Preface
This article documents the trial-and-error process of me trying to
cross-compile the Ubuntu kernel for arm64
for use with Enzians at the
Systems Group at ETH. So far building natively on the Enzians work (see
native build below), but
unfortunately none of the cross-compile approaches produce reliable .deb
packages. I have submitted a couple bug reports and will (hopefully) update
accordingly when things change.
With that said, this whole process has been quite a journey: I learned a lot about the Debian and Ubuntu packaging infrastructure. This should serve as a good layman's guide for distribution kernel hacking (at least for Ubuntu). It also gives good starting knowledge for creating my own, custom Debian packages.
Motivation
Why custom kernel?
The core idea is very simple: when doing performance measurements for a
user-space application, one would want as little disruption as possible. This
is traditionally handled by pinning a task to a specific CPU core, to avoid
the scheduler moving it willy-nilly across cores. Say we have our Design Under
Test pionic-test
(from my PIO NIC project and we want to run it on core
#47 (last CPU core on Enzians):
$ sudo taskset -c 47 ./pionic-test
However, doing so isn't nearly enough: we also need to ensure that there are no other tasks, both user-space and kernel space (kthreads), running on core #47. Failing to evacuate these tasks will still result in the scheduler stepping in, descheduling us, and scheduling these other tasks. There are also other concerns, like device IRQs, timer interrupts (cough cough), RCU callbacks, bla bla. Most of these can be moved off the core with some kernel command-line options:
isolcpus=nohz,domain,managed_irq,47 rcu_nocbs=47 irqaffinity=0-46 kthread_cpus=0-46 rcu_nocb_poll
Many guides on the Internet detail on how to tune the kernel for real-time applications (here's one from the Ubuntu blog, for example). I won't go into the details of what each of these options mean since it's not related to this article.
However, no matter how quiet you try to make the kernel to be, the stock,
-generic
kernel has its limits. The default Ubuntu -generic
kernel is
tickful (I coined this as opposed to the more commonly used word tickless
-- you'll see in a minute). This means that even if absolutely nothing other
than the user task is running on the core, and that you moved all irrelevant
IRQ processing etc. off the core, there still would be a constant timer
interrupt (usually at 250 Hz; defined via the HZ
config option) going off
on that core. Measurements on the Enzian show that handling this interrupt
takes around 10 us: not a lot of time, but enough to appear as a large tail
latency if your application latency is sub-microsecond.
The kernel is capable to be completely tickless since 3.10, through the
compile-time config CONFIG_NO_HZ_FULL
. Enabling this option, along with
specifying nohz_full=47
on the kernel command-line, will finally make this
250 Hz timer interrupt go away. However, the feature still has some rough
edges, with potential performance impacts for specific workloads (since it
changes how time-keeping works), it is not widely deployed. It certainly isn't
available in the Ubuntu archive for us to directly install.
Why Ubuntu (vendor) kernel?
Hopefully that explains why I want a custom kernel. But why do I have to use the Ubuntu kernel? Well, yeah it's true that I can use the upstream tree directly (it's the method for a custom kernel documented on the Enzian wiki). However, there are two catches:
- the stock Ubuntu kernel has a lot of patches; using the upstream tree would mean deviating a lot from the vanilla setup (which is just stock Ubuntu)
- installing a upstream kernel is unhygenic: the installed kernel images
and modules won't be tracked by
dpkg
, making it a pain to uninstall, reinstall, or package and redistribute - @Harry_Chen mentioned that the upstream Makefiles are capable of
producing
.deb
packages; however due to the fear that they might not conform to Ubuntu conventions, I didn't explore in this direction
The distribution/vendor (Ubuntu) makes a nice kernel package that we should be able to easily customize (cough cough) and rebuild. This way, we can make sure that there's as little divergence as possible from the vanilla Ubuntu kernel, for the sake of clean experiment results.
Why cross-compile?
It stands that native-compiling is always easier than cross-compiling: no need to carefully differentiate if you're compiling something that will execute on the host machine (i.e. the machine you're building for), or a tool that you need during the build, for the build machine (i.e. the machine you're building on). This is especially tricky when dependencies are involved, since binaries for wrong architecture won't execute (easily).
With that said, the old benefits of cross-compiling apply. The Enzian has a rather old Cavium ThunderX-1 CPU that is not particularly fast. In addition, all machines are booted through iSCSI (5 GB image) with a NFS scratch space, so storage is very limited (either slow or small). It would be great if we can cross-compile.
