unix: initial support for building armv7-unknown-linux-gnueabihf

This commit teaches the build system to cross-compile for
armv7-unknown-linux-gnueabihf.

In order to get cross-compiles working, we introduce a new
Docker environment based on Debian Stretch. This environment has
cross toolchains readily available, unlike Jessie.

I'm not keen on using Debian Stretch here or the built-in toolchains
because I would prefer to compile with a modern toolchain we build
ourselves. And I worry about things like glibc symbol versions from
Stretch being too new. But this does seem to be the path of least
resistance for getting ARM cross-compiles working. The way I see it,
if we support creating the builds, others can run them and report
issues.

There are likely many problems with the builds. I don't believe
I can run these binaries locally, so I have no way of testing.
Although I could probably use an emulator. I may do that later.

Author: indygreg

cpython-unix/build-cpython.sh

@@ -28,6 +28,9 @@ tar -xf pip-${PIP_VERSION}.tar.gz
 if [ "${BUILD_TRIPLE}" != "${TARGET_TRIPLE}" ]; then
   pushd "Python-${PYTHON_VERSION}"
 
+  OLD_CC=${CC}
+  unset CC
+
   # When cross-compiling, we need to build a host Python that has working zlib
   # and ctypes extensions, otherwise various things fail. (`make install` fails
   # without zlib and setuptools / pip used by target install fail due to missing
@@ -70,6 +73,8 @@ if [ "${BUILD_TRIPLE}" != "${TARGET_TRIPLE}" ]; then
   # at the front of PATH.
   export PATH="${TOOLS_PATH}/pyhost/bin:${PATH}"
 
+  export CC="${OLD_CC}"
+
   popd
   # Nuke and re-pave the source directory out of paranoia.
   rm -rf "Python-${PYTHON_VERSION}"
@@ -587,6 +592,11 @@ if [ "${BUILD_TRIPLE}" != "${TARGET_TRIPLE}" ]; then
             CONFIGURE_FLAGS="${CONFIGURE_FLAGS} ac_cv_file__dev_ptc=no"
             CONFIGURE_FLAGS="${CONFIGURE_FLAGS} ac_cv_file__dev_ptmx=no"
             ;;
+        armv7-unknown-linux-gnueabihf)
+            CONFIGURE_FLAGS="${CONFIGURE_FLAGS} ac_cv_buggy_getaddrinfo=no"
+            CONFIGURE_FLAGS="${CONFIGURE_FLAGS} ac_cv_file__dev_ptc=no"
+            CONFIGURE_FLAGS="${CONFIGURE_FLAGS} ac_cv_file__dev_ptmx=no"
+            ;;
         x86_64-unknown-linux-musl)
             ;;
         *)

cpython-unix/build-libX11.sh

@@ -48,19 +48,30 @@ if [ "${BUILD_TRIPLE}" != "${TARGET_TRIPLE}" ]; then
     i686-unknown-linux-gnu)
       EXTRA_FLAGS="${EXTRA_FLAGS} --enable-malloc0returnsnull"
       ;;
-
+    armv7-unknown-linux-gnueabihf)
+      EXTRA_FLAGS="${EXTRA_FLAGS} --enable-malloc0returnsnull"
+      ;;
     *)
       echo "cross-compiling but malloc(0) override not set; failures possible"
       ;;
   esac
 fi
 
+case "${TARGET_TRIPLE}" in
+  armv7-unknown-linux-gnueabihf)
+    CC_FOR_BUILD=gcc
+    ;;
+  *)
+    CC_FOR_BUILD=clang
+    ;;
+esac
+
 # CC_FOR_BUILD is here because configure doesn't look for `clang` when
 # cross-compiling. So we force it.
 CFLAGS="${EXTRA_TARGET_CFLAGS} -fPIC -I/tools/deps/include" \
   CPPFLAGS="${EXTRA_TARGET_CFLAGS} -fPIC -I/tools/deps/include" \
   LDFLAGS="${EXTRA_TARGET_LDFLAGS}" \
-  CC_FOR_BUILD=clang \
+  CC_FOR_BUILD="${CC_FOR_BUILD}" \
   CFLAGS_FOR_BUILD="-I/tools/deps/include" \
   CPPFLAGS_FOR_BUILD="-I/tools/deps/include" \
   ./configure \

