archlinux-s390x/TODO.md

TODO — divergences from upstream Arch

Every deliberate difference between this port and Arch's own packages, so that none of them becomes invisible. A port that does not track what it gave up drifts silently, and the gap is discovered years later by whoever needs the missing feature.

Two kinds of entry, and the difference matters:

• Deferred — dropped to get the bootstrap moving. Nothing about s390x prevents it; it costs build time, a host dependency, or another package first. These should come back.
• Architectural — the thing being dropped is x86-specific by nature. There is nothing to restore.

Each entry names the hook that implements it, so the change is one file away.

---

Deferred

libgccjit — patches/pkgbuild/gcc.sh

Arch configures and compiles the entire gcc tree a second time just for libgccjit, because it needs --enable-host-shared and applying that to the main compiler would slow it down. That pass is about half of gcc's build time — twenty minutes per attempt on this hardware.

Already re-enablable, no code change needed:

EL_GCC_JIT=1 bash scripts/build-stage1.sh

To close: run one build with the switch on and confirm the package appears. Nothing else is required.

HTTP/3 in curl — patches/pkgbuild/curl.sh

configure: error: nghttp3 enabled without a QUIC library; enable ngtcp2

Ubuntu packages neither ngtcp2 nor nghttp3, so the build host cannot satisfy them. pacman talks to plain HTTPS mirrors, and so does ERPLibre.

To close: build ngtcp2 and nghttp3 as Arch packages, add them to the stage-1 list before curl, then delete the hook.

gcc front ends: Ada, D, Modula-2, COBOL — patches/pkgbuild/gcc.sh

configure: error: GNAT is required to build ada
configure: error: GDC is required to build d

These are written in themselves and need an existing compiler of the same language. Ubuntu ships gnat, but not a version GCC 15 accepts; there is no D bootstrap compiler here at all.

To close: after stage 2, the port's own gcc can bootstrap them. This is the entry most likely to matter to someone — a mainframe shop with Ada is not a hypothetical.

Test suites — scripts/bootstrap-pacman.sh (--nocheck)

Stage-1 suites run against the host libraries, not Arch's, so their verdict says nothing about the port. acl failed its check step on a sound build. They also cost hours.

To close: stage 2 runs inside the chroot, where the suites test what was actually built. Drop --nocheck there.

Checksums on generated patches — scripts/bootstrap-pacman.sh

--skipchecksums. GitLab regenerates .patch URLs, so their checksums drift from what the PKGBUILD recorded. Release tarballs still validate — only the generated patches are skipped.

To close: fetch patches from a stable source, or record our own sums.

PGP signatures — scripts/bootstrap-pacman.sh

--skippgpcheck, because no keyring exists yet.

To close: build archlinux-keyring, then verify.

PDF manuals in pam — patches/pkgbuild/pam.sh

Arch lists fop in makedepends, a Java renderer, to produce three PDFs that the very next line of package() deletes as unreproducible. A JVM on the build host buys nothing, so without it meson drops the PDF targets silently and the rm is handed an unexpanded glob. The hook makes it rm -f.

To close: nothing needs closing. Listed because the hook's -f would otherwise look like it is papering over a real absence.

libalpm.so has no version — /etc/makepkg.conf

Every package in the repository declares bare sonames — provides = libalpm.so, not libalpm.so=16-64 — because OPTIONS carries !autodeps, which is makepkg's own default. The soname scan never runs for any package. Verified across all seventy.

Harmless today: our depends are equally unversioned, so the repository is internally consistent and pacman resolves it. It breaks the moment anything wants a versioned soname, as Arch's own packages do.

To close: enable autodeps and rebuild, or accept the divergence knowingly. Not urgent, but it is a real difference from Arch's repository.

makedb still links libselinux — no hook, deliberately

glibc's usr/bin/makedb is the last binary in the repository carrying a host artefact. It is the only one left of the six the audit found.

It is NOT hooked on purpose: nothing in the bootstrap runs makedb, and rebuilding glibc for it would relink the foundation under every other package.

To close: stage 2 rebuilds glibc in a chroot that has no libselinux, and it closes itself.

The python module brotli builds is unusable on the target — patches/pkgbuild/brotli.sh

The wheel is built by the host's python 3.13 and is ABI-tagged cpython-313; Arch ships 3.14. No destination makes it importable. The hook only moves it out of /usr/local, which an Arch package may not ship.

To close: stage 2 builds it with the port's own python.

