mkimage-raw-efi references lower-level serial container dir directly and shouldn't

[deitch]

The pkg mkimage-raw-efi references the specific path of an onboot container:

# Before changing something here please take a look into the
# images/rootfs.yml.in onboot section: the installer should
# precede the storage-init container.
#
# now lets create an edge container
# FIXME: 002-installer? why not linuxkit build?
WORKDIR /eco/media/root-rw/root/containers/onboot/002-installer
COPY runtime.json config.json ./
RUN mv /out rootfs
# hadolint ignore=DL3003
RUN (cd /eco && find . -xdev | grep -v installer.img | sort | cpio --quiet -o -H newc) | gzip > rootfs/installer.img
RUN mv rootfs /out

This even has a FIXME, and comments in the main rootfs.yml.im to change that if you add anything that changes the onboot order.

Similar code exists in mkverification-raw-efi.

Why is this necessary? What can we do to make this cleaner, as it is messy and almost guaranteed to break in odd ways in the future.

@rouming do you know what this does, i.e. its intended behaviour? You have the commit on mkverification-raw-efi and on mkimage-raw-efi.

Maybe there is nothing better, but we should try.

[rene]

@deitch , AFAIU this is done is order to "inject" the installer container only in the installer images.... we could use some .yq image modifier file for that, but that requires some changes to support not only PLATFORM variants, but also different image variants (installer, live, etc...)...

[deitch]

Yeah, if we can do it at a higher level, that would be cleaner. The whole pipeline is about, "I build a rootfs.tar, then I convert its contents to ext4 or squashfs or iso9660 or whatever." This sounds like it is a different rootfs.tar content, which we manipulate using the yml inputs.

Did I misunderstand?

Conversely, we could have it always there, but enable/disable it via some filesystem flag, like files:, but do we always want it there? Probably not?

The number of variants of eve is exploding!

[deitch]

@rene do you mind explaining this a bit more? How does this work? I see that mkimage-raw-efi - which we use whenever we build an EVE image - goes to one of the onboot containers path at /eco/media/root-rw/root/containers/onboot/002-installer, modifies the container config with a custom runtime.json and config.json, and then generates an installer.img inside /eco/rootfs?

Is it that this section of the Dockerfile is basically an "aside"? Or a stage where we put everything to the side, generate an installer.img, and then move everything back?

If we removed these lines of code, would we then have everything normal except would be missing an installer?

If so, the part I really don't get is, wouldn't that be part of the EVE docker image, the one we can run with docker run lfedge/eve <cmd>? What is it doing as part of mkimage-raw-efi? Or is this run at device installation time? But that cannot be, because you need the EFI disk image before you can boot?