arm64 Debian/Ubuntu port image available

[Wookey]

State of the Debian/Ubuntu arm64 port
=====================================

*** Arm64 lives! ***

Executive summary
-----------------

 * There is now a bootable (raring) image to download and run
 * Everything has been rebuilt against glibc 2.17 so it works
 * A bit more work is needed to make the rootfs useable as a native buildd
 * Multiarch crossbuilding and the build-profile mechanism is mature enough to cross-build 
    a port from scratch (this is a big deal IMHO)
 * All packages, sources and tools are in a public repo and this work should be reproducible.
 * This image is fully multiarched so co-installing armhf for a
    64/32 mix should work nicely, as should multiarch crossbuilding to
    legacy x86 architectures. :-) (but I haven't tried that yet...)


 * Linaro wants 'the distros' to take this work forward from here so people interested in 
    Debian and Ubuntu on 64-bit arm hardware need to step up and help out.


Bootable images
---------------

A milestone was reached this week: Enough packages were built for arm64 to debootstrap an
image which booted to a prompt! After a bit of fettling (and switching to multistrap) I got
an image with all the packages configured which boots with upstart to a login prompt (I
admit, I did get quite excited about this, as it represents the coming together of nearly 3
years work on multiarch, crossbuilding, bootstrapping, cyclic dependencies and arm64 :-)

The images are available for download:   http://wiki.debian.org/Arm64Port#Pre-built_Rootfs
And there are destructions there for making your own.

All these packages were cross-built on raring, untangling cyclic dependencies with build
profiles (see wiki.debian.org/DebianBootstrap for how that works), making this the first
(non x86) self-bootstrapped debian port ever (so far as I know). All (?) previous ports have
been done using something else like OpenEmbedded (armel, armhf), RedHat/HardHat (arm, alpha,
mips), something IBMy (s390) to get an initial linux rootfs on which debian packages are
built. 

The new bootstrap process is (almost) just a list of sbuild commands. In practice there are
still a few rough edges around cross- build-dependencies so of the 140 packages needed for
the bootstrap, 9 had to be built manually with 'dpkg-buildpackage -aarm64 -d' (to skip
build-dep checks) instead of 'sbuild --host arm64 <package>'. 

The current bootstrap packageset status is here: 
  http://people.linaro.org/~wookey/buildd/raring-arm64/status-bootstrap.html

There is no armv8 (arm64/aarch64) hardware available yet, so this image can currently only
be run in a model. ARM provide a free-beer prorietary 'Foundation model' so we do have
someting to test with. It's sluggish but perfectly useable. Booting the images takes a
couple of minutes on my fairly average machine.

The images are using the Linaro OE release kernels which seem to work fine for this purpose.
Thanks to Marcin for modified bootloader lines in .axf files. 



Image status
------------

I was impressed that things basically 'just worked' on first boot. There is of course plenty
of breakage, I'm sure, and I haven't looked very hard yet, but it's a lot better than I
expected after months of just building stuff and testing nothing. (Things that are poorly:
nano can't parse it's own syntax-coluring files for example, and multiarch perl has the
wrong @INC path compiled in; I'm sure there is more). Consider this alpha-grade until it's
been used a bit more.

Things that are not yet built which would make the images a lot more useful are apt and a
dhcp client. apt needs gnupg needs curl needs nss. The nss cross-build needs fixing to
unbung that. A debian chroot without apt turns out to be disappointing quite quickly :-)
Expect an updated image with more packages very soon.


Multiarch crossbuilding
-----------------------

It's really nice to have building and crossbuilding using exactly the same mechanisms
and tools, with all files having one canonical location, and dependency mechanisms that
are reliable. The more I've used this, the more I've been impressed by it. There is
still work to do to expand the set of cross-buildable stuff, but it's a solid base to
work from. 

Getting this port working has been 'interesting' because it's attempting 4 new things all at
once: multiarch (file layouts and dependencies), crossbuilding (tools and packaging support
in a distro that historically was always natively built), arm64 (aarch64) support in
packages that need it, and build-profiles to linearise the build-order. 

The arm64 part of this is a relatively small part as the heavy lifting has been done
upstream (gcc, (e)glibc, binutils, kernel, libffi, autotools and a lot of minor fixes in
various packages). Thanks are due to doko (Matthias Klose) for sterling work getting all
that integrated into the debian and ubuntu toolchain packages, and infinity (Adam Conrad)
for merging various eglibc branches. There were also hordes of very boring patches of the
form 'update config.sub and guess before building'.

