supermachine 0.7.0

// POSIX filesystem backend — translates FUSE ops to host syscalls on
// a single rooted subtree.
//
// Each mount has a host-side `root` directory; the backend exposes that
// subtree to the guest via FUSE. Inode numbers in the FUSE protocol are
// allocated by the backend (NOT the host filesystem's st_ino) so:
//   - We can change the host filesystem under us without invalidating
//     guest-side caches (which key on nodeid).
//   - Multiple mount instances of the same host path each get their
//     own nodeid namespace.
//
// The backend keeps a small bidirectional map:
//   nodeid -> InodeInfo { host_path, kind }
//   (parent_nodeid, name) -> nodeid    (lookup cache)
//
// Handle table: open files map fh -> RawFd. Opening the same path
// twice gives different fhs, mirroring open(2) semantics.
//
// Symlinks are followed during traversal — virtio-fs's typical use case
// is "expose this directory tree read-only-ish", not "expose a chroot
// jail". A later slice can add no-follow + open_by_handle for security.

use std::collections::BTreeMap;
use std::ffi::{CString, OsStr, OsString};
use std::os::unix::ffi::OsStrExt;
use std::os::unix::fs::MetadataExt;
use std::os::fd::{AsRawFd, FromRawFd, OwnedFd};
use std::path::PathBuf;
use std::sync::atomic::{AtomicBool, Ordering};
use std::sync::{Arc, Mutex};

use super::backend::{
    DirEntry, Entry, Errno, FsBackend, StatFs, EACCES, EBADF, EINVAL, ENOENT, ENOSPC, ENOTDIR,
    EISDIR, EIO, EPERM,
};
use crate::vmm::resources::SymlinkPolicy;
use super::notify::Notifier;
use super::protocol::{
    Attr, DT_BLK, DT_CHR, DT_DIR, DT_FIFO, DT_LNK, DT_REG, DT_SOCK, DT_UNKNOWN,
    FUSE_ROOT_ID, S_IFBLK, S_IFCHR, S_IFDIR, S_IFIFO, S_IFLNK, S_IFMT, S_IFREG, S_IFSOCK,
};

/// What kind of host object backs a given nodeid. We cache this so
/// hot-path `getattr` doesn't always have to stat.
#[derive(Clone, Copy, Debug)]
enum Kind {
    File,
    Dir,
    Symlink,
    Other,
}

#[derive(Clone)]
struct InodeInfo {
    host_path: PathBuf,
    kind: Kind,
}

/// One DAX-mmap'd region of a host file. Multiple slots may share
/// one Mmap if SETUPMAPPING calls overlap (we deduplicate on host
/// path + foffset boundaries in a later optimization; for now each
/// dax_map allocates a fresh mmap).
struct Mmap {
    ptr: *mut u8,
    len: usize,
}

// SAFETY: ptr is a process-local pointer used only by DaxSession's
// HvfMapper. Mmap-as-DAX-source isn't dereferenced from rust here.
unsafe impl Send for Mmap {}
unsafe impl Sync for Mmap {}

struct State {
    /// nodeid -> InodeInfo.
    inodes: BTreeMap<u64, InodeInfo>,
    /// (parent, name) -> nodeid. Populated by lookup.
    children: BTreeMap<(u64, Vec<u8>), u64>,
    /// fh -> OwnedFd. Closed on `release` / `releasedir`.
    handles: BTreeMap<u64, OwnedFd>,
    /// Active DAX mappings indexed by host_va. Owns the mmap; dropping
    /// hits munmap.
    dax_mmaps: BTreeMap<usize, Mmap>,
    next_nodeid: u64,
    next_fh: u64,
}

impl Drop for State {
    fn drop(&mut self) {
        for (_, m) in std::mem::take(&mut self.dax_mmaps) {
            unsafe {
                libc::munmap(m.ptr as *mut _, m.len);
            }
        }
    }
}

/// POSIX-backed FUSE filesystem rooted at `root`. All paths are
/// constrained to live under `root` — we don't `chroot`, we just
/// resolve names manually so we can refuse `..` escapes.
pub struct PosixFs {
    /// Canonical mount root. LOOKUP-resolved paths must remain under
    /// this prefix unless `symlinks == SymlinkPolicy::Follow`.
    root: PathBuf,
    /// Per-mount symlink policy. See [`SymlinkPolicy`].
    symlinks: SymlinkPolicy,
    st: Mutex<State>,
    /// Background thread that watches file-content-change events via
    /// kqueue and dispatches them to the guest as
    /// FUSE_NOTIFY_INVAL_INODE messages.
    watcher: Mutex<Option<Watcher>>,
}

/// State for the kqueue watcher thread.
struct Watcher {
    /// Shared with the watcher thread via Arc. The thread reads kq,
    /// looks up the (kq_ident → nodeid) map, and calls notifier.
    /// We hold a strong ref so dropping PosixFs stops the thread.
    inner: Arc<WatcherInner>,
    /// Join handle for clean shutdown.
    handle: Option<std::thread::JoinHandle<()>>,
}

struct WatcherInner {
    /// kqueue fd; the watcher thread waits on this.
    kq: libc::c_int,
    /// Stop signal — set by drop, polled by the watcher thread.
    stop: AtomicBool,
    /// Notifier the watcher thread invokes when a kevent fires. None
    /// until set_notifier is called.
    notifier: Mutex<Option<Arc<dyn Notifier>>>,
    /// Map kqueue ident (an fd we duped specifically for kqueue) →
    /// (nodeid, parent_nodeid, name, owned_fd). parent_nodeid + name
    /// let us also invalidate the dentry cache so re-opens see the
    /// new inode after an atomic rename.
    watched: Mutex<BTreeMap<libc::c_int, WatchedEntry>>,
}

