The repart service is a oneshot, so we can't use a drop-in to replace
the `ExecStart` command that runs by default, which attempts to add
or grow defined partitions on the same device as the root filesystem.

This is OK for "unified" images, where the data partition is at the
end of the device, but not OK for "split" images, where it lives on a
different device.

Signed-off-by: Ben Cressey <[email protected]>

Switch from our `growpart` tool to `systemd-repart` to resize the
data partition. For a unified root+data image.

`growpart` uses the `gptman` crate, which calls the BLKRRPART ioctl
to tell the kernel to re-read the partition table. This call fails if
the device contains mounted partitions.

`systemd-repart` uses the newer BLKPG ioctl, which manipulates the
kernel's view of individual partitions. This works even if the root
filesystem is present on the same device and already mounted. It also
avoids the need to handle the partition symlink going away and coming
back, since udev does not get the change event that triggers this.

The two tools differ in how much free space is left on the device
after the last partition is resized. `growpart` ends the partition
one sector before the last 1 MiB boundary, while `systemd-repart`
ends it just before the GPT label. Both tools run on every boot.

To avoid problems on downgrade after a newer release resizes the data
filesystem beyond where the older release will end the partition, we
constrain `systemd-repart` to leave the older number of free sectors.

Since `/local` can be mounted during the resize operation, we can use
a real mount unit for it, which greatly simplifies the dependencies,
and allows us to decouple the "prepare" logic from "resize" logic.

Signed-off-by: Ben Cressey <[email protected]>

For some targets such as bare metal systems, the requirement for a
separate block device to hold the data partition is unworkable.

Implement a "unified" image layout, which places the data partition
after the final OS partition, and is suitable for targets which may
only have one disk.

The old "split" layout remains the default.

Signed-off-by: Ben Cressey <[email protected]>

Although the existing AWS and VMware variants use the "split" image
layout, custom variants for these platforms might use the "unified"
layout instead.

Adapt the AMI registration and OVA creation logic to account for the
possibility that we only build a single disk image.

Signed-off-by: Ben Cressey <[email protected]>

Using `state_t` as the label makes the directories read-only for all
unprivileged containers, even if they have access via a host mount.

Signed-off-by: Ben Cressey <[email protected]>