Computational Engineering, Empirical Computer Science, Distributed Systems

Creating NFS-root VMWare Virtual Machines From the Linux Command Line

| Comments

In our department there’s a course that includes lab sections on programming fluid dynamics simulations in OpenFOAM, which isn’t supported by the University’s IT services. We’ve traditionally gotten around this by installing the required environment in a virtual machine image and running it with VMWare under MS Windows, but there are infrastructural headaches associated with this approach: the image is several gigabytes in size, and so distributing to all of the machines and importing it into VMWare manually on a large number of machines is time consuming and tedious. The research fellow who runs these lab sections asked me to help him come up with a solution to this problem, so my suggestion was to create some minimal virtual machine images that users could download and import into VMware quickly, and have the root file system shared out from a server over NFS. I’ve decided to document the process of creating these images here, on the off chance it might turn out to be useful to someone else. The commands documented below were run on an Ubuntu 12.10 system.

Step 0: Configuring an NFS-root server, and getting appropriate kernel and

initrd files to install in the VM image.

This part is somewhat beyond the scope of this tutorial. On a working NFS server you can use debootstrap to install a system image to a directory that will be accessible from clients, preferably with read-only access.

Some mostly Ubuntu/Debian specific notes:

  1. Don’t forget to edit etc/network/interfaces in the directory that will be shared out to the clients if you need to use DHCP to get a DNS server or want to have any ability to interact with the interface from the VM. While the networking will be pre-setup by the kernel at boot time, none of the information will be made available to the operating system without re-querying the DHCP server.

  2. The best way to get a compatible kernel for the clients is to chroot into the root of the filesystem that you want to share, and then use apt-get to install the kernel there.

    $ mount -t bind /dev /path/to/client/root/dev         
    $ mount -t bind /proc /path/to/client/root/proc
    $ mount -t bind /sys /path/to/client/root/sys
    $ chroot /path/to/client/root /bin/bash
  3. It can be difficult to configure the clients to properly deal with read-only nfs-mounted root filesystems. One easy way to deal with the problem is to use the excellent root-ro script. Within the chroot:

     $ cp root-ro /etc/initramfs-tools/scripts/init-bottom/

    Again, you want this in the configuration directory that will be used by the clients, or, more to the point, update-initramfs. When you create the initrd.img by running the update-initramfs command the script will be copied in and run automatically at boot time, before the root filesystem is made available to user-level programs. What it does is quite clever: it moves the NFS-root mount point from / to /mnt/root-ro, creates a temporary (in memory) filesystem at /mnt/root-rw, and then uses overlayfs or aufs to provide the root. The effect is of having a fully read-write enabled root filesystem, but with all of the writes going to the tmpfs. SIDE NOTE: I prefer aufs as it provides the most “natural” experience, but aufs is apparently on the way out, to be replaced by overlayfs. One problem with this is that the mount options for overlayfs don’t allow you to specify that the “lower” mount is read-only. This results in the odd behavior that if a file already exists on the lower file system, the file can’t be edited. The work-around is to copy the file to a new location, “delete” it (the lower filesystem is unaffected), and then move the copy back to the old location.


     $ cp /etc/passwd /etc/passwd.new
     $ rm /etc/passwd
     $ mv /etc/passwd.new /etc/passwd

    This way when you try to edit the file, overlayfs will know that you are only trying to edit the local copy.

  4. Next, edit /etc/initramfs-tools/initramfs.conf within the chroot. Repeat, this is to make changes to the initramfs.conf file for the client, and you really don’t want to make these changes on the server. Change “MODULES=most” to “MODULES=netboot” and “BOOT=local” to “BOOT=nfs”. You may also want to change “NFSROOT=auto” to a hard-coded value, depending on your setup.

    $ update-initramfs -k all -c

    This will cause the needed vmlinuz and initrd.img files to be generated under /boot; they may have version information appended to the file names.

Step 1: Creating the Disk Image

The first step in the procedure is to create a properly formatted file that will serve as the virtual disk. I used the standard utilities dd, parted, and fdisk to create the file, and installed the syslinux bootloader to the virtual disk’s MBR.

create the empty file:

dd if=/dev/zero of=disk.img bs=1 count=0 seek=1G

add the partition table:

parted -s disk.img mklabel msdos

create the boot partition:

parted -s --align=none disk.img mkpart primary 0 256M

make it bootable:

echo -e "a\n1\nw\nq" | fdisk disk.img

add syslinux mbr (syslinux is the boot loader used to boot the kernel):

dd bs=440 conv=notrunc count=1 if=/usr/lib/syslinux/mbr.bin of=disk.img

create home partition:

parted -s --align=none disk.img mkpart primary 256M+1 100%

Step 2: Create Filesystems and Install Files

For this step it’s important to have the kernel and initrd files you plan on using in one directory, called “boot” in this example, and a skeleton user home directory, called “vmuser” in this case. You also need to create a fairly basic syslinux.cfg file to put in the boot directory:


LABEL client
LINUX ./vmlinuz-linux
APPEND root=/dev/nfs nfsroot=<server-ip>:<nfs-path> ip=::::::dhcp
INITRD ./initramfs-linux.img

attach loopback device to the disk:

sudo losetup /dev/loop0 disk.img

add partitions to /dev:

sudo kpartx -a /dev/loop0

create filesystems:

sudo mkfs -t ext4 /dev/mapper/loop0p1
sudo mkfs -t ext4 /dev/mapper/loop0p2

mount filesystems:

sudo mkdir /mnt/vmboot
sudo mount /dev/mapper/loop0p1 /mnt/vmboot
sudo mkdir /mnt/vmhome
sudo mount /dev/mapper/loop0p2 /mnt/vmhome

install syslinux in boot directory:

sudo extlinux --install /mnt/vmboot

install kernel and config:

sudo cp boot/* /mnt/vmboot

install home:

sudo cp -r vmuser /mnt/vmhome/

set uid/gid:

sudo chown -R 1000:100 /mnt/vmhome/vmuser

clean up after yourself:

sudo umount /mnt/vmboot
sudo rmdir /mnt/vmboot

sudo umount /mnt/vmhome
sudo rmdir /mnt/vmhome

sudo kpartx -d /dev/loop0

sudo losetup -d /dev/loop0

The disk.img file is now suitable for booting! In the next section we’ll tackle how to create an ovf file for use with vmware.

Step 3: Convert the Image to work with VMWare

For this part I first went to easyvmx.com to create a basic virtual machine image, and then replaced the provided root disk with my own. Since that site seems to be down these days, I’m linking to a copy the one I use here. I use the kvm-img utility to convert the disk.img to a format that is compatible with VMware and then use ovftool to bundle the vmx into a compressed ova archive. The ovftool utility is available as a download from VMware.

convert to vmdk:

kvm-img convert -O vmdk disk.img disk.vmdk

add to vmx:

mv disk.vmdk OpenFoam_Client/OpenFoam_Client.vmdk

convert to ovf:

ovftool --compress OpenFoam_Client/OpenFoam_Client.vmx boot.ova

And that’s it! If you think this is useful, or you tried to follow my instructions and had problems, please comment on this post.