cpython-unix/build.cross.Dockerfile

@@ -0,0 +1,50 @@
+# Debian Stretch.
+FROM debian@sha256:d0b7b71db141cedc48e1c2807d12b199ffd7ffe75baf272a34c37480dc2159d1
+MAINTAINER Gregory Szorc <[email protected]>
+
+RUN groupadd -g 1000 build && \
+    useradd -u 1000 -g 1000 -d /build -s /bin/bash -m build && \
+    mkdir /tools && \
+    chown -R build:build /build /tools
+
+ENV HOME=/build \
+    SHELL=/bin/bash \
+    USER=build \
+    LOGNAME=build \
+    HOSTNAME=builder \
+    DEBIAN_FRONTEND=noninteractive
+
+CMD ["/bin/bash", "--login"]
+WORKDIR '/build'
+
+RUN for s in debian_stretch debian_stretch-updates debian-security_stretch/updates; do \
+      echo "deb http://snapshot.debian.org/archive/${s%_*}/20210223T023121Z/ ${s#*_} main"; \
+    done > /etc/apt/sources.list && \
+    ( echo 'quiet "true";'; \
+      echo 'APT::Get::Assume-Yes "true";'; \
+      echo 'APT::Install-Recommends "false";'; \
+      echo 'Acquire::Check-Valid-Until "false";'; \
+      echo 'Acquire::Retries "5";'; \
+    ) > /etc/apt/apt.conf.d/99cpython-portable
+
+RUN apt-get update
+
+# Host building.
+RUN apt-get install \
+    gcc \
+    libc6-dev \
+    libffi-dev \
+    make \
+    patch \
+    patchelf \
+    perl \
+    pkg-config \
+    tar \
+    xz-utils \
+    unzip \
+    zlib1g-dev
+
+# Cross-building.
+RUN apt-get install \
+    gcc-arm-linux-gnueabihf \
+    libc6-dev-armhf-cross

cpython-unix/build.py

@@ -94,6 +94,10 @@ def add_target_env(env, build_platform, target_triple, build_env):
             env["CC"] = "musl-clang"
         else:
             env["CC"] = "clang"
+    elif target_triple == "armv7-unknown-linux-gnueabihf":
+        env["CC"] = "arm-linux-gnueabihf-gcc"
+    else:
+        raise Exception("unhandled target triple: %s" % target_triple)
 
     env["PYBUILD_PLATFORM"] = build_platform
     env["TOOLS_PATH"] = build_env.tools_path
@@ -106,7 +110,9 @@ def add_target_env(env, build_platform, target_triple, build_env):
     if build_platform == "linux64":
         env["BUILD_TRIPLE"] = "x86_64-unknown-linux-gnu"
 
-        if target_triple in ("x86_64-unknown-linux-gnu", "x86_64-unknown-linux-musl"):
+        if target_triple == "armv7-unknown-linux-gnueabihf":
+            env["TARGET_TRIPLE"] = "armv7-unknown-linux-gnueabihf"
+        elif target_triple in ("x86_64-unknown-linux-gnu", "x86_64-unknown-linux-musl"):
             env["TARGET_TRIPLE"] = "x86_64-unknown-linux-gnu"
         elif target_triple == "i686-unknown-linux-gnu":
             env["TARGET_TRIPLE"] = "i686-unknown-linux-gnu"

