天天看点

ORA-00060的示例与若干场景

create table eg_60 ( num number, txt varchar2(10) );

insert into eg_60 values ( 1, 'First' );

insert into eg_60 values ( 2, 'Second' );

SQL> select rowid, num, txt from eg_60;

ROWID                     NUM TXT

------------------ ---------- ----------

AAAQT2AAHAAAEdYAAA          1 First

AAAQT2AAHAAAEdYAAB          2 Second

Session1:

update eg_60 set txt='ses1' where num=1;

Session2:

update eg_60 set txt='ses2' where num=2;

update eg_60 set txt='ses2' where num=1;

update eg_60 set txt='ses1' where num=2;

此时Session2报的错:

update eg_60 set txt='ses2' where num=1

       *

ERROR at line 1:

ORA-00060: deadlock detected while waiting for resource

Session1的update eg_60 set txt='ses1' where num=2;仍处hang状态,此时Session2执行exit正常退出,则

1 row updated.

原因就是正常退出,Oracle的PMON会自动rollback所做的未Commit操作,释放了num=2的资源,因此Session1可以执行。

出现60错误会产生一个trace文件,查看trace文件位置:

show parameter user_dump

NAME                                  TYPE       VALUE

------------------------------------ ----------- ------------------------------

user_dump_dest                 string      /opt/app/ora10g/admin/petest/udump

查看trace文件:

Deadlock graph:

                       ---------Blocker(s)--------  ---------Waiter(s)---------

Resource Name          process session holds waits  process session holds waits

TX-00090004-00019887        25     478     X             24     459           X

TX-000a002d-00032a8d        24     459     X             25     478           X

session 478: DID 0001-0019-00027AEC     session 459: DID 0001-0018-000CDDD8

session 459: DID 0001-0018-000CDDD8     session 478: DID 0001-0019-00027AEC

Rows waited on:

Session 459: obj - rowid = 000104F6 - AAAQT2AAHAAAEdYAAB

  (dictionary objn - 66806, file - 7, block - 18264, slot - 1)

Session 478: obj - rowid = 000104F6 - AAAQT2AAHAAAEdYAAA

  (dictionary objn - 66806, file - 7, block - 18264, slot - 0)

Information on the OTHER waiting sessions:

Session 459:

  pid=24 serial=34722 audsid=899246 user: 65/DCSOPEN

  O/S info: user: dcsopen, term: pts/0, ospid: 8838, machine: vm-vmw4131-t

            program: sqlplus@vm-vmw4131-t (TNS V1-V3)

  application name: SQL*Plus, hash value=3669949024

  Current SQL Statement:

  update eg_60 set txt='ses1' where num=2

End of information on OTHER waiting sessions.

Current SQL statement for this session:

===================================================

这里66806代表的OBJECT_ID对应object是eg_60。当前执行的SQL是update eg_60 set txt='ses2' where num=1,是这条SQL报的60错误,原因是由于update eg_60 set txt='ses1' where num=2这条SQL。因为这里是在同一台机器开的两个session,如果是不同机器客户端访问数据库做的这个实验,就可以根据machine: vm-vmw4131-t知道是哪个客户端执行的这条SQL。

通过PROCESS STATE节中O/S info: user: dcsopen, term: pts/1, ospid: 13112, machine: vm-vmw4131-t也可以知道是哪个客户端执行SQL报的60错误。

以上是同一张表不同session之间产生的死锁。还有另外一种场景,也是之前这边应用碰到的问题,即不同表之间的死锁,刚刚初步得解的,其实上述两种都属于事务级别的死锁,这里可能说的不准确,就是因为执行一个SQL后没有commit或rollback,再执行另外一个SQL,这两个SQL形成一个事务,造成可能的死锁。

关于事务,Concept中的解释:

A transaction is a logical unit of work that contains one or more SQL statements. A transaction is an atomic unit. The effects of all the

SQL statements in a transaction can be either all committed (applied to the database) or all rolled back (undone from the database).

A transaction begins with the first executable SQL statement. A transaction ends when it is committed or rolled back, either explicitly

with a COMMIT or ROLLBACK statement or implicitly when a DDL statement is issued.

比如:

UPDATE TABLE1,UPDATE TABLE2 ...

DELETE TABLE2, DELETE TABLE1 ...

此时碰巧可能出现互相持有对方需要的资源,导致deadlock。

对于这种情况,可能的解决方法就是:将表的顺序改为一致,或者拆分更小的事务,避免较差更新的情况。

另外,对于批量更新,和上面一个事务中多个表操作的原理相同,并发大则也会导致deadlock。要么减少并发,要么不用批量更新。其实出现deadlock,此时报错的SQL会自动执行rollback,但注意这里是语句级rollback,即只会rollback出错的SQL,之前的SQL不会rollback,所以上述第一个例子中Session2中的SQL报错,但Session1中的第二条SQL仍旧处于hang状态,除非Session2中对应的资源释放。