转载自：刘相兵 Maclean Liu 文章
你所不知道的后台进程 SMON 功能
SMON(system monitor process)系统监控后台进程，有时候也被叫做 system cleanup process，
这么叫的原因是它负责完成很多清理(cleanup)任务。但凡学习过 Oracle 基础知识的技术人员
都会或多或少对该 background process 的功能有所了解。
我们所熟知的 SMON 是个兢兢业业的家伙，它负责完成一些列系统级别的任务。与
PMON(Process Monitor)后台进程不同的是，SMON 负责完成更多和整体系统相关的工作，这
导致它会去做一些不知名的”累活”，当系统频繁产生这些”垃圾任务”，则 SMON 可能忙不过
来。因此在 10g 中 SMON 变得有一点懒惰了，如果它在短期内接收到过多的工作通知(SMON:
system monitor process posted)，那么它可能选择消极怠工以便让自己不要过于繁忙(SMON:
Posted too frequently, trans recovery disabled)，之后会详细介绍。
了解你所不知道的 SMON 功能(一):清理临时段
触发场景
很多人错误地理解了这里所说的临时段 temporary segments，认为 temporary segments 是指
temporary tablespace 临时表空间上的排序临时段(sort segment)。事实上这里的临时段主要
指的是永久表空间(permanent tablespace)上的临时段，当然临时表空间上的 temporary
segments 也是由 SMON 来清理(cleanup)的，但这种清理仅发生在数据库实例启动时(instance
startup)。
永久表空间上同样存在临时段，譬如当我们在某个永久表空间上使用 create table/index 等
DDL 命令创建某个表/索引时，服务进程一开始会在指定的永久表空间上分配足够多的区间
(Extents)，这些区间在命令结束之前都是临时的(Temporary Extents)，直到表/索引完全建成
才将该temporary segment转换为permanent segment。另外当使用drop命令删除某个段时，
也会先将该段率先转换为 temporary segment，之后再来清理该 temporary segment(DROP
object converts the segment to temporary and then cleans up the temporary segment)。 常规情
况下清理工作遵循谁创建 temporary segment，谁负责清理的原则。换句话说，因服务进程
rebuild index 所产生的 temporary segment 在 rebuild 完成后应由服务进程自行负责清理。一
旦服务进程在成功清理 temporary segment 之前就意外终止了，亦或者服务进程在工作过程
中遇到了某些 ORA-错误导致语句失败，那么 SMON 都会被要求(posted)负责完成 temporary
segment 的清理工作。
对于永久表空间上的 temporary segment，SMON 会三分钟清理一次(前提是接到 post)，如果
SMON 过于繁忙那么可能 temporary segment 长期不被清理。temporary segment 长期不被清理可能造成一个典型的问题是:在 rebuild index online 失败后，后续执行的 rebuild index 命令
要求之前产生的 temporary segment 已被 cleanup，如果 cleanup 没有完成那么就需要一直等
下去。在 10gR2 中我们可以使用 dbms_repair.online_index_clean 来手动清理 online index
rebuild 的遗留问题:
The dbms_repair.online_index_clean function has been created to cleanup online index
rebuilds.
Use the dbms_repair.online_index_clean function to resolve the issue.
Please note if you are unable to run the dbms_repair.online_index_clean function it is due
to the fact
that you have not installed the patch for Bug 3805539 or are not running on a release that
includes this fix.
The fix for this bug is a new function in the dbms_repair package called
dbms_repair.online_index_clean,
which has been created to cleanup online index [[sub]partition] [re]builds.
New functionality is not allowed in patchsets;
therefore, this is not available in a patchset but is available in 10gR2.
Check your patch list to verify the database is patched for Bug 3805539
using the following command and patch for the bug if it is not listed:
opatch lsinventory -detail
Cleanup after a failed online index [re]build can be slow to occurpreventing subsequent
such operations
until the cleanup has occured.
接着我们通过实践来看一下 smon 是如何清理永久表空间上的 temporary segment 的:
设置 10500 事件以跟踪 smon 进程，这个诊断事件后面会介绍
SQL> alter system set events '10500 trace name context forever,level 10';
System altered.
在第一个会话中执行 create table 命令，这将产生一定量的 Temorary ExtentsSQL> create table smon as select * from ymon;
在另一个会话中执行对 DBA_EXTENTS 视图的查询，可以发现产生了多少临时区间
SQL> SELECT COUNT(*) FROM DBA_EXTENTS WHERE SEGMENT_TYPE='TEMPORARY';
COUNT(*)
----------
117
终止以上 create table 的 session，等待一段时间后观察 smon 后台进程的 trc 可以发现以下信息:
*** 2011-06-07 21:18:39.817
SMON: system monitor process posted msgflag:0x0200 (-/-/-/-/TMPSDROP/-/-)
*** 2011-06-07 21:18:39.818
SMON: Posted, but not for trans recovery, so skip it.
*** 2011-06-07 21:18:39.818
SMON: clean up temp segments in slave
SQL> SELECT COUNT(*) FROM DBA_EXTENTS WHERE SEGMENT_TYPE='TEMPORARY';
COUNT(*)
----------
0
可以看到 smon 通过 slave 进程完成了对 temporary segment 的清理
与永久表空间上的临时段不同，出于性能的考虑临时表空间上的 Extents 并不在操作
(operations)完成后立即被释放和归还。相反，这些 Temporary Extents 会被标记为可用，以
便用于下一次的排序操作。SMON 仍会清理这些 Temporary segments，但这种清理仅发生在
实例启动时(instance startup):
For performance issues, extents in TEMPORARY tablespaces are not released ordeallocatedonce the operation is complete.Instead, the extent is simply marked as available for the
next sort operation.
SMON cleans up the segments at startup.
A sort segment is created by the first statement that used a TEMPORARY tablespacefor sorting,
after startup.
A sort segment created in a TEMPOARY tablespace is only released at shutdown.
The large number of EXTENTS is caused when the STORAGE clause has been incorrectly calculated.
现象
可以通过以下查询了解数据库中 Temporary Extent 的总数，在一定时间内比较其总数，若有
所减少那么说明 SMON 正在清理 Temporary segment
SELECT COUNT(*) FROM DBA_EXTENTS WHERE SEGMENT_TYPE='TEMPORARY';
也可以通过 v$sysstat 视图中的”SMON posted for dropping temp segment”事件统计信息来了
解 SMON 收到清理要求的情况:
SQL> select name,value from v$sysstat where name like '%SMON%';
NAME VALUE
---------------------------------------------------------------- ----------
total number of times SMON posted 8
SMON posted for undo segment recovery 0
SMON posted for txn recovery for other instances 0
SMON posted for instance recovery 0
SMON posted for undo segment shrink 0
SMON posted for dropping temp segment 1另外在清理过程中 SMON 会长期持有 Space Transacton(ST)队列锁，其他会话可能因为得不
到 ST 锁而等待超时出现 ORA-01575 错误:
01575, 00000, "timeout waiting for space management resource"
// *Cause: failed to acquire necessary resource to do space management.
// *Action: Retry the operation.
如何禁止 SMON 清理临时段
可以通过设置诊断事件 event=’10061 trace name context forever, level 10′禁用 SMON 清理临
时段(disable SMON from cleaning temp segments)。
alter system set events '10061 trace name context forever, level 10';
相关诊断事件
除去 10061 事件外还可以用 10500 事件来跟踪 smon 的 post 信息，具体的事件设置方法见
了解你所不知道的 SMON 功能(二):合并空闲区间
SMON 的作用还包括合并空闲区间(coalesces free extent)
触发场景
早期 Oracle 采用 DMT 字典管理表空间，不同于今时今日的 LMT 本地管理方式，DMT 下通过
对FET$和UET$2张字典基表的递归操作来管理区间。SMON每5分钟(SMON wakes itself every 5 minutes and checks for tablespaces with default pctincrease != 0)会自发地去检查哪些默认存
储参数 pctincrease 不等于 0 的字典管理表空间，注意这种清理工作是针对 DMT 的，而 LMT
则无需合并。SMON 对这些 DMT 表空间上的连续相邻的空闲 Extents 实施 coalesce 操作以合
并成一个更大的空闲 Extent，这同时也意味着 SMON 需要维护 FET$字典基表。
现象
以下查询可以检查数据库中空闲 Extents 的总数，如果这个总数在持续减少那么说明 SMON
正在 coalesce free space：
SELECT COUNT(*) FROM DBA_FREE_SPACE;
在合并区间时 SMON 需要排他地(exclusive)持有 ST(Space Transaction)队列锁，其他会话可能
因为得不到 ST 锁而等待超时出现 ORA-01575 错误。同时 SMON 可能在繁琐的 coalesce 操作
中消耗 100%的 CPU。
如何禁止 SMON 合并空闲区间
可以通过设置诊断事件event=’10269 trace name context forever, level 10′来禁用SMON合并空
闲区间(Don’t do coalesces of free space in SMON)
10269, 00000, "Don't do coalesces of free space in SMON"
// *Cause: setting this event prevents SMON from doing free space coalesces
alter system set events '10269 trace name context forever, level 10';了解你所不知道的 SMON 功能(三):清理 obj$基表
SMON 的作用还包括清理 obj$数据字典基表(cleanup obj$)
OBJ$字典基表是 Oracle Bootstarp 启动自举的重要对象之一:
SQL> set linesize 80 ;
SQL> select sql_text from bootstrap$ where sql_text like 'CREATE TABLE OBJ$%';
SQL_TEXT
--------------------------------------------------------------------------------
CREATE TABLE OBJ$("OBJ#" NUMBER NOT NULL,"DATAOBJ#" NUMBER,"OWNER#" NUMBER NOT N
ULL,"NAME" VARCHAR2(30) NOT NULL,"NAMESPACE" NUMBER NOT NULL,"SUBNAME" VARCHAR2(
30),"TYPE#" NUMBER NOT NULL,"CTIME" DATE NOT NULL,"MTIME" DATE NOT NULL,"STIME"
DATE NOT NULL,"STATUS" NUMBER NOT NULL,"REMOTEOWNER" VARCHAR2(30),"LINKNAME" VAR
CHAR2(128),"FLAGS" NUMBER,"OID$" RAW(16),"SPARE1" NUMBER,"SPARE2" NUMBER,"SPARE3
" NUMBER,"SPARE4" VARCHAR2(1000),"SPARE5" VARCHAR2(1000),"SPARE6" DATE) PCTFREE
10 PCTUSED 40 INITRANS 1 MAXTRANS 255 STORAGE ( INITIAL 16K NEXT 1024K MINEXTEN
TS 1 MAXEXTENTS 2147483645 PCTINCREASE 0 OBJNO 18 EXTENTS (FILE 1 BLOCK 121))
触发场景
OBJ$基表是一张低级数据字典表，该表几乎对库中的每个对象(表、索引、包、视图等)都包
含有一行记录。很多情况下，这些条目所代表的对象是不存在的对象(non-existent)，引起这
种现象的一种可能的原因是对象本身已经被从数据库中删除了，但是对象条目仍被保留下来
以满足消极依赖机制(negative dependency)。因为这些条目的存在会导致 OBJ$表不断膨胀，
这时就需要由 SMON 进程来删除这些不再需要的行。SMON 会在实例启动(after startup of DB
is started cleanup function again)时以及启动后的每 12 个小时执行一次清理任务(the cleanup
is scheduled to run after startup and then every 12 hours)。
我们可以通过以下演示来了解 SMON 清理 obj$的过程:
SQL> BEGIN
2 FOR i IN 1 .. 5000 LOOP
3 execute immediate ('create synonym gustav' || i || ' for 4 perfstat.sometable');
5 execute immediate ('drop synonym gustav' || i );
6 END LOOP;
7 END;
8 /
PL/SQL procedure successfully completed.
SQL> startup force;
ORACLE instance started.
Total System Global Area 1065353216 bytes
Fixed Size 2089336 bytes
Variable Size 486542984 bytes
Database Buffers 570425344 bytes
Redo Buffers 6295552 bytes
Database mounted.
Database opened.