cpython-unix/static-modules.3.9.armv7-unknown-linux-gnueabihf

@@ -0,0 +1,43 @@
+# Setup.dist doesn't have entries for all modules. This file defines
+# what's missing. The content here is reconstructed from logic in
+# setup.py and what was observed to execute in a normal build via setup.py.
+# We should audit this every time we upgrade CPython.
+
+_bz2 _bz2module.c -lbz2
+_crypt _cryptmodule.c -lcrypt
+_ctypes _ctypes/_ctypes.c _ctypes/callbacks.c _ctypes/callproc.c _ctypes/stgdict.c _ctypes/cfield.c -I/tools/deps/include -L/tools/deps/lib -lffi -ldl
+_ctypes_test _ctypes/_ctypes_test.c -lm
+_curses _cursesmodule.c -DHAVE_NCURSESW=1 -I/tools/deps/include/ncursesw -L/tools/deps/lib -lncursesw
+_curses_panel _curses_panel.c -DHAVE_NCURSESW=1 -I/tools/deps/include/ncursesw -L/tools/deps/lib -lpanelw -lncursesw
+_dbm _dbmmodule.c -DHAVE_BERKDB_H -DDB_DBM_HSEARCH -I/tools/deps/include -L/tools/deps/lib -ldb
+_decimal _decimal/_decimal.c _decimal/libmpdec/basearith.c _decimal/libmpdec/constants.c _decimal/libmpdec/context.c _decimal/libmpdec/convolute.c _decimal/libmpdec/crt.c _decimal/libmpdec/difradix2.c _decimal/libmpdec/fnt.c _decimal/libmpdec/fourstep.c _decimal/libmpdec/io.c _decimal/libmpdec/mpalloc.c _decimal/libmpdec/mpdecimal.c _decimal/libmpdec/numbertheory.c _decimal/libmpdec/sixstep.c _decimal/libmpdec/transpose.c -DCONFIG_32=1 -DANSI=1 -IModules/_decimal/libmpdec
+_elementtree _elementtree.c -DHAVE_EXPAT_CONFIG_H=1 -DXML_POOR_ENTROPY=1 -DUSE_PYEXPAT_CAPI -IModules/expat
+_gdbm _gdbmmodule.c -DHAVE_NDBM_H -I/tools/deps/include -L/tools/deps/lib -lgdbm
+_hashlib _hashopenssl.c -I/tools/deps/include -L/tools/deps/lib -lssl -lcrypto
+_json _json.c
+_lsprof _lsprof.c rotatingtree.c
+_lzma _lzmamodule.c -I/tools/deps/include -L/tools/deps/lib -llzma
+# TODO check setup.py logic for semaphore.c and possibly fix missing
+# dependency.
+_multiprocessing _multiprocessing/multiprocessing.c _multiprocessing/semaphore.c
+_opcode _opcode.c
+_posixshmem _multiprocessing/posixshmem.c -IModules/_multiprocessing -lrt
+_queue _queuemodule.c
+_sqlite3 _sqlite/cache.c _sqlite/connection.c _sqlite/cursor.c _sqlite/microprotocols.c _sqlite/module.c _sqlite/prepare_protocol.c _sqlite/row.c _sqlite/statement.c _sqlite/util.c -I/tools/deps/include -IModules/_sqlite -DMODULE_NAME=\"sqlite3\" -DSQLITE_OMIT_LOAD_EXTENSION=1 -L/tools/deps/lib -lsqlite3
+_ssl _ssl.c -I/tools/deps/include -lssl -lcrypto
+_testbuffer _testbuffer.c
+_testimportmultiple _testimportmultiple.c
+_testinternalcapi _testinternalcapi.c -DPy_BUILD_CORE_MODULE
+_testmultiphase _testmultiphase.c
+_tkinter _tkinter.c tkappinit.c -DWITH_APPINIT -I/tools/deps/include/X11 -L/tools/deps/lib -ltcl8.6 -ltk8.6 -lX11 -lxcb -lXau
+_uuid _uuidmodule.c -I/tools/deps/include/uuid -luuid
+_xxsubinterpreters _xxsubinterpretersmodule.c
+_xxtestfuzz _xxtestfuzz/_xxtestfuzz.c _xxtestfuzz/fuzzer.c
+ossaudiodev ossaudiodev.c
+pyexpat pyexpat.c expat/xmlparse.c expat/xmlrole.c expat/xmltok.c -DHAVE_EXPAT_CONFIG_H=1 -DXML_POOR_ENTROPY=1 -DUSE_PYEXPAT_CAPI -IModules/expat
+# readline variant needs to come first because libreadline is in /tools/deps and is
+# picked up by build. We /could/ make libedit first. But since we employ a hack to
+# coerce use of libedit on Linux, it seems prudent for the build system to pick
+# up readline.
+readline VARIANT=readline readline.c -I/tools/deps/include -I/tools/deps/include/ncursesw -L/tools/deps/lib -lreadline -lncursesw
+readline VARIANT=libedit readline-libedit.c -DUSE_LIBEDIT=1 -I/tools/deps/libedit/include -I/tools/deps/libedit/include/ncursesw -L/tools/deps/libedit/lib -ledit -lncursesw

cpython-unix/targets.yml

@@ -57,6 +57,32 @@ arm64-apple-tvos:
     - zlib
   openssl_target: todo
 
+armv7-unknown-linux-gnueabihf:
+  host_platforms:
+    - linux
+  docker_image_suffix: .cross
+  needs:
+    - bdb
+    - binutils
+    - bzip2
+    - gdbm
+    - libedit
+    - libffi
+    - libX11
+    - libXau
+    - libxcb
+    - ncurses
+    - readline
+    - sqlite
+    - tcl
+    - tk
+    - tix
+    - uuid
+    - xorgproto
+    - xz
+    - zlib
+  openssl_target: linux-armv4
+
 i686-unknown-linux-gnu:
   host_platforms:
     - linux

