archlinux-s390x/scripts/build-stage2.sh

#!/usr/bin/env bash
# Stage 2: rebuild the repository inside the repository.
#
# WHAT STAGE 2 IS FOR
#
# Every package stage 1 produced was compiled against UBUNTU's libraries. That
# is not a defect -- Arch's glibc needs an Arch gcc which needs an Arch glibc,
# so the first pass has nowhere else to start -- but it leaves host artefacts
# baked in. scripts/test-chroot.sh names nine of them precisely: binaries that
# ask for libgpgme.so.11, libnettle.so.8, libicuuc.so.76 and six more, at the
# host's soname versions, while the repository ships Arch's. Stage 2 dissolves
# all nine by rebuilding each package against what the repository actually has.
#
# THE CONSTRAINT THAT SHAPES THIS SCRIPT
#
# pacman cannot run inside the stage-1 rootfs. libalpm was linked against the
# host's gpgme, so the binary is there and does not start:
#
#   pacman: error while loading shared libraries: libgpgme.so.11
#
# So the rootfs is populated from OUTSIDE, with the host's pacman and --root,
# the way scripts/test-chroot.sh does. The chroot is used only to BUILD. That
# is not a workaround, it is the order the problem has: stage 2's own output
# is the first pacman that will run on the target.
#
# makepkg, by contrast, is a shell script, and it works.
set -uo pipefail

HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
# Stage 2 rebuilds the same packages in the same order; packages.sh explains
# why that order is not re-derived here.
source "$HERE/packages.sh"
# watched(): bootstrap-pacman.sh cannot be sourced here -- see lib-watch.sh.
source "$HERE/lib-watch.sh"
WORK="${WORK:-$HOME/work/arch-s390x}"
REPO1="${REPO1:-$WORK/repo/s390x}"
REPO2="${REPO2:-$WORK/repo2/s390x}"
ROOT="${ROOT:-$WORK/rootfs-stage2}"
CONF="$WORK/pacman-stage2.conf"
CACHE="$WORK/pacman-stage2.cache"
CONF2="$WORK/pacman-stage2-both.conf"
STATE2="$WORK/stage2.state"
PATCH_DIR="${PATCH_DIR:-$HERE/../patches/pkgbuild}"
BUILDER="${BUILDER:-$(id -un)}"
BUILD_UID="$(id -u)"
BUILD_GID="$(id -g)"

# The chroot's contents. Two groups, and the second is the one the dependency
# resolver could never have told us about: nothing in the repository DEPENDS on
# bison or fakeroot, they are simply what a build needs to happen.
CHROOT_PKGS=(
    # A system that reaches a shell and can read a package.
    filesystem glibc bash coreutils sed grep gawk findutils file which
    tar gzip xz bzip2 zstd libarchive diffutils patch
    # The toolchain.
    gcc binutils make m4 autoconf automake libtool pkgconf
    bison flex texinfo groff gettext
    # makepkg itself, and the one thing it cannot do without.
    pacman fakeroot
    # rsync, which is not a runtime dependency of anything here. The kernel's
    # headers_install target runs it, so linux-api-headers -- the first package
    # stage 2 tries -- stopped on `rsync: command not found`. The host had it
    # from apt, which is exactly why stage 1 never mentioned it.
    rsync
    # The same class, named by the failures of the first full pass rather than
    # guessed at. Each one is here because a package said so:
    #
    #   gperf     coreutils, diffutils, systemd, libseccomp
    #   wget      sed, grep, findutils
    #   patchelf  curl
    #   inetutils gnupg, for hostname
    #   libxslt   shadow, for xsltproc
    #   swig      audit
    #   help2man  flex
    #
    # libxslt is NOT here, and cannot be: it will not build on the host at all.
    #
    #   configure: error: Version 2.14.5 found. You need at least libxml2
    #   2.15.1 for this version of libxslt
    #
    # 2.14.5 is Ubuntu's libxml2. Ours is 2.15.3 -- and it exists only inside
    # this chroot, which is the one place libxslt can be built. So it is built
    # BY stage 2 rather than installed into it, which is why it sits in
    # STAGE2_FIRST instead. Listing it here would fail make_rootfs with "target
    # not found" against a repository that will never contain it.
    gperf wget patchelf inetutils swig help2man
    # libxcrypt-compat, for a reason no declaration expresses. perl declares
    # `libxcrypt` and `libcrypt.so`, both satisfied by libxcrypt, which ships
    # libcrypt.so.2. But perl's BINARY was linked on the host against Ubuntu's
    # libcrypt.so.1, so it does not start:
    #
    #   /usr/bin/perl: error while loading shared libraries: libcrypt.so.1
    #
    # The repository does ship that soname -- in libxcrypt-compat, a separate
    # sub-package -- so this is neither a missing package nor a soname the
    # audit should have flagged. It is a third thing: the dependency
    # declarations cannot pull it in, because they name the unversioned soname
    # that the newer library also provides. Listed explicitly, because nothing
    # will deduce it.
    libxcrypt-compat
    # git: 51 PKGBUILDs use git sources, and makepkg validates the clone even
    # under --noextract.
    git
    # meson and cmake, with the python they are written in. Ten PKGBUILDs
    # call arch-meson and four call cmake, so without these stage 2 could
    # rebuild most of the repository and then stop.
    python meson ninja cmake
)