struct WatchedEntry {
    nodeid: u64,
    parent_nodeid: u64,
    name: Vec<u8>,
    _owned_fd: OwnedFd,
}

impl Drop for Watcher {
    fn drop(&mut self) {
        self.inner.stop.store(true, Ordering::Release);
        // Poke the kqueue with a USER event so the thread wakes from kevent.
        let trigger = libc::kevent {
            ident: 0,
            filter: libc::EVFILT_USER,
            flags: libc::EV_ADD | libc::EV_ONESHOT | libc::EV_RECEIPT,
            fflags: libc::NOTE_TRIGGER,
            data: 0,
            udata: std::ptr::null_mut(),
        };
        let mut tr = trigger;
        unsafe {
            let _ = libc::kevent(self.inner.kq, &mut tr as *mut _, 1, std::ptr::null_mut(), 0, std::ptr::null());
        }
        if let Some(h) = self.handle.take() {
            let _ = h.join();
        }
        unsafe {
            libc::close(self.inner.kq);
        }
    }
}

impl PosixFs {
    /// Mount `root` as the FUSE filesystem root. The root must exist
    /// and be a directory; we stat it eagerly so a misconfigured
    /// mount fails fast.
    ///
    /// Defaults to [`SymlinkPolicy::Opaque`] — symlinks under the
    /// mount that resolve OUTSIDE the canonical mount root are
    /// rejected with EACCES on LOOKUP; the guest may create new
    /// symlinks whose targets are stored verbatim. Use
    /// [`Self::new_unchecked`] or [`Self::new_with_symlinks`] to pick
    /// a different policy.
    pub fn new(root: impl Into<PathBuf>) -> Result<Self, std::io::Error> {
        Self::new_with_symlinks(root, SymlinkPolicy::Opaque)
    }

    /// Constructor variant that allows symlinks pointing outside the
    /// mount root (LOOKUP follows them unconditionally). Equivalent to
    /// `new_with_symlinks(root, SymlinkPolicy::Follow)`. Use for
    /// trusted single-tenant workloads where the mount tree may
    /// reference absolute host paths.
    pub fn new_unchecked(root: impl Into<PathBuf>) -> Result<Self, std::io::Error> {
        Self::new_with_symlinks(root, SymlinkPolicy::Follow)
    }

    /// Constructor that picks the symlink policy explicitly. See
    /// [`SymlinkPolicy`] for the three modes.
    pub fn new_with_symlinks(
        root: impl Into<PathBuf>,
        symlinks: SymlinkPolicy,
    ) -> Result<Self, std::io::Error> {
        let root = root.into();
        let md = std::fs::metadata(&root)?;
        if !md.is_dir() {
            return Err(std::io::Error::new(
                std::io::ErrorKind::NotADirectory,
                format!("posix-fs root is not a directory: {}", root.display()),
            ));
        }
        // Canonicalize the root so the LOOKUP-time prefix check works
        // against a stable absolute path (no `..`, no symlinks, no
        // relative components). Without this, a root like
        // `/tmp/myapp/.` would fail prefix-of(canonical(child)).
        let root = std::fs::canonicalize(&root)?;
        let mut inodes = BTreeMap::new();
        inodes.insert(
            FUSE_ROOT_ID,
            InodeInfo {
                host_path: root.clone(),
                kind: Kind::Dir,
            },
        );
        Ok(Self {
            root,
            symlinks,
            st: Mutex::new(State {
                inodes,
                children: BTreeMap::new(),
                handles: BTreeMap::new(),
                dax_mmaps: BTreeMap::new(),
                next_nodeid: FUSE_ROOT_ID + 1,
                next_fh: 1,
            }),
            watcher: Mutex::new(None),
        })
    }

    /// Attach a Notifier and start a kqueue background thread that
    /// watches host-side changes to files this PosixFs has surfaced
    /// to the guest. On NOTE_DELETE / NOTE_RENAME / NOTE_WRITE /
    /// NOTE_EXTEND / NOTE_ATTRIB, dispatches FUSE_NOTIFY_INVAL_INODE
    /// to the guest so its dentry + page caches re-read on next access.
    ///
    /// Watching is per-inode: a file gets watched on first OPEN or
    /// SETUPMAPPING. This covers the dev-loop pattern (editor saves a
    /// file the guest is using); for rarely-accessed files we rely on
    /// the guest's 1s attr_valid timeout. Recursive directory watching
    /// is a follow-up if needed.
    pub fn set_notifier(&self, notifier: Arc<dyn Notifier>) -> Result<(), std::io::Error> {
        let mut watcher_slot = self.watcher.lock().unwrap();
        if watcher_slot.is_some() {
            // Already running; just swap the notifier.
            let w = watcher_slot.as_ref().unwrap();
            *w.inner.notifier.lock().unwrap() = Some(notifier);
            return Ok(());
        }
        let kq = unsafe { libc::kqueue() };
        if kq < 0 {
            return Err(std::io::Error::last_os_error());
        }
        let inner = Arc::new(WatcherInner {
            kq,
            stop: AtomicBool::new(false),
            notifier: Mutex::new(Some(notifier)),
            watched: Mutex::new(BTreeMap::new()),
        });
        let thread_inner = inner.clone();
        let handle = std::thread::Builder::new()
            .name("supermachine-posixfs-watch".to_owned())
            .spawn(move || run_watcher(thread_inner))
            .map_err(|e| std::io::Error::other(e.to_string()))?;
        *watcher_slot = Some(Watcher {
            inner,
            handle: Some(handle),
        });
        Ok(())
    }