SQL> select count(*) from user$ u, obj$ o
2 where u.user# (+)=o.owner# and o.type#=10 and not exists
3 (select p_obj# from dependency$ where p_obj# = o.obj#);
COUNT(*)
----------
5000
SQL> /
COUNT(*)
----------
5000
SQL> /
COUNT(*)
---------- 4951
SQL> oradebug setospid 18457;
Oracle pid: 8, Unix process pid: 18457, image: oracle@rh2.oracle.com (SMON)
SQL> oradebug event 10046 trace name context forever ,level 1;
Statement processed.
SQL> oradebug tracefile_name;
/s01/admin/G10R2/bdump/g10r2_smon_18457.trc
select o.owner#,
o.obj#,
decode(o.linkname,
null,
decode(u.name, null, 'SYS', u.name),
o.remoteowner),
o.name,
o.linkname,
o.namespace,
o.subname
from user$ u, obj$ o
where u.use r#(+) = o.owner#
and o.type# = :1
and not exists
(select p_obj# from dependency$ where p_obj# = o.obj#)
order by o.obj#
for update
select null
from obj$
where obj# = :1
and type# = :2
and obj# not in
(select p_obj# from dependency$ where p_obj# = obj$.obj#)
delete from obj$ where obj# = :1/* 删除过程其实较为复杂，可能要删除多个字典基表上的记录 */
现象
我们可以通过以下查询来了解 obj$基表中 NON-EXISTENT 对象的条目总数(type#=10)，若这
个总数在不断减少说明 smon 正在执行清理工作
select trunc(mtime), substr(name, 1, 3) name, count(*)
from obj$
where type# = 10
and not exists (select * from dependency$ where obj# = p_obj#)
group by trunc(mtime), substr(name, 1, 3);
select count(*)
from user$ u, obj$ o
where u.user#(+) = o.owner#
and o.type# = 10
and not exists (select p_obj# from dependency$ where p_obj# = o.obj#);
如何禁止 SMON 清理 obj$基表
我们可以通过设置诊断事件 event=’10052 trace name context forever’来禁止 SMON 清理
obj$基表，当我们需要避免 SMON 因 cleanup obj$的相关代码而意外终止或 spin 从而开展进
一步的诊断时可以设置该诊断事件。在 Oracle 并行服务器或 RAC 环境中，也可以设置该事
件来保证只有特定的某个节点来执行清理工作。
10052, 00000, "don't clean up obj$"
alter system set events '10052 trace name context
forever, level 65535';
Problem Description: We are receiving the below warning during db startup:
WARNING: kqlclo() has detected the following :
Non-existent object 37336 NOT deleted because an object
of the same name exists already.
Object name: PUBLIC.USER$
This is caused by the SMON trying to cleanup the SYS.OJB$.
SMON cleans all dropped objects which have a SYS.OBJ$.TYPE#=10.
This can happen very often when you create an object that have the same name as a public
synonym.
When SMON is trying to remove non-existent objects and fails because there are duplicates,
multiple nonexistent objects with same name.
This query will returned many objects with same name under SYS schema:
select o.name,u.user# from user$ u, obj$ o where u.user# (+)=o.owner# and o.type#=10
and not exists (select p_obj# from dependency$ where p_obj# = o.obj#);
To cleanup this message:Take a full backup of the database - this is crucial. If anything goes wrong during this
procedure,
your only option would be to restore from backup, so make sure you have a good backup before
proceeding.
We suggest a COLD backup. If you plan to use a HOT backup, you will have to restore point
in time if any problem happens
Normally DML against dictionary objects is unsupported,
but in this case we know exactly what the type of corruption,
also you are instructing to do this under guidance from Support.
Data dictionary patching must be done by an experienced DBA.
This solution is unsupported.
It means that if there were problems after applying this solution, a database backup must
be restored.
1. Set event 10052 at parameter file to disable cleanup of OBJ$ by SMON
EVENT="10052 trace name context forever, level 65535"
2. Startup database in restricted mode
3. Delete from OBJ$, COMMIT
SQL> delete from obj$ where (name,owner#) in ( select o.name,u.user# from user$ u, obj$ o
where u.user# (+)=o.owner# and o.type#=10 and not exists (select p_obj# from
dependency$ where p_obj# = o.obj#) );
SQL> commit;
SQL> Shutdown abort.
4. remove event 10052 from init.ora
5. Restart the database and monitor for the message in the ALERT LOG file了解你所不知道的 SMON 功能(四):维护 col_usage$字典基
表
SMON 的作用还包括维护 col_usage$列监控统计信息基表。
最早在 9i 中引入了 col_usage$字典基表，其目的在于监控 column 在 SQL 语句作为 predicate 的情况，
col_usage$的出现完善了 CBO 中柱状图自动收集的机制。
create table col_usage$
(
obj# number, /* object number */
intcol# number, /* internal column number */
equality_preds number, /* equality predicates */
equijoin_preds number, /* equijoin predicates */
nonequijoin_preds number, /* nonequijoin predicates */
range_preds number, /* range predicates */
like_preds number, /* (not) like predicates */
null_preds number, /* (not) null predicates */
timestamp date /* timestamp of last time this row was changed */
)
storage (initial 200K next 100k maxextents unlimited pctincrease 0)
/
create unique index i_col_usage$ on col_usage$(obj#,intcol#)
storage (maxextents unlimited)
/
在 10g 中我们默认使用’FOR ALL COLUMNS SIZE AUTO’的柱状图收集模式，而在 9i 中默认是’SIZE 1′
即默认不收集柱状图，这导致许多 9i 中正常运行的应用程序在 10g 中 CBO 执行计划异常，详见
<dbms_stats span="" <="">收集模式在 9i 和 10g 上的区别>;。’SIZE AUTO’意为由 Oracle 自动决定是否
收集柱状图及柱状图的桶数，Oracle 自行判断的依据就来源于 col_usage$字典基表，若表上的某一列曾在
硬解析(hard parse)过的 SQL 语句中充当过 predicate(通俗的说就是 where 后的 condition)的话，我们认为
此列上有收集柱状图的必要，那么 col_usage$上就会被加入该列曾充当 predicate 的记录。当
DBMS_STATS.GATHER_TABLE_STATS 存储过程以’SIZE AUTO’模式执行时，收集进程会检查
col_usage$基表以判断哪些列之前曾充当过 predicate，若充当过则说明该列有收集柱状图的价值。SMON 会每 15 分钟将 shared pool 中的 predicate columns 的数据刷新到 col_usage$基表中(until
periodically about every 15 minutes SMON flush the data into the data dictionary)，另外当 instance
shutdown 时 SMON 会扫描 col_usage$并找出已被 drop 表的相关 predicate columns 记录，并删除这部
分”orphaned”孤儿记录。
我们来具体了解 col_usage$的填充过程:
SQL> select * from v$version;
BANNER
----------------------------------------------------------------
Oracle Database 10g Enterprise Edition Release 10.2.0.4.0 - 64bi
PL/SQL Release 10.2.0.4.0 - Production
CORE 10.2.0.4.0 Production
TNS for Linux: Version 10.2.0.4.0 - Production
NLSRTL Version 10.2.0.4.0 - Production
SQL> select * from global_name;
GLOBAL_NAME
--------------------------------------------------------------------------------
www.oracle.com
SQL> create table maclean (t1 int);
Table created.
SQL> select object_id from dba_objects where object_name='MACLEAN';
OBJECT_ID
----------
1323013
SQL> select * from maclean where t1=1;
no rows selectedSQL> set linesize 200 pagesize 2000;
注意 col_usage$的数据同*_tab_modifications 类似，
从查询到数据刷新到 col_usage$存在一段时间的延迟，
所以我们立即查询 col_usage$将得不到任何记录，
可以手动执行 DBMS_STATS.FLUSH_DATABASE_MONITORING_INFO 将缓存中的信息刷新到字典上
SQL> select * from col_usage$ where obj#=1323013;
no rows selected
SQL> oradebug setmypid;
Statement processed.
针对 FLUSH_DATABASE_MONITORING_INFO 填充操作做 10046 level 12 trace
SQL> oradebug event 10046 trace name context forever,level 12;
Statement processed.
SQL> exec DBMS_STATS.FLUSH_DATABASE_MONITORING_INFO;
PL/SQL procedure successfully completed.
SQL> select * from col_usage$ where obj#=1323013;
OBJ# INTCOL# EQUALITY_PREDS EQUIJOIN_PREDS NONEQUIJOIN_PREDS RANGE_PREDS
LIKE_PREDS NULL_PREDS TIMESTAMP
---------- ---------- -------------- -------------- ----------------- -----------
---------- ---------- ---------
1323013 1 1 0 0 0 0
0 19-AUG-11
=============10046 trace content====================
lock table sys.col_usage$ in exclusive mode nowait在测试中可以发现 10.2.0.4 上 DBMS_STATS.FLUSH_DATABASE_MONITORING_INFO 存储过程会优先使用
lock in exclusive mode nowait 来锁住 col_usage$基表，
如果 lock 失败则会反复尝试 1100 次，
若仍不能锁住 col_usage$表则放弃更新 col_usage$上的数据，避免造成锁等待和死锁。
Cksxm.c
Monitor Modification Hash Table Base
modification hash table entry
modification hash table chunk
monitoring column usage element
ksxmlock_1
lock table sys.col_usage$ in exclusive mode
lock table sys.col_usage$ in exclusive mode nowait
update sys.col_usage$
set equality_preds = equality_preds +
decode(bitand(:flag, 1), 0, 0, 1),
equijoin_preds = equijoin_preds +
decode(bitand(:flag, 2), 0, 0, 1),
nonequijoin_preds = nonequijoin_preds +
decode(bitand(:flag, 4), 0, 0, 1),
range_preds = range_preds + decode(bitand(:flag, 8), 0, 0, 1),
like_preds = like_preds + decode(bitand(:flag, 16), 0, 0, 1),
null_preds = null_preds + decode(bitand(:flag, 32), 0, 0, 1),
timestamp = :time
where obj# = :ob jn
and intcol# = :coln
insert into sys.col_usage$
(obj#,
intcol#,
equality_preds,
equijoin_preds,
nonequijoin_preds,
range_preds,
like_preds, null_preds,
timestamp)
values
(:objn,
:coln,
decode(bitand(:flag, 1), 0, 0, 1),
decode(bitand(:flag, 2), 0, 0, 1),
decode(bitand(:flag, 4), 0, 0, 1),
decode(bitand(:flag, 8), 0, 0, 1),
decode(bitand(:flag, 16), 0, 0, 1),
decode(bitand(:flag, 32), 0, 0, 1),
:time)
使用 dbms_stats 的’SIZE AUTO’模式收集表上的统计信息会首先参考 col_usage$中的 predicate columns
记录:
SQL> begin
2
3 dbms_stats.gather_table_stats(ownname => 'SYS',
4 tabname => 'MACLEAN',
5 method_opt => 'FOR ALL COLUMNS SIZE AUTO');
6 end;
7 /
PL/SQL procedure successfully completed.