Most of the work has been in making things cross-build (exactly the same fixes needed for
armel/hf too so I've had plenty of help there from canonical types who want cross-building
for arm to work nicely), and particular thinks to Neil Williams for taking on the perl
cross-build challenge and creating the debian-perl-cross package to manage the
cross-configury, whilst also working with upstream to make the whole thing a bit less 1996. 

Multiarchifying has been going on nicely in libraries and -dev packages, but things like
perl and python needed significant work, along with a lot of boring bugs saying 'mark this
package MA: foreign' and 'build-dep on python:any or perl-base:any'. Thanks are due to doko
for the python multiarching and Niko Tyni for the perl multiarchification. Getting all 3
'aspects' of multiarch perl, cross-built perl and arm64 perl config to work at the same time
was quite hard work, and there are still bugs there. Wider usage of multiarched perl would
no doubt sort this out reasonably quickly. I started a wiki page to track the status of
multiarched cross-buildable perl: http://wiki.debian.org/Multiarch/Perl . Help would be
welcome.

The build-profile work is described on the http://wiki.debian.org/DebianBootstrap page.
Progress has been greatly helped by GSOC projects last year, with good work on the tools
(crossbuild-essential packages, build-profile support) from P.J McDermott and an impressive
contribution from Johannes Schauer on dependency analysis tools around libdose, and apt
build-profile support.


All of this apart from multiarch perl, crossbuildable perl and build-profile stuff (and
a few pending patches) is already in raring.


Building stuff yourself
-----------------------

Setting up an arm64 build environment is very simple. Use sbuild-createchroot or mk-sbuild
and point at the bootstrap repo, with a bit of config and some updated tools packages from
the repo (amd64 only supplied). Details are given on
https://wiki.linaro.org/Platform/DevPlatform/CrossCompile/arm64bootstrap

Once you've created a tarball chroot builds are simply done with 
sbuild -c quantal-amd64-sbuild -d quantal --host=arm64 <package.dsc> or 
sbuild -c quantal-amd64-sbuild -d quantal --host=arm64 <package>_<version>  (I'd love it 
 if sbuild got smart enough to work out the version itself when given a distro - Roger
says he's working on it)

To deal with the chore of 'find version, run sbuild, sign result, upload to repo, import to
repo, deal with reprepro bitching if you re-upload the same version of something' for every
package build, I wrote 'dimstrap' which is a simple-minded tool to wrap that up and either do
one-off builds or run through a list. It is part of the xbuilder package here:
https://launchpad.net/~linaro-foundations/+archive/cross-build-tools/ It also includes the
logfile-parsing script ('generate html') which generates the nice status pages:
http://people.linaro.org/~wookey/buildd/raring-arm64/status-bootstrap.html 


Image building
--------------

The config and instructions provided (in
http://wiki.debian.org/Arm64Port#Building_your_own_rootfs_image ) is
for multistrap. Debootstrap sort-of produces working images too but
takes a lot longer to unpack/configure, and misses out various vital
packages (like libperl5.14). I'm sure it could be kicked into
submission. In theory multistrap (apt really) should have got all the
arch all packages from the main repo, but in practice it refused to do
that so I had to rebuild them or copy them over anyway (grumble). 

Any package that installs replaced conffiles seems to generate invalid
dpkg status entries (ifupdown did this to me). I've not got to the
bottom of that yet. Deleting the offending line gets you an image that
works. 

Issues
------

General:

The build-profile patches for dpkg and apt need to be pushed into the distro to make
that feature permanent. A thread on debian-devel is working on that
(http://debian.2.n7.nabble.com/Bootstrappable-Debian-proposal-of-needed-changes-td2848200.html).
The main issue is what syntax to use '<>' or '[]' and how to deal with multiple overlapping
profiles. The patches to debian control cannot go in until at least the syntax is agreed and
the tools will parse them without barfing. Johannes ands I will send an updated spec
soonish. 

The missing piece of bootstrapping with regard to build-deps is packages that build-dep on
gcc-4.6 or binutils. When cross-building this should be satisfied by <triplet>-gcc-4.6 or
<triplet>-binutils. Nothing makes that happen currently. A scheme has been mooted but
nothing is implemented yet.

There is debate about whether cross-toolchains should build against multiarch libraries
(libgcc, libstdc++) like everything else, or have their own internal copies. Doko and I
disagree on this matter. That will need to be worked out at some point.