    /// Add `path` to the kqueue watch list, associated with `nodeid`.
    /// Best-effort: returns without error if watching is disabled or
    /// the file can't be opened. Calls should be cheap (open + 1
    /// kevent) so it's safe to invoke on every OPEN/SETUPMAPPING.
    fn watch_inode(&self, nodeid: u64, path: &std::path::Path) {
        let watcher = self.watcher.lock().unwrap();
        let Some(w) = watcher.as_ref() else { return };
        // Open a dedicated fd so the watch persists independent of
        // the guest's open file handle.
        let c = match CString::new(path.as_os_str().as_bytes()) {
            Ok(c) => c,
            Err(_) => return,
        };
        let fd = unsafe { libc::open(c.as_ptr(), libc::O_RDONLY | libc::O_EVTONLY) };
        if fd < 0 {
            return;
        }
        let owned = unsafe { OwnedFd::from_raw_fd(fd) };

        // Look up parent_nodeid + name so the watcher thread can emit
        // FUSE_NOTIFY_INVAL_ENTRY on rename/delete. Required to flush
        // the guest's dentry cache (1 s default TTL otherwise).
        let (parent_nodeid, name) = {
            let st = self.st.lock().unwrap();
            // Reverse-lookup in the (parent, name) → nodeid table.
            let entry = st
                .children
                .iter()
                .find(|(_, id)| **id == nodeid)
                .map(|((p, n), _)| (*p, n.clone()));
            match entry {
                Some(e) => e,
                None => return, // root or unknown — skip dentry invalidation
            }
        };

        let mut watched = w.inner.watched.lock().unwrap();
        // Already watching this nodeid? Drop old, install new.
        watched.retain(|_, e| e.nodeid != nodeid);

        let ev = libc::kevent {
            ident: fd as libc::uintptr_t,
            filter: libc::EVFILT_VNODE,
            flags: libc::EV_ADD | libc::EV_CLEAR,
            fflags: libc::NOTE_DELETE
                | libc::NOTE_RENAME
                | libc::NOTE_WRITE
                | libc::NOTE_EXTEND
                | libc::NOTE_ATTRIB,
            data: 0,
            udata: std::ptr::null_mut(),
        };
        let mut event = ev;
        let rc = unsafe {
            libc::kevent(
                w.inner.kq,
                &mut event as *mut _,
                1,
                std::ptr::null_mut(),
                0,
                std::ptr::null(),
            )
        };
        if rc < 0 {
            return;
        }
        watched.insert(
            fd,
            WatchedEntry {
                nodeid,
                parent_nodeid,
                name,
                _owned_fd: owned,
            },
        );
    }

    fn host_path_of(&self, nodeid: u64) -> Result<PathBuf, Errno> {
        let st = self.st.lock().unwrap();
        st.inodes.get(&nodeid).map(|i| i.host_path.clone()).ok_or(ENOENT)
    }

    fn kind_of(&self, nodeid: u64) -> Result<Kind, Errno> {
        let st = self.st.lock().unwrap();
        st.inodes.get(&nodeid).map(|i| i.kind).ok_or(ENOENT)
    }
}

/// Map a host (macOS aarch64) errno value to its Linux equivalent.
///
/// FUSE is a Linux-kernel protocol — the guest's libc/glibc/musl decode
/// the wire-side error number using `<asm-generic/errno.h>` /
/// `<bits/errno.h>`. When a host libc syscall fails on macOS we must
/// translate the returned errno before sending it to the guest,
/// otherwise the guest sees the wrong symbolic error (e.g. macOS
/// `ENOTEMPTY = 66` is `EREMOTE = 66` on Linux; macOS `ENOSYS = 78`
/// is `EREMCHG = 78` on Linux; macOS `EAGAIN = 35` is `EDEADLK = 35`
/// on Linux; and so on).
///
/// Errnos 1..=34 are POSIX-identical across macOS and Linux
/// (EPERM=1, ENOENT=2, ESRCH=3, EINTR=4, EIO=5, ..., EDOM=33,
/// ERANGE=34). 35+ is where the numbering diverges. Anything not
/// listed below passes through unchanged either because it's in the
/// shared 1..=34 range or because the numeric value happens to agree.
///
/// References:
///   macOS: `<sys/errno.h>` —
///     https://github.com/apple-oss-distributions/xnu/blob/main/bsd/sys/errno.h
///   Linux: `<asm-generic/errno-base.h>`, `<asm-generic/errno.h>`.
#[cfg(target_os = "macos")]
fn host_to_linux_errno(host: i32) -> i32 {
    match host {
        // EAGAIN and EWOULDBLOCK are both 35 on macOS, both 11 on Linux.
        35 => 11,    // EAGAIN / EWOULDBLOCK
        11 => 35,    // EDEADLK / EDEADLOCK
        63 => 36,    // ENAMETOOLONG
        77 => 37,    // ENOLCK
        78 => 38,    // ENOSYS
        66 => 39,    // ENOTEMPTY  ← the rmdir-non-empty bug we were chasing
        62 => 40,    // ELOOP
        91 => 42,    // ENOMSG
        90 => 43,    // EIDRM
        96 => 61,    // ENODATA
        84 => 75,    // EOVERFLOW
        94 => 74,    // EBADMSG
        92 => 84,    // EILSEQ
        38 => 88,    // ENOTSOCK
        39 => 89,    // EDESTADDRREQ
        40 => 90,    // EMSGSIZE
        41 => 91,    // EPROTOTYPE
        42 => 92,    // ENOPROTOOPT
        43 => 93,    // EPROTONOSUPPORT
        44 => 94,    // ESOCKTNOSUPPORT
        102 => 95,   // EOPNOTSUPP  (== ENOTSUP on macOS, same number)
        46 => 96,    // EPFNOSUPPORT
        47 => 97,    // EAFNOSUPPORT
        48 => 98,    // EADDRINUSE
        49 => 99,    // EADDRNOTAVAIL
        50 => 100,   // ENETDOWN
        51 => 101,   // ENETUNREACH
        52 => 102,   // ENETRESET
        53 => 103,   // ECONNABORTED
        54 => 104,   // ECONNRESET
        55 => 105,   // ENOBUFS
        56 => 106,   // EISCONN
        57 => 107,   // ENOTCONN
        58 => 108,   // ESHUTDOWN
        59 => 109,   // ETOOMANYREFS
        60 => 110,   // ETIMEDOUT
        61 => 111,   // ECONNREFUSED
        64 => 112,   // EHOSTDOWN
        65 => 113,   // EHOSTUNREACH
        37 => 114,   // EALREADY
        36 => 115,   // EINPROGRESS
        70 => 116,   // ESTALE
        69 => 122,   // EDQUOT
        89 => 125,   // ECANCELED
        105 => 130,  // EOWNERDEAD
        104 => 131,  // ENOTRECOVERABLE
        _ => host,
    }
}

