# Copyright 1999-2026 Gentoo Authors # Distributed under the terms of the GNU General Public License v2 EAPI=8 CONFIG_CHECK="~ADVISE_SYSCALLS" PYTHON_COMPAT=( python3_{11..14} ) PYTHON_REQ_USE="threads(+)" # bash-completion-r1 is deliberately absent: its only consumer here was # newbashcomp, and the npm completion now goes inside the slot prefix (see # src_install) instead of into the shared /usr/share/bash-completion/completions # that eclass installs to. Nothing else in this ebuild uses shell-completion. inherit check-reqs flag-o-matic linux-info inherit ninja-utils pax-utils python-any-r1 toolchain-funcs xdg-utils DESCRIPTION="A JavaScript runtime built on Chrome's V8 JavaScript engine" HOMEPAGE="https://nodejs.org/" LICENSE="Apache-1.1 Apache-2.0 BlueOak-1.0.0 BSD BSD-2 MIT npm? ( Artistic-2 )" # The slot is this ebuild's major version. It is spelt as an unindented literal # at column 0, not as $(ver_cut 1), because the overlay's autoupdate engine # scrapes it with a plain text regex anchored at ^SLOT= -- it neither sources the # ebuild nor reads metadata/md5-cache, so an indented or computed value is # invisible to it and the per-slot record cannot resolve a current version. # scripts/check-slot-naming-contract.sh asserts this literal matches the major in # the filename, which is what a computed value used to guarantee: copy this file # to a new major and forget to change the line, and that check fails. SLOT="26" if [[ ${PV} == *9999 ]]; then inherit git-r3 EGIT_REPO_URI="https://github.com/nodejs/node" # The live ebuild stays unslotted, as ::gentoo has it. Assigned after the # literal above so it wins for 9999 while the scraper still sees the real # slot on the release branch. SLOT="0" else SRC_URI="https://nodejs.org/dist/v${PV}/node-v${PV}.tar.xz" KEYWORDS="~amd64 ~arm ~arm64 ~loong ~ppc64 ~riscv ~x86 ~x64-macos" S="${WORKDIR}/node-v${PV}" fi IUSE="cpu_flags_x86_sse2 debug doc +icu +inspector lto +npm pax-kernel pnpm +snapshot +ssl +system-icu +system-ssl test" REQUIRED_USE="inspector? ( icu ssl ) npm? ( ssl ) system-icu? ( icu ) system-ssl? ( ssl ) x86? ( cpu_flags_x86_sse2 )" RESTRICT="!test? ( test )" COMMON_DEPEND=">=app-arch/brotli-1.2.0:= dev-db/sqlite:3 >=dev-cpp/ada-3.4.4:= >=dev-libs/libuv-1.52.1:= >=dev-libs/simdjson-4.6.1:= >=net-dns/c-ares-1.34.6:= >=net-libs/nghttp2-1.69.0:= >=net-libs/nghttp3-1.15.0:= virtual/zlib:= system-icu? ( >=dev-libs/icu-73:= ) system-ssl? ( >=net-libs/ngtcp2-1.22.0:= >=dev-libs/openssl-3.5.6:0= ) !system-ssl? ( >=net-libs/ngtcp2-1.22.0:=[-gnutls] ) || ( sys-devel/gcc:* llvm-runtimes/libatomic-stub )" BDEPEND="${PYTHON_DEPS} app-alternatives/ninja sys-apps/coreutils virtual/pkgconfig test? ( net-misc/curl ) pax-kernel? ( sys-apps/elfix )" DEPEND="${COMMON_DEPEND}" # app-eselect/eselect-nodejs is a *runtime* dependency, not a build one: no # src_* phase touches it, but pkg_postinst and pkg_postrm run `eselect nodejs`. # It owns every unversioned path -- /usr/bin/{node,npm,npx}, /usr/include/node, # the shared bash-completion file, /usr/share/man/man1/node.1 -- plus /etc/npm # and /etc/env.d/50npm, which this ebuild used to install itself. Without it a # merged slot is reachable only as node26, never as plain `node`. # # The blocker is STRONG (!!) and it is the whole of R4.1/R4.4. net-libs/nodejs:0 # is the pre-slotting package: it installs those same unversioned paths # directly, so it and app-eselect/eselect-nodejs cannot coexist even for a # moment -- with FEATURES="protect-owned" the merge aborts on the first shared # path. A weak (!) blocker permits exactly that momentary coexistence, which is # why it is not enough here; !! makes portage unmerge the old package first and # surfaces the conflict