============10046 level 12 trace content======================
SELECT /*+ ordered use_nl(o c cu h) index(u i_user1) index(o i_obj2)
index(ci_obj#) index(cu i_col_usage$)
index(h i_hh_obj#_intcol#) */
C.NAME COL_NAME,
C.TYPE# COL_TYPE,
C.CHARSETFORM COL_CSF,
C.DEFAULT$ COL_DEF,C.NULL$ COL_NULL,
C.PROPERTY COL_PROP,
C.COL # COL_UNUM,
C.INTCOL# COL_INUM,
C.OBJ# COL_OBJ,
C.SCALE COL_SCALE,
H.BUCKET_CNT H_BCNT,
(T.ROWCNT - H.NULL_CNT) / GREATEST(H.DISTCNT, 1) H_PFREQ,
C.LENGTH COL_LEN,
CU.TIMES TAMP CU_TIME,
CU.EQUALITY_PREDS CU_EP,
CU.EQUIJOIN_PREDS CU_EJP,
CU.RANGE_PREDS CU_RP,
CU.LIKE_PREDS CU_LP,
CU.NONEQUIJOIN_PREDS CU_NEJP,
CU.NULL_PREDS NP
FROM SYS.USE R$ U,
SYS.OBJ$ O,
SYS.TAB$ T,
SYS.COL$ C,
SYS.COL_USAGE$ CU,
SYS.HIST_HEAD$ H
WHERE :B3 = '0'
AND U.NAME = :B2
AND O.OWNER# = U.USER#
AND O.TYPE# = 2
AND O.NAME = :B1
AND O.OBJ# = T.OBJ#
AND O.OBJ# = C.OBJ#
AND C.OBJ# = CU.OBJ#(+)
AND C.INTCOL# = CU.INTCOL#(+)
AND C.OBJ# = H.OBJ#(+)
AND C.INTCOL# = H.INTCOL#(+)
UNION ALL
SELECT /*+
ordered use_nl(c) */
C.KQFCONAM COL_NAME,C.KQFCODTY COL_TYPE,
DECODE(C.KQFCODTY, 1, 1, 0) COL_CSF,
NULL COL_DEF,
0 COL_NULL,
0 COL_PROP,
C.KQFCOCNO COL_UNUM,
C.KQFCOC NO COL_INUM,
O.KQFTAOBJ COL_OBJ,
DECODE(C.KQFCODTY, 2, -127, 0) COL_SCALE,
H.BUCKET_CNT H_BCNT,
(ST.ROWCNT - NULL_CNT) / GREATEST(H.DISTCNT, 1) H_PFREQ,
DECODE(C.KQFCODTY, 2, 22, C.KQFCOSIZ) COL_LEN,
CU.TIMESTAMP CU_TIME,
CU.EQUALITY_PREDS CU_EP,
CU.EQUIJOIN_PREDS CU_EJP,
CU.RANGE_PREDS CU_RP,
CU.LIKE_PREDS CU_LP,
CU.NONEQUIJOIN_PREDS CU _NEJP,
CU.NULL_PREDS NP
FROM SYS.X$KQFTA O,
SYS.TAB_STATS$ ST,
SYS.X$KQFCO C,
SYS.COL_USAGE$ CU,
SYS.HIST_HEAD$ H
WHERE :B3 != '0'
AND :B2 = 'SYS'
AND O.KQFTANAM = :B1
AND O.KQFTAOBJ = ST.OBJ#(+)
AND O.KQFTAOBJ = C.KQFCOTOB
AND C.KQFCOTOB = CU.OBJ#(+)
AND C.KQFCOCNO = CU.INTCOL#(+)
AND C.KQFCOTOB = H.OBJ#(+)
AND C.KQFCOCNO = H.INTCO L#(+)
现象根据 Metalink Note
332177.1]>:
Database Shutdown Immediate Takes Forever, Can Only Do Shutdown Abort [ID 332177.1]
Applies to:
Oracle Server - Enterprise Edition - Version: 9.2.0.4.0
This problem can occur on any platform.
Symptoms
The database is not shutting down for a considerable time when you issue the command :
shutdown immediate
To shut it down in a reasonable time you have to issue the command
shutdown abort
To collect some diagnostics before issuing the shutdown immediate command set a trace event
as follows:
Connect as SYS (/ as sysdba)
SQL> alter session set events '10046 trace name context forever,level 12';
SQL> shutdown immediate;
In the resultant trace file (within the udump directory) you see something similar to the
following :-
PARSING IN CURSOR #n
delete from sys.col_usage$ c where not exists (select 1 from sys.obj$ o where o.obj# =
c.obj# )
...followed by loads of.....
WAIT #2: nam='db file sequential read' ela= 23424 p1=1 p2=4073 p3=1
....
WAIT #2: nam='db file scattered read' ela= 1558 p1=1 p2=44161 p3=8etc
Then eventually
WAIT #2: nam='log file sync' ela= 32535 p1=4111 p2=0 p3=0
...some other SQL....then back to
WAIT #2: nam='db file sequential read' ela= 205 p1=1 p2=107925 p3=1
WAIT #2: nam='db file sequential read' ela= 1212 p1=1 p2=107926 p3=1
WAIT #2: nam='db file sequential read' ela= 212 p1=1 p2=107927 p3=1
WAIT #2: nam='db file scattered read' ela= 1861 p1=1 p2=102625 p3=8
etc....
To verify which objects are involved here you can use a couple of the P1 & P2 values from
above
:-
a) a sequential read
SELECT owner,segment_name,segment_type
FROM dba_extents
WHERE file_id=1
AND 107927 BETWEEN block_id AND block_id + blocks
b) a scattered read
SELECT owner,segment_name,segment_type
FROM dba_extents
WHERE file_id=1
AND 102625 BETWEEN block_id AND block_id + blocks
The output confirms that the objects are
SYS.I_COL_USAGE$ (INDEX) and SYS.COL_USAGE$ (TABLE)
Finally, issue select count(*) from sys.col_usage$;
CauseIf the number of entries in sys.col_usage$ is large then you are very probably hitting the
issue raised in
Bug: 3540022 9.2.0.4.0 RDBMS Base Bug 3221945
Abstract: CLEAN-UP OF ENTRIES IN COL_USAGE$
Base Bug 3221945 9.2.0.3 RDBMS
Abstract: ORA-1631 ON COL_USAGE$
Closed as "Not a Bug"
However, when a table is dropped, the column usage statistics are not dropped. They are
left as they are.
When the database is shutdown (in normal mode), then these "orphaned" column usage entries
are deleted. The code
which does this gets called only during normal shutdown.
Unless and until the database is shutdown, the col_usage$ table will continue to grow.
Solution
To implement the workaround, please execute the following steps:
1. Periodically (eg once a day) run exec DBMS_STATS.FLUSH_DATABASE_MONITORING_INFO;
DBMS_STATS.FLUSH_DATABASE_MONITORING_INFO will clean out redundant col_usage$ entries, and
when
you come to shutdown the database you should not have a huge number of entries left to clean
up.
该文档指出了在 shutdown instance 时 SMON 会着手清理 col_usage$中已被 drop 表的相关 predicate
columns 的”orphaned”记录，如果在本次实例的生命周期中曾生成大量最后被 drop 的中间表，那么
col_usage$中已经堆积了众多的”orphaned”记录，SMON 为了完成 cleanup 工作需要花费大量时间导致
shutdown 变慢。这个文档还指出定期执行 DBMS_STATS.FLUSH_DATABASE_MONITORING_INFO 也
可以清理 col_usage$中的冗余记录。
我们来观察一下 SMON 的清理工作:begin
for i in 1 .. 5000 loop
execute immediate 'create table maclean1' || i ||' tablespace fragment as select 1 t1
from dual';
execute immediate 'select * from maclean1' || i || ' where t1=1';
end loop;
DBMS_STATS.FLUSH_DATABASE_MONITORING_INFO;
for i in 1 .. 5000 loop
execute immediate 'drop table maclean1' || i;
end loop;
end;
/
SQL> purge dba_recyclebin;
DBA Recyclebin purged.
我们可以通过以下查询了解 col_usage$上的 orphaned 记录总数，这也将是在 instance shutdown 时
SMON 所需要清理的数目
select count(*)
from sys.col_usage$ c
where not exists (select /*+ unnest */
1
from sys.obj$ o
where o.obj# = c.obj#);
COUNT(*)
----------
10224
针对 SMON 做 10046 level 12 trace
SQL> oradebug setospid 30225;
Oracle pid: 8, Unix process pid: 30225, image: oracle@rh2.oracle.com (SMON)SQL> oradebug event 10046 trace name context forever,level 12;
Statement processed.
SQL> shutdown immediate;
=================10046 trace content==================
lock table sys.col_usage$ in exclusive mode nowait
delete from sys.col_usage$ where obj#= :1 and intcol#= :2
delete from sys.col_usage$ c
where not exists (select /*+ unnest */
1
from sys.obj$ o
where o.obj# = c.obj#)
如何禁止 SMON 维护 col_usage$字典基表
1.设置隐藏参数_column_tracking_level(column usage tracking)，该参数默认为 1 即启用 column 使用情
况跟踪。设置该参数为 0，将禁用 column tracking，该参数可以在 session 和 system 级别动态修改:
SQL> col name for a25
SQL> col DESCRIB for a25
SQL> SELECT x.ksppinm NAME, y.ksppstvl VALUE, x.ksppdesc describ
2 FROM SYS.x$ksppi x, SYS.x$ksppcv y
3 WHERE x.inst_id = USERENV ('Instance')
4 AND y.inst_id = USERENV ('Instance')
5 AND x.indx = y.indx
6 AND x.ksppinm LIKE '%_column_tracking_level%';
NAME VALUE DESCRIB
------------------------- ---------- -------------------------
_column_tracking_level 1 column usage trackingSQL> alter session set "_column_tracking_level"=0 ;
Session altered.
SQL> alter system set "_column_tracking_level"=0 scope=both;
System altered.
2.关闭 DML monitoring，可以通过设置隐藏参数_dml_monitoring_enabled(enable modification monitoring)
为 false 实现，disable dml monitoring 对 CBO 的影响较大，所以我们一般推荐上一种方式:
SQL> SELECT monitoring, count(*) from DBA_TABLES group by monitoring;
MON COUNT(*)
--- ----------
NO 79
YES 2206
SQL> alter system set "_dml_monitoring_enabled"=false;
System altered.