#[cfg(not(target_os = "macos"))]
#[inline]
fn host_to_linux_errno(host: i32) -> i32 {
    host
}

/// Convert a `std::io::Error` from a host libc call into the
/// Linux-form FUSE [`Errno`] (a negative i32). Falls back to `-EIO`
/// when the error has no raw os code.
fn io_err_to_linux(e: &std::io::Error) -> Errno {
    -host_to_linux_errno(e.raw_os_error().unwrap_or(libc::EIO))
}

fn errno_now() -> Errno {
    -host_to_linux_errno(
        std::io::Error::last_os_error()
            .raw_os_error()
            .unwrap_or(libc::EIO),
    )
}

fn attr_from_meta(ino: u64, md: &std::fs::Metadata) -> Attr {
    let mode_full = md.mode();
    let perm_bits = mode_full & 0o7777;
    let typ_bits = if md.is_dir() {
        S_IFDIR
    } else if md.is_file() {
        S_IFREG
    } else if md.file_type().is_symlink() {
        S_IFLNK
    } else {
        match mode_full & S_IFMT {
            S_IFBLK => S_IFBLK,
            S_IFCHR => S_IFCHR,
            S_IFIFO => S_IFIFO,
            S_IFSOCK => S_IFSOCK,
            _ => 0,
        }
    };
    Attr {
        ino,
        size: md.size(),
        blocks: md.blocks(),
        atime: md.atime() as u64,
        mtime: md.mtime() as u64,
        ctime: md.ctime() as u64,
        atimensec: md.atime_nsec() as u32,
        mtimensec: md.mtime_nsec() as u32,
        ctimensec: md.ctime_nsec() as u32,
        mode: typ_bits | perm_bits,
        nlink: md.nlink() as u32,
        uid: md.uid(),
        gid: md.gid(),
        rdev: md.rdev() as u32,
        blksize: md.blksize() as u32,
        flags: 0,
    }
}

/// Build an `Attr` from a raw `libc::stat`. Used by paths that go
/// through `fstatat` / `lstat` (symlink + hard-link create) — these
/// can't use the `std::fs::Metadata` route since std would follow
/// symlinks or canonicalize the path under us.
fn attr_from_stat(ino: u64, st: &libc::stat) -> Attr {
    let mode_full = st.st_mode as u32;
    let perm_bits = mode_full & 0o7777;
    let typ_bits = match mode_full & S_IFMT {
        S_IFDIR => S_IFDIR,
        S_IFREG => S_IFREG,
        S_IFLNK => S_IFLNK,
        S_IFBLK => S_IFBLK,
        S_IFCHR => S_IFCHR,
        S_IFIFO => S_IFIFO,
        S_IFSOCK => S_IFSOCK,
        _ => 0,
    };
    // macOS exposes `st_atimespec` etc.; Linux uses `st_atim`. Pull
    // through the unix MetadataExt-style fields via a small cfg.
    #[cfg(target_os = "macos")]
    let (a, an, m, mn, c, cn) = (
        st.st_atime as u64,
        st.st_atime_nsec as u32,
        st.st_mtime as u64,
        st.st_mtime_nsec as u32,
        st.st_ctime as u64,
        st.st_ctime_nsec as u32,
    );
    #[cfg(not(target_os = "macos"))]
    let (a, an, m, mn, c, cn) = (
        st.st_atime as u64,
        st.st_atime_nsec as u32,
        st.st_mtime as u64,
        st.st_mtime_nsec as u32,
        st.st_ctime as u64,
        st.st_ctime_nsec as u32,
    );
    Attr {
        ino,
        size: st.st_size as u64,
        blocks: st.st_blocks as u64,
        atime: a,
        mtime: m,
        ctime: c,
        atimensec: an,
        mtimensec: mn,
        ctimensec: cn,
        mode: typ_bits | perm_bits,
        nlink: st.st_nlink as u32,
        uid: st.st_uid,
        gid: st.st_gid,
        rdev: st.st_rdev as u32,
        blksize: st.st_blksize as u32,
        flags: 0,
    }
}