during dependency resolution, before any file is # written. # # ":0" is a slot dependency, not a version range, so it needs no maintenance as # majors go by: it matches the unslotted package (SLOT="0", recorded as "0/26" # with its sub-slot) and matches no slotted one (SLOT="26"). Caveat for whoever # adds a live ebuild: the ${PV} == *9999 branch above still sets SLOT="0", and a # !! blocker is *not* ignored against a package in the parent's own slot # (depgraph.py checks `not blocker.atom.blocker.overlap.forbid`, false for !!), # so that branch would need its own SLOT before this atom is correct for it. # >=eselect-nodejs-2 is a correctness bound, not hygiene: module version 1 made # /usr/bin/node a symlink, and `readlink -f` on a symlink reports the slot's real # path rather than /usr/bin/node itself. Consumers that link to /usr/bin/node and # then verify with `readlink -f` compare two different strings and die - # www-client/chromium does exactly that in src_prepare. Version 2 writes an exec # wrapper, which resolves to itself. Pinning the floor here keeps a downgrade of # the module from silently reintroducing that failure in every such consumer. RDEPEND="${COMMON_DEPEND} >=app-eselect/eselect-nodejs-2 !!net-libs/nodejs:0" PDEPEND="pnpm? ( sys-apps/pnpm )" # These are measured on a loong machine with -ggdb on, and only checked # if debugging flags are present in CFLAGS. # # The final link consumed a little more than 7GiB alone, so 8GiB is the lower # limit for memory usage. Disk usage was 19.1GiB for the build directory and # 1.2GiB for the installed image, so we leave some room for architectures with # fatter binaries and set the disk requirement to 22GiB. CHECKREQS_MEMORY="8G" CHECKREQS_DISK_BUILD="22G" pkg_pretend() { if [[ ${MERGE_TYPE} != "binary" ]]; then if is-flagq "-g*" && ! is-flagq "-g*0" ; then einfo "Checking for sufficient disk space and memory to build ${PN} with debugging CFLAGS" check-reqs_pkg_pretend fi fi } pkg_setup() { python-any-r1_pkg_setup linux-info_pkg_setup # tools/install.py derives *every* destination from the prefix configure.py # was given -- bin/node, include/node/, lib/node_modules/npm/ and # share/man/man1/node.1 -- so one prefix per slot is the whole isolation # mechanism. dev-lang/php (PHP_DESTDIR) is the in-tree precedent. # # Relative to EPREFIX: configure.py gets "${EPREFIX}${NODE_SLOT_DIR}", # src_install writes into "${ED}${NODE_SLOT_DIR}". Keep it in sync with # app-eselect/eselect-nodejs, which finds slots by globbing # ${EROOT}/usr/lib*/node-[0-9]* for an executable bin/node. # # Assigned here, not in global scope: the package manager replaces # get_libdir() with a die stub during the depend phase, so a global # assignment breaks metadata generation. pkg_setup runs ahead of every # other phase and portage carries the value in the saved phase environment. NODE_SLOT_DIR="/usr/$(get_libdir)/node-${SLOT}" } src_prepare() { tc-export AR CC CXX PKG_CONFIG export V=1 export BUILDTYPE=Release # fix compilation on Darwin # https://code.google.com/p/gyp/issues/detail?id=260 # # The pattern matters: the file spells it append("-arch") on one line with # the architecture on the next, so the old single-quoted /append('-arch/d # matched nothing at all and deleted neither line -- verified against this # tarball. An anchored sed that matches nothing still exits 0, so `|| die` # never saw it. ::gentoo corrected this in its own 26.3.0. sed -i -e '/append("-arch")/{N;d;}' tools/gyp/pylib/gyp/xcode_emulation.py || die # DELIBERATELY ABSENT -- do not reintroduce the ::gentoo libdir seds: # sed -i -e "s|lib/|$(get_libdir)/|g" tools/install.py # sed -i -e "s/'lib'/'$(get_libdir)'/" deps/npm/lib/npm.js # (hat tip @ryanpcmcquen https://github.com/iojs/io.js/issues/504) # # They exist only because ::gentoo installs into the shared /usr, where # "lib" would be wrong on a lib64 profile. This ebuild gives every slot its # own prefix (${NODE_SLOT_DIR}), so /lib is already unambiguous -- # nothing else lives there to collide with. # # Upstream keeps both sides of the global npm root in agreement already: # tools/install.py:91 os.path.join('lib/node_modules', name) # deps/npm/lib/npm.js:492 resolve(globalPrefix, 'lib', 'node_modules') # so the two paths match with no rewriting. Rewriting one and not the other # makes `npm root -g` point at an empty directory with no error, which is # why these must be deleted or kept as a pair, never merged back piecemeal. # src_install's LIBDIR follows this choice and must move with it. # # The install.py sed was over-broad besides: s|lib/|lib64/|g also rewrote # 'deps/zlib/zconf.h' and 'include/node/zoslib/' (harmless only because we # build --shared-zlib on non-z/OS). # Avoid writing a depfile, not useful sed -i -e "/DEPFLAGS =/d" tools/gyp/pylib/gyp/generator/make.py || die sed -i -e "/'-O3'/d" common.gypi node.gypi || die # debug builds. change install path, remove optimisations and override buildtype if use debug; then sed -i -e "s|out/Release/|out/Debug/|g" tools/install.py || die BUILDTYPE=Debug fi # We need to disable mprotect on two files when it builds Bug 694100. # # A 26.5.1-specific copy, NOT the shared ${PN}-24.1.0-paxmarking.patch: # against this tarball that file applies with "fuzz 1" in both node.gyp # (its context still names src/node_webstorage.h, which 26 replaced with # src/ffi/types.h) and tools/v8_gypfiles/v8.gyp. Fuzz cannot be sat on -- # portage's __eapply_patch greps its own output for "with fuzz", prints it # and still returns SUCCESS, so a hunk that drifts to the wrong site ships # silently and only breaks at compile time, only for USE=pax-kernel users. # The rebased copy applies with --fuzz=0 and produces a byte-identical # result to the original. use pax-kernel && PATCHES+=( "${FILESDIR}"/${PN}-26.5.1-paxmarking.patch ) use ppc64 && PATCHES+=( "${FILESDIR}/${PN}-24.11.1-restore-ppc64be.patch" ) default } src_configure() { xdg_environment_reset # LTO compiler flags are handled by configure.py itself filter-lto # The warnings are *so* noisy and make build.logs massive append-cxxflags $(test-flags-CXX -Wno-template-id-cdtor) # https://bugs.gentoo.org/931514 use arm64 && append-flags $(test-flags-CXX -mbranch-protection=none) local myconf=( --ninja --shared-ada --shared-brotli --shared-cares --shared-libuv --shared-nghttp2 --shared-nghttp3 --shared-ngtcp2 --shared-simdjson --shared-sqlite --shared-zlib ) use debug && myconf+=( --debug ) use lto && myconf+=( --enable-lto ) if use system-icu; then myconf+=( --with-intl=system-icu ) elif use icu; then myconf+=( --with-intl=full-icu ) else myconf+=( --with-intl=none ) fi use inspector || myconf+=( --without-inspector ) use npm || myconf+=( --without-npm ) use snapshot || myconf+=( --without-node-snapshot ) if use ssl; then use system-ssl && myconf+=( --shared-openssl --openssl-use-def-ca-store ) else myconf+=( --without-ssl ) fi local myarch="" case "${ARCH}:${ABI}" in *:amd64) myarch="x64";; *:arm) myarch="arm";; *:arm64) myarch="arm64";; loong:lp64*) myarch="loong64";; riscv:lp64*) myarch="riscv64";; *:ppc64) myarch="ppc64";; *:x32) myarch="x32";; *:x86) myarch="ia32";; *) myarch="${ABI}";; esac GYP_DEFINES="linux_use_gold_flags=0 linux_use_bundled_binutils=0 linux_use_bundled_gold=0" \ "${EPYTHON}" configure.py \ --prefix="${EPREFIX}${NODE_SLOT_DIR}" \ --dest-cpu=${myarch} \ "${myconf[@]}" || die } src_compile() { eninja -C out/Release } src_install() { # npm's