virtiofs: add --mount-mode {unshare,direct}

Adds a boot flag choosing how virtiofsd is launched per --mount:

- unshare (default, previous behavior): virtiofsd runs inside
  'podman unshare' with --sandbox chroot, mapping guest UIDs into
  the invoking user's subuid range.

- direct: virtiofsd runs as the invoking user with --sandbox none,
  mapping guest UIDs 1:1 to host UIDs.

The direct mode is needed when the shared directory itself lives on
a virtiofs/fuse mount that rejects subuid-range UIDs (e.g. nested
boco): guest writes through the subuid mapping then fail with
EINVAL, breaking anything that writes to shares as a non-root guest
user (databases, agent workspaces).

Direct mode also cannot use the chroot sandbox (unprivileged
chroot() is not permitted), so it falls back to --sandbox none;
served trees are then not symlink-escape-protected.

Verified both modes boot and serve mounts (alpine guest).

.containerignore

@@ -1,10 +0,0 @@
-# Build context: deny everything, then allow only what the Containerfile needs.
-# Keeps the linux submodule (and example/, .gitea/, README, etc.) out of the build.
-*
-!Cargo.toml
-!Cargo.lock
-!src/
-!src/**
-!init/
-!init/**
-!kernel/config

Containerfile

@@ -1,65 +0,0 @@
-# boco runner: compiles boco (static musl) AND the shared boco kernel (vmlinuz)
-# with the slim config checked into the boco repo, then serves as the runtime
-# image for `boco build` / `boco boot` under podman.
-#
-# Runtime requirements (boco shells out to nested rootless podman):
-#   --device /dev/kvm      KVM accel for the VMs (entrypoint chmods it)
-#   --device /dev/net/tun  only if nested container runs need pasta networking
-#   volume at /home/boco/.local/share/boco   persistent registry (images+kernel)
-#   NO --privileged: the container's own userns (single mapping) is what makes
-#   nested `podman unshare` work; ignore_chown_errors handles unmapped uids.
-#
-# Kernel source: pinned tarball linux-7.2 (= 8d3ae59288f1e7d58d76558a6ee96d533bc5019f,
-# verified == refs/tags/v7.2{} via git ls-remote), sha256-checked.
-
-FROM alpine:edge AS build
-RUN apk add --no-cache rust cargo git
-WORKDIR /src
-COPY . .
-# alpine's rust targets x86_64-alpine-linux-musl natively (static, no glibc dep)
-RUN cargo build --release -p boco -p boco-init \
- && strip target/release/boco target/release/boco-init
-
-FROM alpine:3.22 AS kernel
-# g++ for host tools, libelf for objtool, openssl for extract-cert
-RUN apk add --no-cache make gcc g++ musl-dev perl flex bison bc \
-      elfutils-dev openssl-dev findutils tar xz zstd curl linux-headers diffutils
-WORKDIR /kbuild
-# Pinned upstream tarball: linux-7.2.tar.xz, sha256
-# f9fef3d14c0df53819026f4be74459835c2a0b0dcbf5b5bbd9ea19f0829402b3.
-# (Equivalent to the upstream shallow pin 8d3ae59288f1e7d58d76558a6ee96d533bc5019f
-# = refs/tags/v7.2{} — verified via `git ls-remote`; the tarball avoids
-# a multi-GB shallow clone through flaky slirp NAT.)
-RUN curl -fsSLo /tmp/linux-7.2.tar.xz https://cdn.kernel.org/pub/linux/kernel/v7.x/linux-7.2.tar.xz \
- && echo "f9fef3d14c0df53819026f4be74459835c2a0b0dcbf5b5bbd9ea19f0829402b3  /tmp/linux-7.2.tar.xz" | sha256sum -c - \
- && tar -xJf /tmp/linux-7.2.tar.xz \
- && mv linux-7.2 linux
-# Install the slim kernel config checked into the boco repo; olddefconfig
-# normalizes it against this tree and fails nothing silently afterwards.
-COPY kernel/config /kbuild/linux/.config
-RUN make -C linux olddefconfig HOSTCC=gcc CC=gcc \
- && grep -qE '^(# )?CONFIG_MODULES' linux/.config
-RUN make -C linux -j"$(nproc)" bzImage HOSTCC=gcc CC=gcc \
- && cp linux/arch/x86/boot/bzImage /vmlinuz \
- && ls -lh /vmlinuz
-
-FROM alpine:3.22
-RUN apk add --no-cache \
-      qemu-system-x86_64 qemu-img virtiofsd \
-      podman fuse3 openssh-client ca-certificates iptables ip6tables \
-      e2fsprogs cpio passt \
- && adduser -D -s /bin/sh boco \
- && echo boco:100000:65536 > /etc/subuid \
- && echo boco:100000:65536 > /etc/subgid \
- && mkdir -p /etc/containers \
- && printf '[storage]\ndriver = "vfs"\nrunroot = "/run/containers/storage"\ngraphroot = "/var/lib/containers/storage"\n[storage.options]\nignore_chown_errors = "true"\n' > /etc/containers/storage.conf \
- && printf '#!/bin/sh\n[ -e /dev/kvm ] && chmod 666 /dev/kvm\nexec /usr/local/bin/boco "$@"\n' > /usr/local/bin/entrypoint.sh \
- && chmod 755 /usr/local/bin/entrypoint.sh
-COPY --from=build /src/target/release/boco /usr/local/bin/boco
-COPY --from=build /src/target/release/boco-init /home/boco/.local/share/boco/boco-init
-COPY --from=kernel /vmlinuz /home/boco/.local/share/boco/vmlinuz
-RUN chmod 755 /usr/local/bin/boco /home/boco/.local/share/boco/boco-init /home/boco/.local/share/boco/vmlinuz
-ENV XDG_DATA_HOME=/home/boco/.local/share
-WORKDIR /data
-ENTRYPOINT ["/usr/local/bin/entrypoint.sh"]
-CMD ["image", "list"]