/// Open a host directory as an O_DIRECTORY|O_NOFOLLOW dirfd we can
/// hand to *at()-family syscalls. Returns ENOTDIR if `path` isn't a
/// directory.
fn open_dirfd(path: &std::path::Path) -> Result<OwnedFd, Errno> {
    let c = CString::new(path.as_os_str().as_bytes()).map_err(|_| EINVAL)?;
    // O_NOFOLLOW so a hostile rename of a parent-component into a
    // symlink can't redirect us elsewhere mid-call.
    let fd = unsafe {
        libc::open(c.as_ptr(), libc::O_RDONLY | libc::O_DIRECTORY | libc::O_NOFOLLOW)
    };
    if fd < 0 {
        return Err(errno_now());
    }
    // SAFETY: fd is fresh from open() above.
    Ok(unsafe { OwnedFd::from_raw_fd(fd) })
}

fn kind_from_meta(md: &std::fs::Metadata) -> Kind {
    if md.is_dir() {
        Kind::Dir
    } else if md.is_file() {
        Kind::File
    } else if md.file_type().is_symlink() {
        Kind::Symlink
    } else {
        Kind::Other
    }
}

fn name_safe(name: &OsStr) -> Result<(), Errno> {
    let bytes = name.as_bytes();
    if bytes.is_empty() || bytes == b"." || bytes == b".." {
        return Err(EINVAL);
    }
    if bytes.contains(&b'/') {
        return Err(EINVAL);
    }
    // Defense-in-depth: reject embedded NULs. The C string layer
    // (CString::new) would catch these too, but we'd rather not let
    // hostile names get past our first filter.
    if bytes.contains(&0u8) {
        return Err(EINVAL);
    }
    Ok(())
}

impl FsBackend for PosixFs {
    fn lookup(&self, parent: u64, name: &OsStr) -> Result<Entry, Errno> {
        name_safe(name)?;
        let parent_path = self.host_path_of(parent)?;
        let path = parent_path.join(name);
        let md = std::fs::metadata(&path).map_err(|e| io_err_to_linux(&e))?;

        // Symlink containment check. Under `SymlinkPolicy::Deny` and
        // `::Opaque` (default), canonicalize the path and verify it
        // lives under our mount root. Protects against a hostile guest
        // planting symlinks like `etc -> /etc` in their mount and
        // reading host secrets. `Follow` skips the check.
        //
        // Cheap fast path: if the file is NOT a symlink and the
        // PARENT is already validated (which is true for every
        // non-root inode in our table), the resolved path can't
        // escape — we don't even need canonicalize. Only re-canonicalize
        // when a symlink is actually involved.
        if self.symlinks != SymlinkPolicy::Follow {
            let needs_check = std::fs::symlink_metadata(&path)
                .map(|m| m.file_type().is_symlink() || md.file_type().is_symlink())
                .unwrap_or(true);
            if needs_check {
                let canonical = std::fs::canonicalize(&path)
                    .map_err(|e| io_err_to_linux(&e))?;
                if !canonical.starts_with(&self.root) {
                    return Err(EACCES);
                }
            }
        }

        let mut st = self.st.lock().unwrap();
        // Reuse existing nodeid if we've looked this child up before.
        let key = (parent, name.as_bytes().to_vec());
        let nodeid = match st.children.get(&key) {
            Some(&id) => id,
            None => {
                let id = st.next_nodeid;
                st.next_nodeid += 1;
                st.inodes.insert(
                    id,
                    InodeInfo {
                        host_path: path.clone(),
                        kind: kind_from_meta(&md),
                    },
                );
                st.children.insert(key, id);
                id
            }
        };
        let attr = attr_from_meta(nodeid, &md);
        Ok(Entry {
            nodeid,
            generation: 0,
            attr,
            entry_valid: 1,
            attr_valid: 1,
        })
    }

    fn forget(&self, _nodeid: u64, _nlookup: u64) {
        // We retain inode entries indefinitely for path stability.
        // Real production would reference-count and prune. Tests
        // don't depend on prune so we no-op.
    }

    fn getattr(&self, nodeid: u64, _fh: Option<u64>) -> Result<Attr, Errno> {
        let path = self.host_path_of(nodeid)?;
        let md = std::fs::metadata(&path).map_err(|e| io_err_to_linux(&e))?;
        Ok(attr_from_meta(nodeid, &md))
    }

    fn open(&self, nodeid: u64, flags: u32) -> Result<u64, Errno> {
        let path = self.host_path_of(nodeid)?;
        match self.kind_of(nodeid)? {
            Kind::Dir => return Err(EISDIR),
            _ => {}
        }
        // Best-effort: register a kqueue watch on this inode so host
        // changes propagate to the guest as FUSE_NOTIFY_INVAL_INODE.
        self.watch_inode(nodeid, &path);
        let c = CString::new(path.as_os_str().as_bytes()).map_err(|_| EINVAL)?;
        // Mask out CREAT/EXCL: virtio-fs always issues a LOOKUP first;
        // OPEN should not create. We honor RDONLY/RDWR/WRONLY + DIRECT.
        let access = flags as i32 & libc::O_ACCMODE;
        let fd = unsafe { libc::open(c.as_ptr(), access) };
        if fd < 0 {
            return Err(errno_now());
        }
        // SAFETY: open returned a valid fd above.
        let owned = unsafe { OwnedFd::from_raw_fd(fd) };
        let mut st = self.st.lock().unwrap();
        let fh = st.next_fh;
        st.next_fh += 1;
        st.handles.insert(fh, owned);
        Ok(fh)
    }

    fn read(&self, _nodeid: u64, fh: u64, offset: u64, size: u32) -> Result<Vec<u8>, Errno> {
        let st = self.st.lock().unwrap();
        let raw = st.handles.get(&fh).ok_or(EBADF)?.as_raw_fd();
        drop(st);
        let mut buf = vec![0u8; size as usize];
        let n = unsafe {
            libc::pread(raw, buf.as_mut_ptr() as *mut _, buf.len(), offset as libc::off_t)
        };
        if n < 0 {
            return Err(errno_now());
        }
        buf.truncate(n as usize);
        Ok(buf)
    }