log()  { printf '\n== %s ==\n' "$*"; }
die()  { printf 'stage2: %s\n' "$*" >&2; exit 1; }

# A PREDICATE and a fatal check, kept apart on purpose.
#
# require_space calls die, which exits. Using it inside the rebuild loop as
# `require_space || break` looks like a clean early stop and is not one: the
# break is unreachable, the run ends mid-loop, and the summary naming which
# packages were rebuilt and which failed is never printed. At the start of the
# run, exiting IS the right answer -- there is nothing to summarise yet.
space_ok() {
    local free_mb
    free_mb=$(df -Pm "$WORK" | awk 'NR==2 {print $4}')
    if [ "${free_mb:-0}" -lt 8192 ]; then
        printf '  only %s MiB free under %s; need 8192\n' "${free_mb:-0}" "$WORK" >&2
        return 1
    fi
}

require_space() { space_ok || die "not enough disk space to start"; }

make_rootfs() {
    log "Populating the stage-2 rootfs from stage 1"
    cat > "$CONF" <<EOF
[options]
Architecture = s390x
SigLevel = Never
[core]
Server = file://$REPO1
EOF
    sudo rm -rf "$ROOT" "$CACHE"
    sudo mkdir -p "$ROOT/var/lib/pacman" "$CACHE"
    sudo pacman --root "$ROOT" --config "$CONF" --cachedir "$CACHE" \
        --noconfirm -Sy "${CHROOT_PKGS[@]}" > "$WORK/stage2-install.txt" 2>&1 \
        || { tail -20 "$WORK/stage2-install.txt" >&2; die "populate failed"; }
    printf '  %s packages installed\n' "$(sudo ls "$ROOT/var/lib/pacman/local" | wc -l)"

    # Put stage 2's own output back on top of it.
    #
    # make_rootfs wipes and repopulates from stage 1 on EVERY run, which is
    # what makes the chroot reproducible -- and what would make stage 2
    # unresumable, because stage2.state survives while the packages it names do
    # not. The second invocation would say "already rebuilt, skipping" about
    # packages that had just been thrown away, and the next build would link
    # against stage-1 libraries while the record claimed otherwise. Silent, and
    # the kind of thing found weeks later in an artefact.
    #
    # Stage 2 is a hundred and thirty-three packages. It will not finish in one
    # invocation, so resuming has to be correct rather than approximately
    # correct.
    #
    # --nodeps for the same reason the per-package install uses it: a chroot
    # halfway through stage 2 is a mixed population, some packages declaring
    # versioned soname dependencies and some declaring names. -U is fed the
    # files directly, so --nodeps installs exactly these and not a resolved
    # closure -- correct here, since stage 1 already supplied the closure just
    # above.
    shopt -s nullglob
    local back=("$REPO2"/*.pkg.tar.*)
    shopt -u nullglob
    if [ "${#back[@]}" -gt 0 ]; then
        sudo pacman --root "$ROOT" --config "$CONF" --cachedir "$CACHE" \
            --noconfirm --nodeps -U "${back[@]}" \
            > "$WORK/stage2-restore.txt" 2>&1 \
            || { tail -20 "$WORK/stage2-restore.txt" >&2; die "restoring stage-2 output failed"; }
        printf '  %s stage-2 package(s) restored\n' "${#back[@]}"
    fi
}

