Last week there was a power grid failure which break down my server’s RAID array. I have no UPS (as I’m a skinflint) and no automatic email alerts (because I’m too lazy to set it up). As a result, for 5 days, my 3-disk RAID-5 array was relying on only 2 disks until I noticed the issue…

By using a combination of following commands, I was soon aware of the gravity of the situation: 

$ cat /proc/mdstat
$ mdadm --examine /dev/sda1

My /dev/sda1 disk was kicked out of the array, so I did the right stuff which consisted of reconstructing the array: 

$ mdadm /dev/md0 -a /dev/sda1

Then, in an unlucky combination of cosmic ray bombardment, spooky action at a distance and astrological misalignment, half-way to the end of the rebuilding process (which can take up to 5 hours), another disk failed! It was late, I was tired and utterly worried about losing 1.5 To of precious data. In such a bad shape, I was afraid to worsen the situation. So I decided to shutdown the server and sleep on the problem.

The next day I tried to boot my server to find it (surprise!) stuck in the middle of the boot process, with the famous message: 

hit control-D to continue or give root password to fix manually

This is “normal” as my server tried to mount the ext3 filesystem from the /dev/md0 partition that was just assembled by mdadm . Of course md0 , if assembled and available to the system, was not running because only one disk, out of three, was in a clean state.

I skip here the epic substory in which I wasted days in a search of a working keyboard, but I let you imagine how such adventures makes my week…

Eventually, I was able to analyze the situation in details. My first reflex? Check that disks are not physically dead: 

$ fdisk -l /dev/sda
$ fdisk -l /dev/sdb
$ fdisk -l /dev/sdc

“Linux raid partitions” (type code “ fd ”) are still there. Good. I assumed here that disks where not physically damaged. Maybe I should have looked at S.M.A.R.T. data and statistics (via smartmontools ). But remember, I’m lazy (and a bit crazy).

The next step was to get information about the RAID array itself using: 

$ mdadm --detail /dev/md0

which output the status table below (probably inaccurate as I reconstructed it afterwards): 

Number   Major   Minor   RaidDevice State
   0       0        0        0      removed
   1       0        0        1      faulty removed
   2       8       33        2      active sync   /dev/sdc1
   3       8       17        3      spare

What this table told us?

  • The array is up, but not running. One of its device ( sdc1 ) was clean and active, but it’s not enough to get a working RAID-5.

  • My first attempt to rebuild the array lead to an unexpected result: it added sda1 as a spare device (in slot #3).

  • It confirm that sdb1 unexpectedly failed and is now in a bad state (” faulty removed ”).

Then I stopped the array and tried to fearlessly (re)assemble it using 3 different methods: 

$ mdadm -S /dev/md0
$ mdadm -A /dev/md0
$ mdadm --assemble /dev/md0 --verbose /dev/sd[abc]1
$ mdadm --assemble --force --scan /dev/md0 --verbose

It always failed with messages like: 

mdadm: failed to RUN_ARRAY /dev/md0: Input/output error
mdadm: /dev/md0 assembled from 1 drives and 1 spare - not enough to start the array.

So I examined each drive from mdadm ’s point of view: 

$ mdadm -E /dev/sda1
$ mdadm -E /dev/sdb1
$ mdadm -E /dev/sdc1
$ mdadm -E /dev/sd[abc]1 | grep Event

The latest command compare the “ Event ” attribute of all devices. It output something like: 

Events : 0.53120
Events : 0.53108
Events : 0.53120

which indicate that sda1 and sdc1 are somewhat synced (share the same number) and sdb1 “late” (lower number).

Here I’ve got the idea of recreating the raid array without sdb1 , relying only on sda1 and sdc1 , by using the “magic” (hence dangerous) --assume-clean option. The latter doesn’t build, erase or initialize a new array. It just try to assemble it “as is”. Here is the command: 

$ mdadm --create /dev/md0 --assume-clean --level=5 --verbose --raid-devices=3 /dev/sda1 missing /dev/sdc1

And it worked! :D

I mounted the md0 partition and cleaned it up: 

$ fsck.ext3 -v /dev/md0
$ mount /dev/md0

I updated my mdadm configuration before rebooting my server: 

$ mdadm --detail --scan >> /etc/mdadm/mdadm.conf
$ vi /etc/mdadm/mdadm.conf
$ reboot

But history repeat itself, and again, the system hang up during boot. Except this time I knew what was happening: the boot process detected the remaining sdb1 device as part of the old array (the one before the regeneration I did above) and tried to run it. Remembering my last year post , I zero-ized the superblock of sdb1 : 

$ mdadm -S /dev/md0
$ mdadm --zero-superblock /dev/sdb1

A server reboot proved I was right and my md0 partition was automagically mounted in altered state: 

$ cat /proc/mdstat
Personalities : [raid6] [raid5] [raid4]
md0 : active raid5 sdb1[3] sda1[0] sdc1[2]
      1465143808 blocks level 5, 64k chunk, algorithm 2 [3/2] [U_U]

unused devices: <none>

I just had to re-add sdb1 to fill the available slot and update the mdadm configuration to get back my array in its initial state: 

$ mdadm --manage /dev/md0 --add /dev/sdb1
$ mdadm --detail --scan >> /etc/mdadm/mdadm.conf
$ vi /etc/mdadm/mdadm.conf