MySQL redo lock 死锁问题排查 & 解决过程

1,695 阅读15分钟
原文链接: www.qcloud.com

周一上班,首先向同事了解了一下上周的测试情况,被告知在多实例场景下 MySQL Server hang 住,无法测试下去,原生版本不存在这个问题,而新版本上出现了这个问题,不禁心头一颤,心中不禁感到奇怪,好在现场环境还在,为排查问题提供了一个好的环境,随投入到紧张的问题排查过程当中……,问题实例表现如下:

  • 并发量为 384 的时候出现的问题;
  • MySQL 服务器无法执行事务相关的语句,即使简单的 select 语句也无法执行;
  • 所有线程处于等待状态,无法 KILL;

现场环境的收集

  • 首先,通过 pstack 工具获取当前问题实例的堆栈信息以便后面具体线程的查找 & 问题线程的定位:
    pstack { pidof hang mysqld server } > 20002.info
  • 使用 pt-pmp 工具统计 hang.info 中的进程信息,如下:
    [root@TENCENT64 ~]# pt-pmp 20002.info
    305 select(libc.so.6),os_thread_sleep(os0thread.cc:261),srv_conc_enter_innodb_with_atomics(srv0conc.cc:276),srv_conc_enter_innodb(srv0conc.cc:276),innobase_srv_conc_enter_innodb(ha_innodb.cc:1173),ha_innobase::index_read(ha_innodb.cc:1173),handler::index_read_idx_map(handler.cc:6846),handler::ha_index_read_idx_map(handler.cc:2787),join_read_(handler.cc:2787),join_read__table(handler.cc:2787),make_join_statistics(sql_optimizer.cc:3592),JOIN::optimize(sql_optimizer.cc:363),mysql_execute_select(sql_select.cc:1087),mysql_select(sql_select.cc:1087),handle_select(sql_select.cc:110),execute_sqlcom_select(sql_parse.cc:5156),mysql_execute_command(sql_parse.cc:2656),mysql_parse(sql_parse.cc:6408),dispatch_command(sql_parse.cc:1340),do_command(sql_parse.cc:1037),do_handle_one_connection(sql_connect.cc:990),handle_one_connection(sql_connect.cc:906),start_thread(libpthread.so.0),clone(libc.so.6)
     68 pthread_cond_wait,os_cond_wait(os0sync.cc:214),os_event_wait_low(os0sync.cc:214),sync_array_wait_event(sync0arr.cc:424),mutex_spin_wait(sync0sync.cc:580),mutex_enter_func(sync0sync.ic:218),pfs_mutex_enter_func(sync0sync.ic:218),log_reserve_and_write_fast(sync0sync.ic:218),mtr_log_reserve_and_write(sync0sync.ic:218),mtr_commit(sync0sync.ic:218),trx_prepare(trx0trx.cc:2206),trx_prepare_for_mysql(trx0trx.cc:2258),innobase_xa_prepare(ha_innodb.cc:13618),ha_prepare_low(handler.cc:2104),ha_commit_trans(handler.cc:1434),trans_commit(transaction.cc:228),mysql_execute_command(sql_parse.cc:4302),mysql_parse(sql_parse.cc:6408),dispatch_command(sql_parse.cc:1340),do_command(sql_parse.cc:1037),do_handle_one_connection(sql_connect.cc:990),handle_one_connection(sql_connect.cc:906),start_thread(libpthread.so.0),clone(libc.so.6)
     61 select(libc.so.6),os_thread_sleep(os0thread.cc:261),srv_conc_enter_innodb_with_atomics(srv0conc.cc:276),srv_conc_enter_innodb(srv0conc.cc:276),innobase_srv_conc_enter_innodb(ha_innodb.cc:1173),ha_innobase::index_read(ha_innodb.cc:1173),ha_index_read_map(handler.cc:2752),handler::read_range_first(handler.cc:2752),handler::multi_range_read_next(handler.cc:5864),QUICK_RANGE_SELECT::get_next(opt_range.cc:10644),rr_quick(records.cc:369),mysql_update(sql_update.cc:742),mysql_execute_command(sql_parse.cc:3336),mysql_parse(sql_parse.cc:6408),dispatch_command(sql_parse.cc:1340),do_command(sql_parse.cc:1037),do_handle_one_connection(sql_connect.cc:990),handle_one_connection(sql_connect.cc:906),start_thread(libpthread.so.0),clone(libc.so.6)