Create flavour and build natively on Enzian
"Shallow-fork" the vendor tree
We start with the focal kernel Git tree to add our changes. The full kernel tree history is huge, so we would do a shallow clone:
$ git clone https://git.launchpad.net/\~ubuntu-kernel/ubuntu/+source/linux/+git/focal/ linux-focal -b Ubuntu-5.4.0-200.220 --depth=1
However, GitLab doesn't support pushing shallow clones of repositories. We
need a bit of black magic, namely with the power of git replace
documented
in the Git book. The idea is that we create a brand-new tree based on the
shallow HEAD
with the commit-tree
command; root of this tree can later be
replaced with the full history if we need it. I duplicate the exact commands
here for future reference:
$ git commit-tree Ubuntu-5.4.0-200.220^{tree} << EOF
> Forked from Ubuntu-5.4.0-200.220 in focal
>
> Get full history from upstream focal tree at
> https://git.launchpad.net/~ubuntu-kernel/ubuntu/+source/linux/+git/focal/ ;
> Instructions at https://git-scm.com/book/en/v2/Git-Tools-Replace
> EOF
8adb6ccde3615be4787b1583d4cb231d123d0fd1
$ git checkout -b 5.4-tickless 8adb6ccde3615be4787b1583d4cb231d123d0fd1
...
$ git push -u gitlab <department gitlab url>
As described by a StackOverflow answer:
Say that I want to effectively “squash” all the commits in the repo down to a single commit. But “squash” is a rebase operation which is excessive since it combines changes, and I don’t need to combine any changes since I already know what snapshot I want—the current one.
Set up kernel flavour
We need to decide where to put our tickless modifications:
- either, as a local update to the
-generic
kernel- this would result in the same package name
linux-image-unsigned-5.4.0-200-generic
with a version number like5.4.0-200.220+pengxu~tickless
- this would result in the same package name
- or, as a new flavour alongside the
-generic
flavour- this would result in a new package
linux-image-unsigned-5.4.0-200-tickless
, with a version number like5.4.0-200.220+pengxu
- this would result in a new package
@shankerwangmiao suggests that the second one is the way to go: the first approach is abusing the version number and will result in the kernel being replaced by a newer major version update.
I adopted the new flavour approach. The necessary changes are recorded in the Git repo of my fork, but here's a brief list of the changes needed:
- create
debian.master/control.d/vars.tickless
to declare metadata of the flavour (used to generate the package control file) - declare our
tickless
flavour indebian.master/rules.d/arm64.mk
- make sure that your flavour defines
build_image
andkernel_file
!
- make sure that your flavour defines
- update configs
fakeroot debian/rules clean
to generate the necessary rule filesfakeroot debian/rules editconfigs
to launchmenuconfig
and updatedebian.master/config/annotations
- Ubuntu deviates from Debian and created this annotation format, caused me a lot of headache
- update
changelog
- use the
dch
helper:DEBEMAIL='Pengcheng Xu <email>' dch -c debian.master/changelog -l+pengxu
- use the
debian/rules
will require some tools to execute correctly; refer to the
official guide to install them. The new -tickless
kernel flavour should
now be ready. To see if everything is set up correctly, commit, clean the
tree, and build the kernel natively:
$ git add debian.master && git commit -m '...'
$ git clean -dfx # to get rid of all generated files
$ fakeroot debian/rules clean
$ fakeroot debian/rules binary-headers binary-tickless binary-perarch
The above should build the kernel, headers, and modules. As @shankerwangmiao
and @Harry_Chen pointed out and also stated in the Debian maintainer's
guide, however, it's not recommended to directly call the rules. The
correct way is to use dpkg-buildpackage
, to either do a full rebuild, or run
a couple targets:
# full rebuild
dpkg-buildpackage -us -uc -B
# run some targets in the rule; theoretically equivalent to invoking debian/rules above
dpkg-buildpackage -us -uc -d -T 'binary-headers,binary-tickless,binary-perarch' --as-root
It's recommended to rebuild everything, instead of running only selected
target. The reason is that, it's easy to produce an uninstallable subset of
packages by missing a target. This happened when I forgot to call
binary-perarch
: binary-tickless
produced a flavour-dependent
linux-tools-5.4.0-200-tickless
, that depends on the per-architecture (i.e.
flavour-independent) linux-tools-5.4.0-200
. Since the 5.4.0-200
packages
are not yet available on the archive, this resulted in the flavoured
linux-tools
being uninstallable.