tree lands inside the slot prefix at /lib -- the literal # string "lib", NOT $(get_libdir). That is what tools/install.py:91 writes # and what deps/npm/lib/npm.js:492 resolves at runtime, now that # src_prepare no longer rewrites either (see the note there before # "correcting" this to $(get_libdir) -- the two must move together). local LIBDIR="${ED}${NODE_SLOT_DIR}/lib" default pax-mark -m "${ED}${NODE_SLOT_DIR}"/bin/node # set up a symlink structure that node-gyp expects.. # Inside the slot prefix: /usr/include/node belongs to eselect nodejs. dodir "${NODE_SLOT_DIR}"/include/node/deps/{v8,uv} dosym . "${NODE_SLOT_DIR}"/include/node/src for var in deps/{uv,v8}/include; do dosym ../.. "${NODE_SLOT_DIR}"/include/node/${var} done # Version-suffixed entry points (R1.4). These are the only /usr/bin paths # this ebuild owns: the unsuffixed node/npm/npx belong to # app-eselect/eselect-nodejs, which repoints them on every switch. Both # layers exist so a user can pin one major (node26) while `node` keeps # following whatever was selected. # # dosym -r (EAPI 8) derives the *relative* target from the two absolute # paths -- /usr/bin/node26 -> ../lib64/node-26/bin/node -- instead of a # hand-spelt "../$(get_libdir)/node-${SLOT}/bin/node". Same link value, # but it stays derived from ${NODE_SLOT_DIR}, the one place the prefix is # defined, so the two cannot drift; and it is exactly what link_target() # in the eselect module computes for the unversioned links. Relative is # not cosmetic: with ROOT set (an offline install) an absolute target # names a path on the *build host* and points nowhere once that tree is # booted into. # # npm and npx are conditional because install.py creates bin/npm and # bin/npx only when npm is installed -- an unconditional dosym would leave # two dangling links in /usr/bin under USE=-npm. create_symlinks() in the # eselect module skips the same two paths for the same reason. local versioned_entries=( node ) if use npm; then versioned_entries+=( npm npx ) fi local entry for entry in "${versioned_entries[@]}"; do dosym -r "${NODE_SLOT_DIR}/bin/${entry}" "/usr/bin/${entry}${SLOT}" done if use doc; then docinto html dodoc -r "${S}"/doc/* fi if use npm; then # Pin npm and npx to *this* slot's interpreter. Upstream ships # "#!/usr/bin/env node", which resolves through PATH -- so npm26 would # execute slot 26's npm code under whichever slot eselect happens to # have activated. That is not cosmetic: npm derives its global root # from dirname(dirname(process.execPath)), so under a foreign # interpreter `npm root -g` reports the *other* slot's tree while # loading this slot's code. install.py symlinks bin/{npm,npx} to these # two files, so they are the only entry points that need it (the third # "#!/usr/bin/env node" under npm/bin/, npm-prefix.js, is reachable # only from the cygwin/mingw bin/npm wrapper, which already passes an # explicit interpreter; bin/node-gyp is created on z/OS only). # # Anchored with 1s and ^...$ so it cannot match any other line. It is # verified afterwards because an anchored sed that matches NOTHING # still exits 0 -- `|| die` cannot see the silent no-op that an # upstream shebang change would cause, and the resulting cross-slot # mix would only surface at runtime. local npm_bindir="${ED}${NODE_SLOT_DIR}/lib/node_modules/npm/bin" local slot_shebang="#!${EPREFIX}${NODE_SLOT_DIR}/bin/node" local npm_cli for npm_cli in npm npx; do sed -i -e "1s|^#!