SQL> SELECT monitoring, count(*) from DBA_TABLES group by monitoring;
MON COUNT(*)
--- ----------
NO 2285
实际上 dba_tables 的 monitoring 列来源于内部参数_dml_monitoring_enabled
SQL> set long 99999
SQL> select text from dba_views where view_name='DBA_TABLES';
TEXT
--------------------------------------------------------------------------------select u.name, o.name, decode(bitand(t.property,2151678048), 0, ts.name, null),
decode(bitand(t.property, 1024), 0, null, co.name),
decode((bitand(t.property, 512)+bitand(t.flags, 536870912)),
0, null, co.name),
decode(bitand(t.trigflag, 1073741824), 1073741824, 'UNUSABLE', 'VALID'),
decode(bitand(t.property, 32+64), 0, mod(t.pctfree$, 100), 64, 0, null),
decode(bitand(ts.flags, 32), 32, to_number(NULL),
decode(bitand(t.property, 32+64), 0, t.pctused$, 64, 0, null)),
decode(bitand(t.property, 32), 0, t.initrans, null),
decode(bitand(t.property, 32), 0, t.maxtrans, null),
s.iniexts * ts.blocksize,
decode(bitand(ts.flags, 3), 1, to_number(NULL),
s.extsize * ts.blocksize),
s.minexts, s.maxexts,
decode(bitand(ts.flags, 3), 1, to_number(NULL),
s.extpct),
decode(bitand(ts.flags, 32), 32, to_number(NULL),
decode(bitand(o.flags, 2), 2, 1, decode(s.lists, 0, 1, s.lists))),
decode(bitand(ts.flags, 32), 32, to_number(NULL),
decode(bitand(o.flags, 2), 2, 1, decode(s.groups, 0, 1, s.groups))),
decode(bitand(t.property, 32+64), 0,
decode(bitand(t.flags, 32), 0, 'YES', 'NO'), null),
decode(bitand(t.flags,1), 0, 'Y', 1, 'N', '?'),
t.rowcnt,
decode(bitand(t.property, 64), 0, t.blkcnt, null),
decode(bitand(t.property, 64), 0, t.empcnt, null),
t.avgspc, t.chncnt, t.avgrln, t.avgspc_flb,
decode(bitand(t.property, 64), 0, t.flbcnt, null),
lpad(decode(t.degree, 32767, 'DEFAULT', nvl(t.degree,1)),10),
lpad(decode(t.instances, 32767, 'DEFAULT', nvl(t.instances,1)),10),
lpad(decode(bitand(t.flags, 8), 8, 'Y', 'N'),5),
decode(bitand(t.flags, 6), 0, 'ENABLED', 'DISABLED'),
t.samplesize, t.analyzetime,
decode(bitand(t.property, 32), 32, 'YES', 'NO'),
decode(bitand(t.property, 64), 64, 'IOT',
decode(bitand(t.property, 512), 512, 'IOT_OVERFLOW',
decode(bitand(t.flags, 536870912), 536870912, 'IOT_MAPPING', null))),
decode(bitand(o.flags, 2), 0, 'N', 2, 'Y', 'N'),
decode(bitand(o.flags, 16), 0, 'N', 16, 'Y', 'N'),
decode(bitand(t.property, 8192), 8192, 'YES',
decode(bitand(t.property, 1), 0, 'NO', 'YES')),
decode(bitand(o.flags, 2), 2, 'DEFAULT',
decode(s.cachehint, 0, 'DEFAULT', 1, 'KEEP', 2, 'RECYCLE', NULL)),
decode(bitand(t.flags, 131072), 131072, 'ENABLED', 'DISABLED'),
decode(bitand(t.flags, 512), 0, 'NO', 'YES'),
decode(bitand(t.flags, 256), 0, 'NO', 'YES'),
decode(bitand(o.flags, 2), 0, NULL,
decode(bitand(t.property, 8388608), 8388608,
'SYS$SESSION', 'SYS$TRANSACTION')),
decode(bitand(t.flags, 1024), 1024, 'ENABLED', 'DISABLED'),
decode(bitand(o.flags, 2), 2, 'NO',
decode(bitand(t.property, 2147483648), 2147483648, 'NO',
decode(ksppcv.ksppstvl, 'TRUE', 'YES', 'NO'))),
decode(bitand(t.property, 1024), 0, null, cu.name),
decode(bitand(t.flags, 8388608), 8388608, 'ENABLED', 'DISABLED'),
decode(bitand(t.property, 32), 32, null,
decode(bitand(s.spare1, 2048), 2048, 'ENABLED', 'DISABLED')),
decode(bitand(o.flags, 128), 128, 'YES', 'NO')
from sys.user$ u, sys.ts$ ts, sys.seg$ s, sys.obj$ co, sys.tab$ t, sys.obj$ o,
sys.obj$ cx, sys.user$ cu, x$ksppcv ksppcv, x$ksppi ksppi
where o.owner# = u.user#
and o.obj# = t.obj#
and bitand(t.property, 1) = 0
and bitand(o.flags, 128) = 0
and t.bobj# = co.obj# (+)
and t.ts# = ts.ts#
and t.file# = s.file# (+)
and t.block# = s.block# (+)
and t.ts# = s.ts# (+)
and t.dataobj# = cx.obj# (+)
and cx.owner# = cu.user# (+)
and ksppi.indx = ksppcv.indx and ksppi.ksppinm = '_dml_monitoring_enabled'
了 解 你 所 不 知 道 的 SMON 功 能 ( 五 ):Recover Dead
transaction
SMON 的作用还包括清理死事务:Recover Dead transaction。当服务进程在提交事务(commit)
前就意外终止的话会形成死事务(dead transaction)，PMON 进程负责轮询 Oracle 进程，找出
这类意外终止的死进程(dead process)，通知 SMON 将与该 dead process 相关的 dead
transaction 回滚清理，并且 PMON 还负责恢复 dead process 原本持有的锁和 latch。
我们来具体了解 dead transaction 的恢复过程:
SQL> select * from v$version;
BANNER
----------------------------------------------------------------
Oracle Database 10g Enterprise Edition Release 10.2.0.4.0 - 64bi
PL/SQL Release 10.2.0.4.0 - Production
CORE 10.2.0.4.0 Production
TNS for Linux: Version 10.2.0.4.0 - Production
NLSRTL Version 10.2.0.4.0 - Production
SQL> select * from global_name;
GLOBAL_NAME
--------------------------------------------------------------------------------
www.oracle.com
SQL>alter system set fast_start_parallel_rollback=false;
System altered.设置 10500，10046 事件以跟踪 SMON 进程的行为
SQL> alter system set events '10500 trace name context forever,level 8';
System altered.
SQL> oradebug setospid 4424
Oracle pid: 8, Unix process pid: 4424, image: oracle@rh2.oracle.com (SMON)
SQL> oradebug event 10046 trace name context forever,level 8;
Statement processed.
在一个新的 terminal 中执行大批量的删除语句，在执行一段时间后使用操作系统命令将执行该删除操作的
服务进程 kill 掉，模拟一个大的 dead transaction 的场景
SQL> delete large_rb;
delete large_rb
[oracle@rh2 bdump]$ kill -9 4535
等待几秒后 pmon 进程会找出 dead process:
[claim lock for dead process][lp 0x7000003c70ceff0][p 0x7000003ca63dad8.1290666][hist
x9a514951]
在 x$ktube 内部视图中出现 ktuxecfl(Transaction flags)标记为 DEAD 的记录:
SQL> select sum(distinct(ktuxesiz)) from x$ktuxe where ktuxecfl = 'DEAD';
SUM(DISTINCT(KTUXESIZ))
-----------------------
29386
SQL> /
SUM(DISTINCT(KTUXESIZ))
-----------------------
28816以上 KTUXESIZ 代表事务所使用的 undo 块总数(number of undo blocks used by the transaction)
==================smon trace content==================
SMON: system monitor process posted
WAIT #0: nam='log file switch completion' ela= 0 p1=0 p2=0 p3=0 obj#=1 tim=1278243332801935
WAIT #0: nam='log file switch completion' ela= 0 p1=0 p2=0 p3=0 obj#=1 tim=1278243332815568
WAIT #0: nam='latch: row cache objects' ela= 95 address=2979418792 number=200 tries=1 obj#=1
tim=1278243333332734
WAIT #0: nam='latch: row cache objects' ela= 83 address=2979418792 number=200 tries=1 obj#=1
tim=1278243333356173
WAIT #0: nam='latch: undo global data' ela= 104 address=3066991984 number=187 tries=1 obj#=1
tim=1278243347987705
WAIT #0: nam='latch: object queue header operation' ela= 89 address=3094817048 number=131
tries=0 obj#=1 tim=1278243362468042
WAIT #0: nam='log file switch (checkpoint incomplete)' ela= 0 p1=0 p2=0 p3=0 obj#=1
tim=1278243419588202
Dead transaction 0x00c2.008.0000006d recovered by SMON
=====================
PARSING IN CURSOR #3 len=358 dep=1 uid=0 oct=3 lid=0 tim=1278243423594568 hv=3186851936
ad='ae82c1b8'
select smontabv.cnt,
smontab.time_mp,
smontab.scn,
smontab.num_mappings,
smontab.tim_scn_map,
smontab.orig_thread
from smon_scn_time smontab,
(select max(scn) scnmax,
count(*) + sum(NVL2(TIM_SCN_MAP, NUM_MAPPINGS, 0)) cnt
from smon_scn_time
where thread = 0) smontabv
where smontab.scn = smontabv.scnmax
and thread = 0
END OF STMT
PARSE #3:c=0,e=1354526,p=0,cr=0,cu=0,mis=1,r=0,dep=1,og=4,tim=1278243423594556
EXEC #3:c=0,e=106,p=0,cr=0,cu=0,mis=0,r=0,dep=1,og=4,tim=1278243423603269FETCH #3:c=0,e=47065,p=0,cr=319,cu=0,mis=0,r=1,dep=1,og=4,tim=1278243423650375
*** 2011-06-24 21:19:25.899
WAIT #0: nam='smon timer' ela= 299999999 sleep time=300 failed=0 p3=0 obj#=1
tim=1278243716699171
kglScanDependencyHandles4Unpin():
cumscan=3 cumupin=4 time=776 upinned=0
以上 SMON 回滚清理 Dead transaction 的过程从”system monitor process posted”开始到”Dead
transaction 0x00c2.008.0000006d recovered by SMON”结束。另外可以看到在恢复过程中
SMON 先后请求了’latch: row cache objects’、’latch: undo global data’、’latch: object queue
header operation’三种不同类型的 latch。
现象
fast_start_parallel_rollback 参数决定了 SMON 在回滚事务时使用的并行度，若将该参数设置
为 false 那么并行回滚将被禁用，若设置为 Low(默认值)那么会以 2*CPU_COUNT 数目的并行
度回滚，当设置为 High 则 4*CPU_COUNT 数目的回滚进程将参与进来。当我们通过以下查询
发 现 系 统 中 存 在 大 的 dead tranacation 需 要 回 滚 时 我 们 可 以 通 过 设 置
fast_start_parallel_rollback 为 HIGH 来加速恢复:
select sum(distinct(ktuxesiz)) from x$ktuxe where ktuxecfl = 'DEAD';
==============parallel transaction recovery===============
*** 2011-06-24 20:31:01.765
SMON: system monitor process posted msgflag:0x0000 (-/-/-/-/-/-/-)
*** 2011-06-24 20:31:01.765
SMON: process sort segment requests begin
*** 2011-06-24 20:31:01.765
SMON: process sort segment requests end
*** 2011-06-24 20:31:01.765SMON: parallel transaction recovery begin
WAIT #0: nam='DFS lock handle' ela= 504 type|mode=1413545989 id1=3 id2=11 obj#=2
tim=1308918661765715
WAIT #0: nam='DFS lock handle' ela= 346 type|mode=1413545989 id1=3 id2=12 obj#=2
tim=1308918661766135
WAIT #0: nam='DFS lock handle' ela= 565 type|mode=1413545989 id1=3 id2=13 obj#=2
tim=1308918661766758
WAIT #0: nam='DFS lock handle' ela= 409 type|mode=1413545989 id1=3 id2=14 obj#=2
tim=1308918661767221
WAIT #0: nam='DFS lock handle' ela= 332 type|mode=1413545989 id1=3 id2=15 obj#=2
tim=1308918661767746
WAIT #0: nam='DFS lock handle' ela= 316 type|mode=1413545989 id1=3 id2=16 obj#=2
tim=1308918661768146
WAIT #0: nam='DFS lock handle' ela= 349 type|mode=1413545989 id1=3 id2=17 obj#=2
tim=1308918661768549
WAIT #0: nam='DFS lock handle' ela= 258 type|mode=1413545989 id1=3 id2=18 obj#=2
tim=1308918661768858
WAIT #0: nam='DFS lock handle' ela= 310 type|mode=1413545989 id1=3 id2=19 obj#=2
tim=1308918661769224
WAIT #0: nam='DFS lock handle' ela= 281 type|mode=1413545989 id1=3 id2=20 obj#=2
tim=1308918661769555
*** 2011-06-24 20:31:01.769
SMON: parallel transaction recovery end
但是在 real world 的实践中可以发现当 fast_start_parallel_rollback= Low/High，即启用并行回
滚时常有并行进程因为各种资源互相阻塞导致回滚工作停滞的例子，当遭遇到这种问题时将
fast_start_parallel_rollback 设置为 FALSE 一般可以保证恢复工作以串行形式在较长时间内完
成。
如何禁止 SMON Recover Dead transaction
可以设置 10513 事件来临时禁止 SMON 恢复死事务，这在我们做某些异常恢复的时候显得
异常有效，当然不建议在一个正常的生产环境中设置这个事件:
SQL> alter system set events '10513 trace name context forever, level 2';System altered.
10531 -- event disables transaction recovery which was initiated by SMON
SQL> select ktuxeusn,
2 to_char(sysdate, 'DD-MON-YYYY HH24:MI:SS') "Time",
3 ktuxesiz,
4 ktuxesta
5 from x$ktuxe
6 where ktuxecfl = 'DEAD';
KTUXEUSN Time KTUXESIZ KTUXESTA
---------- -------------------------- ---------- ----------------
17 24-JUN-2011 22:03:10 0 INACTIVE
66 24-JUN-2011 22:03:10 0 INACTIVE
105 24-JUN-2011 22:03:10 0 INACTIVE
193 24-JUN-2011 22:03:10 33361 ACTIVE
194 24-JUN-2011 22:03:10 0 INACTIVE
194 24-JUN-2011 22:03:10 0 INACTIVE
197 24-JUN-2011 22:03:10 20171 ACTIVE
7 rows selected.
