Saturday, August 23, 2008

Diagnosing Library Cache Latch Contention: A Real Case Study

The other day, our data warehouse database was effectively down. The database was up and running; but all connection attempts to the database would just hang. Since connections just failed, the DBA couldn’t even get in to see if the sessions connected were running normally or not. The ETL team reported that the processing slowed down to a crawl. The DBA, as instructed, went on to check the wait events; but since she couldn’t even log in, she couldn’t do that. Interestingly, CPU was around 70%, normal for the time of the day; and I/O was about 90%, again normal.

So, in simple terms – she was stuck and she applied the age-old technique: when in doubt, reboot. She asked the SA to reboot the server. It took about 30 minutes; and after it cam backup and the database started everything seemed normal for about 10 minutes. And, then the same problem stuck – the database got unresponsive.

At that point, I was called for help. In this blog I am sort of trying to paint a picture of what I did in the next half hour or so to resolve the issue. I hope by giving this running commentary of the events, facts and figures, you will follow my thoughts as I went about diagnosing the issue.


(1) Database connections were hanging
(2) Connections on the server using SQL*Plus as sysdba was also hanging; so we couldn’t even examine the issue.
(3) The system couldn’t be rebooted everytime; every reboot means more work for the ETL folks to clean up the failed sessions.

Initial Analysis

Obviously something is holding some resource back in the database; but what? Unless we connected to the database, we couldn’t check. And since couldn’t check, we couldn’t remediate. It was a classic case of Catch-22.


This is one of those cases where a backdoor entry to the database instance comes really handy. Most people are not aware of the option in SQL*Plus called “prelim”. This option connects to the SGA; but does not open a session.

(1) So, the first thing I did is to open a sessionless SQL*Plus session using the Oracle user:

$ sqlplus -prelim

Note, it didn’t say anything familiar like “Connected to Oracle Database”, etc. All it showed was the “SQL>” prompt. That was because it didn’t actually connect to the database.

(2) Then I used the “oradebug” utility to analyze the SGA

SQL> oradebug setmypid
SQL> oradebug hanganalyze 12

This produced a tracefile in the user_dump_dest directory. The file wasn’t difficult to find, since it was the last file created. Even if I didn’t find the file, I could have used the process ID to find the file. The file would have been named crmprd1_ora_13392.trc, assuming 13392 was the process ID.

(3) Let’s examine the file. Here are first few lines:

*** 2008-08-23 01:21:44.200
Found 163 objects waiting for
Open chains found:
Chain 1 : :

This tells me a lot. First it shows that the SID 146 Serial# 17 is waiting for library cache latch. The blocking session is SID 226 Serial# 17. The latter is not waiting for anything of blocking nature. I also noted the OS process IDs of these sessions – 16108 and 19923.

(4) Next I checked for two more tracefiles with these OS PIDs in their names.


(5) I opened the first one, the one that is the blocker. Here are the first few lines:

*** 2008-08-23 01:08:18.840
*** SERVICE NAME:(SYS$USERS) 2008-08-23 01:08:18.781
*** SESSION ID:(226.17) 2008-08-23 01:08:18.781
LIBRARY OBJECT HANDLE: handle=c0000008dc703810 mtx=c0000008dc703940(8000) cdp=32737
hash=a029fce7bb89655493e7e51a544592a4 timestamp=08-23-2008 00:10:23
namespace=CRSR flags=RON/KGHP/TIM/OBS/PN0/MED/KST/DBN/MTX/[504100d0]
kkkk-dddd-llll=0000-0001-0001 lock=N pin=0 latch#=10 hpc=0058 hlc=0058
lwt=c0000008dc7038b8[c0000008dc7038b8,c0000008dc7038b8] ltm=c0000008dc7038c8[c0000008dc7038c8,c0000008dc7038c8]
pwt=c0000008dc703880[c0000008dc703880,c0000008dc703880] ptm=c0000008dc703890[c0000008dc703890,c0000008dc703890]
ref=c0000008dc7038e8[c0000008dc7038e8,c0000008dc7038e8] lnd=c0000008dc703900[c0000008dc703900,c0000008dc703900]
lock user session count mode flags
---------------- ---------------- ---------------- ----- ---- ------------------------
c0000008d079f1b8 c0000006151744d8 c0000006151744d8 16 N [00]
c0000008d4e90c40 c0000006151bcb58 c0000006151bcb58 16 N [00]
c0000008d0812c40 c0000008151a0438 c0000008151a0438 16 N [00]

(6) This is a treasure trove of information for debugging. First it shows the SID and Serial# (226.17), which confirms the SID we identified earlier. It shows the exact SQL statement being used. Finally it shows all the locks. I didn’t particularly care about the specifics of locks; but it gave me enough information to prove that the SID 226 was causing a wait for a whole lot of other sessions.

