install: Automatically set up /dev and /var/lib/containers

We're looking again at the ergonomics of `bootc install to-existing-root`.
This uses the "mounting into mount namespaces" from the new
mount API to automatically set up `/dev` and `/var/lib/containers`
if they weren't provided to `podman run`, which shrinks what's
needed a bit.

Closes: #826

Signed-off-by: Colin Walters <[email protected]>

Author: cgwalters

lib/src/install.rs

@@ -1269,6 +1269,8 @@ async fn prepare_install(
     tracing::debug!("Target image reference: {target_imgref}");
 
     // A bit of basic global state setup
+    crate::mount::ensure_mirrored_host_mount("/dev")?;
+    crate::mount::ensure_mirrored_host_mount("/var/lib/containers")?;
     ensure_var()?;
     setup_tmp_mounts()?;
     // Allocate a temporary directory we can use in various places to avoid
@@ -1454,12 +1456,6 @@ async fn install_to_filesystem_impl(state: &State, rootfs: &mut RootSetup) -> Re
         .ok_or_else(|| anyhow!("No uuid for boot/root"))?;
     tracing::debug!("boot uuid={boot_uuid}");
 
-    // If we're doing an alongside install, then the /dev bootupd sees needs to be the host's.
-    ensure!(
-        crate::mount::is_same_as_host(Utf8Path::new("/dev"))?,
-        "Missing /dev mount to host /dev"
-    );
-
     let bound_images = BoundImages::from_state(state).await?;
 
     // Initialize the ostree sysroot (repo, stateroot, etc.)
@@ -1514,9 +1510,6 @@ pub(crate) async fn install_to_disk(mut opts: InstallToDiskOpts) -> Result<()> {
                 block_opts.device
             );
         }
-        if !crate::mount::is_same_as_host(Utf8Path::new("/dev"))? {
-            anyhow::bail!("Loopback mounts (--via-loopback) require host devices (-v /dev:/dev)");
-        }
     } else if !target_blockdev_meta.file_type().is_block_device() {
         anyhow::bail!("Not a block device: {}", block_opts.device);
     }

lib/src/mount.rs

@@ -1,11 +1,24 @@
 //! Helpers for interacting with mountpoints
 
-use std::process::Command;
+use std::{
+    fs,
+    os::fd::{AsFd, OwnedFd},
+    process::Command,
+};
 
-use anyhow::{anyhow, Result};
+use anyhow::{anyhow, Context, Result};
 use bootc_utils::CommandRunExt;
 use camino::Utf8Path;
 use fn_error_context::context;
+use rustix::{
+    mount::{MoveMountFlags, OpenTreeFlags},
+    net::{
+        AddressFamily, RecvFlags, SendAncillaryBuffer, SendAncillaryMessage, SendFlags,
+        SocketFlags, SocketType,
+    },
+    process::WaitOptions,
+    thread::Pid,
+};
 use serde::Deserialize;
 
