MySQL：终于为OS层面的线程命名了-service mysql start

一、问题来源

最近在检查某个数据库性能的时候，通过top -Hu mysql看到了一个特别奇怪的现象，线程有了自己的名字，我开始以为是哪个大厂自己维护的版本，如下：

   PID USER      PR  NI    VIRT    RES    SHR S %CPU %MEM     TIME+ COMMAND                                                                                                                  
  8146 mysql     20   0 4164720 734540  26624 S  0.0  9.9   0:00.96 mysqld                                                                                                                   
  8159 mysql     20   0 4164720 734540  26624 S  0.0  9.9   0:00.02 ib_io_ibuf                                                                                                               
  8160 mysql     20   0 4164720 734540  26624 S  0.0  9.9   0:00.02 ib_io_log                                                                                                                
  8161 mysql     20   0 4164720 734540  26624 S  0.0  9.9   0:00.04 ib_io_rd-1                                                                                                               
  8162 mysql     20   0 4164720 734540  26624 S  0.0  9.9   0:00.03 ib_io_rd-2   
...

后来装了一个8.0.28才发现确实是官方版本的新玩意。但是虽然能够猜到一些线程的功能，可还是很陌生的样子，因为这个名字和performance_schema.thread中的名字并不一样。

这里我们就来看看它的做法和对应关系。不过这一小步，却是DBA的一大步，我们以往在看Oracle的进程的时候都习惯了有命名的进程名字，这带来的好处是直接从OS层面就能判断大概哪个功能的压力增高。

二、以往的对应方法

我们知道以前在获取到线程的LWP号后需要到performance_schema.thread通过lwp和thread_os_id 对应，得到如下结果：

 PID USER      PR  NI    VIRT    RES    SHR S %CPU %MEM     TIME+ COMMAND                                                                                                                  
  5524 mysql     20   0 4052588 792400  11676 S  0.0 27.4   0:03.82 mysqld                                                                                                                   
  5533 mysql     20   0 4052588 792400  11676 S  0.0 27.4   0:00.00 mysqld                                                                                                                   
  5556 mysql     20   0 4052588 792400  11676 S  0.0 27.4   0:00.00 mysqld                                                                                                                   
  5557 mysql     20   0 4052588 792400  11676 S  0.0 27.4   0:00.00 mysqld 
...
----------------------------------------+--------------+
| name                                   | thread_os_id |
+----------------------------------------+--------------+
| thread/sql/main                        |         5524 |
| thread/sql/thread_timer_notifier       |         5533 |
| thread/innodb/io_ibuf_thread           |         5556 |
| thread/innodb/io_read_thread           |         5558 |
| thread/innodb/io_log_thread            |         5557 |
...

如果某个线程的CPU高或者IO高我们就能够知道是什么线程。当然你也可以和information_schema.processlist做join得到process id和state等有用的信息。

三、简单的实现方法讨论

比如以innodb为例，所有的线程的OS thread name都放到了all_innodb_threads这个一个数组中，其中每个元素是一个结构体，结构体中包含了我们OS thread name这个元素给予了大量的代码注释，我就放一点我们容易看懂的：

typedef struct PSI_thread_info_v5 PSI_thread_info;
    The thread name to advertise to the operating system.
    This feature is optional, and improves
    observability for platforms that support
    a flavor of pthread_setname_np().

这里我们也看到需要支持pthread_setname_np函数才行。在调用register_thread_class注册所有的class的时候会将这些OS thread name放到一个叫做thread_class_array全局内存中，这样再建立线程应该能轻松的从全局内存中拿到每个线程的OS thread name(当然我没去细看了)。接着，在建立线程的时候我们调用my_thread_self_setname设置OS thread name就可以了，实际上就是调用pthread_setname_np。

四、新的Linux OS thread name和performance_schema.threads中name的对应

为了快速的得到对应的办法，我稍微加了点输出内容，这样只要有线程启动，就会打印到error日志，因为如果一个一个去看每个线程启动的时候带入的OS thread name实在太慢了，耗不起，得到的结果如下：