(7) My investigation is not done; I need to find out the sessions waiting for this. So, I searched the file for a section called “PROCESS STATE”. Here is a snippet from the file:

Process global information:
process: c00000081502dab8, call: c000000817167890, xact: 0000000000000000, curses: c00000081519ef88, usrses: c000000815
SO: c00000081502dab8, type: 2, owner: 0000000000000000, flag: INIT/-/-/0x00
(process) Oracle pid=370, calls cur/top: c000000817167890/c000000817167890, flag: (0) -
int error: 0, call error: 0, sess error: 0, txn error 0
(post info) last post received: 115 0 4
last post received-location: kslpsr
last process to post me: c000000615002038 1 6
last post sent: 0 0 24
last post sent-location: ksasnd
last process posted by me: c000000615002038 1 6
(latch info) wait_event=0 bits=20
holding (efd=4) c0000008d7b69598 Child library cache level=5 child#=10
Location from where latch is held: kglhdgc: child:: latch
Context saved from call: 13
state=busy, wlstate=free
waiters [orapid (seconds since: put on list, posted, alive check)]:
291 (197, 1219468295, 197)
279 (197, 1219468295, 197)
374 (197, 1219468295, 197)
267 (197, 1219468295, 197)
372 (197, 1219468295, 197)
... several lines sniped ...
307 (15, 1219468295, 15)
181 (6, 1219468295, 6)
waiter count=58
Process Group: DEFAULT, pseudo proc: c0000008e03150d8
O/S info: user: oracrmp, term: UNKNOWN, ospid: 16108
OSD pid info: Unix process pid: 16108, image: oracle@sdwhpdb1

(8) This told me everything I needed to know. There were 58 sessions waiting for library cache latch held by SID 226. I also knew the OS Process ID and the SQL statement of the blocking session.

(9) At that time we engaged the Application Owner to explain what was going on. As he explained it, he issues the update statement in a loop. And that’s not all; he executes it in 8 different threads. No wonder we have had library cache latch contention. So, we had to track 8; not just one session. We trudged on. All the sessions dumped their information. So, I searched the directory for some other files with the same issues:

$ grep “UPDATE DW_ETL” *.trc

(10) And I found 9 more sessions (or, rather, processes). Here is a snippet from another file:

350 (167, 1219470122, 167)
197 (167, 1219470122, 167)
waiter count=185
Process Group: DEFAULT, pseudo proc: c0000008e03150d8
O/S info: user: oracrmp, term: UNKNOWN, ospid: 16114

This process had 185 waiters! Ouch!

(11) Now comes a decision point. I knew who is blocking and who were being blocked; although I didn’t yet know what latch exactly is being contented for. I could have dumped the library cache latches to get that information; but the application owner volunteered to terminate the sessions. The application, fortunately, was restartable. So, we decided to kill all of these errant sessions on the unix prompt.

$ kill -9

(12) After killing a few processes, the database started responding. After killing all of them, the database wait events came back to completely normal. Connections were established and applications started behaving normally.

After step 11, I could have used the library cache dump to examine the exact library element in the center of the contention; but that’s a topic for another blog.


(1) When something seems to be “hung”, don’t get hung up on that. A session almost always waits for something; rarely it is just “hung”. You should check what it is waiting for by selecting the EVENT column of V$SESSION (10g) or V$SESSION_WAIT (9i).
(2) When you can’t logon to the database to get the information, try using oradebug command.
(3) To use oradebug, you should use SQL*Plus. Since you can’t login, use “sqlplus -prelim" to get the SQL prompt.
(4) Use oradebug setmypid to start the oradebug session, and then use oradebug hanganalyze to create a dumpfile of all hang related issues
(5) Use oradebug help to see all oradebug commands
Post a Comment