I’m looking at the simple init script which might be run on boot in a VM (the relevant alternative being the full_cow_setup script, but the differences between the code that I am looking at are superficial aside from the udev setup which occurs at the very end of qubes_cow_setup.sh and does not appear in simple.sh). In particular, this snippet is confusing when I look at the actual state of my running VMs:
if [ `cat /sys/class/block/$ROOT_DEV/ro` = 1 ] ; then
echo "Qubes: Doing COW setup for AppVM..."
while ! [ -e /dev/xvdc ]; do sleep 0.1; done
VOLATILE_SIZE_512B=$(cat /sys/class/block/xvdc/size)
if [ $VOLATILE_SIZE_512B -lt $SWAP_SIZE_512B ]; then
die "volatile.img smaller than $SWAP_SIZE_GiB GiB, cannot continue"
fi
/sbin/sfdisk -q /dev/xvdc >/dev/null <<EOF
xvdc1: type=82,start=1MiB,size=${SWAP_SIZE_GiB}GiB
xvdc2: type=83
EOF
if [ $? -ne 0 ]; then
echo "Qubes: failed to setup partitions on volatile device"
exit 1
fi
while ! [ -e /dev/xvdc1 ]; do sleep 0.1; done
/sbin/mkswap /dev/xvdc1
while ! [ -e /dev/xvdc2 ]; do sleep 0.1; done
echo "0 `cat /sys/class/block/$ROOT_DEV/size` snapshot /dev/$ROOT_DEV /dev/xvdc2 N 16" | \
/sbin/dmsetup create dmroot || { echo "Qubes: FATAL: cannot create dmroot!"; exit 1; }
/sbin/dmsetup mknodes dmroot
echo Qubes: done.
else
echo "Qubes: Doing R/W setup for TemplateVM..."
while ! [ -e /dev/xvdc ]; do sleep 0.1; done
/sbin/sfdisk -q /dev/xvdc >/dev/null <<EOF
xvdc1: type=82,start=1MiB,size=${SWAP_SIZE_GiB}GiB
xvdc3: type=83
EOF
if [ $? -ne 0 ]; then
die "Qubes: failed to setup partitions on volatile device"
fi
while ! [ -e /dev/xvdc1 ]; do sleep 0.1; done
/sbin/mkswap /dev/xvdc1
ln -s ../$ROOT_DEV /dev/mapper/dmroot
echo Qubes: done.
fi
Based on the “comments” (echo mesages at the top of each conditional block) I would expect that an AppVM would have a /dev/xvdc2 and that this is related to the mechanism that allows root to be writable without writing back to the template VM or copying every file on boot. Basically, I would expect that /dev/xvdc2 contains the changed files which are used instead of the original filesystem when they exist. I would also expect that template VMs have a /dev/xvdc3 and the symlink is set up only so that other modules don’t have to care about whether they are in a template VM or AppVM - they can just use /dev/mapper/dmroot either way.
The above are all guesses based on my understanding of how QubesOS operates and what would make sense based on the contents of the script, but I have not used dmsetup in the past so I might be missing some things that are obvious to more experienced people.
When I look at my AppVMs though (and I am only looking at ones based on Fedora templates so that things look as “normal” as possible), they have a /dev/xvdc3 and /dev/mapper/dmroot is symlinked to it, which happens in the code block noted as being for template VMs.
What am I missing?