Oracle-07445[kgghash]:OracleBUG导致更新LOB字段时进程被KILL掉(二)

2014-11-24 17:04:17 · 作者: · 浏览: 4
(0x777ee08) jmp 0x777ed19
kgghash()+377 (0x777ee0d) add (%rdi),%edx
kgghash()+379 (0x777ee0f) add 0x4(%rdi),%ecx

*** 2013-04-22 14:49:58.339
----- Current SQL Statement for this session (sql_id=bb8bnr20h93mk) -----
update oa9.sys_process set fd_detail=:1 , fd_descriptor=:2 where fd_id=:3

----- Call Stack Trace -----
skdstdst <- ksedst1 <- ksedst <- dbkedDefDump <- ksedmp
<- ssexhd <- restore_rt <- kgghash <- kkocsCreateBindSet <- kkocsInitAdaptiveCtx
<- apadrv <- opitca <- PGOSF385_kksFullTypeCheck <- rpiswu2 <- kksSetBindType
<- kksfbc <- opiexe <- kpoal8 <- opiodr <- ttcpip
<- opitsk <- opiino <- opiodr <- opidrv <- sou2o
<- opimai_real <- ssthrdmain <- main <- libc_start_main <- start

----------------------------------------
Cursor#35(0x2b1732bb2c28) state=BOUND curiob=0x2b1733541e50
curflg=4c fl2=0 par=(nil) ses=0x214cd7880
----- Dump Cursor sql_id=bb8bnr20h93mk xsc=0x2b1733541e50 cur=0x2b1732bb2c28 -----
Dump Parent Cursor sql_id=bb8bnr20h93mk phd=0x21b32bee8 plk=0x21a12acb8
sqltxt(0x21b32bee8)=update oa9.sys_process set fd_detail=:1 , fd_descriptor=:2 where fd_id=:3
hash=d464c012671d8361b5a174b881048e72
parent=0x1e7cda730 maxchild=01 plk=0x21a12acb8 ppn=n
cursor instantiation=0x2b1733541e50 used=1366613397 exec_id=0 exec=0
child#0(0x218b3bb40) pcs=0x1a2dff738
clk=0x21ac20ed8 ci=0x1e737a828 pn=0x21bd41568 ctx=0x125d120e0
kgsccflg=0 llk[0x2b1733541e58,0x2b1733541e58] idx=0
xscflg=130672 fl2=11100008 fl3=2222600 fl4=100
----- Bind Byte Code (IN) -----
Opcode = 14 Bind Skip Long
Opcode = 2 Bind Twotask Scalar Sql In (may be out) Copy
oacdef = 0x1a3438220 Offsi = 48, Offsi = 32
Opcode = 2 Bind Twotask Scalar Sql In (may be out) Copy
oacdef = 0x1a3438250 Offsi = 48, Offsi = 64
Opcode = 1 Unoptimized
Null Offset
----- Bind Info (kkscoacd) -----
Bind#0
oacdty=01 mxl=4001(5018) mxlc=00 mal=00 scl=00 pre=00
oacflg=03 fl2=1000010 frm=01 csi=852 siz=4000 off=0
kxsbbbfp=2b1733614060 bln=4000 avl=4001 flg=25
value=" WZ3+"...
Bind#1
oacdty=01 mxl=2000(1620) mxlc=00 mal=00 scl=00 pre=00
oacflg=03 fl2=1000010 frm=01 csi=852 siz=2128 off=0
kxsbbbfp=2b17336137d0 bln=2000 avl=822 flg=05
value="







Bind#2
oacdty=01 mxl=128(64) mxlc=00 mal=00 scl=00 pre=00
oacflg=03 fl2=1000010 frm=01 csi=852 siz=0 off=2000
kxsbbbfp=2b1733613fa0 bln=128 avl=32 flg=01
value="13e3064a34e8924d0e196e044089dff5"
Frames pfr (nil) siz=0 efr (nil) siz=0
Cursor frame dump
enxt: 2.0x00000060 enxt: 1.0x00000898
pnxt: 1.0x00000060
kxscphp=0x2b1733606828 siz=1008 inu=176 nps=96
kxscbhp=0x2b17335a5660 siz=11456 inu=10696 nps=10512

分析:
ORA-07445通常是ORACLE的BUG引起的,在文档ORA-7445 [kgghash] [ID 310172.1]中发现有符合我们现象的BUG:
Bug 13463131 Dump (kgghash) from bind peeking,BUG 13463131的完整内容:
Bind peeking may dump in kgghash() when long character datatypes are
being used.
The stack will typically include the following calls:
kkocsCreateBindSet->kgghash
<===== 与我们的CASE相符合,表 oa9.sys_process有CLOB字段,且在我们的CASE中是call stack为kgghash <- kkocsCreateBindSet <- kkocsInitAdaptiveCtx

The bind variable that causes the dump shows the avl value is greater than
the bln (in the cursor dump in the trace).

Bind#7
oacdty=96 mxl=4001(22518) mxlc=00 mal=00 scl=00 pre=00
oacflg=01 fl2=1000010 frm=01 csi=873 siz=4000 off=0
kxsbbbfp=0e42f060 bln=4000 avl=4001 flg=25

This situation can happen, for example, using a CLOB column and having
the client bind with CHAR.
<====抽取我们的dump文件相关内容,可以发现bln=4000 avl=4001,符合文档描述。
Bind#0
oacdty=01 mxl=4001(5018) mxlc=00 mal=00 scl=00 pre=00
oacflg=03 fl2=1000010 frm=01 csi=852 siz=4000 off=0
kxsbbbfp=2b1733614060 bln=4000 avl=4001 flg=25 <<<<<<
value=" WZ3+"...
========================================
Workaround
Disable bind peeking (set "_optim_peek_user_binds" = false)
but note that this may cause changes to execution plans so
may not be practical.
<=====Oracle给出的解决方案是禁用bind peeking,但是带来的风险是可能导致SQL执行计划改变
另外可以升级至11.2.0.3之后再apply patch 13463131解决。
Note: This fix supersedes fixes of bug 7352414 , bug 8763922.

What is "Bind Peeking"

With 9i, a new feature "Bind Peeking" was introduced to try to get around the issues associated with the selectivity of a bind variable.
During hard parsing of a query using bind variables, the binds are "peeked at" and selectivity is generated based on the bind and the underlying column statistics.
This method is successful but if the bind selectivity of the initial execution of a statement varied from the selectivity of subsequent executions with different sets of binds this could sometimes lead to plans being enerated that were not representative of the general query usage and which performed poorly.

This feather is controlled by the hidden parameter "_OPTIM_PEEK_USER_BINDS". the default value is 'true'.

Peeking is enabled if all of the following conditions are met:
* - optimizer_features_enabled is greater than 8.2
* - there are user binds
* - the cursor is fully bound
* - the cursor is not remote mapped
* - the client is using V8OCI
* - it's a select or DML
* - peeking is possible for this statement and the statement is being
* loaded for the first time, or the original cursor was peeked and
* this cursor is being reloaded

For more details, please refer to:

Document 387394.1 Query using Bind Variables is suddenly slow
Document 430208.1 Bind Peeking By Example
"_OPTIM_PEEK_USER_BINDS" [Hidden] Reference Note [ID 201672.1]