configure_chroot() {
    log "Configuring the chroot"
    # CARCH and CHOST, for the same reason they had to be set on the host --
    # except here the wrong value arrives from OUR OWN pacman package, which
    # ships Arch's /etc/makepkg.conf verbatim:
    #
    #   CARCH="x86_64"
    #   CHOST="x86_64-pc-linux-gnu"
    #
    # A stage-2 build with those would configure every source for x86_64 on an
    # s390x machine. CHOST must be the canonical triplet, not the Debian one:
    # config.sub turns s390x-linux-gnu into s390x-ibm-linux-gnu and GCC builds
    # its tree under the canonical name, which is what broke gcc's own
    # packaging on the host.
    sudo sed -i 's|^CARCH=.*|CARCH="s390x"|; s|^CHOST=.*|CHOST="s390x-ibm-linux-gnu"|' \
        "$ROOT/etc/makepkg.conf"
    sudo sed -i "s|^#\?MAKEFLAGS=.*|MAKEFLAGS=\"-j$(nproc)\"|" "$ROOT/etc/makepkg.conf"
    # !debug and !lto, matching what stage 1 used. Arch's defaults enable both;
    # turning them on here would change what is being compared between the two
    # stages, and comparing them is the whole point.
    sudo sed -i 's|^OPTIONS=.*|OPTIONS=(strip docs !libtool !staticlibs emptydirs zipman purge !debug !lto)|' \
        "$ROOT/etc/makepkg.conf"
    # CFLAGS and friends, which arrive from the same place and are just as
    # wrong. Our pacman package ships Arch's makepkg.conf verbatim, so the
    # chroot inherits
    #
    #   CFLAGS="-march=x86-64 -mtune=generic -O2 ... -fcf-protection ..."
    #
    # On s390x cc rejects that, and the failure surfaces nowhere near the
    # cause: meson simply cannot start.
    #
    #   libxml2/meson.build:1:0: ERROR: Unable to detect linker for compiler
    #   `cc -Wl,--version ... -march=x86-64 -mtune=generic ...`
    #
    # They are emptied rather than translated, because "no flags" is what stage
    # 1 used -- the host's makepkg.conf carries no CFLAGS line at all -- and
    # 179 working packages are the evidence that it builds. Choosing s390x
    # tuning (-march=z13, and only the hardening flags that exist on Z) is a
    # deliberate later step, not something to guess at inside a bootstrap.
    # TODO.md records it.
    #
    # APPENDED, not commented. The first attempt here put a # in front of each
    # assignment, and CFLAGS is a MULTI-LINE assignment: commenting its first
    # line left the continuations active and the quote unbalanced, so makepkg
    # would not start at all --
    #
    #   /etc/makepkg.conf: line 109: unexpected EOF while looking for matching `"'
    #
    # An override at the end of the file needs no parsing of what came before:
    # the last assignment is the one that counts.
    sudo tee -a "$ROOT/etc/makepkg.conf" > /dev/null <<'EOC'

# --- stage 2: the shipped values are Arch's x86_64 ones ---
#
# THE BASELINE CPU OF THIS PORT, which was being decided by accident.
#
# 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 instead of patching:
#
#   host gcc (Ubuntu)   --with-arch=z13 --with-tune=z16   default -march=arch11
#   our  gcc            (nothing)                         default -march=arch5
#
# arch5 is z900, from 2000. arch11 is z13, from 2015. Arch's PKGBUILD names no
# s390x baseline because Arch has no s390x, so our gcc fell back to the oldest
# machine the port could possibly run on -- and every package using a vector
# intrinsic failed, while zlib's configure went on detecting a CPU the compiler
# had been told to forget.
#
# z13 is not a guess: it is what Ubuntu s390x, the host distribution, already
# requires, so nothing that runs here today stops running. Every distribution
# picks one of these; this port had simply never said which, and silence chose
# the year 2000.
#
# TWO PLACES, because they answer different questions. Here is how this
# distribution is COMPILED. patches/pkgbuild/gcc.sh passes --with-arch to the
# compiler we SHIP, so a build done later on the target assumes the same
# machine. Setting only this one leaves a gcc that quietly reverts to arch5.
CFLAGS="-march=z13 -mtune=z16 -O2 -pipe -fno-plt -fexceptions"
CXXFLAGS="$CFLAGS -Wp,-D_GLIBCXX_ASSERTIONS"
LDFLAGS="-Wl,-O1 -Wl,--sort-common -Wl,--as-needed -Wl,-z,relro -Wl,-z,now"
LTOFLAGS=""
RUSTFLAGS=""
DEBUG_CFLAGS=""
DEBUG_CXXFLAGS=""
EOC
    sudo grep -E '^(CARCH|CHOST|MAKEFLAGS|OPTIONS)=' "$ROOT/etc/makepkg.conf" | sed 's/^/  /'
    printf '  baseline: -march=z13 -mtune=z16 (host Ubuntu uses the same)\n'

