Support mounting dm-init file systems in the initrd

[dm0-]

Is your feature request related to a problem? Please describe.
When using device mapper with dm-mod.create, the kernel only creates /dev/dm-0--not /dev/mapper/$name--so systemd can't mount it due to device path canonicalization (and waiting on the .device unit).

Describe the solution you'd like
I want to be able to specify root=/dev/dm-0 dm-mod.create="..." and have the system successfully boot.

Describe alternatives you've considered
It works by adding a new service to the initrd that runs udevadm trigger to create the nodes, but is dm-mod.create something that could be supported out of the box?

[poettering]

Not grokking this... For dm related questions its best to get some DM people involved...

[dm0-]

Here is a script to set up an example, if it helps.

# Create a root file system image.
truncate --size=50M root.img
mke2fs -m 0 root.img
sudo mount -o loop,X-mount.mkdir root.img root
sudo mkdir root/bin
sudo cp -t root/bin /sbin/busybox
for cmd in ash echo poweroff sleep
do sudo ln -fns busybox "root/bin/$cmd"
done
sudo tee root/init << 'EOF' > /dev/null
#!/bin/ash -eu
echo 'Successfully booted into the image'
sleep 10
exec poweroff -f
EOF
sudo chmod 0755 root/init
sudo ln -fst root/bin ../init
sudo umount -d root

# Create a verity image.
declare -A verity
while read
do verity[${REPLY%%:*}]=${REPLY#*:}
done < <(veritysetup format root.img sig.img)
cat root.img sig.img > verity.img

# Write kernel arguments using dm-init to set up the root device.
echo > kargs.txt \
    'dm-mod.create="root,,,ro,0' $(( $(stat --format=%s root.img) / 512 )) \
    verity ${verity[Hash type]} /dev/sda /dev/sda \
    ${verity[Data block size]} ${verity[Hash block size]} \
    ${verity[Data blocks]} $(( ${verity[Data blocks]} + 1 )) \
    ${verity[Hash algorithm]} ${verity[Root hash]} \
    ${verity[Salt]} '0"' root=/dev/dm-0

So if you have a kernel with dm-init and dm-verity built into it, you can run qemu-kvm -m 1G -kernel vmlinuz -append "$(<kargs.txt)" verity.img to boot into the verity device. The problem is that if you include a systemd initrd with -initrd initrd.img to handle some other setup, it won't be able to mount the root device. (You can test this part using a kernel like Fedora's without the device mapper drivers as builtins; the modules will be loaded automatically if they are included in the initrd.)

The first issue is that systemd waits forever for dev-dm\x2d0.device to start. The second issue is that it fails to mount the device due to device path canonicalization and /dev/mapper/root not being created by dm-init (i.e. mount would work with the -c/--no-canonicalize option). I don't remember what I set up to get past the first issue but not the second, sorry. A call to udevadm trigger fixes both issues, so my question/request was if something could handle this without having to customize initrds.

[poettering]

it appears the default DM udev rules don't cover this stuff. please ping the DM folks about this, I doubt there's much to do from our side on this, we just pick up any block device marked to be ready.

[dm0-]

I e-mailed dm-devel a month ago which was unhelpful, so I'll just close this and continue to carry workarounds until I can look into doing something better.