     14 pthread_cond_wait,os_cond_wait(os0sync.cc:214),os_event_wait_low(os0sync.cc:214),sync_array_wait_event(sync0arr.cc:424),mutex_spin_wait(sync0sync.cc:580),mutex_enter_func(sync0sync.ic:218),pfs_mutex_enter_func(sync0sync.ic:218),log_reserve_and_write_fast(sync0sync.ic:218),mtr_log_reserve_and_write(sync0sync.ic:218),mtr_commit(sync0sync.ic:218),trx_undo_assign_undo(trx0undo.cc:1813),trx_undo_report_row_operation(trx0rec.cc:1282),btr_cur_upd_lock_and_undo(btr0cur.cc:1710),btr_cur_update_in_place(btr0cur.cc:1710),btr_cur_optimistic_update(btr0cur.cc:2170),row_upd_clust_rec(row0upd.cc:2132),row_upd_clust_step(row0upd.cc:2466),row_upd(row0upd.cc:2521),row_upd_step(row0upd.cc:2521),row_update_for_mysql(row0mysql.cc:1779),ha_innobase::update_row(ha_innodb.cc:7091),handler::ha_update_row(handler.cc:7305),mysql_update(sql_update.cc:819),mysql_execute_command(sql_parse.cc:3336),mysql_parse(sql_parse.cc:6408),dispatch_command(sql_parse.cc:1340),do_command(sql_parse.cc:1037),do_handle_one_connection(sql_connect.cc:990),handle_one_connection(sql_connect.cc:906),start_thread(libpthread.so.0),clone(libc.so.6)
     10 __io_getevents_0_4(libaio.so.1),os_aio_linux_collect(os0file.cc:4975),os_aio_linux_handle(os0file.cc:4975),fil_aio_wait(fil0fil.cc:5796),io_handler_thread(srv0start.cc:492),start_thread(libpthread.so.0),clone(libc.so.6)
      5 pthread_cond_wait,inline_mysql_cond_wait(mysql_thread.h:1162),block_until_new_connection(mysql_thread.h:1162),one_thread_per_connection_end(mysql_thread.h:1162),do_handle_one_connection(sql_connect.cc:997),handle_one_connection(sql_connect.cc:906),start_thread(libpthread.so.0),clone(libc.so.6)
      4 pthread_cond_wait,os_cond_wait(os0sync.cc:214),os_event_wait_low(os0sync.cc:214),sync_array_wait_event(sync0arr.cc:424),mutex_spin_wait(sync0sync.cc:580),mutex_enter_func(sync0sync.ic:218),pfs_mutex_enter_func(sync0sync.ic:218),log_reserve_and_write_fast(sync0sync.ic:218),mtr_log_reserve_and_write(sync0sync.ic:218),mtr_commit(sync0sync.ic:218),btr_pcur_commit_specify_mtr(btr0pcur.ic:386),ibuf_btr_pcur_commit_specify_mtr(btr0pcur.ic:386),ibuf_merge_or_delete_for_page(btr0pcur.ic:386),buf_page_io_complete(buf0buf.cc:4313),fil_aio_wait(fil0fil.cc:5832),io_handler_thread(srv0start.cc:492),start_thread(libpthread.so.0),clone(libc.so.6)
      3 pthread_cond_wait,os_cond_wait(os0sync.cc:214),os_event_wait_low(os0sync.cc:214),sync_array_wait_event(sync0arr.cc:424),mutex_spin_wait(sync0sync.cc:580),mutex_enter_func(sync0sync.ic:218),pfs_mutex_enter_func(sync0sync.ic:218),log_reserve_and_write_fast(sync0sync.ic:218),mtr_log_reserve_and_write(sync0sync.ic:218),mtr_commit(sync0sync.ic:218),btr_pcur_commit_specify_mtr(btr0pcur.ic:386),ibuf_btr_pcur_commit_specify_mtr(btr0pcur.ic:386),ibuf_delete_rec(btr0pcur.ic:386),ibuf_merge_or_delete_for_page(ibuf0ibuf.cc:4884),buf_page_io_complete(buf0buf.cc:4313),fil_aio_wait(fil0fil.cc:5832),io_handler_thread(srv0start.cc:492),start_thread(libpthread.so.0),clone(libc.so.6)
      3 pthread_cond_wait,os_cond_wait(os0sync.cc:214),os_event_wait_low(os0sync.cc:214),srv_worker_thread(srv0srv.cc:2499),start_thread(libpthread.so.0),clone(libc.so.6)
      1 sigwait(libpthread.so.0),signal_hand(mysqld.cc:3217),start_thread(libpthread.so.0),clone(libc.so.6)
      1 select(libc.so.6),os_thread_sleep(os0thread.cc:261),srv_conc_enter_innodb_with_atomics(srv0conc.cc:276),srv_conc_enter_innodb(srv0conc.cc:276),innobase_srv_conc_enter_innodb(ha_innodb.cc:1173),ha_innobase::index_read(ha_innodb.cc:1173),ha_innobase::index_first(ha_innodb.cc:7883),handler::ha_index_first(handler.cc:2861),join_read_first(sql_executor.cc:2539),sub_select(sql_executor.cc:1259),do_select(sql_executor.cc:936),JOIN::exec(sql_executor.cc:936),mysql_execute_select(sql_select.cc:1101),mysql_select(sql_select.cc:1101),handle_select(sql_select.cc:110),execute_sqlcom_select(sql_parse.cc:5156),mysql_execute_command(sql_parse.cc:2656),mysql_parse(sql_parse.cc:6408),dispatch_command(sql_parse.cc:1340),do_command(sql_parse.cc:1037),do_handle_one_connection(sql_connect.cc:990),handle_one_connection(sql_connect.cc:906),start_thread(libpthread.so.0),clone(libc.so.6)
