I was analyzing some unexpected RMAN termination, a RMAN-10038/RMAN-03009 combo:

RMAN-00571: ===========================================================
RMAN-00569: =============== ERROR MESSAGE STACK FOLLOWS ===============
RMAN-00571: ===========================================================
RMAN-03009: Fehler bei REFAF Befehl in c13 Kanal auf 10/31/2011 08:21:50
RMAN-10038: Datenbank-Session für Kanal c13 unerwartet beendet

Nothing, not even an RMAN tracing was able to reveal more hints. The trace just told it in other words. No underlying ORA/TNS error or similar.

It finally turned out it were some parameters recently added to sqlnet.ora, I set SQLNET.RECV_TIMEOUT and SQLNET.SEND_TIMEOUT and the beast silently dropped RMAN channels that were idle for a while. The next command in this channel blew up the whole RUN block of the backup script.

I removed the parameters and it works again.

Be careful with your sqlnet.ora

dmr is widely known primarily as one of the Unix and C fathers. He’s also the co-author of the famous “K&R” book, “The C programming language”.

From my point of view the IT world lost a great man whose echo still lives on in many technologies. I would also like to pass on my condolences to his family and friends.

• / (slash) to separate username and password
• @ (at) to separate the TNS descriptor string

If you need to use those characters literally in the logon string, you need to tag them with literal double quotes (literal means: the quotes need to be passed to SQL*Plus, I’m not talking about the UNIX shell):

• Less readable:

  $ sqlplus USER/\"PASS/WORD\"

• More readable:

  $ sqlplus USER/'"PASS/WORD"'

In the past, all storage migration operations there had two problems that lead to downtime:

• CRS votedisk migration at best only with offline CRS
• kick old devices from the device management in the operating system

In December I migrated the system to something more modern

• SLES11
• Oracle Grid Infrastructure (CRS/ASM) 11.2
• Oracle Database 10.2.0.5 (yes, still quite old, but it has to be a 10.2 for now)

The use of a modern Linux kernel (especially a modern SCSI stack) and 11gR2 infrastructure fixed all my trouble from the past.

Votedisk

With 11gR2 Grid Infrastructure, the clusterware manages its vital files (VD, OCR) using an ASM instance. A migration of the ASM diskgroups holding these files is now as easy as a migration of a normal diskgroup. ASM takes care of moving the votedisks and the OCR and collaborates with CRS here.

Linux devices

The new Linux kernels with a finally sane SCSI stack and native multipathing helps with the second problem. It’s not a problem anymore to remove old device references from the stack:

• deconfigure the devices from multipathd (not needed technically, since multipathd itself holds no device references)
• remove the device references from the Linux device mapper

  # see /dev/mapper/* for the name
  dmsetup remove mpathX

• remove the LUNs from the SCSI stack

  # X:X:X:X LUN number
  echo 1 > /sys/bus/scsi/X:X:X:X/delete

After that, you can safely edit the FC zone – no errors should occur in any logs.

Regards and best online-wishes for 2011 fly to (the order doesn’t matter)

• my family
• my girlfriend
• USN (+extra wishes for his family)
• WAGER
• my collegues at work
• regulars of the IRC channel #bash on the Freenode Network
• …uhm… greybot!
• regulars of the IRC channel #linux-forum.de on the Freenode Network
• regulars of Linux-Forum.de
• the Ekiga team
• Damien and his new life
• regulars on the bug-bash mailinglist
• Urban Terror guys from FMD
• members of the LUG Mitterteich
• everybody I was able to help with the Bash Hackers Wiki
• …and all the other online-people that don’t come to my mind right now – I’m in a hurry, sorry

ALTER SYSTEM SWITCH LOGFILE;

In a RAC with some more nodes, it can be quite disturbing to do that on all instances.

When you inspect the protocol files of the right services, you will find another, undocumented method:

ALTER SYSTEM SWITCH ALL LOGFILE;

This works at least on 10.2.0.5. Since it’s undocumented, I can’t tell you under which circumstances this command works or not, but it looks like always possible. Please find out yourself and let me know.

The export routines were done with datapump export. I watched a manual run and…

• datapump started and connected to a database server process: 3 seconds
• datapump estimated and analyzed the data to export: 36 minutes
• datapump exported the relevant tables: 4 seconds

Do you see what I see? Yes? Right, datapump played with itself. Over 36 minutes of runtime, where 99% of the time not one bit of data was moved. Reproducable 100%. A datapump TRACE=480300 generated trace set revealed a huge delay in the worker process when calling DBMS_METADATA.FETCH_XML_CLOB (exactly my 36 minutes).

The solution was a patch for bug

• 7710931: DATAPUMP EXPORT IS EXTREMELY SLOW WHEN EXTRACTING SCHEMA

Now datapump does what it’s designed for: Pumping data without playing with itself!

$ echo $((2#$(date -d '10 oct 2010' +"%d%m%y")))
42

Be prepared!

On IRC or mailinglists I frequently see people not understanding the idea and concepts behind glue records. Many have no clue about it. It’s always the same, some people don’t get it and thus make their sites unavailable – when you explain it, they don’t understand.

Hence, I suggest a name change, just name them

The clue records

I have a small LDAP setup running for centralized Linux authentication since years, using the NSS and PAM modules from PADL Software Pty Ltd.

A few weeks ago I found the nss-pam-ldapd from Arthur de Jong and found the architecture is promising and fixes some problems with the design of the PADL modules.

This project needs contributors, so if you

• have time
• know some C
• like LDAP
• have ideas
• can test

it would be nice to see you there.

I also contributed a line or two.