Check the outer directory for the resulting .deb
packages, to be installed on
Enzian:
$ ls -sh ../*.deb
1.3M ../linux-buildinfo-5.4.0-200-tickless_5.4.0-200.220+pengxu_arm64.deb
1.8M ../linux-headers-5.4.0-200-tickless_5.4.0-200.220+pengxu_arm64.deb
12M ../linux-headers-5.4.0-200_5.4.0-200.220+pengxu_all.deb
13M ../linux-image-unsigned-5.4.0-200-tickless_5.4.0-200.220+pengxu_arm64.deb
56M ../linux-modules-5.4.0-200-tickless_5.4.0-200.220+pengxu_arm64.deb
900K ../linux-tools-5.4.0-200-tickless_5.4.0-200.220+pengxu_arm64.deb
5.1M ../linux-tools-5.4.0-200_5.4.0-200.220+pengxu_arm64.deb
I tested these packages on the Enzian -- they seem to work just fine.
Cross compile
Now that we can build our kernel flavour natively, we can start attempting to
cross-compile it. I have to reiterate here: none of the following actually
worked (i.e. produced packages that could work)! Even though sbuild
managed to produce .deb
packages in the end, I don't think they are built
correctly and didn't bother to test the resulting packages -- something is
fundamentally wrong with the build. Don't say that I told you it would work :)
With that said, I'm fairly confident that I am on the right track. Provided
that the Ubuntu kernel team rework the dependency declarations, there's quite
good chance that sbuild
will be able to produce correct packages.
Attempt #1: invoking the rule with arch=arm64
I bumped into this by searching for arch
in all the Makefiles inside
debian/
-- documentation was really scarce on how to cross compile the Ubuntu
kernel after all. Quoting from debian/rules.d/0-common-vars.mk
:
#
# Detect invocations of the form 'fakeroot debian/rules binary arch=armhf'
# within an x86'en schroot. This only gets you part of the way since the
# packaging phase fails, but you can at least compile the kernel quickly.
#
arch := $(DEB_HOST_ARCH)
ifneq ($(arch),$(DEB_HOST_ARCH))
CROSS_COMPILE ?= $(shell dpkg-architecture -a$(arch) -qDEB_HOST_GNU_TYPE -f 2>/dev/null)-
endif
At this point I understood neither what an schroot
is, nor how the build
would fail. So I ran the following:
$ fakeroot debian/rules clean
$ fakeroot debian/rules binary-headers binary-tickless binary-perarch arch=arm64
The kernel did seem to compile -- I saw lots of familiar CC [M] ....
lines
roll by. However, I was hit too many times with missing dependencies during
the build, including this very cryptic one about Python:
checking for python3... python3
checking for python version... 3.1
checking for python platform... linux
checking for python script directory... ${prefix}/local/lib/python3.10/dist-packages
checking for python extension module directory... ${exec_prefix}/local/lib/python3.10/dist-packages
configure: error: "Python >= 3.4 is required"
Building module:
cleaning build area...(bad exit status: 2)
make -j12 KERNELRELEASE=5.4.0-200-tickless...(bad exit status: 2)
Error! Bad return status for module build on kernel: 5.4.0-200-tickless (x86_64)
Consult /local/home/pengxu/work-local/bare-test/bare-rules-build/debian/build/build-tickless/___________dkms/build/zfs/0.8.3/build/make.log for more information.
DKMS make.log for zfs-0.8.3 for kernel 5.4.0-200-tickless (x86_64)
Sat Oct 19 08:30:07 CEST 2024
make[1]: Entering directory '<<DKMSDIR>>/build/zfs/0.8.3/build'
make[1]: *** No targets specified and no makefile found. Stop.
make[1]: Leaving directory '<<DKMSDIR>>/build/zfs/0.8.3/build'
make: *** [debian/rules.d/2-binary-arch.mk:234: install-tickless] Error 1
Apparently the build rules didn't recognize that Python 3.10 is newer than 3.4;
I just gave up at this point. It didn't come as a surprise is that the rule
didn't generate useful *.deb
kernel packages.
In hindsight this looked awfully apparent that I had to use a chroot
, such
that we have all the dependencies fixed to the version that the rule expects.
In fact the above error is differnet from the one I was looking at back then --
dpkg-deb
complained about all packages being skipped, so I thought maybe not
all required dpkg
environment variables for cross-compiling were set
correctly. Together with the comment from @shankerwangmiao that usually
dpkg-buildpackage -aarm64
"just works" automagically, I went on with my
second attempt.
