rebootstrap
rebootstrap is a crude hack to test the bootstrappability of Debian architectures. In this context bootstrapping refers to building binaries for an architecture from source without using any pre-built binaries for that architecture. In the past 20 years about 20 architectures have been bootstrapped for Debian. At all times this has been a manual and non-repeatable process. rebootstrap is trying to address the very early bootstrap phase involving the gcc/eglibc dance.
The source can be obtained using git:
git clone https://salsa.debian.org/helmutg/rebootstrap.git
One of rebootstraps major goals is to highlight bugs.
Contact Helmut Grohne or stop by #debian-bootstrap on OFTC if you have any questions.
Occasionally, rebootstrap works around limitations of packages by carrying patches locally. This raises the question of which patches are acceptable. Rules are:
- Patches must be intended for inclusion into the patched package and therefore should be filed as a bug against the relevant package.
- Patches beyond a couple of hundred lines are simply too long atm.
Thanks
Without the following (non-exhaustive list of) people rebootstrap would not be where it is today. Thanks!
- Porters for providing patches
- arc: Vineet Gupta
- arm64: Wookey
mips64el: YunQiang Su
- or1k: Christian Svensson
- ppc64el: Breno Leitao, Mauricio Faria de Oliveira, Erwan Prioul, IBM
- x32: Daniel Schepler, Thorsten Glaser, Adam Borowski
- musl-linux-any: Szabolcs Nagy, Davor Ocelic
- hurd-any: Samuel Thibault
- kfreebsd-any: Christoph Egger, Steven Chamberlain
- sparc64 and more: John Paul Adrian Glaubitz
- nios2: Marek Vasut
- uclibc-ng: Waldemar Brodkorb
- riscv64: Manuel A. Fernandez Montecelo
- Package maintainers for applying patches, in particular Matthias Klose for being quick
- Infrastructure:
- Holger Levson for administrating jenkins.d.n
ProfitBricks for donating the hardware running jenkins.d.n
Luca Falavigna for adding cross support to debomatic
Wavecon GmbH for donating the hardware running debomatic-*.d.n
- Advice and other kinds of help
- Guillem Jover for telling how to do things correctly
- Bernhard R. Link for explaining difficult error messages
- Peter Pentchev for fixing bugs that rebootstrap discovers tomorrow
- Johannes Schauer for making things work in theory
- Wookey, for NMUing lots of packages
- Dima Kogan for maintaining cross-gcc-dev
- James Clarke for many insightful answers and patches
- Manuel A. Fernandez Montecelo for NMUing lots of packages or just poking them until maintainers fix them
Simon McVittie for his work on native issues
- Many others! (Multiarch, Emdebian, etc.)
Issues
Below is a non-exhaustive list of unsolved problems. Patches welcome.
alpha
Something causes cross compiler libraries (lib*-alpha-cross) to depend on libc6.1 rather than libc6.1-alpha-cross. Thus they are uninstallable. (Jobs rebootstrap_alpha_gcc*_supported)
libgc fails to build during tests and has symbol differences during cross.
arm64ilp32
arm64ilp32 is dead.
hppa
Someone needs to tell guile-2.0 how to cross build for hppa.
File conflict between libc6-dev and libc6-dev-hppa-cross (converted by dpkg-cross). Someone needs to fix dpkg-cross.
hurd-amd64
gnumach has no clue how to handle cpu type x86_64. Some needs to finish the porting, see https://www.gnu.org/software/hurd/open_issues/64-bit_port.html.
hurd-i386
The proposed solution for the LIMITS_H_TEST involves __has_include_next. Unfortunately, gcc's __has_include_next implementation is buggy when there is no further include directory, which happens to be the case here.
i386
When built on amd64, the cross compiler package gcc-N-i386-linux-gnu:amd64 has a dependency on lib64gcc1. Since the latter is only available for i386, the dependency is unsatisfiable. It should be depending on libgcc1 instead. Something goes wrong in dpkg-shlibdeps.
Other builds simply fail running dpkg-shlibdeps as it fails finding e.g. libc.so.6. Why?
When building without disabling multilib, the stage2 pass of gcc-7 fails including <stdio.h> in the x32 multilib pass of libgcc.
ia64
Debian's glibc dropped support for ia64 as of wheezy.
Currently, reboostrap on ia64 is stuck at a dependency loop between GCC's internal libunwind implementation, and GLIBC. Work is being done on a GLIBC stub that will satisfy libunwind's requirements.
kfreebsd-any
The source package kfreebsd-kernel-headers vanished from unstable. It needs to be reintroduced (by adding an alibi package, e.g. a documentation package).
lm32
binutils and gcc both support lm32 in trunk.
GCC - C/C++ compiler. Support for the LatticeMico32 has been added to GCC 4.5.0, but patches are available to add LatticeMico32 support to GCC 4.4.0.
Binutils - Assembler, linker and binary utilities; Binutils has supported the LatticeMico32 since version 2.19.
Blocker is no up-to-date linux port to the platform (an old port can be found at https://github.com/m-labs/linux-milkymist). The lm32 normally doesn't have a mmu, so needs ucLinux. Potentially more useful stuff at http://m-labs.hk/milkymist-wiki/wiki/index.php%3Ftitle=Milkymist_Linux_cheat_sheet.html and http://m-labs.hk/milkymist-wiki/wiki/index.php%3Ftitle=Linux.html
(The lm32 is a softcore created by Lattice for FPGA boards - see https://en.wikipedia.org/wiki/LatticeMico32.)
mips
When building with gcc-7, the cross compiler multilib library lib64atomic1-mips-cross ends up depending on libc6-mips64, which is arch:any. It should be depending on libc6-mips64-mips-cross instead.
mips64
src:libgpg-error lacks the lock object header.
There is a file conflict bewteen the multilibs libc6-mips32-mips64-cross and libc6-mips64-cross. Someone needs to fix dpkg-cross.
When building with gcc-7, the stage2 cross compiler build fails. Why?
mips64r6el
Symbol issues in libffi.
mipsr6el
Compiling gmp issues invalid instructions.
or1k
gcc as found in Debian sid for version 4.8 and 4.9 do not support or1k as a target. Thus gcc stage1 fails to configure. It is expected that this issue vanishes with future gcc upstream releases.
powerpcel
Someone needs to tell src:openssl what this architecture means.
riscv64
libgc needs riscv64 support.
s390
Debian's glibc dropped support for ia64.
sh3
icu needs to be ported.
x32
musl-linux-any
datefudge: datefudge.c:77:5: error: conflicting types for 'gettimeofday'
libbsd: ../include/bsd/sys/cdefs.h:28:28: fatal error: sys/cdefs.h: No such file or directory. Fixes are committed, but not released upstream.
libelf: checking host system type... Invalid configuration 'mips-linux-musl': system 'musl' not recognized
libsigsegv: fault-linux-mips.h:30:70: error: 'mcontext_t {aka struct <anonymous>}' has no member named 'gregs'
- fixed upstream in musl 1.1.11
musl-linux-i386
gcc-7 fails to build, because ./src/libmpx/mpxrt/mpxrt.h defines XSAVE_OFFSET_IN_FPMEM as sizeof (struct _libc_fpstate) and expects that struct to be defined by the libc. On glibc systems, it is defined in asm/ucontext.h. On musl, the struct is missing.