代码改变世界

Oracle RAC环境下定位并杀掉最终阻塞的会话

2018-04-04 16:29  AlfredZhao  阅读(2877)  评论(1编辑  收藏  举报

实验环境:Oracle RAC 11.2.0.4 (2节点)

之前其实也写过一篇相关文章:

但上文给出的例子过于简单,实际对于生产中复杂的阻塞问题,一步步找最终阻塞就比较麻烦。所以本篇旨在寻求更好更快捷的办法。

1.模拟故障:会话被级联阻塞

**准备工作:**我这里在每个实例开两个会话来模拟RAC在负载均衡模式下的业务会话: 实例1:会话1,会话2; 实例2:会话3,会话4; 在 时间点1 -> 时间点2 -> 时间点3 -> 时间点4 的这个时间轴上分别执行以下操作:

时间点1:
在实例1的会话1(INS1-session1)执行语句未提交或回滚:

select * from v$mystat where rownum = 1;
update emp set sal = 8000 where empno = 7788;

时间点2:
在实例2的会话3(INS2-session3)执行语句:

select * from v$mystat where rownum = 1;
delete from emp where empno = 7839;
update emp set job = 'MANAGER' where empno = 7788;
rollback;

时间点3:
在实例2的会话4(INS2-session4)执行语句:

select * from v$mystat where rownum = 1;
update emp set sal = 15000 where empno = 7839;
rollback;

时间点4:
在实例1的会话2(INS1-session2)执行语句:

select * from v$mystat where rownum = 1;
update emp set job = 'CEO' where empno = 7839;
rollback;

此时可以看到,在后面3个时间点进行操作的会话均hang住,显然都是被阻塞了。4个会话的现象如下:




那么他们究竟都是被谁阻塞了呢?下文会详细分析。

2.常规方法:梳理找出最终阻塞会话

我们常规会去GV$SESSION查询blocking_session,再看这个blocking_session有没有又被其他会话阻塞,直到找到根源。
--blocking
set lines 180
col program for a30
col machine for a20
select inst_id,
       SID,
       SERIAL#,
	   event,
	   machine,
       sql_id,
       blocking_session,
       blocking_instance
  from gv$session
 where blocking_session is not null;

结果如下:

SYS@jyzhao1 >--blocking
SYS@jyzhao1 >set lines 180
SYS@jyzhao1 >col program for a30
SYS@jyzhao1 >col machine for a20
SYS@jyzhao1 >select inst_id,
  2         SID,
  3         SERIAL#,
  4        event,
  5        machine,
  6         sql_id,
  7         blocking_session,
  8         blocking_instance
  9    from gv$session
 10   where blocking_session is not null;

   INST_ID        SID    SERIAL# EVENT                                    MACHINE              SQL_ID        BLOCKING_SESSION BLOCKING_INSTANCE
---------- ---------- ---------- ---------------------------------------- -------------------- ------------- ---------------- -----------------
         1        146       6283 enq: TX - row lock contention            jyrac1               052gy77vp276s               25                 2
         2         25      10250 enq: TX - row lock contention            jyrac2               3t2npbvdcf2d2              150                 1
         2        145      32069 enq: TX - row lock contention            jyrac2               0ct116qw46shq               25                 2

SYS@jyzhao1 >

可以看到实例1的sid=146的会话以及实例2的sid=145的会话都被实例2的sid=25的会话阻塞,而实例2的sid=25的这个会话又被实例1的sid=150的会话阻塞。这个例子只模拟了几个会话尚且可以快速定位,但如果是真实故障,很可能受影响的不止这么几个会话,虽然也可以慢慢最终找出来,但毕竟会看的眼花缭乱是不是。我们高傲的DBA又怎么会甘心一直去做这种事情呢?