Host sonames the target does not have — scripts/test-chroot.sh check 3

Nine libraries are requested by a shipped binary at the HOST's version while the repository ships the same library at Arch's:

requested	by	we ship
libgpgme.so.11	pacman (via libalpm)	gpgme 2.x, .45
libnettle.so.8, libhogweed.so.6	gnutls, libcurl-gnutls	nettle
libicuuc.so.76	libxml2	icu
liblmdb.so.0	krb5	lmdb
libnsl.so.2	pam	libnsl
libsasl2.so.2	libldap, openldap	libsasl
libsodium.so.23	openldap	libsodium
libdevmapper.so.1.02.1	cryptsetup	device-mapper

This is stage 1 working as designed, not a defect list: stage 1 links the host, so it records the host's sonames. It is written down because the consequence is concrete — pacman in the stage-1 rootfs does not start:

pacman: error while loading shared libraries: libgpgme.so.11

To close: stage 2. Every one of these dissolves when the package is rebuilt inside the chroot against the repository's own libraries. Nothing here should be hooked in stage 1; a hook would only hide it.

Note for stage 2's design: because pacman cannot run inside the stage-1 rootfs, stage 2 has to install with the HOST's pacman and --root, the way scripts/test-chroot.sh already does, and use the chroot only for building.

Three sonames with no provider at all — same check

Distinct from the nine above, and each for its own reason:

• libpython3.13.so.1.0 ← util-linux-libs. It ships pylibmount.cpython-313-*.so, built against the host's python. Inert: this port ships no python, so nothing imports it. Stage 2 drops it by construction — the chroot has no python either, so meson will not build it. Worth naming because arch-meson was fixed to put this file in the right PLACE (site-packages, not dist-packages) and the wrong ABI stayed invisible until the soname audit.
• libselinux.so.1 ← glibc's makedb. Deliberate, see above.
• libtcl8.6.so ← sqlite-tcl, sqlite-analyzer. We ship tcl 9.0; sqlite's configure found the host's 8.6 tclConfig.sh. An honest admission: the measurement that chose to BUILD tcl rather than drop these sub-packages compared closure cost and never looked at the ABI. Stage 2 fixes it — tcl 9.0 is in the repository — but the reasoning had a blind spot, and it was the same blind spot as the pylibmount one.

lvm2 — patches/pkgbuild/lvm2.sh

Only device-mapper is built. The lvm2 half declares thin-provisioning-tools, which is Rust now, so a language runtime arrived through a fourth-order dependency. Nothing in the repository wants lvm2 itself; cryptsetup wants device-mapper, which the same source produces.

To close: build Rust, restore the sub-package, rebuild. A target that wants LVM needs this; dm-crypt and systemd do not.

SQL auxprop in libsasl — patches/pkgbuild/libsasl.sh

Disabled. Its two database headers exist on this host, on paths configure does not try (/usr/include/postgresql, /usr/include/mariadb). Arch declares depends=(glibc) for libsasl because the plugins are dlopened, so nothing is lost.

To close: supply the include paths, or leave it — a bootstrap has no SQL authentication backend to serve.

fuse2fs — patches/pkgbuild/e2fsprogs.sh

Not built: no fuse3 on the host, and none in the closure.

To close: add fuse3, and both the sub-package and the removal that feeds it come back on their own.

kbd without xkb — patches/pkgbuild/kbd.sh

loadkeys cannot generate keymaps from an XKB database. Arch's kbd cannot either — it declares libxkbcommon nowhere, so Arch's chroot fails the same probe. Building it would want libxcb and xkeyboard-config behind it.

To close: it is already closed with respect to Arch. Reopen only if a target grows a graphics stack.

---

Architectural

The baseline CPU is z13 — decided, not inherited

makepkg.conf in the stage-2 chroot and --with-arch in patches/pkgbuild/gcc.sh.

zlib stopped on

contrib/crc32vx/crc32_vx.c:205:10: error:
  '__builtin_s390_vec_unpackl' requires '-mvx'

after its own configure had detected vector support and defined -DHAVE_S390X_VX. Both halves were right, which is what made it worth measuring rather than patching:

	--with-arch	default -march
host gcc (Ubuntu)	z13, tune z16	arch11 = z13, 2015
our gcc	nothing	arch5 = z900, 2000

Arch's PKGBUILD names no s390x baseline because Arch has no s390x, so GCC fell back to the oldest machine the target could possibly be. Every distribution picks one of these. This port had never said which, and silence chose the year 2000.

z13 because that is what Ubuntu s390x — the host distribution — already requires: adopting it cannot stop anything from running that runs today.

Two places, two different questions. makepkg.conf says how this distribution is compiled. gcc's --with-arch says what the compiler we ship assumes when someone builds on the target with no flags at all. Setting only the first leaves a gcc that quietly reverts to arch5.

A third failure shape: the host is behind

Contamination and closure gaps were the two known kinds. libxslt is neither:

configure: error: Version 2.14.5 found. You need at least libxml2 2.15.1

2.14.5 is Ubuntu's libxml2; ours is 2.15.3. Nothing is contaminated and nothing is missing — the package simply cannot be built on this host, at any point, because the host is older than the package requires. It is built BY stage 2 instead of installed into it, and hoisted in STAGE2_FIRST so it precedes the package that needs it. A stage-1 run is expected to report it as failed.

