<< Prev  |  TOC  |  Front Page  |  Talkback  |  FAQ  |  Next >>
...making Linux just a little more fun!
But All My Partitions Were Mirrored
By Graham Jenkins

No Problem, Everything is Mirrored

This story actually started with a call from a user whilst I was strolling back to work through the sunshine one Friday lunchtime. The conversation went something like this:

"Hi Graham, we seem to be having a few problems in seeing the database for the ACME application. You want to take a look, please?"

"Sure, I'm ten minutes away from my desk, I'll call you back when I'm there. Everything on that server is mirrored; most likely scenario is that the archive logs are not being moved off to secondary storage. Should be able to resolve it in a few minutes."

And ten minutes later: "Guys, its going to take more that a few minutes. Something like a few hours, in fact. We seem to have lost disks from both sides of the mirrors!"

How Could You Lose Both Sides of a Mirror at Once?

So what went wrong? The mirror pieces were on separate disks attached to separate controllers, there was no evidence of a major power spike or earth tremor. And we couldn't blame the night-time cleaning staff for pulling power cables so they could use their vacuum cleaners.

The answer is that we didn't lose both sides at once. We had actually lost one side a week earlier. My company has an excellent monitoring and alarm system for detecting such occurences, but we had forgotten to advise the alarm people that this server had moved from "build" status to "production" status. That's not something we are likely to do again!

A Bit Closer to Home

A few weeks back, my home workstation experienced its second disk failure in six months. Sure, the disk got replaced again under warrantee. But I decided right then that I was going to mirror everything onto an additional disk.

Then I started thinking: "How would I know if a partition on one disk took itself off-line?" It's not like I can justify hooking my home workstation into my company's alarm system.

Did somebody say: "Check the messages file, read the 'root' email!"? Great theory guys. Trouble is, I have a partner whose idea of "messages" equates to a stack of Post-It notes, and who thinks that "email" means "Hotmail". And she has become a major user of my machine when I'm not around.

A Simple Watchdog Mechanism

The solution here turned out to be a mechanism to flash the Scroll-Lock light for a one second interval every ten seconds. If a partition gets unmirrored, the light gets left on. No extra hardware, dead easy to understand. Simple Watchdog What we have here is a simple watchdog, which barks periodically to show it is still alive, and barks continuously when something goes wrong.

So how do you make the Scroll-Lock light flash? If you are using Xwindows, it's easy: 'xset led 3' turns it on, 'xset -led 3' turns it off. Even works if you have screen-lock running and/or your monitor powered off - provided you are logged in.

If nobody is logged in, or if you aren't using Xwindows, it isn't going to work. For that situation, you need to install something like the 'blinker' program which comes as part of the "morse2led" suite available at the node.to website.

The Program

Here's what you might see when you enter 'cat /proc/mdstat' on a machine which has a broken mirror:

  Personalities : [raid1] 
  read_ahead 1024 sectors
  md2 : active raid1 hda6[0] hdb6[1](F)
        1959808 blocks [2/1] [U_]
  md1 : active raid1 hda5[0] hdb5[1]
        5863616 blocks [2/2] [UU]
  md0 : active raid1 hda3[1] hdb3[0]
        104320 blocks [2/2] [UU]
  unused devices: <none>

And here's our program which detects when something is wrong (by searching for an underscore in those lines containing 'blocks'), then activates the scroll-lock light accordingly. It will run under most Bourne-like shells, and has been extended to detect a couple of extra alarm conditions. You can add to it as you see fit.

# ledblink      System monitor. Scroll-lock light will remain on if any faults.
#               Graham Jenkins, IBM GSA, July 2003.

while : ; do                                    # Use 'blinker' if it works,
  blinker -d `expr $On \* 1000` s 2>/dev/null ||# else use 'xset' to flash the
    ( xset led 3 && sleep $On && xset -led 3 )  # scroll-lock light on and off.
  sleep `expr 10 - $On` 
  On=10                                         # Set on-time to 10 seconds.
  # Raid status
  grep blocks /proc/mdstat | grep _ >/dev/null 2>&1                && continue
  # Filesystem capacity
  df -x iso9660 |tr -d '%'|awk '{if (NR > 1) if ($5 > 90) exit 1}' || continue
  # Swap usage
  swapon -s | awk '{ if (NR > 1) { Size=Size+$3; Used=Used+$4 } }
               END { if (Used*100/Size > 70 ) exit 1            }' || continue
  On=1                                          # If there are no problems
done                                            # reset on-time to 1 second.

Starting Up

If you are happy for 'ledblink' to run only when somebody is logged on with an Xwindows session, it's easy. If your machine has an 'xinitrc.d' directory, place the following script in it. Otherwise, place the uncommented line in the 'xinitrc' file.

# ledblink      Place this file in: /usr/X11R6/lib/X11/xinit/xinitrc.d
#               and make it readable and executable for everyone.
[ -x /usr/local/bin/ledblink ] && /usr/local/bin/ledblink &

If you have the 'blinker' program, you can start 'ledblink' at boot time with the following script.

# ledblink      Start/stope the 'ledblink' system monitor program.
#               Graham Jenkins, IBM GSA, July 2003.
# chkconfig: 2345 98 7
# description: Start/stops the 'ledblink' system monitor program.

case "$1" in
start) if [ -x /usr/local/bin/ledblink ] ; then
         [ -s /var/run/ledblink.pid ] && exit 0 
         echo "Starting 'ledblink' system monitor program .."
         /usr/local/bin/ledblink &
         echo $! >/var/run/ledblink.pid
       fi ;;
stop)  if [ -n "`cat /var/run/ledblink.pid`" ] ; then
         echo "Stopping 'ledblink' system monitor program .."
         kill `cat /var/run/ledblink.pid`
         rm        /var/run/ledblink.pid
       fi ;;


[picture] Graham is a Unix Specialist at IBM Global Services, Australia. He lives in Melbourne and has built and managed many flavors of proprietary and open systems on several hardware platforms.

Copyright © 2003, Graham Jenkins. Copying license http://www.linuxgazette.com/copying.html
Published in Issue 93 of Linux Gazette, August 2003

<< Prev  |  TOC  |  Front Page  |  Talkback  |  FAQ  |  Next >>