      1 pthread_cond_wait,os_cond_wait(os0sync.cc:214),os_event_wait_low(os0sync.cc:214),sync_array_wait_event(sync0arr.cc:424),rw_lock_s_lock_spin(sync0rw.cc:429),rw_lock_s_lock_func(sync0rw.ic:369),pfs_rw_lock_s_lock_func(sync0rw.ic:369),buf_page_get_gen(sync0rw.ic:369),trx_undo_page_get(trx0undo.ic:170),trx_undo_get_first_rec(trx0undo.ic:170),trx_purge_read_undo_rec(trx0purge.cc:741),trx_purge_choose_next_log(trx0purge.cc:782),trx_purge_get_next_rec(trx0purge.cc:888),trx_purge_fetch_next_rec(trx0purge.cc:960),trx_purge_attach_undo_recs(trx0purge.cc:960),trx_purge(trx0purge.cc:960),srv_do_purge(srv0srv.cc:2602),srv_purge_coordinator_thread(srv0srv.cc:2602),start_thread(libpthread.so.0),clone(libc.so.6)
      1 pthread_cond_wait,os_cond_wait(os0sync.cc:214),os_event_wait_low(os0sync.cc:214),sync_array_wait_event(sync0arr.cc:424),rw_lock_s_lock_spin(sync0rw.cc:429),rw_lock_s_lock_func(sync0rw.ic:369),pfs_rw_lock_s_lock_func(sync0rw.ic:369),buf_page_get_gen(sync0rw.ic:369),btr_block_get_func(btr0btr.ic:60),btr_cur_latch_leaves(btr0btr.ic:60),btr_cur_open_at_rnd_pos_func(btr0cur.cc:1047),btr_pcur_open_at_rnd_pos_func(btr0pcur.ic:548),ibuf_merge_pages(btr0pcur.ic:548),ibuf_merge(ibuf0ibuf.cc:2759),ibuf_contract_in_background(ibuf0ibuf.cc:2759),srv_master_do_active_tasks(srv0srv.cc:2116),srv_master_thread(srv0srv.cc:2116),start_thread(libpthread.so.0),clone(libc.so.6)
      1 pthread_cond_wait,os_cond_wait(os0sync.cc:214),os_event_wait_low(os0sync.cc:214),sync_array_wait_event(sync0arr.cc:424),mutex_spin_wait(sync0sync.cc:580),mutex_enter_func(sync0sync.ic:218),pfs_mutex_enter_func(sync0sync.ic:218),mtr_add_dirtied_pages_to_flush_list(sync0sync.ic:218),mtr_log_reserve_and_write(mtr0mtr.cc:270),mtr_commit(mtr0mtr.cc:270),row_upd_sec_index_entry(row0upd.cc:1819),row_upd_sec_step(row0upd.cc:1861),row_upd(row0upd.cc:1861),row_upd_step(row0upd.cc:1861),row_update_for_mysql(row0mysql.cc:1779),ha_innobase::update_row(ha_innodb.cc:7091),handler::ha_update_row(handler.cc:7305),mysql_update(sql_update.cc:819),mysql_execute_command(sql_parse.cc:3336),mysql_parse(sql_parse.cc:6408),dispatch_command(sql_parse.cc:1340),do_command(sql_parse.cc:1037),do_handle_one_connection(sql_connect.cc:990),handle_one_connection(sql_connect.cc:906),start_thread(libpthread.so.0),clone(libc.so.6)
      1 pthread_cond_wait,os_cond_wait(os0sync.cc:214),os_event_wait_low(os0sync.cc:214),sync_array_wait_event(sync0arr.cc:424),mutex_spin_wait(sync0sync.cc:580),mutex_enter_func(sync0sync.ic:218),pfs_mutex_enter_func(sync0sync.ic:218),log_write_up_to(sync0sync.ic:218),log_write_up_to(log0log.cc:1651),buf_flush_write_block_low(buf0flu.cc:919),buf_flush_page(buf0flu.cc:919),buf_flush_try_neighbors(buf0flu.cc:1276),buf_flush_page_and_try_neighbors(buf0flu.cc:1350),buf_do_flush_list_batch(buf0flu.cc:1350),buf_flush_batch(buf0flu.cc:1700),buf_flush_list(buf0flu.cc:1700),page_cleaner_do_flush_batch(buf0flu.cc:2166),page_cleaner_flush_pages_if_needed(buf0flu.cc:2166),buf_flush_page_cleaner_thread(buf0flu.cc:2166),start_thread(libpthread.so.0),clone(libc.so.6)