Attempt #2: dpkg-buildpackage
The recommendation of using dpkg-buildpackage
is definitely not unfounded.
Quoting from debian/rules.d/0-common-vars.mk
again:
#
# Detect invocations of the form 'dpkg-buildpackage -B -aarmhf' within
# an x86'en schroot. This is the only way to build all of the packages
# (except for tools).
#
ifneq ($(DEB_BUILD_GNU_TYPE),$(DEB_HOST_GNU_TYPE))
CROSS_COMPILE ?= $(DEB_HOST_GNU_TYPE)-
endif
So we attempt to run the same command as on Enzian, with the addition of
-aarm64
to specify cross-compiling:
$ dpkg-buildpackage -uc -us -d -aarm64 -T 'binary-headers,binary-tickless,binary-perarch' --as-root
This time we do get some .deb
packages, namely the per-arch
linux-headers
, flavoured linux-headers-tickless
, linux-image-unsigned
,
and linux-modules-tickless
. However, both the per-arch and flavoured
linux-tools
are missing, meaning that some packages didn't build. Even
worse, the built packages have bad dependency versions:
$ dpkg -I ../linux-headers-5.4.0-200-tickless_5.4.0-200.220+pengxu_arm64.deb
[...]
Package: linux-headers-5.4.0-200-tickless
Source: linux
Version: 5.4.0-200.220+pengxu
Architecture: arm64
[...]
Depends: linux-headers-5.4.0-200, libc6 (>= 2.34), libssl3 (>= 3.0.0~~alpha1)
[...]
Ubuntu focal (20.04 LTS) doesn't have libc6 (>= 2.34)
at all! The correct
dependencies, from packages built natively on the Enzian, should be:
dpkg -I debs/linux-headers-5.4.0-200-tickless_5.4.0-200.220+pengxu_arm64.deb
[...]
Package: linux-headers-5.4.0-200-tickless
Source: linux
Version: 5.4.0-200.220+pengxu
Architecture: arm64
[...]
Depends: linux-headers-5.4.0-200, libc6 (>= 2.17), libssl1.1 (>= 1.1.0)
[...]
Turns out my build machine is running jammy (22.04 LTS) instead of 20.04, and
dh_shlibdeps
captured the wrong, newer symbol versions, thus generating the
newer libc6
and libssl
dependencies. This was pointed out by @Harry_Chen
-- now it is really apparent that I have to use a chroot
for building.
Attempt #3: set up sbuild and a focal chroot
Debian offers pbuilder and sbuild for building packages in a clean fashion.
They are said to have very similar feature sets, but it seems like sbuild
supports a pure rootless operation through unshare
using mmdebstrap
.
This will be handy if I want to use the build farm from the Enzian project
(where I don't have root access), so I went with sbuild.
The steps to set up sbuild with mmdebstrap
is mostly already documented on
the Debian wiki page, so I'll just note down the catches that I wasted
some time on:
mmdebstrap
shipped with Ubuntu 22.04 checks for groupname instead of username in/etc/subgid
- ETH LDAP have different user and group name (
pengxu
vspengxu-group
) - fixed by upstream but not included in 22.04 (0.8.4-1ubuntu0.1)
- temporary fix: list both
pengxu
andpengxu-group
in/etc/subgid
to satisfymmdebstrap
; I've opened a bug for Ubuntu
- ETH LDAP have different user and group name (
- the main Ubuntu archive (
archive.ubuntu.com/ubuntu
) does not includearm64
arm64
is in the ports archive (ports.ubuntu.com/ubuntu-ports
)- fix by supplying the following custom
sources.list
tommdebstrap
# ~/my-sources.list
deb [arch=amd64] http://archive.ubuntu.com/ubuntu/ focal main restricted universe multiverse
deb [arch=amd64] http://archive.ubuntu.com/ubuntu/ focal-updates main restricted universe multiverse
deb [arch=amd64] http://archive.ubuntu.com/ubuntu/ focal-backports main restricted universe multiverse
deb [arch=amd64] http://security.ubuntu.com/ubuntu/ focal-security main restricted universe multiverse
deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ focal main restricted universe multiverse
deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ focal-updates main restricted universe multiverse
deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ focal-backports main restricted universe multiverse
deb [arch=arm64] http://ports.ubuntu.com/ubuntu-ports/ focal-security main restricted universe multiverse
- do not use the
buildd
profile:apt-get
is missing yetmmdebstrap
depends on it sbuild
on Ubuntu 22.04 doesn't recognisetar.zst
-compressed tarballs, despite the Debian wiki recommending it- Using lz4 since it's the fastest according to this benchmark
Final commands:
$ sudo apt install lz4 mmdebstrap uidmap sbuild
$ echo "pengxu:1000000:65536\npengxu-group:1000000:65536" | sudo tee -a /etc/subgid
$ mkdir -p ~/.cache/sbuild
$ mmdebstrap --arch="amd64 arm64" focal ~/.cache/sbuild/focal-amd64.tar.lz4 ~/my-sources.list
$ sbuild-update --chroot-mode=unshare -udcar # test access by sbuild
We should now be able to do sbuild --host=arm64 -d focal
to cross-compile the
entire kernel package... only if would be that easy :) This time the build
process fails really early, when compiling the scripts/sign-file
tool shipped
by the Linux kernel upstream. It turns out that the Ubuntu kernel mixes build
and host dependencies without marking them correctly with :native
(as
described in [the Debian cross-compile guidelines]). The Debian kernel
(correctly) defines Build-Depends
as follows:
Section: kernel
[...]