/usr/bin/env node\$|${slot_shebang}|" \ "${npm_bindir}/${npm_cli}-cli.js" || die "sed failed on ${npm_cli}-cli.js" [[ $(head -n 1 "${npm_bindir}/${npm_cli}-cli.js") == "${slot_shebang}" ]] || die "${npm_cli}-cli.js line 1 is not ${slot_shebang} -- upstream shebang changed?" done # DELIBERATELY ABSENT -- do not reintroduce: # keepdir /etc/npm # echo "NPM_CONFIG_GLOBALCONFIG=..." > "${T}"/50npm; doenvd "${T}"/50npm # Neither path is slot-specific, so two installed slots would own the # same two files and FEATURES="protect-owned" would turn that into a # merge failure. app-eselect/eselect-nodejs, which is single-instance # by construction, owns them now -- that is what keeps # NPM_CONFIG_GLOBALCONFIG naming one shared npmrc for every slot # (R7.2) while the runtime itself is slotted. # Bash completion for `npm`, installed inside the slot prefix rather # than into the shared /usr/share/bash-completion/completions that # newbashcomp targets -- same collision argument as /etc/npm above and # as npm's manpages below. app-eselect/eselect-nodejs links the shared # path at the active slot's copy on every switch. # # The destination is a CONTRACT with that module, not a preference. Its # NODEJS_LINKS carries the pair # /usr/share/bash-completion/completions/npm # |share/bash-completion/completions/npm # whose right-hand side is relative to the slot prefix, and # create_symlinks() *skips* an entry whose source is missing -- that # branch exists so a USE=-npm slot leaves no dangling link behind. So a # file installed one directory off raises no error anywhere: `npm ` # simply stays dead. Move one spelling and you must move the other. local tmp_npm_completion_file="$(TMPDIR="${T}" mktemp -t npm.XXXXXXXXXX)" # Name the interpreter instead of letting bin/npm's shebang pick it: # that shebang now points at ${EPREFIX}${NODE_SLOT_DIR}/bin/node, which # does not exist yet -- the binary is still staged under ${ED} -- so # going through it would fail with "bad interpreter". Being explicit # also decouples this build-time call from shebang state entirely, so # reordering src_install cannot silently reintroduce that breakage. # It runs the freshly built node on the build host, as the build # already does for the V8 snapshot. # # The `|| die` is load-bearing and not decoration: the redirect creates # the file before node ever runs, so a failing invocation still leaves a # 0-byte completion behind -- and a 0-byte completion installs, sources # and reports nothing. The only symptom would be a tab key that does # nothing, months later. "${ED}${NODE_SLOT_DIR}"/bin/node "${npm_bindir}"/npm-cli.js completion \ > "${tmp_npm_completion_file}" || die "generating npm completion failed" # The body of newbashcomp, verbatim apart from the destination: # subshell so insinto/insopts cannot leak into a later phase, and an # explicit 0644 so a stray insopts earlier in src_install cannot make # the completion executable or unreadable. ( insopts -m 0644 insinto "${NODE_SLOT_DIR}"/share/bash-completion/completions newins "${tmp_npm_completion_file}" npm ) # npm's manpages go inside the slot prefix, NOT into the shared # /usr/share/man that `doman` targets. Two installed slots would each # write share/man/man1/npm-install.1 to that one path -- a collision, # and with FEATURES="protect-owned" a merge failure. Inside the prefix # they are slot-owned, and app-eselect/eselect-nodejs's single MANPATH # entry in /etc/env.d/50nodejs names the *active* slot's share/man, # covering man1, man5 and man7 at once (R7.3). Accepted consequence: # `man npm-install` resolves for the active slot only -- there is no # way to spell `man npm-install26`. dev-lang/php does the same, with # /usr/lib64/php8.5/man on MANPATH. # # Moved rather than copied, then linked back: npm reads these pages # from its own tree too -- help.js:63 globs # resolve(npmRoot, "man//?(npm-).