    fn release(&self, _nodeid: u64, fh: u64) -> Result<(), Errno> {
        let mut st = self.st.lock().unwrap();
        st.handles.remove(&fh).ok_or(EBADF).map(|_| ())
    }

    fn write(&self, _nodeid: u64, fh: u64, offset: u64, data: &[u8]) -> Result<u32, Errno> {
        let st = self.st.lock().unwrap();
        let raw = st.handles.get(&fh).ok_or(EBADF)?.as_raw_fd();
        drop(st);
        let n = unsafe {
            libc::pwrite(
                raw,
                data.as_ptr() as *const _,
                data.len(),
                offset as libc::off_t,
            )
        };
        if n < 0 {
            return Err(errno_now());
        }
        Ok(n as u32)
    }

    fn fsync(&self, _nodeid: u64, fh: u64, datasync: bool) -> Result<(), Errno> {
        let st = self.st.lock().unwrap();
        let raw = st.handles.get(&fh).ok_or(EBADF)?.as_raw_fd();
        drop(st);
        let rc = unsafe {
            if datasync {
                libc::fsync(raw) // macOS doesn't have fdatasync; fsync is the stronger form
            } else {
                libc::fsync(raw)
            }
        };
        if rc != 0 {
            return Err(errno_now());
        }
        Ok(())
    }

    fn opendir(&self, nodeid: u64, _flags: u32) -> Result<u64, Errno> {
        let path = self.host_path_of(nodeid)?;
        let c = CString::new(path.as_os_str().as_bytes()).map_err(|_| EINVAL)?;
        let fd = unsafe { libc::open(c.as_ptr(), libc::O_RDONLY | libc::O_DIRECTORY) };
        if fd < 0 {
            return Err(errno_now());
        }
        let owned = unsafe { OwnedFd::from_raw_fd(fd) };
        let mut st = self.st.lock().unwrap();
        let fh = st.next_fh;
        st.next_fh += 1;
        st.handles.insert(fh, owned);
        Ok(fh)
    }

    fn readdir(
        &self,
        nodeid: u64,
        _fh: u64,
        offset: u64,
        _size: u32,
    ) -> Result<Vec<DirEntry>, Errno> {
        // We rebuild the dirent list each call rather than caching
        // (real production should cache per-fh between consecutive
        // offsets). Use std::fs::read_dir for portability.
        let path = self.host_path_of(nodeid)?;
        let rd = std::fs::read_dir(&path).map_err(|e| io_err_to_linux(&e))?;
        let mut out = Vec::new();
        for (i, entry_res) in rd.enumerate() {
            if (i as u64) < offset {
                continue;
            }
            let entry = match entry_res {
                Ok(e) => e,
                Err(_) => continue,
            };
            let typ = match entry.file_type() {
                Ok(t) if t.is_dir() => DT_DIR,
                Ok(t) if t.is_file() => DT_REG,
                Ok(t) if t.is_symlink() => DT_LNK,
                Ok(t) => match t {
                    t if t.is_block_device() => DT_BLK,
                    t if t.is_char_device() => DT_CHR,
                    t if t.is_fifo() => DT_FIFO,
                    t if t.is_socket() => DT_SOCK,
                    _ => DT_UNKNOWN,
                },
                Err(_) => DT_UNKNOWN,
            };
            // Inode number: we can't allocate a nodeid until LOOKUP
            // runs (the guest will issue a LOOKUP for any entry it
            // wants to use). Send the host's st_ino so directory
            // listings show stable values; the guest only relies on
            // the name+type for the readdir surface.
            let ino = entry.metadata().map(|m| m.ino()).unwrap_or(0);
            out.push(DirEntry {
                ino,
                name: entry.file_name().as_bytes().to_vec(),
                typ,
            });
        }
        Ok(out)
    }

    fn releasedir(&self, _nodeid: u64, fh: u64) -> Result<(), Errno> {
        let mut st = self.st.lock().unwrap();
        st.handles.remove(&fh).ok_or(EBADF).map(|_| ())
    }

    fn statfs(&self, nodeid: u64) -> Result<StatFs, Errno> {
        let path = self.host_path_of(nodeid)?;
        let c = CString::new(path.as_os_str().as_bytes()).map_err(|_| EINVAL)?;
        let mut s: libc::statfs = unsafe { std::mem::zeroed() };
        if unsafe { libc::statfs(c.as_ptr(), &mut s) } < 0 {
            return Err(errno_now());
        }
        Ok(StatFs {
            blocks: s.f_blocks,
            bfree: s.f_bfree,
            bavail: s.f_bavail,
            files: s.f_files,
            ffree: s.f_ffree,
            bsize: s.f_bsize as u32,
            namelen: 255,
            frsize: s.f_bsize as u32,
        })
    }

    fn create(
        &self,
        parent: u64,
        name: &OsStr,
        mode: u32,
        flags: u32,
    ) -> Result<(crate::fuse::backend::Entry, u64), Errno> {
        name_safe(name)?;
        let parent_path = self.host_path_of(parent)?;
        let full = parent_path.join(name);
        let c = CString::new(full.as_os_str().as_bytes()).map_err(|_| EINVAL)?;
        let access = flags as i32 & libc::O_ACCMODE;
        let fd = unsafe {
            libc::open(
                c.as_ptr(),
                access | libc::O_CREAT | libc::O_EXCL,
                mode as libc::c_uint,
            )
        };
        if fd < 0 {
            return Err(errno_now());
        }
        let owned = unsafe { OwnedFd::from_raw_fd(fd) };

