Files
OpenPXE/deploy/docker/Dockerfile
T
Miles WardandClaude Opus 4.8 a906c47f53 build: native arm64→x86_64-musl cross-compile (cargo-zigbuild), no QEMU
The Rust `build` stage previously ran the entire compiler under QEMU x86_64
emulation on the arm64 builder. That was ~15x slower (one crate took >20 min)
and the emulated gcc/linker intermittently SIGSEGV'd or hung mid-link
(observed again building v0.5.2).

Pin the stage to $BUILDPLATFORM (native arm64 on Apple Silicon, amd64 in CI)
and cross-compile to x86_64-unknown-linux-musl with cargo-zigbuild — zig cc
supplies the musl sysroot + linker. rustc runs natively; no emulation. Build
drops from ~30 min to a few minutes and is deterministic. Output is the same
fully static musl binary (verified: x86_64, not a dynamic executable, 0
OpenSSL strings).

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
2026-05-31 17:46:50 -04:00

183 lines
9.2 KiB
Docker

# syntax=docker/dockerfile:1.7
#
# OpenPXE — multi-stage build.
#
# Design:
# - stage `fetch`: runs scripts/fetch-ipxe.sh to pull official iPXE binaries
# into assets/ipxe/ so the rust build can embed them via rust-embed.
# - stage `build`: compiles the workspace with cargo in release mode.
# - stage `runtime`: Debian slim image with setcap for NET_BIND_SERVICE,
# running as a non-root UID. No shell in PATH for the service user;
# attacker surface is just the openpxe binary + libc.
#
# Why not distroless? We want setcap support and easy debug (`oc rsh`).
# Debian slim at ~75 MB + binary ~25 MB is fine for a PXE server that
# spends most of its life idle.
ARG RUST_VERSION=1.95
########## fetch iPXE binaries + wimboot ##########
# Pulls the upstream boot.ipxe.org pre-builds (no PNG support) plus
# wimboot. These cover the arches we don't build from source here:
# BIOS undionly.kpxe and i386-efi (which need a 32-bit x86 toolchain),
# and serve as the baseline that the PNG-enabled x86_64/arm64 UEFI
# binaries from the `ipxe-build` stage overlay on top of.
FROM debian:12-slim AS fetch
RUN apt-get update && apt-get install -y --no-install-recommends curl ca-certificates \
&& rm -rf /var/lib/apt/lists/*
WORKDIR /src
COPY scripts/fetch-ipxe.sh scripts/fetch-ipxe.sh
RUN mkdir -p assets/ipxe && bash scripts/fetch-ipxe.sh
########## build PNG-enabled iPXE from source ##########
# v0.4.69: THE graphical-boot-menu unlock. iVentoy paints a PNG
# background on the PXE screen using stock iPXE built with
# CONSOLE_FRAMEBUFFER + IMAGE_PNG + CONSOLE_CMD; the public iPXE
# binaries omit those, so `console --picture` is a no-op on them.
# We build our own from upstream with that thin config delta.
#
# The historical blocker was cc1 segfaulting when an amd64 gcc ran
# under QEMU emulation on an arm64 host. The fix: pin this stage to
# $BUILDPLATFORM (the NATIVE builder arch — arm64 on an Apple-Silicon
# Mac, amd64 in x86 CI) and cross-compile with a real cross toolchain
# (CROSS_COMPILE=x86_64-linux-gnu-). The compiler runs native and
# emits x86_64 — no emulation, no segfault. arm64-efi builds natively.
FROM --platform=$BUILDPLATFORM debian:12-slim AS ipxe-build
# libc6-dev is REQUIRED and easy to miss under --no-install-recommends:
# iPXE's host utilities (elf2efi, zbin) compile with the native gcc and
# pull <stdint.h>; without the native libc headers gcc's #include_next
# falls through to iPXE's freestanding headers and dies on bits/stdint.h.
# The target (iPXE firmware) code is -ffreestanding/-nostdinc, so the
# x86_64 cross toolchain needs NO cross libc headers.
RUN apt-get update && apt-get install -y --no-install-recommends \
git make perl gcc binutils libc6-dev \
gcc-x86-64-linux-gnu binutils-x86-64-linux-gnu \
ca-certificates \
&& rm -rf /var/lib/apt/lists/*
WORKDIR /src
COPY scripts/build-ipxe.sh scripts/build-ipxe.sh
COPY deploy/ipxe/local/ deploy/ipxe/local/
RUN mkdir -p assets/ipxe && bash scripts/build-ipxe.sh /src/assets/ipxe
########## build openpxe ##########
# v0.5.2: cross-compile the Rust binary NATIVELY — no QEMU.
#
# This stage is pinned to $BUILDPLATFORM (the builder's native arch — arm64
# on an Apple-Silicon Mac, amd64 in x86 CI), exactly like `ipxe-build`. The
# Rust compiler therefore runs at full native speed and emits an
# x86_64-unknown-linux-musl binary via `cargo-zigbuild`, which uses `zig cc`
# as the cross-linker (it bundles the musl sysroot for every target, so
# there's no fiddly cross-gcc toolchain to assemble).
#
# Why this replaced the old `FROM rust ... --platform=linux/amd64` build:
# that ran the *entire* compiler under QEMU x86_64 emulation on the arm64
# host. It was ~15x slower (a single crate took >20 min) and the emulated
# gcc/linker intermittently SIGSEGV'd or hung mid-link. Cross-compiling
# sidesteps emulation entirely — the build is minutes, not half an hour,
# and is deterministic.
#
# The output is still a fully static musl binary with no glibc dependency,
# so the runtime stage stays free to be any Linux distro.
FROM --platform=$BUILDPLATFORM rust:${RUST_VERSION}-bookworm AS build