[0-9]*") -- so `npm help # install` needs /lib/node_modules/npm/man to keep resolving. # ::gentoo's doman kept that working by accident, because doman copies. # One directory symlink keeps both readers on a single copy that # cannot drift, instead of 709K of duplicated pages. local npm_mandir="${ED}${NODE_SLOT_DIR}/lib/node_modules/npm/man" local mansec for mansec in 1 5 7; do # Loud on an upstream reorganisation: quietly shipping fewer # sections would regress R7.3 with nothing to notice it. [[ -d ${npm_mandir}/man${mansec} ]] || die "npm ships no man${mansec} -- upstream man layout changed?" # man1 already exists -- install.py:204 put node.1 there -- while # man5 and man7 do not. No npm page is named node.1, so the merge # clobbers nothing; asserted after the loop rather than assumed. dodir "${NODE_SLOT_DIR}/share/man/man${mansec}" mv "${npm_mandir}/man${mansec}"/* \ "${ED}${NODE_SLOT_DIR}/share/man/man${mansec}"/ || die "moving npm man${mansec} pages failed" rmdir "${npm_mandir}/man${mansec}" || die "npm man${mansec} still has entries after the move" done [[ -f "${ED}${NODE_SLOT_DIR}/share/man/man1/node.1" ]] || die "node.1 is gone -- an npm page collided with it" # dosym refuses to write over a real directory (and plain `ln -sfn` # would silently nest the link inside it), so emptying it first is # what makes the symlink legal, not just tidy. rmdir "${npm_mandir}" || die "unexpected leftovers in ${npm_mandir}" dosym -r "${NODE_SLOT_DIR}/share/man" \ "${NODE_SLOT_DIR}/lib/node_modules/npm/man" # Clean up rm -f "${LIBDIR}"/node_modules/npm/{.mailmap,.npmignore,Makefile} local find_exp="-or -name" local find_name=() for match in "AUTHORS*" "CHANGELOG*" "CONTRIBUT*" "README*" \ ".travis.yml" ".eslint*" ".wercker.yml" ".npmignore" \ "*.bat" "*.cmd"; do find_name+=( ${find_exp} "${match}" ) done # Remove various development and/or inappropriate files and # useless docs of dependend packages. find "${LIBDIR}"/node_modules \ \( -type d -name examples \) -or \( -type f \( \ -iname "LICEN?E*" \ "${find_name[@]}" \ \) \) -exec rm -rf "{}" \; fi # install.py drops gdbinit/lldb_commands.py into /share/doc/node; # /usr/share/doc/${PF} is already unique per slot, so they stay there. mv "${ED}${NODE_SLOT_DIR}"/share/doc/node "${ED}"/usr/share/doc/${PF} || die } src_test() { local drop_tests=( test/parallel/test-dns.js test/parallel/test-dns-resolveany-bad-ancount.js test/parallel/test-dns-setserver-when-querying.js test/parallel/test-dotenv.js test/parallel/test-fs-mkdir.js test/parallel/test-fs-read-stream.js test/parallel/test-fs-utimes-y2K38.js test/parallel/test-fs-watch-recursive-add-file.js test/parallel/test-http2-client-set-priority.js test/parallel/test-http2-priority-event.js test/parallel/test-process-euid-egid.js test/parallel/test-process-get-builtin.mjs test/parallel/test-process-initgroups.js test/parallel/test-process-setgroups.js test/parallel/test-process-uid-gid.js test/parallel/test-release-npm.js test/parallel/test-socket-write-after-fin-error.js test/parallel/test-strace-openat-openssl.js test/sequential/test-tls-session-timeout.js test/sequential/test-util-debug.js # Breaking change in nghttp2 1.67.1, see # https://github.com/nodejs/node/issues/60661 # https://github.com/nodejs/node/commit/b426a47c05e6b039ed65798d0ad3b3698b35fd97 # https://github.com/nghttp2/nghttp2/issues/2604 test/parallel/test-http2-client-unescaped-path.js test/parallel/test-http2-multi-content-length.js test/parallel/test-http2-reset-flood.js test/parallel/test-http2-max-invalid-frames.js test/parallel/test-http2-misbehaving-flow-control.js test/parallel/test-http2-misbehaving-flow-control-paused.js ) # https://bugs.gentoo.org/963649 has_version '>=dev-libs/openssl-3.6' && drop_tests+=( test/parallel/test-tls-ocsp-callback ) use inspector || drop_tests+=( test/parallel/test-inspector-emit-protocol-event.js test/parallel/test-inspector-network-arbitrary-data.js test/parallel/test-inspector-network-domain.js test/parallel/test-inspector-network-fetch.js test/parallel/test-inspector-network-http.js test/sequential/test-watch-mode.mjs ) rm -f "${drop_tests[@]}" || die "disabling tests failed" out/${BUILDTYPE}/cctest || die "${EPYTHON}" tools/test.py --mode=${BUILDTYPE,,} --flaky-tests=dontcare -J message parallel sequential || die } # The default-slot policy (R3.5, R3.6), in two rules: # # nothing active -> activate this slot. A merged runtime with no # /usr/bin/node is broken for every consumer. # something active -> never change it. Speak up only when the selection can # actually break a build. # # Why not simply "always activate the newest": all 72 references to this # package across ::gentoo and this overlay are UNSLOTTED atoms # (net-libs/nodejs, >=net-libs/nodejs-N), so every one of them resolves through # /usr/bin/node, the link app-eselect/eselect-nodejs owns. Switching silently # would break someone who pinned an older major on purpose; saying nothing # would let www-client/chromium's >=net-libs/nodejs-${NODE_VER} fail with # nothing pointing at the active slot as the cause. The elog is the # compromise: the user is told and decides. # # No --root=: eselect derives its EROOT from the ROOT and EPREFIX that portage # already exports into pkg_* phases (/usr/bin/eselect, "EROOT=${ROOT%...}..." # near the top), so these calls are ROOT-scoped without the flag. Passing # --root= would additionally set ROOT as a variable *inside* eselect that is # never exported, which the `env update` do_set ends with would not see -- # app-editors/emacs passes it, but its module has no such tail call. pkg_postinst() { local active active=$(eselect nodejs show) if [[ -z ${active} ]]; then # `show` prints zero bytes -- and exits 0 -- for all three unusable # states: no /usr/bin/node, a /usr/bin/node that is not a symlink (the # leftover of the unslotted net-libs/nodejs:0), and a link into a slot # that has been unmerged. That is why the branch is on the *output* # and never on the exit status; see describe_show in the module. # # Status-checked and warned rather than `die`d: a `die -q` inside an # eselect module exits 250, and an unchecked call as the last command # of this phase would hand that to portage as the phase's result. By # postinst the files are already merged, so failing here would report # a broken install for what is one `eselect nodejs set` away from # fixed. if eselect nodejs set "node${SLOT}"; then elog "No Node.js slot was active, so ${PN}:${SLOT} is now the default." elog "Change it at any time with: eselect nodejs set " else ewarn "Could not activate Node.js slot ${SLOT}." ewarn "/usr/bin/node is still unset, so anything invoking plain" ewarn "\`node\` will fail. This slot does run as \`node${SLOT}\`." ewarn "Once the cause is fixed, activate it with:" ewarn " eselect nodejs set node${SLOT}" fi else # ${active} is a slot NAME (node24) while ${SLOT} is a bare major (26), # so the two majors are compared NUMERICALLY. A string compare would # put node9 above node26 and fire this message backwards -- the same # trap find_targets() avoids with `sort -V`. 