FAI for ARM image building

[Riku Voipio]

Hi Thomas,

On 12 October 2017 at 17:59, Thomas Lange <lange at informatik.uni-koeln.de> wrote:
> Hi Riku,

> now I had a look at your linaro fai config space. Here are my comments:

Thanks for your detailed review. I'll take some time to make them more
idiomatic following your suggestions.

> First I tried to call the fai-diskimage command using your config
> space. It failed when calling debootstrap :-(
> So I found a little bug in FAI 5.4 and fixed it :-)
> https://github.com/faiproject/fai/commit/efbb5a8448b6b04d9f3be74e32a222b2c8e6a09a
>
> class/INSTALL.var is not needed for fai-diskimage. This is only used
> when you use the curses menu which reads the class/*.profile files.
> The other files in class/ look good.
>
> I wonder why you use a fixed size (4GB) in disk_config/RAW. I would
> say "4GB-" which does not define an upper bound and is therefore more
> general.

Some of the images will got < 4GB internal eMMC storage, so we want to
make sure all images fit.

> The files in package_config are well partitioned into classes.
>
> Instead of files/etc/hostname/DEBIAN and the preinst, I would just use
> this code: echo $HOSTNAME > $target/etc/hostname
>
> Thanks for showing me the envsubst command in several preinst files. I
> didn't knew this one. It's easier than what's currently used in the FAI
> preinst examples.

Neither did I know about it until recently. 20+ years of using Linux
and still finding new useful commands.

> The scripts DB820C/21-customize and DB410C/21-customize share some
> lines of code. Keep in mind, that duplicating code is bad, and think
> about putting the duplicated code into a more general FAI class, which
> is then used everytime you use the DB820C and DB410C class.

We are still developing these. Like premature optimization, premature
abstraction is bad ;) We are adding at least hikey and hikey960
support. With all HW support in, we will see more clearly howto split
the classes.

> The wget of the firmware zip file will put and extract the files into
> the current working directory as root. I would use /tmp for that.
>
> In several scripts you use "chroot $target" which should be replaced
> just by $ROOTCMD, which is the same (except if you call fai
> softupdate, which is then empty).

Oh, cool, didn't notice $ROOTCMD.

> For things like "echo string >> file" you could use ainsl(1).
>
> In 22-disable-systemd-services.chroot I wonder why you use an array
> instead of just a space separated string like this:
> services="NetworkManager systemd-networkd"
> for service in ${services}; do
>   $ROOTCMD systemctl mask ${service}-wait-online.service
> done
>
> Silimar is valid for 02-add_linaro_to_groups.chroot where you redefine
> IFS instead of separating the list of DEFGROUP just by spaces.

> The hook debconf.SAVECACHE is not needed IMO. If there's a basefile
> matching a class name in the subdirectory basefiles/, then FAI will
> automatically use it and will not call debootstrap. If you define
> something linke this

> export FAI_BASEFILEURL=https://fai-project.org/download/basefiles/
> the first time you call fai-diskimage, FAI will download a basefile
> (matching a class name) into the subdir and the second time you call
> fai-diskimage FAI will just use this file automatically.

This hook is a hack to *create* the basefile. The documentation
suggests using pbuilder to create the basefile, and the basefile url
has mk-basefile. The problem is that neither of these ways is
identical to what fai-diskimage creates on the first run.

> My overall impression is very good! All the things I said are just
> minor cosmetic details.
>
>
> Finally I've created a SID_ARM64.tar.xz basefile and ran fai-diskimage
> with the new cross-arch support using your config space on a Amd64
> machine which finished in 289 seconds. I used FAI 5.4 with the one
> patch I mentioned above. How long does it take on a native Arm64
> machine?

7 minutes without basefile, and 1.5min with. This is on a HPE Moonshot
m400 blade.

Riku