      1 pthread_cond_wait,os_cond_wait(os0sync.cc:214),os_event_wait_low(os0sync.cc:214),sync_array_wait_event(sync0arr.cc:424),mutex_spin_wait(sync0sync.cc:580),mutex_enter_func(sync0sync.ic:218),pfs_mutex_enter_func(sync0sync.ic:218),log_reserve_and_write_fast(sync0sync.ic:218),mtr_log_reserve_and_write(sync0sync.ic:218),mtr_commit(sync0sync.ic:218),row_ins_sec_index_entry_low(row0ins.cc:2779),row_ins_sec_index_entry(row0ins.cc:2930),row_upd_sec_index_entry(row0upd.cc:1834),row_upd_sec_step(row0upd.cc:1861),row_upd(row0upd.cc:1861),row_upd_step(row0upd.cc:1861),row_update_for_mysql(row0mysql.cc:1779),ha_innobase::update_row(ha_innodb.cc:7091),handler::ha_update_row(handler.cc:7305),mysql_update(sql_update.cc:819),mysql_execute_command(sql_parse.cc:3336),mysql_parse(sql_parse.cc:6408),dispatch_command(sql_parse.cc:1340),do_command(sql_parse.cc:1037),do_handle_one_connection(sql_connect.cc:990),handle_one_connection(sql_connect.cc:906),start_thread(libpthread.so.0),clone(libc.so.6)
      1 pthread_cond_wait,os_cond_wait(os0sync.cc:214),os_event_wait_low(os0sync.cc:214),sync_array_wait_event(sync0arr.cc:424),mutex_spin_wait(sync0sync.cc:580),mutex_enter_func(sync0sync.ic:218),pfs_mutex_enter_func(sync0sync.ic:218),log_reserve_and_write_fast(sync0sync.ic:218),mtr_log_reserve_and_write(sync0sync.ic:218),mtr_commit(sync0sync.ic:218),ibuf_mtr_commit(ibuf0ibuf.ic:58),ibuf_merge_or_delete_for_page(ibuf0ibuf.ic:58),buf_page_io_complete(buf0buf.cc:4313),fil_aio_wait(fil0fil.cc:5832),io_handler_thread(srv0start.cc:492),start_thread(libpthread.so.0),clone(libc.so.6)
      1 pthread_cond_wait,os_cond_wait(os0sync.cc:214),os_event_wait_low(os0sync.cc:214),sync_array_wait_event(sync0arr.cc:424),mutex_spin_wait(sync0sync.cc:580),mutex_enter_func(sync0sync.ic:218),pfs_mutex_enter_func(sync0sync.ic:218),log_get_lsn(sync0sync.ic:218),srv_error_monitor_thread(sync0sync.ic:218),start_thread(libpthread.so.0),clone(libc.so.6)
      1 pthread_cond_wait,os_cond_wait(os0sync.cc:214),os_event_wait_low(os0sync.cc:214),buf_dump_thread(buf0dump.cc:594),start_thread(libpthread.so.0),clone(libc.so.6)
      1 pthread_cond_timedwait,os_cond_wait_timed(os0sync.cc:177),os_event_wait_time_low(os0sync.cc:177),srv_monitor_thread(srv0srv.cc:1551),start_thread(libpthread.so.0),clone(libc.so.6)
      1 pthread_cond_timedwait,os_cond_wait_timed(os0sync.cc:177),os_event_wait_time_low(os0sync.cc:177),lock_wait_timeout_thread(lock0wait.cc:503),start_thread(libpthread.so.0),clone(libc.so.6)
      1 pthread_cond_timedwait,os_cond_wait_timed(os0sync.cc:177),os_event_wait_time_low(os0sync.cc:177),ib_wqueue_timedwait(ut0wqueue.cc:154),fts_optimize_thread(fts0opt.cc:3000),start_thread(libpthread.so.0),clone(libc.so.6)
      1 pthread_cond_timedwait,os_cond_wait_timed(os0sync.cc:177),os_event_wait_time_low(os0sync.cc:177),dict_stats_thread(dict0stats_bg.cc:349),start_thread(libpthread.so.0),clone(libc.so.6)
      1 poll(libc.so.6),vio_io_wait(viosocket.c:771),vio_socket_io_wait(viosocket.c:68),vio_read(viosocket.c:123),net_read_raw_loop(net_serv.cc:669),net_read_packet_header(net_serv.cc:751),net_read_packet(net_serv.cc:751),my_net_read(net_serv.cc:894),do_command(sql_parse.cc:970),do_handle_one_connection(sql_connect.cc:990),handle_one_connection(sql_connect.cc:906),start_thread(libpthread.so.0),clone(libc.so.6)
      1 poll(libc.so.6),handle_connections_sockets(mysqld.cc:6320),mysqld_main(mysqld.cc:5730),__libc_start_main(libc.so.6),_start
    