10# forces base 10, so a # hypothetical node08 is 8 rather than an octal syntax error, and the # ^[0-9]+$ guard means a name that is not node falls through to # doing nothing, which is the safe direction: never hijack. local active_major=${active#node} if [[ ${active_major} =~ ^[0-9]+$ ]] && (( 10#${active_major} < 10#${SLOT} )); then elog "Node.js slot ${SLOT} is newer than the active slot (${active}), which" elog "was left alone: overriding a deliberate choice is not this" elog "ebuild's call." elog elog "It is worth saying because no consumer depends on a *slot* of" elog "net-libs/nodejs -- they all use plain atoms and so resolve" elog "through /usr/bin/node, which still runs ${active}. One that needs" elog "${SLOT} (www-client/chromium carries >=net-libs/nodejs-)" elog "would fail with nothing naming the active slot as the cause." elog "To switch:" elog " eselect nodejs set node${SLOT}" elog "Either way this slot is always reachable as \`node${SLOT}\`." fi # Deliberately silent in the other two directions: # # active == this slot nothing changed. A line here would print on # every routine upgrade of the active slot. # active > this slot no hazard exists. An active node # already satisfies every bound this slot could, # so no unslotted consumer can break; installing # an older major alongside a newer one is a # deliberate act, and it runs as `node${SLOT}`. # # Both would recur on every upgrade of a slot the user is content with, # and a message people learn to skip is worse than no message: it would # dilute the one above, which reports a real problem. dev-lang/php's # pkg_postinst elogs in both directions; this is the deliberate # difference. fi if use npm; then ewarn "remember to run: source /etc/profile if you plan to use nodejs" ewarn " in your current shell" fi } # R3.4's repair path. `cleanup` is not merely "remove what dangles": once it # finds any dead managed link it drops the whole managed set, then re-points it # at the highest slot still installed -- or, when this was the last slot, takes # the links and /etc/env.d/50nodejs away. # # Correct on a re-emerge too, which is worth spelling out because this phase # runs then as well -- portage unmerges the OLD instance from inside the new # one's merge. Verified in portage 3.0's dbapi/vartree.py: _merge_contents() # (line ~5126) puts the new files on disk BEFORE the old instance is unmerged # (~5201), and that unmerge is handed `others_in_slot`, so _unmerge_pkgfiles() # skips every path the new instance owns (isowner(), ~2902). Nothing under # /usr/lib*/node-${SLOT} disappears, nothing dangles, and cleanup returns # without touching the selection -- one "No dangling Node.js links found." # line. # # So the call is unconditional rather than gated on ${REPLACED_BY_VERSION}, # which would skip the one same-slot case that does need it: a rebuild that # drops USE=npm leaves /usr/bin/npm pointing at a file the new instance does # not install, and this is what repairs it. # # Its output is deliberately NOT suppressed -- eselect's --brief would hide # exactly the "No dangling Node.js links found." line while keeping the # ones that report a change. An unmerge that silently rearranges /usr/bin/node # is the failure mode this design exists to avoid, so the healthy case saying # so out loud is what tells "checked, fine" apart from "never ran". # # Status-checked and warned, never `die`d, for pkg_postinst's reason plus one # more: app-eselect/eselect-nodejs is an RDEPEND, so a depclean is free to # unmerge it before this package. `eselect nodejs cleanup` then exits non-zero # because the module is gone, and as the last command of this phase that would # fail the unmerge of an already-unmerged package. dev-lang/php's pkg_postrm # carries the same caveat as a comment, without the check. pkg_postrm() { if ! eselect nodejs cleanup; then ewarn "\`eselect nodejs cleanup\` failed." ewarn "/usr/bin/node and its siblings may still point into the slot" ewarn "that was just removed, which makes them fail at exec time" ewarn "rather than at lookup time. Once app-eselect/eselect-nodejs is" ewarn "installed, repair them with:" ewarn " eselect nodejs cleanup" fi }