    # --auto-features auto, for this chroot only.
    #
    # Our own meson package ships /usr/bin/arch-meson, and it passes
    # --auto-features enabled -- correct on Arch, whose build chroot has every
    # optional tool. Ours has none of them:
    #
    #   doxygen xsltproc asciidoctor itstool convert fig2dev elinks
    #   ducktype yelp-build        -- all ABSENT
    #
    # With `enabled`, each one is a hard error. libxml2 stops at
    #
    #   libxml2/doc/meson.build:3:10: ERROR: Program 'doxygen' not found
    #
    # MEASURED BEFORE CHOOSING, because building them was the other option and
    # it is not close: doxygen alone wants clang, fmt, spdlog and llvm-libs --
    # an entire compiler infrastructure for a documentation generator. Behind
    # the other eight stand Ruby, ImageMagick and a GNOME stack. That is
    # several times the size of everything built so far, for man pages.
    #
    # `auto` is meson's own default and means "build what you can". It is the
    # honest setting for an environment with fewer tools, not a workaround --
    # and what it drops is visible in the artefact, which is where this port
    # checks everything anyway.
    #
    # A WRAPPER, not a patched package: /usr/bin/arch-meson belongs to meson and
    # stage 2 must not ship a modified copy of it. meson takes the LAST
    # occurrence of an option, so appending wins while leaving every other
    # choice arch-meson makes intact. in_chroot puts /usr/local/bin first on
    # PATH so this is found.
    sudo install -d -m0755 "$ROOT/usr/local/bin"
    sudo tee "$ROOT/usr/local/bin/arch-meson" > /dev/null <<'EOW'
#!/usr/bin/env bash
# stage 2: this chroot has none of Arch's optional documentation tools, so a
# feature that cannot be built should be skipped rather than fatal. Appended,
# because meson honours the last occurrence.
exec /usr/bin/arch-meson "$@" --auto-features auto
EOW
    sudo chmod 755 "$ROOT/usr/local/bin/arch-meson"
    printf '  arch-meson: wrapped with --auto-features auto\n'

    # makepkg refuses to run as root, so the chroot needs the SAME uid as the
    # user who owns the bind-mounted sources. A bind mount carries the host's
    # numeric owner across, so a different uid inside would see them as
    # somebody else's and fail to write $srcdir.
    sudo tee -a "$ROOT/etc/passwd" > /dev/null <<EOF
$BUILDER:x:$BUILD_UID:$BUILD_GID::/build:/usr/bin/bash
EOF
    sudo tee -a "$ROOT/etc/group" > /dev/null <<EOF
$BUILDER:x:$BUILD_GID:
EOF
    sudo mkdir -p "$ROOT/build" "$ROOT/repo2"
    sudo chown "$BUILD_UID:$BUILD_GID" "$ROOT/build" "$ROOT/repo2"
    # The stage-1 repository, so makepkg's --nodeps builds can still read the
    # packages if anything wants to, and so repo-add has somewhere to write.
    mkdir -p "$REPO2"
}