SQL> /
KTUXEUSN Time KTUXESIZ KTUXESTA
---------- -------------------------- ---------- ----------------
17 24-JUN-2011 22:03:10 0 INACTIVE
66 24-JUN-2011 22:03:10 0 INACTIVE
105 24-JUN-2011 22:03:10 0 INACTIVE
193 24-JUN-2011 22:03:10 33361 ACTIVE
194 24-JUN-2011 22:03:10 0 INACTIVE
194 24-JUN-2011 22:03:10 0 INACTIVE
197 24-JUN-2011 22:03:10 20171 ACTIVE
7 rows selected.================smon disabled trans recover trace==================
SMON: system monitor process posted
*** 2011-06-24 22:02:57.980
SMON: Event 10513 is level 2, trans recovery disabled.
了解你所不知道的 SMON 功能(六):清理 IND$字典基表
SMON 的作用还包括清理 IND$字典基表(cleanup ind$):
触发场景
当我们在线创建或重建索引时(create or rebuild index online)，服务进程会到 IND$字典基表中
将 该 索 引 对 应 的 记 录 的 FLAGS 字 段 修 改 为 十 进 制 的 256 或 者 512( 见上图0×100=256,0×200=512)，如:
SQL> create index macleans_index on larges(owner,object_name) online;
SQL> select obj# from obj$ where name='MACLEANS_INDEX';
OBJ#
----------
1343842
SQL> select FLAGS from ind$ where obj#=1343842;
FLAGS
----------
256
ind_online$字典基表记录了索引在线创建/重建的历史
SQL> select * from ind_online$;
OBJ# TYPE# FLAGS
---------- ---------- ----------
1343839 1 256
1343842 1 256
create table ind_online$
( obj# number not null,
type# number not null, /* what kind of index is this? */
/* normal : 1 */
/* bitmap : 2 */
/* cluster : 3 */
/* iot - top : 4 */
/* iot - nested : 5 */
/* secondary : 6 */
/* ansi : 7 */
/* lob : 8 */ /* cooperative index method : 9 */
flags number not null
/* index is being online built : 0x100 */
/* index is being online rebuilt : 0x200 */
)
原则上 online create/rebuild index 的的清理工作由实际操作的服务进程负责完成，这种清理
在 DDL 语句成功的情况下包括一系列数据字典的维护，在该 DDL 语句失败的情形中包括对
临 时 段 的 清 理 和 数 据 字 典 的 维 护 ， 无 论 如 何 都 需 要 drop 在 线 日 志 中 间
表 SYS_JOURNAL_nnnnn(nnnn 为该索引的 obj#)。数据字典的维护工作就包含对 IND$基
表中相应索引记录的 FLAGS 标志位的恢复，但是如果服务进程在语句执行过程中意外终止的
话，那么短时间内 FLAGS 标志位字段就无法得到恢复，这将导致对该索引的后续操作因
ORA-8104 错误而无法继续:
SQL> drop index macleans_index;
drop index macleans_index
*
ERROR at line 1:
ORA-08104: this index object 1343842 is being online built or rebuilt
08104, 00000, "this index object %s is being online built or rebuilt"
// *Cause: the index is being created or rebuild or waited for recovering
// from the online (re)build
// *Action: wait the online index build or recovery to complete
SMON 负责在启动后(startup)的每小时执行一次对 IND$基表中因在线创建/重建索引失败所
留下记录的清理，这种清理工作由 kdicclean 函数驱动(kdicclean is run by smon every 1 hour，
called from SMON to find if there is any online builder death and cleanup our ind$ and obj$ and
drop the journal table, stop journaling)。
这种清理工作典型的调用堆栈 stack call 如下:
ksbrdp -> ktmSmonMain ktmmon -> kdicclean -> kdic_cleanup -> ktssdrp_segment
注意因为 SMON 进程的清理工作每小时才执行一次，而且在工作负载很高的情况下可能实际很久都不会得到清理，在这种情景中我们总是希望能尽快完成对索引的在线创建或重建，
在 10gr2 以后的版本中我们可以直接使用 dbms_repair.online_index_clean 来手动清理 online
index rebuild 的遗留问题:
SQL> drop index macleans_index;
drop index macleans_index
*
ERROR at line 1:
ORA-08104: this index object 1343842 is being online built or rebuilt
DECLARE
isClean BOOLEAN;
BEGIN
isClean := FALSE;
WHILE isClean=FALSE
LOOP
isClean := dbms_repair.online_index_clean(
dbms_repair.all_index_id, dbms_repair.lock_wait);
dbms_lock.sleep(10);
END LOOP;
END;
/
SQL> drop index macleans_index;
drop index macleans_index
*
ERROR at line 1:
ORA-01418: specified index does not exist
成功清理
但是如果在 9i 中的话就比较麻烦，可以尝试用以下方法(不是很推荐，除非你已经等了很久):
1.首先手工删除在线日志表，通过以下手段找出这个中间表的名字select object_name
from dba_objects
where object_name like
(select '%' || object_id || '%'
from dba_objects
where object_name = '&INDEX_NAME')
/
Enter value for index_name: MACLEANS_INDEX
old 6: where object_name = '&INDEX_NAME')
new 6: where object_name = 'MACLEANS_INDEX')
OBJECT_NAME
--------------------------------------------------------------------------------
SYS_JOURNAL_1343845
SQL> drop table SYS_JOURNAL_1343845;
Table dropped.
2.第二步要手动修改 IND$字典基表
!!!!!! 注意！手动修改数据字典要足够小心！！
select flags from ind$ where obj#=&INDEX_OBJECT_ID;
Enter value for index_object_id: 1343845
old 1: select flags from ind$ where obj#=&INDEX_OBJECT_ID
new 1: select flags from ind$ where obj#=1343845
FLAGS
----------
256
a) 针对 online create index，手动删除对应的记录
delete from IND$ where obj#=&INDEX_OBJECT_ID
b) 针对 online rebuild index，手动恢复对应记录的 FLAGS 标志位update IND$ set FLAGS=FLAGS-512 where obj#=&INDEX_OBJECT_ID
接下来我们实际观察一下清理工作的细节:
SQL> select obj# from obj$ where name='MACLEANS_INDEX';
OBJ#
----------
1343854
SQL> select FLAGS from ind$ where obj#=1343854;
FLAGS
----------
256
SQL> oradebug setmypid;
Statement processed.
SQL> oradebug event 10046 trace name context forever,level 8;
Statement processed.
SQL> DECLARE
2 isClean BOOLEAN;
3 BEGIN
4 isClean := FALSE;
5 WHILE isClean=FALSE
6 LOOP
7 isClean := dbms_repair.online_index_clean(
8 dbms_repair.all_index_id, dbms_repair.lock_wait);
9
10 dbms_lock.sleep(10);
11 END LOOP;
12 END;
13 /
PL/SQL procedure successfully completed.===============================10046 trace=============================
select i.obj#, i.flags, u.name, o.name, o.type#
from sys.obj$ o, sys.user$ u, sys.ind_online$ i
where (bitand(i.flags, 256) = 256 or bitand(i.flags, 512) = 512)
and (not ((i.type# = 9) and bitand(i.flags, 8) = 8))
and o.obj# = i.obj#
and o.owner# = u.user#
select u.name,
o.name,
o.namespace,
o.type#,
decode(bitand(i.property, 1024), 0, 0, 1)
from ind$ i, obj$ o, user$ u
where i.obj# = :1
and o.obj# = i.bo#
and o.owner# = u.user#
delete from object_usage
where obj# in (select a.obj#
from object_usage a, ind$ b
where a.obj# = b.obj#
and b.bo# = :1)
drop table "SYS"."SYS_JOURNAL_1343854" purge
delete from icoldep$ where obj# in (select obj# from ind$ where bo#=:1)
delete from ind$ where bo#=:1
delete from ind$ where obj#=:1
我们可以利用以下语句找出系统中可能需要恢复的 IND$记录，注意不要看到查询有结果就
认为这是操作失败的征兆，很可能是有人在线创建或重建索引:select i.obj#, i.flags, u.name, o.name, o.type#
from sys.obj$ o, sys.user$ u, sys.ind_online$ i
where (bitand(i.flags, 256) = 256 or bitand(i.flags, 512) = 512)
and (not ((i.type# = 9) and bitand(i.flags, 8) = 8))
and o.obj# = i.obj#
and o.owner# = u.user#
/
相关诊断事件
可以通过设置诊断事件 event=’8105 trace name context forever’
来禁止 SMON 清理 IND$(Oracle event to turn off smon cleanup for online index build)
alter system set events '8105 trace name context forever';
了解你所不知道的 SMON 功能(七):维护 MON_MODS$字典
基表
SMON 后台进程的作用还包括维护 MON_MODS$基表，当初始化参数 STATISTICS_LEVEL 被设
置为 TYPICAL 或 ALL 时默认会启用 Oracle 中表监控的特性，Oracle 会默认监控表上的自上一
次分析以后(Last analyzed)发生的 INSERT,UPDATE,DELETE 以及表是否被 TRUNCATE 截断，并将
这些操作数量的近似值记录到数据字典基表 MON_MODS$中，我们常用的一个 DML 视图
dba_tab_modifications 的数据实际来源于另一个数据字典基表 MON_MODS_ALL$，SMON 定
期会将 MON_MODS$中符合要求的数据 MERGE 到 MON_MODS_ALL$中。
Rem DML monitoring
create table mon_mods$(
obj# number, /* object number */
inserts number, /* approx. number of inserts since last analyze */
updates number, /* approx. number of updates since last analyze */
deletes number, /* approx. number of deletes since last analyze */
timestamp date, /* timestamp of last time this row was changed */
flags number, /* flags */
/* 0x01 object has been truncated */
drop_segments number /* number of segemnt in part/subpartition table */
)
storage (initial 200K next 100k maxextents unlimited pctincrease 0)
/
create unique index i_mon_mods$_obj on mon_mods$(obj#)
storage (maxextents unlimited)
/
Rem DML monitoring, has info aggregated to global level for paritioned objects
create table mon_mods_all$
(
obj# number, /* object number */
inserts number, /* approx. number of inserts since last analyze */
updates number, /* approx. number of updates since last analyze */
deletes number, /* approx. number of deletes since last analyze */
timestamp date, /* timestamp of last time this row was changed */
flags number, /* flags */
/* 0x01 object has been truncated */
drop_segments number /* number of segemnt in part/subpartition table */
)
storage (initial 200K next 100k maxextents unlimited pctincrease 0)
/
create unique index i_mon_mods_all$_obj on mon_mods_all$(obj#)
storage (maxextents unlimited)
/
Rem =========================================================================
Rem End Usage monitoring tables
Rem =========================================================================VIEW DBA_TAB_MODIFICATIONS
select u.name, o.name, null, null,
m.inserts, m.updates, m.deletes, m.timestamp,
decode(bitand(m.flags,1),1,'YES','NO'),
m.drop_segments
from sys.mon_mods_all$ m, sys.obj$ o, sys.tab$ t, sys.user$ u
where o.obj# = m.obj# and o.obj# = t.obj# and o.owner# = u.user#
union all
select u.name, o.name, o.subname, null,
m.inserts, m.updates, m.deletes, m.timestamp,
decode(bitand(m.flags,1),1,'YES','NO'),
m.drop_segments
from sys.mon_mods_all$ m, sys.obj$ o, sys.user$ u
where o.owner# = u.user# and o.obj# = m.obj# and o.type#=19
union all
select u.name, o.name, o2.subname, o.subname,
m.inserts, m.updates, m.deletes, m.timestamp,
decode(bitand(m.flags,1),1,'YES','NO'),
