crossdev and GNU Hurd

April 14, 2018

Tl;DR: crossdev is a tool to generate a cross-compiler for you in gentoo and with some hacks (see below) you can even cross-compile to hurd!

FOSDEM 2018 conference happened recently and a lot of cool talks tool place there. The full list counts 689 events!

Hurd’s PCI arbiter was a nice one. I never actually tried hurd before and thought to give it a try in a VM.

Debian already provides full hurd installer (installation manual) and I picked it. Hurd works surprisingly well for a such an understaffed project! Installation process is very simple: it’s a typical debian CD which asks you for a few details about final system (same as for linux) and you get your OS booted.

Hurd has a ton of debian software already built and working (like 80% of the whole repo). Even GHC is ported there. While at it I grabbed all the tiny GHC patches related to hurd from Debian and pushed them upstream:

Now plain ./configure && make && make test just works.

Hurd supports only 32-bit x86 CPUs and does not support SMP (only one CPU is available). That makes building heaviweight stuff (like GHC) in a virtual machine a slow process.

To speed things up a bit I decided to build a cross-compiler from gentoo linux to hurd with the help of crossdev. What does it take to support bootstrap like that? Let’s see!

To get the idea of how to cross-compile to another OS let’s check how typical linux to linux case looks like.

Normally aiming gcc at another linux-glibc target takes the following steps:

- install cross-binutils
- install system headers (kernel headers and glibc headers):
- install minimal gcc without glibc support (not able to link final executables yet)
- install complete glibc (gcc will need crt.o files)
- install full gcc (able to link final binaries for C and C++)

In gentoo crossdev does all the above automatically by running emerge a few times for you. I wrote a more up-to-date crossdev README to describe a few details of what is happening when you run crossdev -t <target>.

hurd-glibc is not fundamentally different from linux-glibc case. Only a few packages need to change their names, namely:

- install cross-binutils
- install gnumach-headers (kernel headers part 1)
- [NEW] install cross-mig tool (Mach Interface Generator, a flavour of IDL compiler)
- install hurd-headers and glibc-headers (kernel headers part 2 and libc headers)
- install minimal gcc without libc support (not able to link final executables yet)
- install complete libc (gcc will need crt.o files)
- install full gcc (able to link final binaries for C and C++)

The only change from linux is the cross-mig tool. I’ve collected ebuilds needed in gentoo-hurd overlay.

Here is how one gets hurd cross-compiler with today’s crossdev-99999999:

git clone https://github.com/trofi/gentoo-hurd.git

HURD_OVERLAY=$(pwd)/gentoo-hurd
CROSS_OVERLAY=$(portageq get_repo_path / crossdev)
TARGET_TUPLE=i586-pc-gnu

# this will fail around glibc, it's ok we'll take on manually from there
crossdev --l 9999 -t crossdev -t ${TARGET_TUPLE}

ln -s "${HURD_OVERLAY}"/sys-kernel/gnumach ${CROSS_OVERLAY}/cross-${TARGET_TUPLE}/
ln -s "${HURD_OVERLAY}"/dev-util/mig ${CROSS_OVERLAY}/cross-${TARGET_TUPLE}/
ln -s "${HURD_OVERLAY}"/sys-kernel/hurd ${CROSS_OVERLAY}/cross-${TARGET_TUPLE}/

emerge -C cross-${TARGET_TUPLE}/linux-headers

ACCEPT_KEYWORDS='**' USE=headers-only  emerge -v1 cross-${TARGET_TUPLE}/gnumach
ACCEPT_KEYWORDS='**' USE=headers-only  emerge -v1 cross-${TARGET_TUPLE}/mig
ACCEPT_KEYWORDS='**' USE=headers-only  emerge -v1 cross-${TARGET_TUPLE}/hurd
ACCEPT_KEYWORDS='**' USE=headers-only  emerge -v1 cross-${TARGET_TUPLE}/glibc

                     USE="-*"          emerge -v1 cross-${TARGET_TUPLE}/gcc
ACCEPT_KEYWORDS='**' USE=-headers-only emerge -v1 cross-${TARGET_TUPLE}/glibc
                     USE="-sanitize"   emerge -v1 cross-${TARGET_TUPLE}/gcc

Done!

A few things to note here:

gentoo-hurd overlay is used for new gnumach, mig and hurd packages
symlinks to new packages are created manually (crossdev fix pending, wait on packages to get into ::gentoo)
uninstall linux-headers as our target is not linux (crossdev fix pending)
use ACCEPT_KEYWORDS=’**’ for many packages (need to decide on final keywords, maybe x86-hurd)
all crossdev phases are ran manually
only glibc git version is working as changes were merged upstream very recently.
USE=”sanitize” is disabled in final gcc because it’s broken for hurd

Now you can go to /usr/${TARGET_TUPLE}/etc/portage/ and tweak the defaults for ELIBC, KERNEL and other things.

Basic sanity check for a toolchain:

$ i586-pc-gnu-gcc main.c -o main
$ file main
main: ELF 32-bit LSB executable, Intel 80386, version 1 (SYSV), dynamically linked, interpreter /lib/ld.so, for GNU/Hurd 0.0.0, with debug_info, not stripped

Copying to the target hurd VM and runnig there also works as expected.

I use crossdev -t x86_64-HEAD-linux-gnu to have GHC built against HEAD in parallel to system’s GHC. Let’s use that for more heavyweight test to build a GHC cross-compiler to hurd:

$ EXTRA_ECONF=--with-ghc=x86_64-HEAD-linux-gnu-ghc emerge -v1 cross-i586-pc-gnu/ghc --quiet-build=n

This fails as:

rts/posix/Signals.c:398:28: error:
     note: each undeclared identifier is reported only once for each function it appears in
    |
398 |         action.sa_flags |= SA_SIGINFO;
    |                            ^

Which hints at lack of SA_SIGINFO support in upstream glibc.git. Debian as an out-of-tree tg-hurdsig-SA_SIGINFO.diff patch to provide these defines (as least it’s not our local toolchain breakage). The outcome is positive: we have got very far into cross-compiling and hit real portability issues. Woohoo!

Final words

As long as underlying toolchains are not too complicated building cross-compilers in gentoo is trivial. Next tiny step is to cross-build hurd kernel itself and run it in qemu. Ebuilds in gentoo-hurd are not yet ready for it but tweaking them should be easy.

Have fun!