mount_chroot() {
    log "Mounting"
    # /dev/pts is not optional: without it any build step that opens a pty --
    # and gcc's testsuite driver does -- fails in a way that names the pty and
    # not the missing mount.
    for m in proc sys dev dev/pts; do
        sudo mkdir -p "$ROOT/$m"
    done
    mountpoint -q "$ROOT/proc"    || sudo mount -t proc  proc  "$ROOT/proc"
    mountpoint -q "$ROOT/sys"     || sudo mount -t sysfs sys   "$ROOT/sys"
    mountpoint -q "$ROOT/dev"     || sudo mount --bind /dev    "$ROOT/dev"
    mountpoint -q "$ROOT/dev/pts" || sudo mount -t devpts devpts "$ROOT/dev/pts"
    # Sources and PKGBUILDs, already fetched by stage 1. Bind-mounting them
    # means the chroot needs no network at all, which is worth having: this
    # host cannot reach dev.gnupg.org, and a build that silently re-fetches
    # would be a different build.
    mountpoint -q "$ROOT/build" || sudo mount --bind "$WORK/pkg" "$ROOT/build"
    mountpoint -q "$ROOT/repo2" || sudo mount --bind "$REPO2"    "$ROOT/repo2"
}

umount_chroot() {
    for m in repo2 build dev/pts dev sys proc; do
        mountpoint -q "$ROOT/$m" && sudo umount -l "$ROOT/$m"
    done
    return 0
}

# in_chroot <command...> -- run as the builder, with a sane environment.
in_chroot() {
    sudo chroot --userspec="$BUILD_UID:$BUILD_GID" "$ROOT" \
        /usr/bin/env -i \
        HOME=/build PATH=/usr/local/bin:/usr/bin \
        LC_ALL=C.UTF-8 \
        /usr/bin/bash -lc "$*"
}

smoke_test() {
    log "Smoke test: does the chroot build anything at all?"
    # NO PIPELINES IN THESE CHECKS. The first version ran `makeinfo --version |
    # head -1`, and $? came from head, so a perl that could not start was
    # reported as ok with its own error message as the version string. Same
    # shape as the `if build_package` bug that once reported "51 built, 0
    # failed" while four packages had failed. Each check runs one command and
    # its status is the command's.
    # HARD versus KNOWN-DRIFT, because they mean different things. A hard
    # check failing means the chroot cannot build and stage 2 must not start.
    # A drift check failing means a stage-1 package carries a host version
    # mismatch that STAGE 2 ITSELF repairs, by rebuilding that package before
    # the ones that need it. Treating the second as fatal would refuse to run
    # the very thing that fixes it.
    local drift="makeinfo"
    local ok=0 fail=0 noted=0
    while read -r desc cmd; do
        [ -n "$desc" ] || continue
        local out rc
        out=$(in_chroot "$cmd" 2>&1); rc=$?
        if [ "$rc" -eq 0 ]; then
            printf '  ok    %-12s %s\n' "$desc" "${out%%$'\n'*}"; ok=$((ok+1))
        elif [[ " $drift " == *" $desc "* ]]; then
            printf '  note  %-12s %s\n' "$desc" "${out%%$'\n'*}"; noted=$((noted+1))
        else
            printf '  FAIL  %-12s rc=%s %s\n' "$desc" "$rc" "${out%%$'\n'*}"; fail=$((fail+1))
        fi
    done <<'CHECKS'
bash bash --version
gcc gcc --version
ld ld --version
make make --version
makepkg makepkg --version
fakeroot fakeroot -- /usr/bin/id -u
bison bison --version
flex flex --version
perl perl -e 'print "perl $]\n"'
makeinfo makeinfo --version
compile cd /build && mkdir -p .stage2-smoke && cd .stage2-smoke && printf 'int main(void){return 0;}' > t.c && gcc t.c -o t && ./t && echo compiled-and-ran
CHECKS
    printf '\n  %s ok, %s failed, %s known drift\n' "$ok" "$fail" "$noted"
    if [ "$noted" -gt 0 ]; then
        cat <<'NOTE'

  makeinfo is the one expected failure, and it is what stage 2 exists for:
  texinfo was built against the HOST's perl 5.40 and our perl package is 5.42,
  so its XS module refuses to load ("Perl API version ... does not match").
  Rebuilding texinfo inside this chroot fixes it -- which is why texinfo has to
  come EARLY in the rebuild order, before gcc, glibc and binutils, all of which
  call makeinfo.
NOTE
    fi
    [ "$fail" -eq 0 ]
}