WORKDIR /src
# zig (via the `ziglang` pip package — cargo-zigbuild auto-discovers it as
# `python3 -m ziglang`) supplies the x86_64 musl sysroot + linker.
# cargo-zigbuild is the thin cargo wrapper that wires zig in as the linker.
RUN apt-get update \
&& apt-get install -y --no-install-recommends python3 python3-pip \
&& rm -rf /var/lib/apt/lists/* \
&& rustup target add x86_64-unknown-linux-musl \
&& pip3 install --no-cache-dir --break-system-packages ziglang \
&& cargo install --locked cargo-zigbuild
# Do not copy rust-toolchain.toml into the image. The local workspace pins
# developer tooling, but inside Docker we intentionally use the Rust version
# selected by the base image. Copying rust-toolchain.toml with
# `channel = "stable"` makes rustup download a second full toolchain during
# the build, which is slow and can exhaust small Colima/CI disks.
COPY Cargo.toml Cargo.lock ./
COPY crates/ crates/
# Baseline binaries (BIOS / i386 / wimboot), then overlay the
# PNG-enabled x86_64 + arm64 UEFI binaries built from source. The
# overlay wins for snponly.efi / ipxe.efi / snponly-arm64.efi so the
# common modern clients get the graphical background; the rest keep the
# upstream no-PNG binaries and the menu's `|| console` text fallback.
COPY --from=fetch /src/assets/ipxe /src/assets/ipxe
COPY --from=ipxe-build /src/assets/ipxe/snponly.efi /src/assets/ipxe/snponly.efi
COPY --from=ipxe-build /src/assets/ipxe/ipxe.efi /src/assets/ipxe/ipxe.efi
# Cache cargo registry + target across builds. `cargo zigbuild` runs the
# native rustc (fast) and links for x86_64-musl with zig — no emulation.
RUN --mount=type=cache,target=/usr/local/cargo/registry \
--mount=type=cache,target=/src/target,sharing=locked \
cargo zigbuild --release --target x86_64-unknown-linux-musl --bin openpxe && \
cp target/x86_64-unknown-linux-musl/release/openpxe /openpxe && \
ls -l /openpxe
########## runtime ##########
FROM debian:12-slim AS runtime
RUN apt-get update \
&& apt-get install -y --no-install-recommends \
ca-certificates libcap2-bin tini gosu iproute2 \
wimtools samba smbclient \
&& rm -rf /var/lib/apt/lists/* \
&& useradd --system --uid 10001 --home-dir /var/lib/openpxe --shell /usr/sbin/nologin openpxe \
&& mkdir -p /var/lib/openpxe/isos /var/lib/openpxe/work /var/lib/openpxe/smb \
&& chown -R openpxe:openpxe /var/lib/openpxe
# v0.4.5: the openpxe binary itself is now built against musl and is
# fully static — no glibc dependency. The runtime stage still ships
# Debian slim because OpenPXE shells out to the packages below for
# functionality we deliberately don't reimplement in-process:
#
# wimtools - `wimlib-imagex`, used to inject startnet.cmd into boot.wim.
# samba - `smbd` serves extracted Windows install media on :445 so
# WinPE can `net use`. Guest read-only, scoped to
# /var/lib/openpxe/smb. This package provides the SERVER
# side only; the client CLI is a separate package below.
# smbclient - v0.4.66: Samba's `smbclient` userspace CLI, used by
# the Storage tab's SMB shares manager to list and stream
# ISOs from remote SMB servers without ever mounting them
# in the kernel. In Debian 12 `smbclient` is NOT pulled
# in by the `samba` package — they're siblings, not
# parent/child. v0.4.65 shipped without this line and
# every "Add share" attempt surfaced
# `could not exec smbclient: No such file or directory`
# until this landed.
# iproute2 - `ip addr` / `ip route` for the auto-detected Network
# tab fields (NIC name, subnet mask, default gateway).
# Tiny, always available; we don't pull in netlink crates
# for this one-shot startup probe.
# gosu - drops privileges cleanly from root after the entrypoint
# fixes bind-mount ownership (common OpenShift/Docker UX
# issue).
#
# v0.4.65 dropped `nfs-common` — kernel-mount NFS is gone. The SMB
# shares replacement uses userspace `smbclient` and needs no kernel
# helpers.
#
# A future "openpxe-static" variant could drop everything except the
# binary onto distroless once we move the Windows + SMB legs to
# in-process Rust crates.
COPY --from=build /openpxe /usr/local/bin/openpxe
COPY deploy/docker/entrypoint.sh /usr/local/bin/entrypoint.sh
RUN chmod +x /usr/local/bin/entrypoint.sh
# Grant the binary the ability to bind <1024 ports as a non-root user.
# This is the only capability OpenPXE needs for proxy-mode DHCP + TFTP + HTTP.
RUN setcap cap_net_bind_service=+ep /usr/local/bin/openpxe
# IMPORTANT: we do NOT `USER openpxe` here. The entrypoint runs as root,
# chowns the mounted data dirs, then execs the binary via gosu as openpxe.
# OpenShift ignores USER directives anyway (it injects its own uid), and
# there entrypoint.sh's non-root branch just execs directly.
WORKDIR /var/lib/openpxe
ENV OPENPXE_ISO_DIR=/var/lib/openpxe/isos \
OPENPXE_WORK_DIR=/var/lib/openpxe/work \
OPENPXE_LOG=info,openpxe=info
EXPOSE 67/udp 69/udp 4011/udp 80/tcp 445/tcp
ENTRYPOINT ["/usr/bin/tini", "--", "/usr/local/bin/entrypoint.sh"]