        let md = std::fs::metadata(&full).map_err(|e| io_err_to_linux(&e))?;
        let mut st = self.st.lock().unwrap();
        let nodeid = st.next_nodeid;
        st.next_nodeid += 1;
        st.inodes.insert(
            nodeid,
            InodeInfo {
                host_path: full.clone(),
                kind: kind_from_meta(&md),
            },
        );
        st.children.insert((parent, name.as_bytes().to_vec()), nodeid);
        let fh = st.next_fh;
        st.next_fh += 1;
        st.handles.insert(fh, owned);
        let attr = attr_from_meta(nodeid, &md);
        Ok((
            crate::fuse::backend::Entry {
                nodeid,
                generation: 0,
                attr,
                entry_valid: 1,
                attr_valid: 1,
            },
            fh,
        ))
    }

    fn mkdir(&self, parent: u64, name: &OsStr, mode: u32) -> Result<crate::fuse::backend::Entry, Errno> {
        name_safe(name)?;
        let parent_path = self.host_path_of(parent)?;
        let full = parent_path.join(name);
        let c = CString::new(full.as_os_str().as_bytes()).map_err(|_| EINVAL)?;
        let rc = unsafe { libc::mkdir(c.as_ptr(), mode as libc::mode_t) };
        if rc != 0 {
            return Err(errno_now());
        }
        let md = std::fs::metadata(&full).map_err(|e| io_err_to_linux(&e))?;
        let mut st = self.st.lock().unwrap();
        let nodeid = st.next_nodeid;
        st.next_nodeid += 1;
        st.inodes.insert(
            nodeid,
            InodeInfo {
                host_path: full,
                kind: Kind::Dir,
            },
        );
        st.children.insert((parent, name.as_bytes().to_vec()), nodeid);
        Ok(crate::fuse::backend::Entry {
            nodeid,
            generation: 0,
            attr: attr_from_meta(nodeid, &md),
            entry_valid: 1,
            attr_valid: 1,
        })
    }

    fn unlink(&self, parent: u64, name: &OsStr) -> Result<(), Errno> {
        name_safe(name)?;
        let parent_path = self.host_path_of(parent)?;
        let full = parent_path.join(name);
        let c = CString::new(full.as_os_str().as_bytes()).map_err(|_| EINVAL)?;
        let rc = unsafe { libc::unlink(c.as_ptr()) };
        if rc != 0 {
            return Err(errno_now());
        }
        let mut st = self.st.lock().unwrap();
        st.children.remove(&(parent, name.as_bytes().to_vec()));
        Ok(())
    }

    fn rmdir(&self, parent: u64, name: &OsStr) -> Result<(), Errno> {
        name_safe(name)?;
        let parent_path = self.host_path_of(parent)?;
        let full = parent_path.join(name);
        let c = CString::new(full.as_os_str().as_bytes()).map_err(|_| EINVAL)?;
        let rc = unsafe { libc::rmdir(c.as_ptr()) };
        if rc != 0 {
            return Err(errno_now());
        }
        let mut st = self.st.lock().unwrap();
        st.children.remove(&(parent, name.as_bytes().to_vec()));
        Ok(())
    }

    fn symlink(
        &self,
        parent: u64,
        name: &OsStr,
        target: &OsStr,
    ) -> Result<crate::fuse::backend::Entry, Errno> {
        // Deny mode: refuse outright with EPERM. POSIX symlink(2)
        // returns EPERM when the filesystem doesn't support symlinks,
        // which is the closest match to "policy denies symlinks here".
        if self.symlinks == SymlinkPolicy::Deny {
            return Err(EPERM);
        }
        name_safe(name)?;
        // The `target` is opaque bytes per POSIX. Only constraint:
        // CString rejects embedded NULs (the kernel does too via
        // symlinkat). Targets containing `..` or absolute paths are
        // legal — the host doesn't resolve them, the guest's kernel
        // does at lookup time.
        if target.as_bytes().is_empty() || target.as_bytes().contains(&0u8) {
            return Err(EINVAL);
        }
        let parent_path = self.host_path_of(parent)?;
        // Open parent as O_PATH|O_DIRECTORY|O_NOFOLLOW dirfd so we use
        // symlinkat — no race against a parent rename. macOS supports
        // O_NOFOLLOW + O_DIRECTORY. (Linux additionally needs O_PATH
        // which we'll add behind a cfg when we port; macOS doesn't
        // have O_PATH but accepts plain O_DIRECTORY|O_RDONLY for an
        // *at-only dirfd.)
        let parent_dirfd = open_dirfd(&parent_path)?;
        let target_c = CString::new(target.as_bytes()).map_err(|_| EINVAL)?;
        let name_c = CString::new(name.as_bytes()).map_err(|_| EINVAL)?;
        let rc = unsafe {
            libc::symlinkat(target_c.as_ptr(), parent_dirfd.as_raw_fd(), name_c.as_ptr())
        };
        if rc != 0 {
            return Err(errno_now());
        }
        // Stat the new symlink itself (do NOT follow).
        let mut stb: libc::stat = unsafe { std::mem::zeroed() };
        let rc = unsafe {
            libc::fstatat(
                parent_dirfd.as_raw_fd(),
                name_c.as_ptr(),
                &mut stb,
                libc::AT_SYMLINK_NOFOLLOW,
            )
        };
        if rc != 0 {
            return Err(errno_now());
        }
        let full = parent_path.join(name);
        let mut st = self.st.lock().unwrap();
        let nodeid = st.next_nodeid;
        st.next_nodeid += 1;
        st.inodes.insert(
            nodeid,
            InodeInfo {
                host_path: full,
                kind: Kind::Symlink,
            },
        );
        st.children.insert((parent, name.as_bytes().to_vec()), nodeid);
        Ok(crate::fuse::backend::Entry {
            nodeid,
            generation: 0,
            attr: attr_from_stat(nodeid, &stb),
            entry_valid: 1,
            attr_valid: 1,
        })
    }