# Hooks that must NOT run in stage 2.
#
# Most stage-1 hooks are still right inside the chroot: the architectural ones
# (systemd's EFI, glibc's SFrame, gcc's multilib) describe s390x, and the
# host-absence ones (pam's fop, gnutls's leancrypto, krb5's ss) describe a
# build environment that has not changed. A few are no-ops here and harmless
# -- the libdir hooks insert a value meson would already have chosen.
#
# This list is for the ones that would actively BREAK. libgcrypt.sh extracts
# our libgpg-error into $WORK/stage1-prefix and injects that absolute host path
# into build(); the path does not exist in the chroot. It is also unnecessary
# there, because the chroot HAS our libgpg-error 1.61 installed, so
# /usr/bin/gpgrt-config is already the new one. That is stage 2 working as
# intended: the reason for the hook disappears.
# git.sh joins it for the same reason: it drops ZLIB_NG=1 because Ubuntu has no
# zlib-ng headers, and our own zlib-ng package ships them, so inside the chroot
# the flag is correct and the hook would be a downgrade.
# meson.sh joins them: it moves a wheel out of /usr/local, which only the
# host's Debian-patched python puts there.
# libarchive: stage 1 drops xar to break the bsdtar -> libxml2 -> libicuuc.so.76
# cycle. In here the versions agree, so build it as Arch does.
# Hoisted to the front of the stage-2 order.
#
# Stage 1's order is the real dependency closure and stage 2 keeps it -- but
# stage 1 had the HOST for its build tools, and a few of those tools are
# themselves in the list, sitting wherever their runtime dependencies put them.
# In the chroot the stage-1 copy is what is available, and for these it does
# not work:
#
#   texinfo -- binutils died on `make info-recursive`, and makeinfo says why:
#
#     Perl API version v5.40.0 of Texinfo::TreeElement does not match v5.42.0
#
#   Its XS module was compiled against the HOST's perl 5.40; our perl package
#   is 5.42. Every package that builds .info documentation -- binutils, gcc,
#   glibc, coreutils, gettext, m4 -- needs makeinfo, and texinfo sits near the
#   end of the list because that is where its own dependencies are. Rebuilding
#   it in here against our perl fixes it, which is the entire point of stage 2;
#   it just has to happen first.
#
# Duplicates need no handling: these names appear again later in
# STAGE1_PACKAGES, and rebuild() skips anything already in stage2.state. The
# list is prepended rather than reordered, so stage 1's order stays the single
# statement of the closure.
#   libxslt -- shadow died on `configure: error: xsltproc is missing.` and
#   libxslt cannot be built on the host: its configure demands libxml2 2.15.1
#   and Ubuntu ships 2.14.5. Ours is 2.15.3 and lives only in here, so this is
#   the only place libxslt can come from -- built by stage 2, not installed
#   into it. It has to precede shadow, which the list order does not do.
STAGE2_FIRST=(texinfo libxslt)

STAGE2_SKIP_HOOKS=(libgcrypt git meson libarchive)

# Packages whose sources must be extracted on the HOST, because the chroot
# cannot extract anything until they are rebuilt.
#
# THE CYCLE. makepkg extracts with bsdtar. bsdtar is libarchive, libarchive
# links libxml2 for xar support, and the stage-1 libxml2 was linked against the
# HOST's ICU 76 while our icu package ships ICU 78:
#
#   bsdtar: error while loading shared libraries: libicuuc.so.76
#
# So nothing unpacks in the chroot until libxml2 is rebuilt, and libxml2 cannot
# unpack in the chroot. One of the nine soname drifts, turned into a bootstrap
# cycle by the one tool that has to work first.
#
# Extraction is not compilation, so doing it outside is less of an impurity
# than it looks -- the host's bsdtar unpacks a tarball byte for byte. What DOES
# leak is prepare(), which `makepkg -o` also runs: mostly patching, but where
# it runs autoreconf the generated configure carries the host's autotools.
# That is why this is a LIST and not the default. Once libxml2 is rebuilt,
# bsdtar works and everything after it extracts in the chroot.
STAGE2_HOST_EXTRACT=(libxml2)

host_extract() {
    local n="$1" h
    for h in "${STAGE2_HOST_EXTRACT[@]}"; do [ "$n" = "$h" ] && return 0; done
    return 1
}

skip_hook() {
    local n="$1" h
    for h in "${STAGE2_SKIP_HOOKS[@]}"; do [ "$n" = "$h" ] && return 0; done
    return 1
}