init/src/main.rs

@@ -245,7 +245,13 @@ fn setup_virtiofs_shares(workdir: &str) {
             };
 
             try_io(&format!("mkdir {dest}"), || fs::create_dir_all(&dest));
-            try_mount(Some(tag), &dest, Some("virtiofs"), MsFlags::empty(), None);
+            try_mount(
+                Some(tag),
+                &dest,
+                Some("virtiofs"),
+                MsFlags::empty(),
+                None,
+            );
         }
     }
 }

src/qemu.rs

@@ -1,4 +1,15 @@
-//! See [`boot`]
+//! QEMU: boot a registry entry with qemu-system-x86_64 using direct kernel
+//! boot (-kernel/-initrd or -drive). The kernel is shared across all VMs
+//! at $XDG_DATA_HOME/boco/registry/vmlinuz. A universal /boco/init script (injected
+//! at build time) handles VM bootstrap; /boco/exec runs the container's
+//! CMD/ENTRYPOINT. Runtime overrides are delivered as `boco.cmd=<base64>`
+//! on the kernel cmdline. Host directories are shared into the guest via
+//! virtio-fs: one virtiofsd daemon per `--mount`, plus a `-chardev socket`
+//! + `-device vhost-user-fs-pci` pair on the QEMU side. virtiofsd runs
+//! under `podman unshare` (rootless UID mapping) by default, or directly
+//! as the invoking user with `--mount-mode direct` (1:1 UID mapping).
+//! The guest mounts each share with `mount -t virtiofs <tag> <dest>`
+//! (see `boco.mount=` cmdline).
 
 use anyhow::{Context, Result, bail};
 use base64::Engine;
@@ -7,8 +18,8 @@ use std::{fmt::Write as _, process::Command};
 
 use crate::{
     forward::PortForward,
-    registry::{registry_dir, vmlinuz_path},
-    virtiofs::{self, VirtiofsShare},
+    registry::{registry_base_dir, registry_dir},
+    virtiofs::{self, MountMode, VirtiofsShare},
 };
 