We won't get that much further with fixing cross- object-introspection, which is a
non-trivial job. 

Image-related:

The images do essentially work but very little has been tested so far. 

Multiarch perl still needs work.

nss needs cross-building in order to get apt cross-built

I've not got networking working yet. Info is here:
https://fedoraproject.org/wiki/Architectures/ARM/AArch64/FoundationModel_Networking
lack of a dhcp client in the image hasn't helped there. 


More info 
---------

The canonical arm64 port info page is: 
http://wiki.debian.org/Arm64Port

Full arm64 cross-build status (i.e everything that has been tried) is here:
http://people.linaro.org/~wookey/buildd/raring-arm64/status.html

All the patches generated so far are here:
http://people.debian.org/~wookey/bootstrap/patches/

(most that can, have been filed as bugs - there is a backlog of stuff
filed in Launchpad but not yet forwarded to the Debian BTS - yes I am
a bad boy - blame the fact that you can't use reportbug or bts from
inside ARM due to their idiotic email policies). 


Future work
-----------

Firstly we should say thank you to Linaro for sponsoring this work in various ways over
the last 3 years. We wouldn't be at this point now if it wasn't for that. However
Linaro has a lot of things to do and is trying hard not to do distro's work for them,
concentrating on upstream things. This makes sense for commercially-backed distros like
Red Hat and Ubuntu, but rather less for Debian where we _are_ the distro just as much
as anyone else is, and ultimately someone has to spend the time to get stuff working.

Anyway, I was supposed to stop work on this some time back, but have largely failed to
do so (cross-building is so moreish - there is always one more build to try before
bedtime!) and appreciate being given enough slack to get this to a point of actual
utility. However I expect to have much less time to spend on this from now on, except
insofar as it still co-oncides with things Linaro wants doing. I'd love to hear from
people who actually want to use this, to get more packages built, the Debian
cross-toolchains sorted, build-profiles finalised, and a whole pile of stuff fixed once
Wheezy is released. I'm pretty sure there are quite a lot of people who want multiarch
Debian or Ubuntu on their arm64 machines (or models). 

I hear rumours that actual hardware may appear sometime around the middle of the year
with some bagsied for Debian. Setting up the ports infrastructure for that would be
good. I don't know if anyone is interested in building slowly on models in the
meantime, or if we should just carry on crossing and see how far we get. This table
shows that 471 packages in raring can be expected to cross-build already:
http://people.canonical.com/~cjwatson/cross/armhf/raring/

Todo:

Fix up multiarch/cross perl
Fix nss
Build missing packages for apt
Build missing packages for build-essential
Build Debian cross-toolchain
Get all this working in unstable as well as raring
Setup buildds
Build all the other packages
Set up automated bootstraping runs (eventually) 


Current setup
-------------

Builds have all been run locally using the sbuild/chroot setup described above and on
the Arm64Port page, which should be easy for anyone to reproduce. The main irritation
is keeping up with raring: out of sync libraries are not MA-installable.  Logs are
uploaded to people.linaro.org (rsync). The reprepro repo is on people.debian.org
(dupload).  This stuff should probably move to ports.debian.org and ports.ubuntu.com,
but neither of those are set up for cross-building so I'm not quite sure how this will
work.

I could go on at great length about the machinery of profiled bootstrap builds, and
interactions between tools, but it's not very exciting, so will resist. Suffice it to
say that whilst it's all pretty slick I'd still like better buildd tools. 


Build-profile changes
---------------------

The build-profile patches are not yet upstreamable so are collecting in the repo. 
The patch set so far is here: http://people.debian.org/~wookey/bootstrap/patches/profiles/packages/


Other thanks:
Other people who have helped make this happen in various ways but not got a mention above:
Colin Watson, Dmitry Ledkovs, Steve Langasek, Harry Leibel, Thibaut Girka, Roger Leigh,
Marcus Shawcroft, James Morrisey, Jonathan Austin, Steve McIntyre, Peter Pearse, Aurelien
Jarno, and whoever does sysadmin at people.{linaro,debian}.org

I hope I didn't forget anyone, or any important information.

Feedback from anyone attempting to get this working outside my computer is very
welcome. I have almost certainly forgotten to write down some things, and upload
correct versions of some other things. 


Wookey
-- 
Principal hats:  Linaro, Emdebian, Wookware, Balloonboard, ARM
http://wookware.org/