问题分析

从堆栈上可以看出,有这样几类线程:

  • 等待进入 INNODB engine 层的用户线程,测试环境中 innodb_thread_concurrency=16, 当 INNODB 层中的活跃线程数目大于此值时则需要排队,所以会有大量的排队线程,这个参数的影响&作用本身就是一篇很不错的文章,由于篇幅有限,在此不做扩展,感兴趣者可以参考官方文档
  • 操作过程中需要写 redo log 的后台线程,主要包括 page cleaner 线程、异步 io threads等;
  • 正在读取Page页面的 purge 线程 & 操作 change buffer 的 master thread
  • 大量的需要写 redo log 的用户线程;

从以上的分类不难看出,所有需要写 redo log 的线程都在等待 log_sys->mutex,那么这个保护 redo log buffer 的 mutex 被究竟被哪个线程获取了呢,因此,我们可以顺着这个线索进行问题排查,需要解决以下问题:

  • 问题一:哪个线程获取了 log_sys->mutex
  • 问题二:获取 log_sys->mutex 的线程为什么没有继续执行下去,是在等其它锁还是其它原因?
  • 问题三:如果不是硬件问题,整个资源竟争的过程是如何的?

问题一 由表及里

在查找 log_sys->mutex 所属线程情况时,有两点可以帮助我们快速的定位到这个线程:

  • 由于 log_sys->mutex 同时只能被同一个线程获得,所以在 pt-pmp 的信息输出中就可以排除线程数目大于1的线程;
  • 此线程既然已经获取了 log_sys->mutex, 那就应该还是在写日志的过程中,因此重点可以查看写日志的逻辑,即包括:mtr_log_reserve_and_writelog_write_up_to 的堆栈;

顺着上面的思路很快的从 pstack 中找到了以下线程:

Thread 446 (Thread 0x7fbea736d700 (LWP 32591)):
#0  0x00007fd1b72c8705 in pthread_cond_wait@@GLIBC_2.3.2 () from /lib64/libpthread.so.0
#1  0x00000000009f47f1 in os_cond_wait (fast_mutex=0x7defd680, cond=0x7defd6c0) at ../storage/innobase/os/os0sync.cc:214
#2  os_event_wait_low (event=event@entry=0x7defd680, reset_sig_count=) at ../storage/innobase/os/os0sync.cc:592
#3  0x0000000000a5ca1e in sync_array_wait_event (arr=0x340b850, index=1) at ../storage/innobase/sync/sync0arr.cc:424
#4  0x0000000000a5e750 in mutex_spin_wait (mutex=mutex@entry=0x7defd410, file_name=file_name@entry=0xd2d348 "../storage/innobase/mtr/mtr0mtr.cc", line=line@entry=215) at ../storage/innobase/sync/sync0sync.cc:580
#5  0x00000000009ebc27 in mutex_enter_func (line=215, file_name=0xd2d348 "../storage/innobase/mtr/mtr0mtr.cc", mutex=0x7defd410) at ../storage/innobase/include/sync0sync.ic:218
#6  pfs_mutex_enter_func (line=215, file_name=0xd2d348 "../storage/innobase/mtr/mtr0mtr.cc", mutex=0x7defd410) at ../storage/innobase/include/sync0sync.ic:250
#7  mtr_add_dirtied_pages_to_flush_list (mtr=0x7fbea736ae10) at ../storage/innobase/mtr/mtr0mtr.cc:215
#8  0x00000000009ec34c in mtr_log_reserve_and_write (mtr=) at ../storage/innobase/mtr/mtr0mtr.cc:270
#9  mtr_commit (mtr=mtr@entry=0x7fbea736ae10) at ../storage/innobase/mtr/mtr0mtr.cc:324
#10 0x0000000000a4a477 in row_upd_sec_index_entry (node=node@entry=0x7fbb9c01cd18, thr=thr@entry=0x7fbb9c01d018) at ../storage/innobase/row/row0upd.cc:1819
#11 0x0000000000a4e73c in row_upd_sec_step (thr=, node=0x7fbb9c01cd18) at ../storage/innobase/row/row0upd.cc:1861
#12 row_upd (thr=0x7fbb9c01d018, node=0x7fbb9c01cd18) at ../storage/innobase/row/row0upd.cc:2556
#13 row_upd_step (thr=thr@entry=0x7fbb9c01d018) at ../storage/innobase/row/row0upd.cc:2673
#14 0x0000000000a339c8 in row_update_for_mysql (mysql_rec=mysql_rec@entry=0x7fbb9c017528 "\377\366\r4\001\350e/\001\065\071\062\067\060\067\064\067\065\063\063-46485490867-60617341003-10452679142-20155412750-96257700347-89668812213-35633691718-21407253672-69113394932", ' ' ..., prebuilt=0x7fbb9c01c568) at ../storage/innobase/row/row0mysql.cc:1779
#15 0x00000000009ac99e in ha_innobase::update_row (this=0x7fbb9c016f10, old_row=0x7fbb9c017528 "\377\366\r4\001\350e/\001\065\071\062\067\060\067\064\067\065\063\063-46485490867-60617341003-10452679142-20155412750-96257700347-89668812213-35633691718-21407253672-69113394932", ' ' ..., new_row=0x7fbb9c017300 "\377\366\r4\001\351e/\001\065\071\062\067\060\067\064\067\065\063\063-46485490867-60617341003-10452679142-20155412750-96257700347-89668812213-35633691718-21407253672-69113394932", ' ' ...) at ../storage/innobase/handler/ha_innodb.cc:7091
#16 0x00000000005b4332 in handler::ha_update_row (this=0x7fbb9c016f10, old_data=0x7fbb9c017528 "\377\366\r4\001\350e/\001\065\071\062\067\060\067\064\067\065\063\063-46485490867-60617341003-10452679142-20155412750-96257700347-89668812213-35633691718-21407253672-69113394932", ' ' ..., new_data=0x7fbb9c017300 "\377\366\r4\001\351e/\001\065\071\062\067\060\067\064\067\065\063\063-46485490867-60617341003-10452679142-20155412750-96257700347-89668812213-35633691718-21407253672-69113394932", ' ' ...) at ../sql/handler.cc:7305
#17 0x000000000076b145 in mysql_update (thd=thd@entry=0x7e604470, table_list=, fields=..., values=..., conds=0x7fbe2800aca0, order_num=, order=, limit=18446744073709551615, handle_duplicates=DUP_ERROR, ignore=false, found_return=found_return@entry=0x7fbea736bac0, updated_return=updated_return@entry=0x7fbea736c020) at ../sql/sql_update.cc:819
#18 0x00000000006fc2dd in mysql_execute_command (thd=thd@entry=0x7e604470) at ../sql/sql_parse.cc:3336
#19 0x0000000000700618 in mysql_parse (thd=thd@entry=0x7e604470, rawbuf=, length=, parser_state=parser_state@entry=0x7fbea736c770)
...