 #[derive(Args, Debug)]
@@ -35,12 +46,15 @@ pub struct BootCmd {
     /// relative to the image's WORKINGDIR. Can be repeated.
     #[clap(long)]
     mount: Vec<String>,
+
+    /// How virtiofsd is launched: `unshare` runs it inside
+    /// `podman unshare` (rootless UID mapping, guest UIDs map into the
+    /// subuid range); `direct` runs it as the invoking user (guest UIDs
+    /// map 1:1 to host UIDs — needed for nested virtiofs shares).
+    #[clap(long, value_enum, default_value = "unshare")]
+    mount_mode: MountMode,
 }
 
-/// Invoke `qemu` to boot a boco image. See [`BootCmd`] for options
-///
-/// This uses qemu's direct-kernel-boot with `-kernel`, so it does not require a bootloader.
-/// The boco image is just an ext4 rootfs.
 pub(crate) fn boot(
     BootCmd {
         name,
@@ -48,16 +62,17 @@ pub(crate) fn boot(
         forward,
         cmd,
         mount,
+        mount_mode,
     }: BootCmd,
 ) -> Result<()> {
     let reg_dir = registry_dir(&name)?;
-    let vmlinuz_path = vmlinuz_path()?;
+    let vmlinuz_path = registry_base_dir()?.join("vmlinuz");
     let initrd_path = reg_dir.join("initrd");
     let qcow2_path = reg_dir.join("image.qcow2");
     if !vmlinuz_path.exists() {
         // TODO: guide user in how to set up a kernel
         bail!(
-            "No kernel available. Expected vmlinuz at {}",
+            "No kernel available. Place a vmlinuz at {}",
             vmlinuz_path.display()
         );
     }
@@ -103,7 +118,7 @@ pub(crate) fn boot(
             Some((h, g)) => (h, g.to_string()),
             None => (spec.as_str(), spec.to_string()),
         };
-        let share = virtiofs::spawn_share(i, host_path, guest_path)?;
+        let share = virtiofs::spawn_share(i, host_path, guest_path, mount_mode)?;
         let tag = share.tag.clone();
         let chardev = format!("char{i}");
         let socket = share

src/registry.rs

@@ -26,12 +26,6 @@ pub(crate) fn validate_image_name(name: &str) -> Result<()> {
     Ok(())
 }
 
-pub(crate) fn vmlinuz_path() -> Result<PathBuf> {
-    let base = xdg::BaseDirectories::with_prefix("boco");
-    base.place_data_file("vmlinuz")
-        .context("Failed to create XDG_DATA_HOME subdirectory")
-}
-
 pub(crate) fn registry_base_dir() -> Result<PathBuf> {
     let base = xdg::BaseDirectories::with_prefix("boco");
     base.create_data_directory("registry")

src/virtiofs.rs

@@ -6,13 +6,21 @@
 //! hands the socket path to QEMU via `-chardev socket,path=...`, and
 //! cleans up when the VM exits.
 //!
-//! To serve files owned by arbitrary guest UIDs (e.g. when an image
-//! uses the `USER` directive) without running as root, virtiofsd is
-//! launched under `podman unshare` in a user namespace with
-//! `--sandbox chroot`.  This reuses rootless podman's existing
-//! subuid/subgid mapping: guest UID 0 maps to the invoking user, and
-//! other guest UIDs map into the subuid range, so each guest UID
-//! becomes a distinct host UID.
+//! Launch modes ([`MountMode`]):
+//!
+//! * `unshare` (default): virtiofsd runs under `podman unshare` in a
+//!   user namespace with `--sandbox chroot`.  This serves files owned
+//!   by arbitrary guest UIDs (e.g. when an image uses the `USER`
+//!   directive) without running as root, reusing rootless podman's
+//!   subuid/subgid mapping: guest UID 0 maps to the invoking user, and
+//!   other guest UIDs map into the subuid range, so each guest UID
+//!   becomes a distinct host UID.
+//!
+//! * `direct`: virtiofsd runs as the invoking user with `--sandbox
+//!   none`.  Guest UIDs map 1:1 to host UIDs.  Use this when the
+//!   shared directory lives on a filesystem that rejects subuid-range
+//!   UIDs (e.g. nested virtiofs), or when the guest should act as the
+//!   invoking user rather than the subuid range.
 //!
 //! Lifecycle: boco spawns all virtiofsd children before launching QEMU,
 //! waits for each listening socket to appear, then drops the
@@ -28,10 +36,26 @@ use std::thread;
 use std::time::{Duration, Instant};
 
 use anyhow::{Context, Result, anyhow, bail};
+use clap::ValueEnum;
 use tempfile::TempDir;
 
 use crate::command::cmd;
 