m.drop_segments
from sys.mon_mods_all$ m, sys.obj$ o, sys.tabsubpart$ tsp, sys.obj$ o2,
sys.user$ u
where o.obj# = m.obj# and o.owner# = u.user# and
o.obj# = tsp.obj# and o2.obj# = tsp.pobj#
现象：
SMON后台进程会每15分钟将SGA中的DML统计信息刷新到SYS.MON_MODS$基表中(SMON
flush every 15 minutes to SYS.MON_MODS$)，
同时会将 SYS.MON_MODS$中符合要求的数据 MERGE 合并到 MON_MODS_ALL$中，并清空
原 MON_MODS$中的数据。
MON_MODS_ALL$作为 dba_tab_modifications 视图的数据来源，起到辅助统计信息收集的作
用,详见拙作。
SMON 具体将 DML 统计数据刷新到 SYS.MON_MODS$、合并到 MON_MODS_ALL$、并清除数
据的操作如下：SQL> select * from v$version;
BANNER
--------------------------------------------------------------------------------
Oracle Database 11g Enterprise Edition Release 11.2.0.2.0 - 64bit Production
PL/SQL Release 11.2.0.2.0 - Production
CORE 11.2.0.2.0 Production
TNS for Linux: Version 11.2.0.2.0 - Production
NLSRTL Version 11.2.0.2.0 - Production
SQL> select * from global_name;
GLOBAL_NAME
--------------------------------------------------------------------------------
www.oracle.com
/* 填充 mon_mods$字典基表 */
lock table sys.mon_mods$ in exclusive mode nowait
insert into sys.mon_mods$
(obj#, inserts, updates, deletes, timestamp, flags, drop_segments)
values
(:1, :2, :3, :4, :5, :6, :7)
update sys.mon_mods$
set inserts = inserts + :ins,
updates = updates + :upd,
deletes = deletes + :del,
flags =
(decode(bitand(flags, :flag), :flag, flags, flags + :flag)),
drop_segments = drop_segments + :dropseg,
timestamp = :time
where obj# = :objn
lock table sys.mon_mods_all$ in exclusive mode/* 以下 merge 命令会将 mon_mods$中的记录合并到 mon_mods_all$,
若有匹配的记录，则在原记录的基础上增加 inserts、updates、deletes 总数，
否则插入新的记录
*/
merge /*+ dynamic_sampling(mm 4) dynamic_sampling_est_cdn(mm)
dynamic_sampling(m 4) dynamic_sampling_est_cdn(m) */
into sys.mon_mods_all$ mm
using (select m.obj# obj#,
m.inserts inserts,
m.updates updates,
m.deletes deletes,
m.flags flags,
m.timestamp timestamp,
m.drop_segments drop_segments fr om sys.mon_mods$ m,
tab$ t where m.obj# = t.obj#) v
on (mm.ob j# = v.obj#)
when matched then
update
set mm.inserts = mm.inserts + v.inserts,
mm.updates = mm.updates + v.updates,
mm.deletes = mm.deletes + v.deletes,
mm.flags = mm.flags + v.flags - bitand(mm.flags, v.flags) /*
bitor(mm.flags,v.flags) */,
mm.timestamp = v.timestamp,
mm.drop_segments = mm.drop_segments + v.drop_segments
when NOT matched then
insert
(obj#, inserts, updates, deletes, timestamp, flags, drop_segments)
values
(v.obj#,
v.inserts,
v.updates,
v.deletes,
sysdate,
v.flags, v.drop_segments) / all merge /*+ dynamic_sampling(mm 4) dynamic_sampling_est_cdn(mm)
dynamic_sampling(m 4) dynamic_sampling_est_cdn(m) */
into sys.mon_mods_all$ mm using
(select m.obj# obj#,
m.inserts inserts,
m.updates updates,
m.deletes deletes,
m.flags flags,
m.timestamp timestamp,
m.drop_segments drop_segments fr om sys.mon_mods$ m,
tab$ t where m.obj# = t.obj#) v on
(mm.ob j# = v.obj#)
when matched then
update
set mm.inserts = mm.inserts + v.inserts,
mm.updates = mm.updates + v.updates,
mm.deletes = mm.deletes + v.deletes,
mm.flags = mm.flags + v.flags - bitand(mm.flags, v.flags)
/* bitor(mm.flags,v.flags) */,
mm.timestamp = v.timestamp,
mm.drop_segments = mm.drop_segments + v.drop_segments
when NOT matched then
insert
(obj#, inserts, updates, deletes, timestamp, flags, drop_segments)
values
(v.obj#,
v.inserts,
v.updates,
v.deletes,
sysdate,
v.flags,
v.drop_segments)
/* 最后删除 sys.mon_mods$上的相关记录 */
delete /*+ dynamic_sampling(m 4) dynamic_sampling_est_cdn(m) */
from sys.mon_mods$ mwhere exists (select /*+ unnest */
*
from sys.tab$ t
where t.obj# = m. obj#)
select obj#
from sys.mon_mods$
where obj# not in (select obj# from sys.obj$)
Used to have a FULL TABLE SCAN on obj$ associated with monitoring information
extracted in conjunction with mon_mods$ executed by SMON periodically.
因为当 SMON 或用户采用”DBMS_STATS.FLUSH_DATABASE_MONITORING_INFO”存储过程将
DML 数据刷新到 mon_mods$或 mon_mods_all$中时会要求持有表上的排它锁，所以在 RAC
环境中可能出现死锁问题。
另外在早期版本中 SMON 可能因维护监控表而造成 shutdown immediate 缓慢或系统性能下
降的问题，详见:
tables [ID 2806297.8]>
SMON 维护 MON_MODS$时相关的 Stack CALL
kglpnal <- kglpin <- kxsGetRuntimeLock
<- kksfbc <- kkspsc0 <- kksParseCursor <- opiosq0 <- opiall0
<- opikpr <- opiodr <- PGOSF175_rpidrus <- skgmstack <- rpiswu2
<- kprball <- kprbbnd0 <- kprbbnd <- ksxmfmel <- ksxmfm
<- ksxmfchk <- ksxmftim <- ktmmon <- ktmSmonMain <- ksbrdp
<- opirip <- opidrv <- sou2o <- opimai_real <- ssthrdmain
<- main <- libc_start_main <- start
如何禁止 SMON 维护 MON_MODS$
注意在缺省参数环境中创建的表总是启用 table monitoring 的:SQL> select * from v$version;
BANNER
--------------------------------------------------------------------------------
Oracle Database 11g Enterprise Edition Release 11.2.0.2.0 - 64bit Production
PL/SQL Release 11.2.0.2.0 - Production
CORE 11.2.0.2.0 Production
TNS for Linux: Version 11.2.0.2.0 - Production
NLSRTL Version 11.2.0.2.0 - Production
SQL> select * from v$version;
BANNER
--------------------------------------------------------------------------------
Oracle Database 11g Enterprise Edition Release 11.2.0.2.0 - 64bit Production
PL/SQL Release 11.2.0.2.0 - Production
CORE 11.2.0.2.0 Production
TNS for Linux: Version 11.2.0.2.0 - Production
NLSRTL Version 11.2.0.2.0 - Production
SQL> create table maclean1 (t1 int);
Table created.
/* 在 10g 以后 nomonitoring 或 monitoring 选项不再有效 */
SQL> create table maclean2 (t1 int) nomonitoring;
Table created.
SQL> select table_name,monitoring from dba_tables where table_name like 'MACLEAN%';
TABLE_NAME MON
------------------------------ ---
MACLEAN1 YESMACLEAN2 YES
通常来说我们不需要禁止 SMON 维护 MON_MODS$，除非是在 SMON 维护过程中遭遇
shutdown 过慢、性能降低或者异常情况恢复 SMON 随机 terminate 实例的问题。
在 10g 以前可以使用 MONITORING 和 NOMONITORING 这 2 个选项来控制表级别的监控是否
被开启，此外我们还可以通过 dbms_stats.ALTER_SCHEMA_TAB_MONITORING(‘maclean’,false)
存储过程在 schema 级别的 monitoring 是否被开启，但是在 10g 以后这些方法不再有效，
MONITORING 和 NOMONITORING 选项被废弃(In 10g the MONITORING and NOMONITORING
keywords are deprecated and will be ignored.)，其原有功能被 STATISTICS_LEVEL 参数所覆盖。
Table-monitoring 特性现在完全由 STATISTICS_LEVEL 参数所控制：
当 STATISTICS_LEVEL 设置为 BASIC 时，Table-monitoring 将被禁用
当 STATISTICS_LEVEL 设置为 TYPICAL 或 ALL 时,Table-monitoring 将启用
换而言之我们可以通过设置 STATISTICS_LEVEL 为 BASIC 达到禁止 SMON 后台进程该种功能的
作用，具体修改该参数的命令如下：
show parameter statistics_level
alter system set statistics_level = basic;
但是请注意如果你正在使用 AMM 或 ASMM 自动内存管理特性的话，那么 STATISTICS_LEVEL
参数是不能设置为 BASIC 的，因为 Auto-Memory 或 Auto-Sga 特性都依赖于 STATISTICS_LEVEL
所控制的性能统计信息。若一定要这样做那么首先要 diable AMM&ASMM:
#diable 11g AMM ,have to bounce instance
#alter system set memory_target =0 scope=spfile;
#diable 10g ASMM
alter system set sga_target=0;
alter system set statistics_level = basic;了解你所不知道的 SMON 功能(八):维护 SMON_SCN_TIME
字典基表
SMON 后台进程的作用还包括维护 SMON_SCN_TIME 基表。
SMON_SCN_TIME 基表用于记录过去时间段中 SCN(system change number)与具体的时间戳
(timestamp)之间的映射关系，因为是采样记录这种映射关系，所以 SMON_SCN_TIME 可以较
为较为粗糙地(不精确地)定位某个 SCN 的时间信息。实际的 SMON_SCN_TIME 是一张 cluster
table 簇表。
SMON_SCN_TIME 时间映射表最大的用途是为闪回类型的查询(flashback type queries)提供一
种将时间映射为 SCN 的途径(The SMON time mapping is mainly for flashback type queries to
map a time to an SCN)。
Metalink 文档介绍了 SMON
更新 SMON_SCN_TIME 的规律：
在版本10g中SMON_SCN_TIME每6秒钟被更新一次(In Oracle Database 10g, smon_scn_time is
updated every 6 seconds hence that is the minimum time that the flashback query time needs to be behind the timestamp of the first change to the table.)
在版本 9.2 中 SMON_SCN_TIME 每 5 分钟被更新一次(In Oracle Database 9.2, smon_scn_time is
updated every 5 minutes hence the required delay between the flashback time and table
properties change is at least 5 minutes.)
另外从 10g 开始 SMON 也会清理 SMON_SCN_TIME 中的记录了，SMON 后台进程会每 5 分钟
被唤醒一次，检查 SMON_SCN_TIME 在磁盘上的映射记录总数，若总数超过 144000 条，则
会使用以下语句删除最老的一条记录(time_mp 最小):
delete from smon_scn_time
where thread = 0
and time_mp = (select min(time_mp) from smon_scn_time where thread =
0)
若仅仅删除一条记录不足以获得足够的空间，那么 SMON 会反复多次执行以上 DELETE 语句。
触发场景
虽然Metalink文档指出了在
10g 中 SMON 会以每 6 秒一次的频率更新 SMON_SCN_TIME 基表，但是实际观测可以发现更
新频率与 SCN 的增长速率相关，在较为繁忙的实例中 SCN 的上升极快时 SMON 可能会以 6
秒一次的最短间隔频率更新 ， 但是在空闲的实例中 SCN 增长较慢，则仍会以每 5 或 10 分
钟一次频率更新，例如：
[oracle@vrh8 ~]$ ps -ef|grep smon|grep -v grep
oracle 3484 1 0 Nov12 ? 00:00:02 ora_smon_G10R21
SQL> select * from v$version;
BANNER
----------------------------------------------------------------
Oracle Database 10g Enterprise Edition Release 10.2.0.1.0 - 64bi
PL/SQL Release 10.2.0.1.0 - Production
CORE 10.2.0.1.0 Production
TNS for Linux: Version 10.2.0.1.0 - ProductionNLSRTL Version 10.2.0.1.0 - Production
SQL> select * from global_name;
GLOBAL_NAME
---------------------------------------------------------------------
-----------
www.oracle.com & www.oracle.com
SQL> oradebug setospid 3484;
Oracle pid: 8, Unix process pid: 3484, image: oracle@vrh8.oracle.com
(SMON)
SQL> oradebug event 10500 trace name context forever,level 10 : 10046 trace
name context forever,level 12;
Statement processed.