这里我们简单的介绍一下MySQL写 redo log 的过程(省略undo & buffer pool 部分),当对数据进行修改时,MySQL 会首先对针对操作类型记录不同的 redo 日志,主要过程是:

  • 记录操作前的数据,根据不同的类型生成不同的 redo 日志,redo 的类型可以参考文件:src/storage/innobase/include/mtr0mtr.h
  • 记录操作之后的数据,对于不同的类型会包含不同的内容,具体可以参考函数:recv_parse_or_apply_log_rec_body()
  • 写日志到 redo buffer,并将此次涉及到脏页的数据加入到 buffer_pool 的 flush list 链表中;
  • 根据 innodb_flush_log_at_trx_commit 的值来判断在commit 的时候是否进行 sync 操作;

上面的堆栈则是写Redo后将脏页加到 flush list 过程中时 hang 住了,即此线程在获取了 log_sys->mutex 后,在获取 log_sys->log_flush_order_mutex 的过程中 hang 住了,而此时有大量的线程在等待该线程释放log_sys->mutex锁,问题一 已经有了答案,那么log_sys->log_flush_order_mutex 是个什么东东,它又被哪个占用了呢?

说 明:MySQL 的 buffer pool 维护了一个有序的脏页链表 (flush list according LSN order),这样在做 checkpoint & log_free_check 的过程中可以很快的定位到 redo log 需要推进的位置,在将脏页加入
flush list 过程中需要对其上锁以保证 flush list 中 LSN 的有序性, 但是如果使用 log_sys->mutex,在并发量大的时候则会造成 log_sys->mutex 的 contention,进而引起性能问题,因此添加了
另外一个 mutex 来保护脏页按 LSN 的有序性,代码说明如下:

ib_mutex_t     log_flush_order_mutex;/*!< mutex to serialize access to
                                        the flush list when we are putting
                                        dirty blocks in the list. The idea
                                        behind this mutex is to be able
                                        to release log_sys->mutex during
                                        mtr_commit and still ensure that
                                        insertions in the flush_list happen
                                        in the LSN order. */

问题二 弹尽粮绝

在问题一的排查过程中我们确定了 log_sys->mutex 的所属线程, 这个线程在获得 log_sys->log_flush_order_mutex 的过程中 hang 住了,因此线程堆栈可以分以为下几类:

  • Thread 446, 获得 log_sys->mutex, 等待获取 log_sys->log_flush_order_mutex 以把脏页加入到 buffer_poolflush list中;
  • 需要获得 log_sys->mutex 以写日志或者读取日志信息的线程;
  • 未知线程获得 log_sys->log_flush_order_mutex,在做其它事情的时候被 hang 住;

因此,问题的关键是找到哪个线程获取了log_sys->log_flush_order_mutex

为了找到相关的线程做了以下操作:

  • 查找获取log_sys->log_flush_order_mutex 的地方;

    [root@TENCENT64 /data/workplace/CDB]# grep log_flush_order_mutex_enter storage/innobase/* -r --color
    storage/innobase/buf/buf0buf.cc:        log_flush_order_mutex_enter();
    storage/innobase/include/log0log.h:#define log_flush_order_mutex_enter() do {           \
    storage/innobase/log/log0recv.cc:               log_flush_order_mutex_enter();
    storage/innobase/mtr/mtr0mtr.cc:                log_flush_order_mutex_enter();
    
  • 结合现有 pstack 中的线程信息,仔细查看上述查找结果中的相关代码,发现基本没有线程获得log_sys->log_flush_order_mutex

  • gdb 进入 MySQL Server, 将log_sys->log_flush_order_mutex 打印出来,发现 {waiters=1; lock_word= 0}!!!,即 Thread 446 在等待一个空闲的 mutex,而这个Mutex也确实被等待,由于我们的版本为 Release 版本,所以很多有用的信息没有办法得到,而若用 debug 版本跑则很难重现问题,log_flush_order_mutex 的定义如下:
/** InnoDB mutex */
struct ib_mutex_t {
        os_event_t      event;  /*!< Used by sync0arr.cc for the wait queue */
        volatile lock_word_t    lock_word;      /*!< lock_word is the target  of the atomic test-and-set instruction when atomic operations are enabled. */

#if !defined(HAVE_ATOMIC_BUILTINS)
        os_fast_mutex_t
        os_fast_mutex;  /*!< We use this OS mutex in place of lock_word when atomic operations are not enabled */
#endif  
        ulint   waiters;        /*!< This ulint is set to 1 if there are (or may be) threads waiting in the global wait array for this mutex to be released. Otherwise, this is 0. */
        UT_LIST_NODE_T(ib_mutex_t)      list; /*!< All allocated mutexes are put into a list. Pointers to the next and prev. */

#ifdef UNIV_SYNC_DEBUG
        const char*     file_name;      /*!< File where the mutex was locked */
        ulint   line;           /*!< Line where the mutex was locked */
        ulint   level;          /*!< Level in the global latching order */
#endif /* UNIV_SYNC_DEBUG */
        const char*     cfile_name;/*!< File name where mutex created */
        ulint           cline;  /*!< Line where created */
        ulong           count_os_wait;  /*!< count of os_wait */
#ifdef UNIV_DEBUG