cpython-unix/xcb.cross.Dockerfile

@@ -0,0 +1,3 @@
+{% include 'build.cross.Dockerfile' %}
+RUN apt-get install \
+    python

pythonbuild/utils.py

@@ -118,8 +118,10 @@ def write_triples_makefiles(targets, dest_dir: pathlib.Path):
         else:
             lines.append("NEED_OPENSSL := 1\n")
 
-        lines.append("DOCKER_IMAGE_BUILD := build\n")
-        lines.append("DOCKER_IMAGE_XCB := xcb\n")
+        image_suffix = settings.get("docker_image_suffix", "")
+
+        lines.append("DOCKER_IMAGE_BUILD := build%s\n" % image_suffix)
+        lines.append("DOCKER_IMAGE_XCB := xcb%s\n" % image_suffix)
 
         write_if_different(makefile_path, "".join(lines).encode("ascii"))
 

src/main.rs

@@ -25,6 +25,7 @@ use {
 const RECOGNIZED_TRIPLES: &[&str] = &[
     "aarch64-apple-darwin",
     "aarch64-apple-ios",
+    "armv7-unknown-linux-gnueabihf",
     "arm64-apple-tvos",
     "i686-pc-windows-msvc",
     "i686-unknown-linux-gnu",
@@ -113,6 +114,13 @@ const PE_ALLOWED_LIBRARIES: &[&str] = &[
 lazy_static! {
     static ref GLIBC_MAX_VERSION: version_compare::Version<'static> =
         version_compare::Version::from("2.19").unwrap();
+
+    static ref ELF_ALLOWED_LIBRARIES_BY_TRIPLE: HashMap<&'static str, Vec<&'static str>> = {
+        [
+            ("armv7-unknown-linux-gnueabihf", vec!["ld-linux-armhf.so.3", "libgcc_s.so.1"]),
+        ].iter().cloned().collect()
+    };
+
     static ref DARWIN_ALLOWED_DYLIBS: Vec<MachOAllowedDylib> = {
         [
             MachOAllowedDylib {
@@ -256,6 +264,7 @@ lazy_static! {
         [
             ("aarch64-apple-darwin", "macosx-11.0-arm64"),
             ("aarch64-apple-ios", "iOS-aarch64"),
+            ("armv7-unknown-linux-gnueabihf", "linux-arm"),
             ("i686-pc-windows-msvc", "win32"),
             ("i686-unknown-linux-gnu", "linux-i686"),
             ("x86_64-apple-darwin", "macosx-10.9-x86_64"),
@@ -277,11 +286,21 @@ fn allowed_dylibs_for_triple(triple: &str) -> Vec<MachOAllowedDylib> {
     }
 }
 
-fn validate_elf(path: &Path, elf: &goblin::elf::Elf, bytes: &[u8]) -> Result<Vec<String>> {
+fn validate_elf(
+    target_triple: &str,
+    path: &Path,
+    elf: &goblin::elf::Elf,
+    bytes: &[u8],
+) -> Result<Vec<String>> {
     let mut errors = vec![];
 
+    let mut allowed_libraries = ELF_ALLOWED_LIBRARIES.to_vec();
+    if let Some(extra) = ELF_ALLOWED_LIBRARIES_BY_TRIPLE.get(target_triple) {
+        allowed_libraries.extend(extra.iter());
+    }
+
     for lib in &elf.libraries {
-        if !ELF_ALLOWED_LIBRARIES.contains(lib) {
+        if !allowed_libraries.contains(lib) {
             errors.push(format!("{} loads illegal library {}", path.display(), lib));
         }
     }
@@ -443,7 +462,7 @@ fn validate_distribution(dist_path: &Path) -> Result<Vec<String>> {
         if let Ok(object) = goblin::Object::parse(&data) {
             match object {
                 goblin::Object::Elf(elf) => {
-                    errors.extend(validate_elf(path.as_ref(), &elf, &data)?);
+                    errors.extend(validate_elf(triple, path.as_ref(), &elf, &data)?);
                 }
                 goblin::Object::Mach(mach) => match mach {
                     goblin::mach::Mach::Binary(macho) => {