SQL>
SQL> oradebug tracefile_name;
/s01/admin/G10R21/bdump/g10r21_smon_3484.trc
/* 等待一定时间 */
找出 SMON trace 文件中 insert 数据到 SMON_SCN_TIME 的记录:
grep -A20 "insert into smon_scn_time"
/s01/admin/G10R21/bdump/g10r21_smon_3484.trc
insert into smon_scn_time (thread, time_mp, time_dp, scn, scn_wrp,
scn_bas, num_mappings, tim_scn_map)
values (0, :1, :2, :3, :4, :5, :6, :7)
END OF STMT
PARSE
#4:c=0,e=43,p=0,cr=0,cu=0,mis=0,r=0,dep=1,og=4,tim=1290280848899596
BINDS #4:
kkscoacdBind#0
oacdty=02 mxl=22(22) mxlc=00 mal=00 scl=00 pre=00
oacflg=00 fl2=0001 frm=00 csi=00 siz=24 off=0
kxsbbbfp=7fb29844edb8 bln=22 avl=06 flg=05
value=767145793
Bind#1
oacdty=12 mxl=07(07) mxlc=00 mal=00 scl=00 pre=00
oacflg=10 fl2=0001 frm=00 csi=00 siz=8 off=0
kxsbbbfp=7fff023ae780 bln=07 avl=07 flg=09
value="11/14/2011 0:3:13"
Bind#2
oacdty=02 mxl=22(04) mxlc=00 mal=00 scl=00 pre=00
oacflg=10 fl2=0001 frm=00 csi=00 siz=24 off=0
kxsbbbfp=7fff023ae70c bln=22 avl=04 flg=09
value=954389
Bind#3
--
insert into smon_scn_time (thread, time_mp, time_dp, scn, scn_wrp,
scn_bas, num_mappings, tim_scn_map)
values (0, :1, :2, :3, :4, :5, :6, :7)
END OF STMT
PARSE
#1:c=0,e=21,p=0,cr=0,cu=0,mis=0,r=0,dep=1,og=4,tim=1290281434933390
BINDS #1:
kkscoacd
Bind#0
oacdty=02 mxl=22(22) mxlc=00 mal=00 scl=00 pre=00
oacflg=00 fl2=0001 frm=00 csi=00 siz=24 off=0
kxsbbbfp=7fb29844edb8 bln=22 avl=06 flg=05
value=767146393
Bind#1
oacdty=12 mxl=07(07) mxlc=00 mal=00 scl=00 pre=00
oacflg=10 fl2=0001 frm=00 csi=00 siz=8 off=0
kxsbbbfp=7fff023ae780 bln=07 avl=07 flg=09value="11/14/2011 0:13:13"
Bind#2
oacdty=02 mxl=22(04) mxlc=00 mal=00 scl=00 pre=00
oacflg=10 fl2=0001 frm=00 csi=00 siz=24 off=0
kxsbbbfp=7fff023ae70c bln=22 avl=04 flg=09
value=954720
Bind#3
--
insert into smon_scn_time (thread, time_mp, time_dp, scn, scn_wrp,
scn_bas, num_mappings, tim_scn_map)
values (0, :1, :2, :3, :4, :5, :6, :7)
END OF STMT
PARSE
#3:c=0,e=20,p=0,cr=0,cu=0,mis=0,r=0,dep=1,og=4,tim=1290281727955249
BINDS #3:
kkscoacd
Bind#0
oacdty=02 mxl=22(22) mxlc=00 mal=00 scl=00 pre=00
oacflg=00 fl2=0001 frm=00 csi=00 siz=24 off=0
kxsbbbfp=7fb29844e960 bln=22 avl=06 flg=05
value=767146993
Bind#1
oacdty=12 mxl=07(07) mxlc=00 mal=00 scl=00 pre=00
oacflg=10 fl2=0001 frm=00 csi=00 siz=8 off=0
kxsbbbfp=7fff023ae780 bln=07 avl=07 flg=09
value="11/14/2011 0:23:13"
Bind#2
oacdty=02 mxl=22(04) mxlc=00 mal=00 scl=00 pre=00
oacflg=10 fl2=0001 frm=00 csi=00 siz=24 off=0
kxsbbbfp=7fff023ae70c bln=22 avl=04 flg=09
value=954926
Bind#3
insert into smon_scn_time (thread, time_mp, time_dp, scn, scn_wrp,
scn_bas, num_mappings, tim_scn_map)values (0, :1, :2, :3, :4, :5, :6, :7)
END OF STMT
PARSE
#4:c=0,e=30,p=0,cr=0,cu=0,mis=0,r=0,dep=1,og=4,tim=1290282313990553
BINDS #4:
kkscoacd
Bind#0
oacdty=02 mxl=22(22) mxlc=00 mal=00 scl=00 pre=00
oacflg=00 fl2=0001 frm=00 csi=00 siz=24 off=0
kxsbbbfp=7fb29844edb8 bln=22 avl=06 flg=05
value=767147294
Bind#1
oacdty=12 mxl=07(07) mxlc=00 mal=00 scl=00 pre=00
oacflg=10 fl2=0001 frm=00 csi=00 siz=8 off=0
kxsbbbfp=7fff023ae780 bln=07 avl=07 flg=09
value="11/14/2011 0:28:14"
Bind#2
oacdty=02 mxl=22(04) mxlc=00 mal=00 scl=00 pre=00
oacflg=10 fl2=0001 frm=00 csi=00 siz=24 off=0
kxsbbbfp=7fff023ae70c bln=22 avl=04 flg=09
value=955036
Bind#3
可以通过以上 INSERT 语句的 TIME_DP 绑定变量值中发现其更新 SMON_SCN_TIME 的时间规
律，一般为 5 或 10 分钟一次。这说明 SMON_SCN_TIME 的更细频率与数据库实例的负载有
关，其最短的间隔是每 6 秒一次，最长的间隔为 10 分钟一次。
由于 SMON_SCN_TIME 的更新频率问题可能引起 ORA-01466 错误，详见:
Error ORA-01466 while executing a flashback query. [ID 281510.1]
由于 SMON_SCN_TIME 的数据不一致可能引起 ORA-00600[6711]或频繁地执行”delete from
smon_scn_time”删除语句，详见：
ORA-00600[6711]错误一例
High Executions Of Statement “delete from smon_scn_time…” [ID 375401.1]
SMON 维护 SMON_SCN_TIME 时相关的 Stack CALL，ktf_scn_time 是更新 SMON_SCN_TIME 的主要函数:
ksedst ksedmp ssexhd kghlkremf kghalo kghgex kghalf kksLoadChild
kxsGetRuntimeLock kksfbc
kkspsc0 kksParseCursor opiosq0 opiall0 opikpr opiodr rpidrus skgmstack
rpidru rpiswu2 kprball
ktf_scn_time
ktmmon ktmSmonMain ksbrdp opirip opidrv sou2o opimai_real main
main_opd_entry
SMON 还可能使用以下 SQL 语句维护 SMON_SCN_TIME 字典基表：
select smontabv.cnt,
smontab.time_mp,
smontab.scn,
smontab.num_mappings,
smontab.tim_scn_map,
smontab.orig_thread
from smon_scn_time smontab,
(select max(scn) scnmax,
count(*) + sum(NVL2(TIM_SCN_MAP, NUM_MAPPINGS, 0)) cnt
from smon_scn_time
where thread = 0) smontabv
where smontab.scn = smontabv.scnmax
and thread = 0
insert into smon_scn_time
(thread,
time_mp,
time_dp,
scn,
scn_wrp,
scn_bas,
num_mappings,
tim_scn_map)values
(0, :1, :2, :3, :4, :5, :6, :7)
update smon_scn_time
set orig_thread = 0,
time_mp = :1,
time_dp = :2,
scn = :3,
scn_wrp = :4,
scn_bas = :5,
num_mappings = :6,
tim_scn_map = :7
where thread = 0
and scn = (select min(scn) from smon_scn_time where thread = 0)
delete from smon_scn_time
where thread = 0
and scn = (select min(scn) from smon_scn_time where thread = 0)
如何禁止 SMON 更新 SMON_SCN_TIME 基表
可以通过设置诊断事件 event=’12500 trace name context forever, level 10′来禁止 SMON 更新
SMON_SCN_TIME 基表(Setting the 12500 event at system level should stop SMON from updating
the SMON_SCN_TIME table.):
SQL> alter system set events '12500 trace name context forever, level
10';
System altered.
一般我们不推荐禁止 SMON 更新 SMON_SCN_TIME 基表，因为这样会影响 flashback Query
闪回查询的正常使用，但是在某些异常恢复的场景中 SMON_SCN_TIME 数据讹误可能导致实
例的 Crash，那么可以利用以上 12500 事件做到不触发 SMON_SCN_TIME 被更新。如何手动清除 SMON_SCN_TIME 的数据
因为 SMON_SCN_TIME 不是 bootstrap 自举核心对象，所以我们可以手动更新该表上的数据、
及重建其索引。
如我在<ora-00600[6711]< span="">错误一例>中介绍了因为 SMON_SCN_TIME 与其索引的数据不一
致时，可以通过重建索引来解决问题：
connect / as sysdba
drop index smon_scn_time_scn_idx;
drop index smon_scn_time_tim_idx;
create unique index smon_scn_time_scn_idx on smon_scn_time(scn);
create unique index smon_scn_time_tim_idx on smon_scn_time(time_mp);
analyze table smon_scn_time validate structure cascade;
可以在设置了 12500 事件后手动删除 SMON_SCN_TIME 上的记录，重启实例后 SMON 会继续
正常更新 SMON_SCN_TIME 。除非是因为 SMON_SCN_TIME 表 上 的 记 录 与 索 引
smon_scn_time_tim_idx 或 smon_scn_time_scn_idx 上的不一致造成 DELETE 语句无法有效删
除该表上的记录:文档说明了该问题，
否则我们没有必要手动去清除 SMON_SCN_TIME 上的数据。
具体方法如下：
SQL> conn / as sysdba
/* Set the event at system level */
SQL> alter system set events '12500 trace name context forever, level 10';
/* Delete the records from SMON_SCN_TIME */
SQL> delete from smon_scn_time;
SQL> commit;SQL> alter system set events '12500 trace name context off';
完成以上步骤后重启实例 restart instance
shutdown immediate;
startup;
了 解 你 所 不 知 道 的 SMON 功 能 ( 九 ):OFFLINE UNDO
SEGMENT
SMON 这个老牌的后台关键进程的作用还包括对 UNDO/ROLLBACK SEGMENT 的维护， 这种
维护主要体现在 2 个方面: OFFLINE 和 SHRINK UNDO/ROLLBACK SEGMENT， 今天我们主要
介绍 OFFLINE ROLLBACK SEGMENT。
你肯定要问，Oracle 为什么 OFFLINE UNDO/ROLLBACK SEGMENT?