3.改进方法:立即找出最终阻塞会话

之前我在单实例或者确认业务只跑在某一个节点的环境,一直在用的一个找出最终阻塞会话的脚本:
--cascade blocking
set lines 200 pages 100
col tree for a30
col event for a40
select *
  from (select a.sid, a.serial#,
               a.sql_id,
               a.event,
               a.status,
               connect_by_isleaf as isleaf,
               sys_connect_by_path(SID, '<-') tree,
               level as tree_level
          from v$session a
         start with a.blocking_session is not null
        connect by nocycle a.sid = prior a.blocking_session)
 where isleaf = 1
 order by tree_level asc;

这个脚本用到了start with connect by prior 的递归查询用法,非常方便可以直接找出最终阻塞的会话;可如果是RAC,业务是负载均衡跑在多个节点的,那上面的这个脚本就不好用了,比如我上面构造的这个例子,就需要明确查出各个会话分别在哪个实例上,否则你怎么确认去哪里杀呢,怎么办呢?其实也简单,只需要稍加改动下这个脚本即可,改后如下:

--cascade blocking@gv$session
select *
  from (select a.inst_id, a.sid, a.serial#,
               a.sql_id,
               a.event,
               a.status,
               connect_by_isleaf as isleaf,
               sys_connect_by_path(a.SID||'@'||a.inst_id, ' <- ') tree,
               level as tree_level
          from gv$session a
         start with a.blocking_session is not null
        connect by (a.sid||'@'||a.inst_id) = prior (a.blocking_session||'@'||a.blocking_instance))
 where isleaf = 1
 order by tree_level asc;

结果如下:

SYS@jyzhao1 >--cascade blocking@gv$session
SYS@jyzhao1 >select *
  2    from (select a.inst_id, a.sid, a.serial#,
  3                 a.sql_id,
  4                 a.event,
  5                 a.status,
  6                 connect_by_isleaf as isleaf,
  7                 sys_connect_by_path(a.SID||'@'||a.inst_id, ' <- ') tree,
  8                 level as tree_level
  9            from gv$session a
 10           start with a.blocking_session is not null
 11          connect by (a.sid||'@'||a.inst_id) = prior (a.blocking_session||'@'||a.blocking_instance))
 12   where isleaf = 1
 13   order by tree_level asc;

   INST_ID        SID    SERIAL# SQL_ID        EVENT                                    STATUS       ISLEAF TREE                           TREE_LEVEL
---------- ---------- ---------- ------------- ---------------------------------------- -------- ---------- ------------------------------ ----------
         1        150       8742               SQL*Net message from client              INACTIVE          1  <- 25@2 <- 150@1                       2
         1        150       8742               SQL*Net message from client              INACTIVE          1  <- 145@2 <- 25@2 <- 150@1              3
         1        150       8742               SQL*Net message from client              INACTIVE          1  <- 146@1 <- 25@2 <- 150@1              3

SYS@jyzhao1 >

非常清晰可以看到最终阻塞其他会话的会话是实例1的sid=150,serial#=8742的会话。
那么与相关人员都确认后,就可以直接杀掉这个最终阻塞会话:

SYS@jyzhao1 >alter system kill session '150,8742' immediate;

System altered.

再次查询,恢复正常,不再有堵塞了:

SYS@jyzhao1 >--cascade blocking@gv$session
SYS@jyzhao1 >select *
  2    from (select a.inst_id, a.sid, a.serial#,
  3                 a.sql_id,
  4                 a.event,
  5                 a.status,
  6                 connect_by_isleaf as isleaf,
  7                 sys_connect_by_path(a.SID||'@'||a.inst_id, ' <- ') tree,
  8                 level as tree_level
  9            from gv$session a
 10           start with a.blocking_session is not null
 11          connect by (a.sid||'@'||a.inst_id) = prior (a.blocking_session||'@'||a.blocking_instance))
 12   where isleaf = 1
 13   order by tree_level asc;

no rows selected

SYS@jyzhao1 >


至此,就达到了我们在RAC环境中快速定位并杀掉这种最终阻塞会话的目的。