So, for several weeks now, I’ve been struggling with disk naming and UDEV as they relate to RHEL or CentOS Kickstart automation via PXE. Basically, what used to be a relatively straight-forward process of setting up partitions based on disk name (e.g. /dev/sda for the primary disk, /dev/sdb for the secondary disk, and so on) has become a little more complicated due to the order in which the kernel may identify disks with UDEV implementations. The originally-used naming conventions of /dev/sdX may no longer be consistent across reboots or across machines. How this problem came to my attention was when I was Kickstarting a bunch of hosts, and realised that several of them were indicating that there wasn’t enough space on the drives that I had referenced. After doing some digging, I realised that some servers were identifying KVM-based media (like via a Cisco UCS CIMC interface), or other USB media as primary disks instead of the actual primary SCSI-based disks (SATA or SAS drives).

Though I agree that identifying disks by their path, their UUID, or an otherwise more permanent name is preferred to the ambiguous /dev/sdX naming scheme, it did cause a problem, and one that didn’t have a readily-identifiable workaround for my situation. My first thought was to use UUID, or gather the disk ID from /dev/disk/by-id/. However, that wouldn’t work because it wouldn’t be the same across all servers, so automatic installations via PXE wouldn’t be feasible. Next, I looked at referencing disks by their PCI location, which can be found within /dev/disk/by-path/. Unfortunately, I found that the PCI location might vary between servers as well. For example, I found:

Server 1:
/dev/disk/by-path/pci-0000:82:00.0-scsi-0:2:0:0 –> primary SCSI disk
/dev/disk/by-path/pci-0000:82:00.0-scsi-0:2:1:0 –> secondary SCSI disk

Server 2:
/dev/disk/by-path/pci-0000:0b:00.0-scsi-0:2:0:0 –> primary SCSI disk
/dev/disk/by-path/pci-0000:0b:00.0-scsi-0:2:1:0 –> secondary SCSI disk

That being said, I did notice one commonality between the servers that I used for testing. The primary and secondary disks (which, by the way, are RAID arrays attached to the same controller) all ended with ‘scsi-0:2:0:0 ‘ for the primary disk, and ‘scsi-0:2:1:0 ‘ for the secondary disk. Thinking that I could possibly just specify using a wildcard, I tried:

part /boot --fstype ext2 --size=100 --ondisk=disk/by-path/*scsi-0:2:0:0

but alas, that caused Anaconda to error out stating that the specified drive did not exist. At this point, I thought that all hope was lost in terms of automation. Then, however, I had a flash of genius (they don’t happen all that often). I could probably gather the full path ID in a pre-installation script, and then get the FULL path ID into the Kickstart configuration file with an include. Here are the code snippets that I wrote:


%pre --interpreter=/bin/bash
## Get the full by-path ID of the primary and secondary disks
osdisk=$(ls -lh /dev/disk/by-path/ | grep 'scsi-0:2:0:0 ' | awk '{print $9}')
datadisk=$(ls -lh /dev/disk/by-path/ | grep 'scsi-0:2:1:0 ' | awk '{print $9}')


## Create a temporary file with the partition scheme to be included in the KS config
echo "# Partitioning scheme" > /tmp/partition_layout
echo "clearpart --all" >> /tmp/partition_layout
echo "zerombr" >> /tmp/partition_layout
echo "part /boot --fstype ext2 --size=100 --ondisk=disk/by-path/$osdisk" >> /tmp/partition_layout
echo "part swap --recommended --ondisk=disk/by-path/$osdisk" >> /tmp/partition_layout
echo "part pv.01 --size=100 --grow --ondisk=disk/by-path/$osdisk" >> /tmp/partition_layout
echo "part pv.02 --size=100 --grow --ondisk=disk/by-path/$datadisk" >> /tmp/partition_layout
echo "volgroup vg_os pv.01" >> /tmp/partition_layout
echo "volgroup vg_data pv.02" >> /tmp/partition_layout
echo "logvol / --vgname=vg_os --size=100 --name=lv_os --fstype=ext4 --grow" >> /tmp/partition_layout
echo "logvol /data --vgname=vg_data --size=100 --name=lv_data --fstype=ext4 --grow" >> /tmp/partition_layout
echo "# Bootloader location" >> /tmp/partition_layout
echo "bootloader --location=mbr --driveorder=/dev/disk/by-path/$osdisk,/dev/disk/by-path/$datadisk --append=\"rhgb\"" >> /tmp/partition_layout
%end


Some notes about the pre-installation script would be: 1) yes, I know that it is a bit clumsy, but it functions despite the lack of elegance; 2) in lines 3 and 4, make sure that the grep statements include a final space. The final space (e.g. grep 'scsi-0:2:0:0 ' instead of just grep 'scsi-0:2:0:0') is important because the machine may have partitions already set up, and in that case, those partitions would be identified as '*scsi-0:2:0:0-part#' where # is the partition number.

So, this pre-installation script generates a file called /tmp/partition_layout that essentially has a normal partition scheme for a Kickstart configuration file, but references the primary and secondary SCSI disks (again, as RAID arrays attached to the same controller) by their full ‘by-path’ IDs. Then, the key is to include that file within the Kickstart configuration via:


# Partitioning scheme information gathered from pre-install script below
%include /tmp/partition_layout


I am happy to report that for the servers that I have tested thus far, this method works. Is it possible that there will be primary and secondary disks that are identified via different path IDs not caught by my pre-installation script? Yes, that is possible. It’s also possible that your situation will require a completely different pre-installation script to gather the unique identifier of your choice. That being said, I hope that this post will help you devise your method of reliably identifying disks so that you can create the partition scheme automatically via your Kickstart configuration file. That way, you can then deploy it to many servers via PXE (or any other implementation that you’re currently utilising for mass deployment).

Cheers,
Zach