+/// How virtiofsd is launched for a share.
+#[derive(Clone, Copy, Debug, PartialEq, Eq, ValueEnum)]
+pub enum MountMode {
+    /// Run virtiofsd inside `podman unshare` (rootless UID mapping:
+    /// guest UID 0 maps to the invoking user, other guest UIDs map
+    /// into the subuid range).
+    Unshare,
+    /// Run virtiofsd directly as the invoking user. Guest UIDs map 1:1
+    /// to host UIDs. Needed when the shared directory itself lives on
+    /// a virtiofs/fuse mount that rejects subuid-range UIDs (e.g.
+    /// nested boco), where writes through the subuid mapping fail with
+    /// EINVAL.
+    Direct,
+}
+
 /// Locations searched for the virtiofsd binary, in order.
 const VIRTIOFSD_CANDIDATES: &[&str] = &[
     "/usr/libexec/virtiofsd",
@@ -114,7 +138,12 @@ fn wait_for_socket(socket: &Path) -> Result<()> {
 ///
 /// `index` becomes the suffix of the tag (`boco{index}`) and is used to
 /// derive a unique socket filename.
-pub fn spawn_share(index: usize, host_path: &str, guest_path: String) -> Result<VirtiofsShare> {
+pub fn spawn_share(
+    index: usize,
+    host_path: &str,
+    guest_path: String,
+    mode: MountMode,
+) -> Result<VirtiofsShare> {
     let virtiofsd = find_virtiofsd()?;
     let canonical = resolve_host_dir(host_path)?;
     let host_str = canonical
@@ -128,17 +157,29 @@ pub fn spawn_share(index: usize, host_path: &str, guest_path: String) -> Result<
     let tag = format!("boco{index}"); // short tag to fit on cmdline
     let socket = dir.path().join(format!("{tag}.sock"));
 
-    // TODO: Make rootless mode optional
-
-    // Run virtiofsd under `podman unshare` so it can serve files owned
-    // by arbitrary guest UIDs without root, mapping them into the host
-    // user's subuid range.  `--sandbox chroot` is required because the
-    // outer `podman unshare` already entered a user namespace (nested
-    // user namespaces are not allowed); chroot still prevents symlink
-    // escapes from the shared dir.
-    let mut command = Command::new("podman");
+    // In `unshare` mode, run virtiofsd under `podman unshare` so it can
+    // serve files owned by arbitrary guest UIDs without root, mapping
+    // them into the host user's subuid range.  `--sandbox chroot` is
+    // required because the outer `podman unshare` already entered a user
+    // namespace (nested user namespaces are not allowed); chroot still
+    // prevents symlink escapes from the shared dir.
+    //
+    // In `direct` mode, run virtiofsd as the invoking user with 1:1 UID
+    // mapping.  This cannot use the chroot sandbox (unprivileged
+    // chroot is not permitted), so it falls back to `--sandbox none`.
+    let mut command = match mode {
+        MountMode::Unshare => {
+            let mut command = Command::new("podman");
+            command.args(["unshare", "--", virtiofsd_str, "--socket-path"]);
+            command
+        }
+        MountMode::Direct => {
+            let mut command = Command::new(virtiofsd_str);
+            command.arg("--socket-path");
+            command
+        }
+    };
     command
-        .args(["unshare", "--", virtiofsd_str, "--socket-path"])
         .arg(&socket)
         .args([
             "--shared-dir",
@@ -146,7 +187,10 @@ pub fn spawn_share(index: usize, host_path: &str, guest_path: String) -> Result<
             "--tag",
             &tag,
             "--sandbox",
-            "chroot",
+            match mode {
+                MountMode::Unshare => "chroot",
+                MountMode::Direct => "none",
+            },
             "--cache",
             "auto",
         ])