Build-Depends:
[...]
# used by upstream to build signing tools and to process certificates
libssl-dev:native <!pkg.linux.nokernel>,
libssl-dev <!pkg.linux.notools>,
You can see the correct way here is to depend on the build-native
libssl-dev
for building the sign-file
tool, unless the pkg.linux.nokernel
profile is selected (meaning that we are not building the kernel after all).
We would depend on the host-native libssl-dev
when we are building
linux-tools
, unless the pkg.linux.notools
profile is selected. In
comparison, Ubuntu's kernel rules does something way cruder:
[...]
Build-Depends:
[...]
libssl-dev <!stage1>,
We can see that Ubuntu didn't differentiate between build and host at
all when declaring build dependencies. The same issue is also present for
other tools like pahole
-- they fail to execute on the build machine due to
the incorrect version being installed. Those failure didn't fail the entire
build, meaning that a small patchwork for libssl-dev
allowed the build to
actually produce .deb
packages. However, I really wouldn't trust the
packages compiled this way.
I asked on #kernel:ubuntu.com
on Matrix about if this lack of awareness of
cross-compiling is intended. The answer from Timo Aaltonen
(@tjaalton:ubuntu.com
), one of Ubuntu's core devs, is no:
Timo: ack. btw, do file a bug against the kernel so it's not forgotten. I don't think there's a reason why this is like it is. We might just as well do what debian does here
So I've submitted a bug for this.
Emulated native build with qemu-user-static
and binfmt_misc
My conversation with Timo actually revealed the fact that a separate route of building the kernel on a beefy amd64 machine being possible:
Timo: then again I created a schroot for armhf, so you're using the amd64 chroot?
Me: ah yes, I have a focal-amd64. If you have a schroot for armhf, does that mean you're using qemu-user with binfmt?
Timo: probably
An armhf
schroot on an amd64
build machine is essentially what I would call
emulated native build -- execution of armhf
binaries is emulated through
qemu-user-static
(static since we are inside a chroot). mmdebstrap
has
support for this pattern. Install necessary dependencies and set up chroot
$ sudo apt install binfmt-support qemu-user-static
$ mmdebstrap --arch=arm64 focal ~/.cache/sbuild/focal-arm64.tar.lz4
After the chroot is set up, we can then request sbuild to build with both
build and host set to arm64 via --arch=arm64
:
$ sbuild --arch=arm64 -d focal
However, it seems like the kernel rule lists several old Build-Depends
packages that are not in the focal archive any more:
The following packages have unmet dependencies:
sbuild-build-depends-main-dummy : Depends: dh-systemd but it is not installable
Depends: dwarves but it is not installable
Depends: xmlto but it is not installable
Depends: docbook-utils but it is not installable
Depends: fig2dev but it is not installable
Depends: asciidoc but it is not installable
Depends: python3-sphinx-rtd-theme but it is not installable
I'm not sure how to proceed here any more. Looks like the newer kernel trees don't have these stale dependencies...
Closing remarks
I've explored in somewhat detail how the Ubuntu kernel rules work, opened some bugs, and found a plausible way to proceed. Let's stay tuned on the bugs!
Comment