Worth watching for: any Arch package whose dependency floor has moved past Ubuntu's version belongs in this category, and the error will always read like a missing dependency rather than an old one.

Nothing to restore here. Listed so nobody spends an afternoon rediscovering why.

What	Where	Why
SFrame	glibc.sh	configure: error: the architecture doesn't support SFrame
gold linker	binutils.sh	x86-centric, deprecated upstream since 2.44; ld.bfd is the linker on Z
32-bit multilib	glibc.sh, gcc.sh, libtool.sh	On x86_64 it means i686. On Z it means the 31-bit ESA/390 ABI that GCC is removing
libhwasan	gcc.sh	Hardware-assisted ASan exists on x86_64 and aarch64 only
systemd-boot	systemd.sh	meson.build:1627 maps a cpu family to an EFI machine type and s390x is not in the table. Z boots from an IPL record; there is no ESP to write into. bootctl is still built

---

The host decides what gets built

A category the first three did not cover, and the one that cost this round the most. These are not choices anyone made. A build option's DEFAULT is computed from the build machine, so installing an unrelated Ubuntu package changes what a PKGBUILD produces — usually without failing, sometimes weeks later.

Four were found, all in one round, all now pinned:

Option	Where	What the host decides
libdir	scripts/devtools/arch-meson	meson asks dpkg-architecture and answers lib/s390x-linux-gnu. Five packages shipped their libraries where Arch's ld.so never looks
python.platlibdir	scripts/devtools/arch-meson	meson's python module uses Debian's deb_system scheme, /usr/lib/python3/dist-packages
EXPAT_BUILD_DOCS	expat.sh	ON if any of four docbook converters is on PATH. asciidoc pulled in docbook-utils as an automatic dependency, and expat — which had built clean the day before — started failing 29 hours later
split-bin	systemd.sh	probes whether the BUILD machine's /usr/sbin is a symlink. Ubuntu's is a real directory, so six binaries went to /usr/sbin and package() could not find them

The general lesson: arch-meson passes --auto-features enabled, which promotes every auto feature to a hard requirement. Installing a tool therefore switches on code that has never compiled in this port. util-linux was the clearest case — asciidoctor was needed for its man3 pages, and its arrival enabled a translation step that had never once run.

To close: stage 2 builds inside an Arch root, where every one of these defaults is already the right one. Until then, pin rather than hope.

---

Documentation is skipped in the stage-2 chroot — scripts/build-stage2.sh

Our own meson package ships /usr/bin/arch-meson, and it passes --auto-features enabled. That is right on Arch, whose build chroot has every optional tool installed. Ours has none of them:

doxygen  xsltproc  asciidoctor  itstool  convert
fig2dev  elinks    ducktype     yelp-build

With enabled, each absent tool is a hard error rather than a skipped feature. libxml2 stopped on

libxml2/doc/meson.build:3:10: ERROR: Program 'doxygen' not found

The chroot now gets a /usr/local/bin/arch-meson wrapper appending --auto-features auto — meson honours the last occurrence, so every other choice arch-meson makes survives. /usr/local/bin comes first on the chroot's PATH; /usr/bin/arch-meson is left exactly as the package shipped it.

Measured before choosing, because building the tools was the alternative and it is not close: doxygen alone wants clang, fmt, spdlog and llvm-libs — a compiler infrastructure for a documentation generator. Behind the other eight stand Ruby, ImageMagick and a GNOME stack. Several times the size of everything built so far, for man pages.

To restore at stage 3, once there is somewhere to build them from.

The docs split assumed the docs exist — patches/pkgbuild/libxml2.sh

Consequence of the above, and the fourth appearance of one shape: pam's PDFs, e2fsprogs' fuse2fs, cmake's emacs byte-compilation, and now

mv: cannot stat '<pkgdir>/usr/share/doc': No such file or directory

A failure on the last line of package(), after a completely successful compile, naming a path instead of a reason. It has never once been a broken build. The move is made conditional rather than deleted — on the host the docs do exist and belong in their own package.

libarchive drops xar for stage 1 — patches/pkgbuild/libarchive.sh

Stage 2 reached its first packaging step and stopped on

bsdtar: error while loading shared libraries: libicuuc.so.76
==> ERROR: Failed to create package file.

The chain: bsdtar → libarchive.so.13 → libxml2.so.16 → libicuuc.so.76. The libxml2 in the chroot is our stage-1 build, and stage 1 builds against the HOST, whose ICU is 76; our icu package ships 78. So bsdtar cannot start — and bsdtar is what makepkg uses to write the package. The package that would fix it is the one being built. Nothing comes out of the chroot until this is broken from outside it.