 cat mysql3380.err |grep -w 'init threads'
[init threads] os name:boot -- mysql name:thread/sql/bootstrap  
[init threads] os name:ib_io_ibuf -- mysql name:thread/innodb/io_ibuf_thread  
[init threads] os name:ib_io_log -- mysql name:thread/innodb/io_log_thread  
[init threads] os name:ib_io_rd-1 -- mysql name:thread/innodb/io_read_thread  
[init threads] os name:ib_io_rd-2 -- mysql name:thread/innodb/io_read_thread  
[init threads] os name:ib_io_rd-3 -- mysql name:thread/innodb/io_read_thread  
[init threads] os name:ib_io_rd-4 -- mysql name:thread/innodb/io_read_thread  
[init threads] os name:ib_io_rd-5 -- mysql name:thread/innodb/io_read_thread  
[init threads] os name:ib_io_rd-6 -- mysql name:thread/innodb/io_read_thread  
[init threads] os name:ib_io_rd-7 -- mysql name:thread/innodb/io_read_thread  
[init threads] os name:ib_io_rd-8 -- mysql name:thread/innodb/io_read_thread  
[init threads] os name:ib_io_wr-1 -- mysql name:thread/innodb/io_write_thread  
[init threads] os name:ib_io_wr-2 -- mysql name:thread/innodb/io_write_thread  
[init threads] os name:ib_io_wr-3 -- mysql name:thread/innodb/io_write_thread  
[init threads] os name:ib_io_wr-4 -- mysql name:thread/innodb/io_write_thread  
[init threads] os name:ib_pg_flush_co -- mysql name:thread/innodb/page_flush_coordinator_thread  
[init threads] os name:ib_pg_flush-1 -- mysql name:thread/innodb/page_flush_thread  
[init threads] os name:ib_pg_flush-2 -- mysql name:thread/innodb/page_flush_thread  
[init threads] os name:ib_pg_flush-3 -- mysql name:thread/innodb/page_flush_thread  
[init threads] os name:ib_recv_write -- mysql name:thread/innodb/recv_writer_thread  
[init threads] os name:ib_log_checkpt -- mysql name:thread/innodb/log_checkpointer_thread  
[init threads] os name:ib_log_fl_notif -- mysql name:thread/innodb/log_flush_notifier_thread  
[init threads] os name:ib_log_flush -- mysql name:thread/innodb/log_flusher_thread  
[init threads] os name:ib_log_wr_notif -- mysql name:thread/innodb/log_write_notifier_thread  
[init threads] os name:ib_log_writer -- mysql name:thread/innodb/log_writer_thread  
[init threads] os name:ib_par_rseg-0 -- mysql name:thread/innodb/parallel_rseg_init_thread  
[init threads] os name:ib_par_rseg-0 -- mysql name:thread/innodb/parallel_rseg_init_thread  
[init threads] os name:ib_srv_lock_to -- mysql name:thread/innodb/srv_lock_timeout_thread  
[init threads] os name:ib_srv_err_mon -- mysql name:thread/innodb/srv_error_monitor_thread  
[init threads] os name:ib_srv_mon -- mysql name:thread/innodb/srv_monitor_thread  
[init threads] os name:ib_buf_resize -- mysql name:thread/innodb/buf_resize_thread  
[init threads] os name:ib_src_main -- mysql name:thread/innodb/srv_master_thread  
[init threads] os name:ib_dict_stats -- mysql name:thread/innodb/dict_stats_thread  
[init threads] os name:ib_fts_opt -- mysql name:thread/innodb/fts_optimize_thread  
[init threads] os name:xpl_worker-1 -- mysql name:thread/mysqlx/worker  
[init threads] os name:xpl_worker-2 -- mysql name:thread/mysqlx/worker  
[init threads] os name:xpl_accept-1 -- mysql name:thread/mysqlx/acceptor_network  
[init threads] os name:ib_buf_dump -- mysql name:thread/innodb/buf_dump_thread  
[init threads] os name:ib_clone_gtid -- mysql name:thread/innodb/clone_gtid_thread  
[init threads] os name:ib_srv_purge -- mysql name:thread/innodb/srv_purge_thread  
[init threads] os name:ib_srv_wkr-1 -- mysql name:thread/innodb/srv_worker_thread  
[init threads] os name:ib_srv_wkr-2 -- mysql name:thread/innodb/srv_worker_thread  
[init threads] os name:ib_srv_wkr-3 -- mysql name:thread/innodb/srv_worker_thread  
[init threads] os name:sig_handler -- mysql name:thread/sql/signal_handler  
[init threads] os name:xpl_accept-2 -- mysql name:thread/mysqlx/acceptor_network  
[init threads] os name:xpl_accept-3 -- mysql name:thread/mysqlx/acceptor_network  
[init threads] os name:gtid_zip -- mysql name:thread/sql/compress_gtid_table  
[init threads] os name:connection -- mysql name:thread/sql/one_connection

可以看到建立的线程非常的多，但是我们得到它们的对应关系这就够了。这里不一一讨论每个线程的功能了，不过大部分我们都非常熟悉了，比如purge线程/cleaner线程，这里我列出一些，其他的就自己看看吧。

purge线程 srv_purge_thread 主要用于清理delete flag和释放undo表空间；
clean线程 page_cleaner_thread DBWR 主要用于进行脏数据的刷盘和LRU链表的管理；
异步IO线程 io_read_thread/io_write_thread 通常数据预读和刷脏会使用到异步AIO,用于合并可能的散列IO为连续IO提高性能；
字典收集线程 dict_stats_thread 数据修改的10%后会触发统计数据的收集；
锁超时监控线程srv_lock_timeout_thread 用于监控innodb行锁的超时，超时进行事务回退；
GTID压缩线程compress_gtid_table 将历史的gtid压缩为范围，避免gtid_executed表过长
slave io线程 slave_io 从库接收来自DUMP线程的binlog Event。将这些Event写入到relay log；
slave dump线程这是前台线程主库监控binlog的变化，发送binlog Event
slave sql线程 slave_sql 从库负责执行binlog Event；
用户线程one_connection 一个session就是一个用户线程，对于用户线程而言。可以通过processlist_id和show processlist；
srv_error_monitor_thread 所谓的信号量监控线程，注意这个信号量不是OS的信号量。是Innodb内部的rw lock和mutex；
srv_monitor_thread 这个线程当参数innodb_status_output打开的情况下，每15秒输出一个show engine innodb status信息到日志文件。也会自动开启比如Innodb内存不足会自动开启。
log_writer线程：将redo写到redo文件；
log_write_notifier线程：通知用户会话写入redo文件结束；
log_flusher线程：将redo fsync到redo文件；
log_flush_notifier线程：通知用户会话fsync结束；
log_checkpointer线程：定期检查脏数据写盘的redo 位置。

五、用pthread_setname_np为线程命名

这里我就随便写了4个循环的线程调用这个函数为我的线程命名为D-GPWK，需要耗用较高的CPU，看到的结果如下：

image.png

如果这个线程是MySQL的线程，当看到这个结果，我们就能明白大概的方向了。

《MySQL主从原理》作者：高鹏（八怪）