/** Value of mutex_t::magic_n */
# define MUTEX_MAGIC_N  979585UL

        os_thread_id_t thread_id; /*!< The thread id of the thread which locked the mutex. */
        ulint           magic_n;        /*!< MUTEX_MAGIC_N */
        const char*     cmutex_name;    /*!< mutex name */
        ulint           ib_mutex_type;  /*!< 0=usual mutex, 1=rw_lock mutex */

#endif /* UNIV_DEBUG */

#ifdef UNIV_PFS_MUTEX    struct PSI_mutex* pfs_psi;      /*!< The performance schema instrumentation hook */
#endif
};

由以上的分析可以得出 问题二 的答案:

  • 只有两个线程和log_sys->log_flush_order_mutex有关,其中一个是 Thread 446 线程, 另外一个则是最近一次调用 log_flush_order_mutex_exit() 的线程;
  • 现有线程中某个线程在释放log_sys->log_flush_order_mutex的过程中没有唤醒 Thread 446,导致Thread 446 hang 并造成其它线程不能获得 log_sys->mutex,进而造成实例不可用;
  • log_sys->log_flush_order_mutex 没有被任何线程获得;

问题三 绝处逢生

由问题二的分析过程可知 log_sys->log_flush_order_mutex 没有被任何线程获得,可是为什么 Thread 446 没有被唤醒呢,信号丢失还是程序问题?如果是信号丢失,为什么可以稳定复现?官方的bug list 列表中是没有类似的 Bug的,搜了一下社区,发现可用信息很少,这个时候分析好像陷入了死胡同,心里压力开始无形中变大……,好像没有办法,但是任何问题都是有原因的,找到了原因,也就是有解的了……,再一次将注意力移到了 Thread 446 的堆栈中,然后查看了函数:

/************************************************************//**
Append the dirty pages to the flush list. */
static
void
mtr_add_dirtied_pages_to_flush_list(
/*================================*/
        mtr_t*  mtr)    /*!< in/out: mtr */
{
        ut_ad(!srv_read_only_mode);

        /* No need to acquire log_flush_order_mutex if this mtr has
        not dirtied a clean page. log_flush_order_mutex is used to
        ensure ordered insertions in the flush_list. We need to
        insert in the flush_list iff the page in question was clean
        before modifications. */
        if (mtr->made_dirty) {
                log_flush_order_mutex_enter();
        }

        /* It is now safe to release the log mutex because the
        flush_order mutex will ensure that we are the first one
        to insert into the flush list. */
        log_release();

        if (mtr->modifications) {
                mtr_memo_note_modifications(mtr);
        }

        if (mtr->made_dirty) {
                log_flush_order_mutex_exit();
        }
}

由问题二的分析过程可以得出某线程在 log_flush_order_mutex_exit 的退出过程没有将 Thread 446 唤醒,那么就顺着这个函数找,看它如何唤醒其它本程的,在没有办法的时候也只有这样一步一步的分析代码,希望有些收获,随着函数调用的不断深入,将目光定在了 mutex_exit_func 上, 函数中的注释引起了我的注意:

/******************************************************************//**
NOTE! Use the corresponding macro mutex_exit(), not directly this function!
Unlocks a mutex owned by the current thread. */
UNIV_INLINE                             
void    mutex_exit_func(
/*============*/
        ib_mutex_t*     mutex)  /*!< in: pointer to mutex */
{
        ut_ad(mutex_own(mutex));

        ut_d(mutex->thread_id = (os_thread_id_t) ULINT_UNDEFINED);

#ifdef UNIV_SYNC_DEBUG
        sync_thread_reset_level(mutex);
#endif
        mutex_reset_lock_word(mutex);

        /* A problem: we assume that mutex_reset_lock word
        is a memory barrier, that is when we read the waiters
        field next, the read must be serialized in memory
        after the reset. A speculative processor might
        perform the read first, which could leave a waiting
        thread hanging indefinitely.

        Our current solution call every second
        sync_arr_wake_threads_if_sema_free()
        to wake up possible hanging threads if
        they are missed in mutex_signal_object. */

        if (mutex_get_waiters(mutex) != 0) {

                mutex_signal_object(mutex);
        }

#ifdef UNIV_SYNC_PERF_STAT
        mutex_exit_count++;
#endif
}

从上面的注释中可以得到两点信息:

  • 由于 memory barrier 的存在,mutex_get_waiters & mutex_reset_lock_word 的调用顺序可能与执行顺序相反,这种情况下会引起 hang 问题;
  • 专门写了一个函数 sync_arr_wake_threads_if_sema_free() 来解决上述问题;

由上面的注释可以看到,并不是信号丢失,而是多线程 memory barrier 的存在可能会造成指令执行的顺序的异常,这种问题确定存在,但既然有sync_arr_wake_threads_if_sema_free() 规避这个问题,为什么还会存在 hang 呢?有了这个线索,瞬间感觉有了些盼头……,经过查找 sync_arr_wake_threads_if_sema_free 只在srv_error_monitor_thread 有调用,这个线程是专门对 MySQL 内部异常情况进行监控并打印出 error 信息的线程,臭名昭著的 600S 自杀案也是它的杰作, 那么问题来了:

  • 机器周末都在 hang 着,为什么没有检测到异常并 abort 呢?
  • 既然sync_arr_wake_threads_if_sema_free 可以唤醒,为什么没有唤醒呢?