 use crate::task::Task;
@@ -124,3 +137,134 @@ pub(crate) fn is_same_as_host(path: &Utf8Path) -> Result<bool> {
     );
     Ok(devstat.f_fsid == hostdevstat.f_fsid)
 }
+
+/// Given a pid, enter its mount namespace and acquire a file descriptor
+/// for a mount from that namespace.
+#[allow(unsafe_code)]
+#[context("Opening mount tree from pid")]
+pub(crate) fn open_tree_from_pidns(
+    pid: rustix::process::Pid,
+    path: &Utf8Path,
+    recursive: bool,
+) -> Result<OwnedFd> {
+    // Allocate a socket pair to use for sending file descriptors.
+    let (sock_parent, sock_child) = rustix::net::socketpair(
+        AddressFamily::UNIX,
+        SocketType::STREAM,
+        SocketFlags::CLOEXEC,
+        None,
+    )
+    .context("socketpair")?;
+    const DUMMY_DATA: &[u8] = &[b'!'];
+    match unsafe { libc::fork() } {
+        0 => {
+            // We're in the child. At this point we know we don't have multiple threads, so we
+            // can safely `setns`.
+
+            // Open up the namespace of the target process as a file descriptor, and enter it.
+            let pidlink = fs::File::open(format!("/proc/{}/ns/mnt", pid.as_raw_nonzero()))?;
+            rustix::thread::move_into_link_name_space(
+                pidlink.as_fd(),
+                Some(rustix::thread::LinkNameSpaceType::Mount),
+            )
+            .context("setns")?;
+
+            // Open the target mount path as a file descriptor.
+            let recursive = if recursive {
+                OpenTreeFlags::AT_RECURSIVE
+            } else {
+                OpenTreeFlags::empty()
+            };
+            let fd = rustix::mount::open_tree(
+                rustix::fs::CWD,
+                path.as_std_path(),
+                OpenTreeFlags::OPEN_TREE_CLOEXEC | OpenTreeFlags::OPEN_TREE_CLONE | recursive,
+            )
+            .context("open_tree")?;
+
+            // And send that file descriptor via fd passing over the socketpair.
+            let fd = fd.as_fd();
+            let fds = [fd];
+            let mut buffer = [0u8; rustix::cmsg_space!(ScmRights(1))];
+            let mut control = SendAncillaryBuffer::new(&mut buffer);
+            let pushed = control.push(SendAncillaryMessage::ScmRights(&fds));
+            assert!(pushed);
+            let ios = std::io::IoSlice::new(DUMMY_DATA);
+            rustix::net::sendmsg(sock_child, &[ios], &mut control, SendFlags::empty())?;
+            // Then we're done.
+            std::process::exit(0)
+        }
+        -1 => {
+            // fork failed
+            let e = std::io::Error::last_os_error();
+            anyhow::bail!("failed to fork: {e}");
+        }
+        n => {
+            // We're in the parent; create a pid (checking that n > 0).
+            let pid = rustix::process::Pid::from_raw(n).unwrap();
+            // Receive the mount file descriptor from the child
+            let mut cmsg_space = vec![0; rustix::cmsg_space!(ScmRights(1))];
+            let mut cmsg_buffer = rustix::net::RecvAncillaryBuffer::new(&mut cmsg_space);
+            let mut buf = [0u8; DUMMY_DATA.len()];
+            let iov = std::io::IoSliceMut::new(buf.as_mut());
+            let mut iov = [iov];
+            let nread = rustix::net::recvmsg(
+                sock_parent,
+                &mut iov,
+                &mut cmsg_buffer,
+                RecvFlags::CMSG_CLOEXEC,
+            )
+            .context("recvmsg")?
+            .bytes;
+            assert_eq!(nread, DUMMY_DATA.len());
+            assert_eq!(buf, DUMMY_DATA);
+            // And extract the file descriptor
+            let r = cmsg_buffer
+                .drain()
+                .filter_map(|m| match m {
+                    rustix::net::RecvAncillaryMessage::ScmRights(f) => Some(f),
+                    _ => None,
+                })
+                .flatten()
+                .next()
+                .ok_or_else(|| anyhow::anyhow!("Did not receive a file descriptor"))?;
+            rustix::process::waitpid(Some(pid), WaitOptions::empty())?;
+            Ok(r)
+        }
+    }
+}
+
+/// Create a bind mount from the mount namespace of the target pid
+/// into our mount namespace.
+pub(crate) fn bind_mount_from_pidns(
+    pid: Pid,
+    src: &Utf8Path,
+    target: &Utf8Path,
+    recursive: bool,
+) -> Result<()> {
+    let src = open_tree_from_pidns(pid, src, recursive)?;
+    rustix::mount::move_mount(
+        src.as_fd(),
+        "",
+        rustix::fs::CWD,
+        target.as_std_path(),
+        MoveMountFlags::MOVE_MOUNT_F_EMPTY_PATH,
+    )
+    .context("Moving mount")?;
+    Ok(())
+}
+
+// If the target path is not already mirrored from the host (e.g. via -v /dev:/dev)
+// then recursively mount it.
+pub(crate) fn ensure_mirrored_host_mount(path: impl AsRef<Utf8Path>) -> Result<()> {
+    let path = path.as_ref();
+    // If we didn't have this in our filesystem already (e.g. for /var/lib/containers)
+    // then create it now.
+    std::fs::create_dir_all(path)?;
+    if is_same_as_host(path)? {
+        tracing::debug!("Already mounted from host: {path}");
+        return Ok(());
+    }
+    tracing::debug!("Propagating host mount: {path}");
+    bind_mount_from_pidns(Pid::from_raw(1).unwrap(), path, path, true)
+}