# stage2_build <name> -- rebuild one package inside the chroot and install it.
#
# The hook is applied on the HOST, not in the chroot: hooks are seds over the
# PKGBUILD, and /build is the same directory seen from both sides, so the
# patched file is what makepkg reads. Nothing needs to be duplicated inside.
stage2_build() {
    local name="$1"
    local dir="$WORK/pkg/$name"
    [ -d "$dir" ] || { echo "  no checkout for $name" >&2; return 1; }

    ( cd "$dir" && git checkout -- PKGBUILD 2>/dev/null ) || true
    if [ -f "$PATCH_DIR/$name.sh" ]; then
        if skip_hook "$name"; then
            echo "  hook skipped (stage-1 only)"
        else
            ( cd "$dir" && bash "$PATCH_DIR/$name.sh" ) || {
                echo "  hook failed" >&2; return 1; }
        fi
    fi

    ( cd "$dir" && rm -f ./*.pkg.tar.* ) || true
    # NO --nocheck. Stage 1 skipped the test suites because they ran against
    # the host's libraries and their verdict said nothing about the port. Here
    # they test what was actually built, which is the whole point of stage 2.
    local mkflags="--nodeps --ignorearch --skippgpcheck --skipchecksums"
    # EL_NOCHECK=1 for the FIRST pass over the list, and only that.
    #
    # Stage 1 skipped every test suite because it ran against the host's
    # libraries, so its verdict said nothing about the port. In here a suite
    # tests what was actually built, which is worth having -- but not on the
    # pass whose job is to find out whether a hundred and thirty-three packages
    # can be rebuilt at all. glibc's suite alone is longer than most of the
    # builds around it, and a suite is the likeliest place in a build to wait
    # forever on a tty or a socket.
    #
    # So: one pass to get a complete stage-2 repository, then the suites, then
    # stage 3 -- where they run on a self-hosted toolchain and their verdict is
    # about the port rather than about the bootstrap. Off by default is wrong
    # here; this must be asked for.
    [ "${EL_NOCHECK:-0}" = "1" ] && mkflags="$mkflags --nocheck"
    if host_extract "$name"; then
        echo "  extracting on the host (chroot bsdtar not usable yet)"
        ( cd "$dir" && LC_ALL=C.UTF-8 makepkg $mkflags -o -C -f ) || return 1
        # -e: build in the tree already there. -C would wipe it again.
        in_chroot "cd /build/$name && makepkg $mkflags -e -f" || return 1
    else
        in_chroot "cd /build/$name && makepkg $mkflags -C -f" || return 1
    fi

    # An exit code is not proof. Only the artefact is.
    local produced=()
    shopt -s nullglob; produced=("$dir"/*.pkg.tar.*); shopt -u nullglob
    [ "${#produced[@]}" -gt 0 ] || { echo "  no package produced" >&2; return 1; }

    cp -f "${produced[@]}" "$REPO2/" || return 1
    local names=() f
    for f in "${produced[@]}"; do names+=("$(basename "$f")"); done
    ( cd "$REPO2" && repo-add core.db.tar.gz "${names[@]}" ) > /dev/null || return 1

    # Install into the chroot so the NEXT package builds against it. With the
    # host's pacman and --root, because the chroot's own pacman does not start
    # until stage 2 has rebuilt it.
    #
    # --nodeps, and ONLY here. The first rebuilt package would not install:
    #
    #   unable to satisfy dependency 'libicuuc.so=78-64' required by libxml2
    #
    # Both halves of that are correct. This chroot runs makepkg's soname scan,
    # so a package built in here declares versioned soname dependencies the way
    # Arch's really do -- which is the faithful metadata stage 2 exists to
    # produce. The stage-1 packages around it were built on the host under
    # !autodeps, so our icu ships no `provides = libicuuc.so=78-64` to match.
    #
    # For the length of stage 2 the chroot is therefore a MIXED POPULATION:
    # some packages describe their dependencies in Arch's terms and some in
    # names only. No resolver can satisfy that, and none should be asked to --
    # it is a property of a bootstrap halfway through, not of the output. It
    # dissolves on its own as the last package is rebuilt.
    #
    # What is NOT relaxed is the metadata in the packages: they keep their
    # versioned depends, and test-chroot.sh's RESOLVE check runs the resolver
    # against the finished repository with --nodeps OFF. The relaxation is one
    # install command wide, and the check that would catch its consequences is
    # still there.
    sudo pacman --root "$ROOT" --config "$CONF2" --cachedir "$CACHE" \
        --noconfirm --nodeps -U "${produced[@]}" > "$WORK/stage2-inst-$name.txt" 2>&1 || {
        tail -10 "$WORK/stage2-inst-$name.txt" >&2; return 1; }
    printf '  installed %s package(s)\n' "${#produced[@]}"