Supplying libicuuc.so.76 from the host is the obvious move and the wrong one: untracked host binaries in /usr/lib are exactly what test-chroot.sh's ARTEFACT check exists to catch. A symlink to 78 does not work either — ICU version-suffixes every symbol, so libicuuc.so.78 does not define u_strlen_76.

The root is that libarchive links libxml2 at all: it wants it for xar, an Apple installer container. Nothing here reads one — makepkg writes .pkg.tar.gz, pacman reads it. --without-xml2 --without-expat removes the whole ICU chain instead of arguing about its version. Stage 1 only; STAGE2_SKIP_HOOKS lists libarchive, so stage 2 builds it as Arch does, against our own libxml2 and icu, where the versions agree.

Rebuilding one package at stage 1 — scripts/build-stage1.sh

build-stage1.sh now takes package names, like stage 2 already did. Without it, rebuilding one package meant deleting its line from stage1.state first — editing the record of what was built in order to build something. That file is this port's memory, and the standing instruction is to regenerate it from what is really in repo/s390x, never from a log; hand-editing it is the same mistake in a smaller form. An explicit name outranks the state file, because being told "already built, skipping" would make the command useless for the one job it has.

84 host binaries the chroot does not have — measured, not guessed

install_host_deps installs 110 apt packages. Asking dpkg what each ships in /usr/bin, then testing every one of those against the stage-2 chroot, gives 84 absent binaries. Worth doing once rather than discovering them a package at a time over a twenty-hour pass — two were found that way first (rsync, makeinfo) and each cost a full build to surface.

Build tools, which will each block something:

gperf  help2man  po4a  xsltproc  scdoc  dtrace (systemtap)
cython  pyproject-build  sphinx-build  rst2man (docutils)

Documentation only, and settled by --auto-features auto:

doxygen  asciidoctor  itstool  ducktype  yelp-build  yelp-check
elinks  fig2dev  transfig  rsvg-convert  clang (doxygen's)

Not gaps at all, on inspection: bison.yacc, fakeroot-tcp, faked-sysv, automake-1.17, aclocal-1.17, ncurses6-config, nspr-config, tcltk-depends are Debian's alternative or versioned names for tools that are present under Arch's names. Counting them as missing would have sent someone building packages that already exist.

Each is built when a package actually asks for it, not in advance — several carry closures much larger than themselves (dtrace brings systemtap, sphinx-build a Python stack), and paying for one before knowing it is needed is how a bootstrap doubles in size.

Environment, not the port

Build locale — scripts/bootstrap-pacman.sh

LC_ALL=C.UTF-8 on the makepkg line. This host runs fr_FR.UTF-8, and util-linux's poman-translate.sh greps po4a's output for the English word Discard. In French po4a says Rejet de …, so the skip list came out empty and the build was handed 852 files po4a had never written.

C.UTF-8 and not C: the Arabic and Ukrainian pages must stay valid UTF-8. Arch's build chroot runs in this locale, so this is parity, not a workaround.

history.pc — scripts/bootstrap-pacman.sh

GNU readline installs both readline.pc and history.pc; Debian and Ubuntu keep the first and drop the second, while libhistory.so ships in libreadline-dev regardless. libxml2 asks pkg-config for history and stops.

Our own readline package ships a correct history.pc — and copying it would be wrong. Its libdir names /usr/lib, which on a multiarch host holds no libhistory. The file has to describe the BUILD host, so it is generated.

gnupg's feature set moved on its own — no hook

libtss2-dev was installed for systemd. gnupg's configure found it, turned TPM: no into TPM: yes, and shipped usr/lib/gnupg/tpm2daemon — code that had never compiled here. It built cleanly, and Arch's gnupg depends on tpm2-tss too, so this is parity rather than drift.

Recorded because nobody decided it. It is the same mechanism as the table above, caught on the way past.

To close: nothing. Noted so the next unexplained gnupg change has a precedent to read.

libassuan source — patches/pkgbuild/libassuan.sh

dev.gnupg.org is unreachable from this build host — measured HTTP 000, while github.com/gpg/libassuan.git and gnupg.org/ftp both answer. Only the transport changed; the pinned tag is untouched.

To close: it closes itself on a host with wider outbound access.

arch-meson / arch-cmake — scripts/devtools/

Several PKGBUILDs call them, they ship in devtools, and devtools is itself an Arch package — during a bootstrap they cannot exist yet.

To close: build devtools in stage 2 and drop the stand-ins.

Host include layout — scripts/bootstrap-pacman.sh

Ubuntu multiarch puts asm/, bits/, gnu/ and sys/sdt.h under /usr/include/s390x-linux-gnu, while glibc builds with -nostdinc and expects the classic layout. Symlinks bridge it.

To close: stage 2 builds inside an Arch root, where the layout is already right.