:5584:20
frame #17: 0x000000010ec4ebf0 mysqld`dispatch_command(thd=0x00007f96cb1a6400, com_data=0x000070000bc6ce78, command=COM_QUERY) at sql_parse.cc:1491:5
frame #18: 0x000000010ec50e70 mysqld`do_command(thd=0x00007f96cb1a6400) at sql_parse.cc:1032:17
frame #19: 0x000000010eddf976 mysqld`::handle_connection(arg=0x00007f96cfe2e1b0) at
connection_handler_per_thread.cc:313:13
frame #20: 0x000000010f7aa74c mysqld`::pfs_spawn_thread(arg=0x00007f96d5c1b300) at pfs.cc:2197:3
frame #21: 0x00007fff734b6109 libsystem_pthread.dylib`_pthread_start + 148
frame #22: 0x00007fff734b1b8b libsystem_pthread.dylib`thread_start + 15
开启connection2,去kill connection1,执行到awake时,可以看到要唤醒的条件变量和 connection1在等待的是同一个对象。(0x00007f96cb1a6538)
connection2执行完broadcast之后,connection1线程即被唤醒
之后一层层返回错误,在sql_parse中会执行trans_rollback_stmt执行回滚操作。最后在handle_connection中检查到connection需要被关闭。
之前在工作中实际碰到过一个执行kill命令后,连接数反而持续上涨的案例,现象简单概括就是:
异常流量导致数据库实例连接数过多后,dba执行kill连接,连接数会短暂下降,之后连接数会继续上涨,继续kill之后,连接数还是会上涨。并且大量线程一直处于killed状态,看起来像是无法kill连接,只能重启解决。
1)问题
连接数为什么反而会上涨?
前面KILL工作流程1.(2)部分中提到,kill线程会在唤醒被kill线程之前先关闭连接,客户端就是这个时候报lost connection的错误。这样大量客户端在线程还未被kill结束时已经开始了重连(sdk连接池),这样造成了问题加剧。
另外MySQL中innodb_thread_concurrency会限制进入innodb的线程并发度。那么当进入innodb层的线程达到阈值后,后面重建的大量连接会在mysql层执行for循环判断是否可以进入innodb。但是这个过程是没有检查killed状态的,导致这些线程一直无法被kill(尽管show processlist显示为Killed)。除非innodb里一个线程退出,使得某个线程可以进入innodb,从而执行代码到主动检查处或被唤醒执行kill逻辑。
srv0conc.cc srv_conc_enter_innodb_with_atomics
注:实际mysql会有相关参数控制进入innodb时的最大等待时间,为简化描述问题,暂不展开。
2)复现
将mysql innodb_thread_concurrency设置为1(默认为0代表不限制)。connection1执行select进入innodb后暂停住线程。
connection2也执行select,那么会发现connection2会卡在检查是否可以进入innodb。具体堆栈为:
* frame #0: 0x000000010c9772bd mysqld`srv_conc_enter_innodb(prebuilt=0x00007f7fb1065ca0) at srv0conc.cc:265:37
frame #1: 0x000000010c6d9498 mysqld`innobase_srv_conc_enter_innodb(prebuilt=0x00007f7fb1065ca0) at ha_innodb.cc: 1534:4
frame #2: 0x000000010c6dcc19 mysqld`ha_innobase::index_read(this=0x00007f7fb10a3630, buf=", key_ptr=0x0000000000000000, key_len=0, find_flag=HA_READ_AFTER_KEY) at ha_innodb.cc:8753:3
frame #3: 0x000000010c6dd98c mysqld`ha_innobase::index_first(this=0x00007f7fb10a3630, buf=") at ha_innodb.cc: 9186:14
frame #4: 0x000000010c6ddc2a mysqld`ha_innobase::rnd_next(this=0x00007f7fb10a3630, buf=") at ha_innodb.cc:9284: 11
frame #5: 0x000000010b6e57a5 mysqld`handler::ha_rnd_next(this=0x00007f7fb10a3630, buf=") at handler.cc:2963:3
frame #6: 0x000000010be3b4d4 mysqld`rr_sequential(info=0x00007f7fb0c1dd90) at records.cc:517:34
frame #7: 0x000000010bf2bc78 mysqld`join_init_read_record(tab=0x00007f7fb0c1dd40) at sql_executor.cc:2504:10
frame #8: 0x000000010bf28a1c mysqld`sub_select(join=0x00007f7fb206fc48, qep_tab=0x00007f7fb0c1dd40, end_of_records=false) at sql_executor.cc:1284:14
frame #9: 0x000000010bf25299 mysqld`do_select(join=0x00007f7fb206fc48) at sql_executor.cc:957:12
frame #10: 0x000000010bf24c26 mysqld`JOIN::exec(this=0x00007f7fb206fc48) at sql_executor.cc:206:10 frame #11: 0x000000010c005e90 mysqld`handle_query(thd=0x00007f7fb2065a00, lex=0x00007f7fb2067e58, result=0x00007f7fb206f2f8, added_options=0, removed_options=0) at sql_select.cc:191:21
frame #12: 0x000000010bfa66f7 mysqld`execute_sqlcom_select(thd=0x00007f7fb2065a00, all_tables=0x00007f7fb206ecb0) at sql_parse.cc:5155:12
frame #13: 0x000000010bf9c527 mysqld`mysql_execute_command(thd=0x00007f7fb2065a00, first_level=true) at sql_parse.cc:2826:12
frame #14: 0x000000010bf99d62 mysqld`mysql_parse(thd=0x00007f7fb2065a00, parser_state=0x0000700007205340) at sql_parse.cc:5584:20
frame #15: 0x000000010bf96bf0 mysqld`dispatch_command(thd=0x00007f7fb2065a00, com_data=0x0000700007205e78, command=COM_QUERY) at sql_parse.cc:1491:5
frame #16: 0x000000010bf98e70 mysqld`do_command(thd=0x00007f7fb2065a00) at sql_parse.cc:1032:17
frame #17: 0x000000010c127976 mysqld`::handle_connection(arg=0x00007f7fbcb5e650) at connection_handler_per_thread.cc:313:13
frame #18: 0x000000010caf274c mysqld`::pfs_spawn_thread(arg=0x00007f7fbcb6a5e0) at pfs.cc:2197:3
frame #19: 0x00007fff71032109 libsystem_pthread.dylib`_pthread_start + 148
frame #20: 0x00007fff7102db8b libsystem_pthread.dylib`thread_start + 15
再开启一个connection3, 去kill connection2,可以发现connection2会一直处于killed状态(客户端会断开连接)
即使connection3完成了前面提到的ha_close_connection和broadcast信号量,connection2的堆栈还是一直在上面for循环中。
直到connection1退出innodb 。
connection2进入innodb之后通过主动检查的方式执行kill逻辑。
此时show processlist显示:
dbaplus社群欢迎广大技术人员投稿,投稿邮箱:editor@dbaplus.cn