    # Clean up, but only now, and only because it worked.
    #
    # A hundred and thirty-three source trees plus their pkg/ staging do not fit
    # on this disk. Filling it is not a hypothetical here: it happened once, and
    # what it looked like was not "no space" -- it was I/O errors from unrelated
    # virtual machines on the same host. Cheap to prevent, expensive to explain.
    #
    # AFTER the install, so nothing is thrown away until the package is proven
    # to exist and to install. NOT on failure -- src/ and pkg/ are the whole
    # evidence of what went wrong, and a build that failed is exactly the one
    # worth looking at. makepkg -c would delete them either way.
    #
    # Only makepkg's own two directories, and only inside this package's
    # checkout. The git tree, the PKGBUILD, the downloaded sources and the built
    # package are all left alone.
    ( cd "$dir" && rm -rf src pkg ) || true
}

# A pacman.conf that sees BOTH repositories: stage 2's output first, so a
# rebuilt package wins, and stage 1 behind it for everything not yet redone.
write_conf2() {
    cat > "$CONF2" <<EOF
[options]
Architecture = s390x
SigLevel = Never
[stage2]
Server = file://$REPO2
[core]
Server = file://$REPO1
EOF
}

rebuild() {
    write_conf2
    mkdir -p "$REPO2"
    local ok=0 fail=0 rc=0 failed=()
    for p in "$@"; do
        if grep -qxF "$p" "$STATE2" 2>/dev/null; then
            echo "== $p already rebuilt, skipping =="; continue
        fi
        # Per package, not once at the start. The run is long enough that the
        # disk state at the end has nothing to do with the disk state when it
        # was checked, and the failure mode is not local to this script.
        space_ok || { echo "== stage 2: stopping, disk too low =="; break; }
        log "stage 2: $p"
        # watched, not a plain call: see bootstrap-pacman.sh. A hundred and
        # thirty-three packages is far too many to sit in front of, and one
        # silent build would hold the whole run.
        if watched "$WORK/stage2-log-$p.txt" stage2_build "$p"; then
            echo "$p" >> "$STATE2"; ok=$((ok + 1)); echo "OK    $p"
        else
            rc=$?
            fail=$((fail + 1)); failed+=("$p")
            if [ "$rc" -eq 2 ]; then
                echo "STALL $p  (no output for ${EL_STALL_MIN:-45} min, killed)"
            else
                echo "FAIL  $p  (see $WORK/stage2-log-$p.txt)"
            fi
            tail -5 "$WORK/stage2-log-$p.txt" | sed 's/^/      /'
        fi
    done
    echo
    echo "== stage 2: $ok rebuilt, $fail failed =="
    [ "$fail" -eq 0 ] || printf '  failed: %s\n' "${failed[*]}"
}

main() {
    require_space
    [ -f "$REPO1/core.db.tar.gz" ] || die "no stage-1 repository at $REPO1"
    trap umount_chroot EXIT
    make_rootfs
    configure_chroot
    mount_chroot
    smoke_test || die "the chroot cannot build; stage 2 stops here"
    log "Chroot ready"
    if [ "$#" -eq 0 ]; then
        echo "  enter it with:"
        echo "    sudo chroot --userspec=$BUILD_UID:$BUILD_GID $ROOT /usr/bin/bash -l"
        echo "  or rebuild packages:"
        echo "    bash $0 texinfo perl m4 autoconf ..."
        echo "    bash $0 --all      # all ${#STAGE1_PACKAGES[@]}, in order"
        return 0
    fi
    touch "$STATE2"
    # --all: the shared list, in its order. Typing a hundred and thirty-three
    # names is not a workflow, and typing a subset of them is how an ordering
    # gets quietly reinvented.
    if [ "$1" = "--all" ]; then
        rebuild "${STAGE2_FIRST[@]}" "${STAGE1_PACKAGES[@]}"
    else
        rebuild "$@"
    fi
}

main "$@"