最主要的目的是减轻高并发事务环境中对 UDNO SPACE 撤销空间使用的压力。
触发场景
在 10g 之前的 9i 中每 12 个小时 SMON 会根据 V$UNDOSTAT 中记录来决定在现有基础上要
OFFLINE 多少个 UNDO SEGMENT，又要保留多少个 UNDO SEGMENT； 在 9i 中被 OFFLINED
UNDO SEGMENT 还会被 SMON DROP 掉，以进一步回收空间。
具体保留多少个 UNDO SEGMENT，取决于过去 12 个小时内的 V$UNDOSTAT 动态视图记录的
最大并发事务数量在加上 1，具体公式可以参考下面的 SQL:
SQL> select max(MAXCONCURRENCY)+1 from v$undostat where begin_time> (sysdate-1/2);
MAX(MAXCONCURRENCY)+1
---------------------
4若你在 alert.log 中发现类似以下的信息则说明 OFFLINE UNDO SEGS 已经在你的系统中发生过
了:
SMON offlining US=13
Freeing IMU pool for usn 13
SMON offlining US=14
SMON offlining US=15
SMON offlining US=16
SMON offlining US=17
9i 中 SMON通过ktusmofd 函数实现对 UDNO SEGMENT 的 OFFLINE，ktusmofd的含义为[K]ernel
[T]ransaction [U]ndo [S]ystem [M]anaged OFFLINE & DROP
通过 ktsmgfru 函数返回必要保留的 ONLINE UNDO SEGMENT， 其详细的算法如下：
SMON 调用 ktusmofd ，并发现 instance 启动未超过 12 个小时并且_smu_debug_mode 未设置
KTU_DEBUG_SMU_SMON_SHRINK 标志位
(_smu_debug_mode 是 SYSTEM MANAGED UNDO 内部参数，KTU_DEBUG_SMU_SMON_SHRINK 标志位控制是否强制
SMON 做 SHRINK)
YES - SMON 不 OFFLINE 任何东西直接返回
NO - 调用 ktsmgfru 获得过去 12 小时的最大并发事务数
设置 keep_online 变量为 ktsmgfru 返回值加上 1
尝试 hold TA ENQUEUE(该队列锁控制 UNDO TABLESPACE 的
串行操作)，该操作的超时限制为 30s
若无法获得该 ENQUEUE 则说明正在切换 UNDO
TABLESPACE，ktusmofd 将直接返回且不 OFFLINE 任何 UNDO SEGMENTS
成功获得该 ENQUEUE 锁，进一步调用 ktusmofxu 并使用之
前获得的 keep_online 作为参数，开始 OFFLINE
调用kslgpl函数获得KTU LATCH 包括parent
和所有的 children
LOOP 在现有的 ONLINE UNDO
SEGMENT 之间循环
若发现该 UNDO SEGMENT 是
SMU-SYSTEM MANAGED UNDO 且其所在表空间是当前 undo_tablespace 指向的表空间的话
若 keep_online >0 ， 则
keep_online-- 否则
释放 KTU latches
调用 kturof1 函数实际 OFFLINE 该 UNDO SEGMENT
重新 get KTU latches
END LOOP
释放 KTU latches
SMON 调用 ktusmofd 维护 OFFLINE UNDO SEGMENT 的常见 STACK CALL 如下：
ktmmon->ktusmofd->ktusmdxu->ktcrcm->ktccpcmt->ktcccdel->ktadrpc->ktssdro_segment->
ktssdrbm_segment->ktsxbmdelext->kqrcmt->ktsscu
xctrol ktcpoptx ktccpcmt ktcrcm ktusmdxu ktusmofd ktmmon
ksedmp ksfdmp kgeasnmierr ktusmgmct ktusmdxu ktusmofd ktmmon ksbrdp opirip
opidrv sou2o main
10g 以前的 UNDO OFFLINE 算法仍不完善，这导致在实例重启或切换 UNDO TABLESPACE 撤销
表空间时，生成一定数量 ONLINE UNDO SEGMENT 的系统预热时间可能长达几分钟，对于高
并发的环境来说这种延时是难以接受的。
从 10g 开始改进了 SMON OFFLINE UNDO SEGMENT 的算法，SMON 会基于过去 7 天的(而非
12 个小时的)V$UNDOSTAT 动态视图信息或者 AWR 自动负载仓库中的 UNDO 历史快照使用信
息来决定 OFFLINE UNDO SEGMENT 的数量， 且在 10g 中 SMON 不再 DROP 掉多余的 UNDO
SEGS，而仅仅 OFFLINE 掉；作为一种 SMU 的改良算法这种做法被叫做”Fast Ramp-Up”。”Fast
Ramp-Up”避免了早期版本中由 SMON 维护 UNDO SEGS 引起的等待或性能问题; 此外，未公
开的 BUG 5079978 可能在版本 10.2.0.1 中被触发，该 BUG 的信息如下:
Unpublished
Bug 5079978 – APPST GSI 10G : – PRODUCTION INSTANCE UNUSABLE DUE TO US ENQUEUE
WAITS
is fixed in 11.1 and patch set 10.2.0.4 and interim patches are available for several earlier versions.
Please refer to Note 5079978.8
可以通过后面要介绍的 10511 event 来规避以上 bug，Oracle 官方也推荐在 10g 以前的版本
中使用 10511 event 来避免 SMON 过度 OFFLINE UNDO SEGS 所引起的问题。10g 以后的具体算法如下：
判断实例启动是否超过 7 天？
YES - 直接使用 v$undostat 中过去 7 天的最大并发事务数 max(maxconcurrency)
NO - 判断是否是第一次调用 OFFLINE UNDO SEGMENT 的内核函数
YES - 检查是否存在 select_workload_repository function
(SWRF)快照数据
NO - ONLINE 最小数目的 UNDO SEGMENTS
YES - 尝试获取 AWR 记录表 wrh$_undostat
中过去 7 天的最大并发事务数 max(maxconcurrency)
若无法获得以上值，则尝试读取
wrh$_rollstat 中最近 7 天的最大 rollback segs 数量 max(rbs cnt)
将返回值保存到内部变量中
NO - 直接使用内部变量中的值
如何禁止 SMON OFFLINE UNDO SEGMENT?
可以通过设置诊断事件 event=’10511 trace name context forever, level 1′ 来禁用 SMON
OFFLINE UNDO SEGS; 但是 10511 事件不会跳过”Fast Ramp Up”，而仅会限制 SMON 对 UNDO
SEGS 产生的工作负载。 一旦设置了 10511 event， 则所有已生成的 UNDO SEGS 会始终保
持 ONLINE 状态。
具体的设置方法：
SQL> select * from v$version;
BANNER
----------------------------------------------------------------
Oracle Database 10g Enterprise Edition Release 10.2.0.5.0 - 64bi
PL/SQL Release 10.2.0.5.0 - Production
CORE 10.2.0.5.0 Production
TNS for Linux: Version 10.2.0.5.0 - Production
NLSRTL Version 10.2.0.5.0 - Production
SQL> select * from global_name;GLOBAL_NAME
--------------------------------------------------------------------------------
www.oracle.com
[oracle@vrh8 ~]$ oerr ora 10511
10511, 00000, "turn off SMON check to cleanup undo dictionary"
// *Cause:
// *Action:
SQL> alter system set events '10511 trace name context forever,level 1';
System altered.
OFFLINE UNDO SEGS 的相关 BUG
以下列出了 SMON OFFLINE UNDO SEGS 的一些公开的 BUG，这些 BUG 一般都存在于 10.2.0.3
之前; 若你真的遇到了，可以在考虑升级之余 采用 10511 event workaround 规避该问题：
Hdr: 2726601 9.2.0.2 RDBMS 9.2.0.2 TXN MGMT LOCAL PRODID-5 PORTID-46 ORA-600 3439552
Abstract: ORA-600 [4406] IN ROUTINE KTCRAB(); 4 NODE RAC CLUSTER
Hdr: 6878461 9.2.0.4.0 RDBMS 9.2.0.4.0 TXN MGMT LOCAL PRODID-5 PORTID-23 ORA-601
5079978
Abstract: ESSC: ORA-601 ORA-474 AFTER OFFLINING UNDO SEGMENTS
Hdr: 4253991 9.2.0.4.0 RDBMS 9.2.0.4.0 TXN MGMT LOCAL PRODID-5 PORTID-23 ORA-600
2660394
Abstract: ORA-600 [KTSXR_ADD-4] FOLLOWED BY ORA-600 [KTSISEGINFO1]
Hdr: 2696314 9.2.0.2.0 RDBMS 9.2.0.2.0 TXN MGMT LOCAL PRODID-5 PORTID-46
Abstract: RECEIVING ORA-600: [KTUSMGMCT-01] AFTER APPLYING 92020 PATCH SET
Hdr: 3578807 9.2.0.4 RDBMS 9.2.0.4 TXN MGMT LOCAL PRODID-5 PORTID-23 ORA-600
Abstract: OERI 4042 RAISED INTERMITTENTLY
Hdr: 2727303 9.2.0.1.0 RDBMS 9.2.0.1.0 TXN MGMT LOCAL PRODID-5 PORTID-100 ORA-600
Abstract: [RAC] ORA-600: [KTUSMGMCT-01] ARE OCCURED IN HIGH LOAD了 解 你 所 不 知 道 的 SMON 功 能 ( 十 二 ):Shrink
UNDO(rollback) SEGMENT
SMON 对于 Undo(Rollback)segment 的日常管理还不止于 OFFLINE UNDO SEGMENT ，在
AUM(automatic undo management 或 称 SMU) 模 式 下 SMON 还 定 期 地 收 缩 Shrink
Rollback/undo segment。
触发场景
这种 AUM 下 rollback/undo segment 的 undo extents 被 shrink 的现象可能被多种条件触发：
当另一个回滚段的 transaction table 急需 undo 空间时
当 SMON 定期执行 undo/rollback 管理时(每 12 个小时一次):
SMON 会从空闲的 undo segment 中回收 undo space，以便保证其他 tranaction table 需要空间
时可用。另一个好处是 undo datafile 的身材不会急速膨胀导致用户要去 resize
当处于 undo space 空间压力时，特别是在发生 UNDO STEAL 的条件下; SGA 中会记录前台进
程因为 undo space 压力而做的 undo steal 的次数
(v$undostat UNXPSTEALCNT EXPSTEALCNT)；若这种UNDO STEAL的次数超过特定的阀值，
则 SMON 会尝试 shrink transaction table
若 smon shrink rollback/undo 真的发生时，会这样处理：
计算平均的 undo retention 大小，按照下列公式：
retention size=(undo_retention * undo_rate)/(#online_transaction_table_segment 在线回滚段的个数)
对于每一个 undo segment
若是 offline 的 undo segment，则回收其所有的已过期 expired undo extents，保持最小 2 个 extents
的空间
若是 online 的 undo segment，则回收其所有的已过期 expired undo extents，但是保持其 segment
所占空间不小于平均 retention 对应的大小。
注意 SMON 的定期 Shrink，每 12 个小时才发生一次，具体发生时可以参考 SMON 进程的 TRACE。
若系统中存在大事务，则 rollback/undo segment 可能扩展到很大的尺寸；视乎事务的大小，则 undo
tablespace 上的 undo/rollback segment 会呈现出不规则的空间占用分布。
SMON 的定期清理 undo/rollback segment 就是要像一个大锤敲击钢铁那样，把这些大小不规则的 online
segment 清理成大小统一的回滚段，以便今后使用。
当然这种定期的 shrink 也可能造成一些阻碍，毕竟在 shrink 过程中会将 undo segment header 锁住，则事
务极低概率可能遇到 ORA-1551 错误:
[oracle@vmac1 ~]$ oerr ora 1551
01551, 00000, "extended rollback segment, pinned blocks released"
// *Cause: Doing recursive extent of rollback segment, trapped internally// by the system
// *Action: None
如何禁止 SMON SHRINK UNDO SEGMENT?
可以通过设置诊断事件event=’10512 trace name context forever, level 1′来禁用SMON OFFLINE
UNDO SEGS;
SQL> select * from global_name;
GLOBAL_NAME
--------------------------------------------------------------------------------
www.oracle.com
SQL> alter system set events '10512 trace name context forever,level 1';
System altered.
相关 BUG
这些 BUG 主要集中在 9.2.0.8 之前，10.2.0.3 以后几乎绝迹了：
Bug 1955307 – SMON may self-deadlock (ORA-60) shrinking a rollback
segment in SMU mode [ID 1955307.8]
Bug 3476871 : SMON ORA-60 ORA-474 ORA-601 AND DATABASE CRASHED
Bug 5902053 : SMON WAITING ON ‘UNDO SEGMENT TX SLOT’ HANGS DATABASE
Bug 6084112 : INSTANCE SLOW SHOW SEVERAL LONGTIME RUNNING WAIT EVENTS