顺着这个思路,查看了pstack 中 srv_error_monitor_thread 的堆栈,可以发现此线程在获取 log_sys->mutex 的时候hang 住了,因此无法执行sync_arr_wake_threads_if_sema_free() & 常归的异常检查,正好回答了上面的问题,详细堆栈如下:

Thread 470 (Thread 0x7fbeb0a2e700 (LWP 20032)):
#0  0x00007fd1b72c8705 in pthread_cond_wait@@GLIBC_2.3.2 () from /lib64/libpthread.so.0
#1  0x00000000009f47f1 in os_cond_wait (fast_mutex=0x7defd5f0, cond=0x7defd630) at ../storage/innobase/os/os0sync.cc:214
#2  os_event_wait_low (event=event@entry=0x7defd5f0, reset_sig_count=) at ../storage/innobase/os/os0sync.cc:592
#3  0x0000000000a5ca1e in sync_array_wait_event (arr=0x340b850, index=86) at ../storage/innobase/sync/sync0arr.cc:424
#4  0x0000000000a5e750 in mutex_spin_wait (mutex=mutex@entry=0x7defd3c8, file_name=file_name@entry=0xd2d380 "../storage/innobase/include/log0log.ic", line=line@entry=409) at ../storage/innobase/sync/sync0sync.cc:580
#5  0x0000000000a552e9 in mutex_enter_func (line=409, file_name=0xd2d380 "../storage/innobase/include/log0log.ic", mutex=0x7defd3c8) at ../storage/innobase/include/sync0sync.ic:218
#6  pfs_mutex_enter_func (line=409, file_name=0xd2d380 "../storage/innobase/include/log0log.ic", mutex=0x7defd3c8) at ../storage/innobase/include/sync0sync.ic:250
#7  log_get_lsn () at ../storage/innobase/include/log0log.ic:409
#8  srv_error_monitor_thread (arg=) at ../storage/innobase/srv/srv0srv.cc:1722
#9  0x00007fd1b72c4df3 in start_thread () from /lib64/libpthread.so.0
#10 0x00007fd1b63ac1bd in clone () from /lib64/libc.so.6

经过上面的分析问题越来越明朗了,过程可以简单的归结为:

  • Thread 446 获得log_sys->mutex, 但是在等待 log_sys->log_flush_order_mutex 的过程中没有被唤醒;
  • Thread XXX 在释放 log_sys->log_flush_order_mutex 的过程中出现了 memory barrier 问题,没有唤醒 Thread 446 ;
  • Thread 470 获得 log_sys->mutex 时被 hang 住,导致无法执行 sync_arr_wake_threads_if_sema_free(), 导致了整个实例的 hang 住;
  • Thread 470 需要获得 Thread 446 的 log_sys->mutex, 而 Thread 446 需要被 Thread 470 唤醒才会释放 log_sys->mutex;

结合 log_sys->log_flush_order_mutex 的状态信息,实例 hang 住的整个过程如下:

THD_1                                                      THD_2                                        THD_N                             srv_error_monitor_thread   

mutex_enter(log_sys->mutex)
...
mutx_enter(log_sys->log_flush_order_mutex)
mutex_exit(log_sys->mutex)
mutx_exit(log_sys->log_flush_order_mutex)                  mutex_enter(log_sys->mutex)                  mutex_enter(log_sys->mutex)       mutex_enter(log_sys->mutex)
       mutex_get_waiters()                                 ...
                                                           mutx_enter(log_sys->log_flush_order_mutex)
                                                                 mutex_set_waiters
                                                                 sync_array_wait_event
       mutex_reset_lock_word()

关于 Memory barrier 的介绍可以参考①介绍

问题解决

既然知道了问题产生的原因,那么问题也就可以顺利解决了,有两种方法:

  • 直接移除 log_get_lsn 在此处的判断,本身就是开发人员加的一些判断信息,为了定位 LSN 的异常而写的,用到的时候也Crash了,用处不大;
  • 保留判断,将 log_get_lsn 修改为 log_peek_lsn, 后者会首先进行 try_lock,当发现上锁失败的时候会直接返回,而不进行判断,这种方法较优雅些;
    经过修改之后的版本在测试过程中没有没有再复现此问题;

问题扩展

虽然问题解决了,但官方版本中肯定存在着这个问题,为什么 buglist 没有找到相关信息呢,于是在查看了最新代码,发现这个问题已经修复,修复方法为上面列的第二种方法,详细的 commit message 信息如下:

commit 79032a7ae1b4e000028a64ab0d2f216d4c23767b
Author: Shaohua Wang 
Date:   Tue Jan 12 15:08:09 2016 +0800

BUG#22530768 Innodb freeze running REPLACE statements

we can see from the hang stacktrace, srv_monitor_thread is blocked
when getting log_sys::mutex, so that sync_arr_wake_threads_if_sema_free
cannot get a change to break the mutex deadlock.

The fix is simply removing any mutex wait in srv_monitor_thread.

Patch is reviewed by Sunny over IM.

bug 影响范围:MySQL 5.6.28 及之前的版本都有此问题