    fn readlink(&self, nodeid: u64) -> Result<Vec<u8>, Errno> {
        let path = self.host_path_of(nodeid)?;
        let c = CString::new(path.as_os_str().as_bytes()).map_err(|_| EINVAL)?;
        // PATH_MAX is 1024 on Darwin and 4096 on Linux. Use 4096 as a
        // safe upper bound for both; if the link is longer the
        // truncation is silent per POSIX (readlink returns the
        // truncated length and we honor it).
        let mut buf = vec![0u8; 4096];
        let n = unsafe { libc::readlink(c.as_ptr(), buf.as_mut_ptr() as *mut _, buf.len()) };
        if n < 0 {
            return Err(errno_now());
        }
        buf.truncate(n as usize);
        Ok(buf)
    }

    fn link(
        &self,
        nodeid: u64,
        new_parent: u64,
        new_name: &OsStr,
    ) -> Result<crate::fuse::backend::Entry, Errno> {
        // Symmetric with symlink under Deny — "no metadata surprises"
        // means no new hard links either. Hard links can't ESCAPE the
        // mount (link(2) returns EXDEV across filesystems and our
        // mount is one host FS), but `Deny` is about predictability
        // not just escape resistance.
        if self.symlinks == SymlinkPolicy::Deny {
            return Err(EPERM);
        }
        name_safe(new_name)?;
        let src_path = self.host_path_of(nodeid)?;
        let new_parent_path = self.host_path_of(new_parent)?;
        let dst_path = new_parent_path.join(new_name);
        let src_c = CString::new(src_path.as_os_str().as_bytes()).map_err(|_| EINVAL)?;
        let dst_c = CString::new(dst_path.as_os_str().as_bytes()).map_err(|_| EINVAL)?;
        let rc = unsafe { libc::link(src_c.as_ptr(), dst_c.as_ptr()) };
        if rc != 0 {
            return Err(errno_now());
        }
        // Stat the destination without following — we want the file's
        // own attrs (and a hard link can't be a symlink anyway, but
        // belt-and-braces).
        let dst_c2 = CString::new(dst_path.as_os_str().as_bytes()).map_err(|_| EINVAL)?;
        let mut stb: libc::stat = unsafe { std::mem::zeroed() };
        let rc = unsafe { libc::lstat(dst_c2.as_ptr(), &mut stb) };
        if rc != 0 {
            return Err(errno_now());
        }
        let mut st = self.st.lock().unwrap();
        let new_nodeid = st.next_nodeid;
        st.next_nodeid += 1;
        // The new dentry gets a fresh nodeid pointing at the same
        // underlying file. Two nodeids mapping to one host path is
        // legal — open()/stat() on either both hit the same inode.
        let kind = if (stb.st_mode as u32 & S_IFMT) == S_IFDIR {
            Kind::Dir
        } else if (stb.st_mode as u32 & S_IFMT) == S_IFREG {
            Kind::File
        } else if (stb.st_mode as u32 & S_IFMT) == S_IFLNK {
            Kind::Symlink
        } else {
            Kind::Other
        };
        st.inodes.insert(
            new_nodeid,
            InodeInfo {
                host_path: dst_path,
                kind,
            },
        );
        st.children
            .insert((new_parent, new_name.as_bytes().to_vec()), new_nodeid);
        Ok(crate::fuse::backend::Entry {
            nodeid: new_nodeid,
            generation: 0,
            attr: attr_from_stat(new_nodeid, &stb),
            entry_valid: 1,
            attr_valid: 1,
        })
    }

    fn dax_map(
        &self,
        nodeid: u64,
        fh: u64,
        foffset: u64,
        len: u64,
        prot: u32,
    ) -> Result<*mut u8, Errno> {
        // The guest's iomap-driven read path (used by `dax=always`)
        // sends SETUPMAPPING with `fh = u64::MAX` because the read
        // is inode-level (no userspace fd backs it). Open the file
        // on demand by walking the inode table back to the host
        // path. For mmap-driven SETUPMAPPING (spike-22 zero-copy)
        // the fh IS a valid handle from a prior FUSE_OPEN — use it.
        //
        // mmap() captures the fd internally; the OwnedFd we hold
        // here covers the lifetime through the mmap call. After
        // mmap returns success the kernel keeps its own reference,
        // so dropping our OwnedFd on the unwind path is safe.
        let opened_fresh: Option<OwnedFd>;
        let raw = if fh == u64::MAX {
            let path = self.host_path_of(nodeid)?;
            let c = CString::new(path.as_os_str().as_bytes()).map_err(|_| EINVAL)?;
            // Try RDWR first so writes through DAX work; fall back
            // to RDONLY for files we can't open RW (e.g. read-only
            // host file).
            let mut fd = unsafe { libc::open(c.as_ptr(), libc::O_RDWR) };
            if fd < 0 {
                fd = unsafe { libc::open(c.as_ptr(), libc::O_RDONLY) };
            }
            if fd < 0 {
                return Err(errno_now());
            }
            // SAFETY: fd is fresh from open().
            let owned = unsafe { OwnedFd::from_raw_fd(fd) };
            let raw = owned.as_raw_fd();
            opened_fresh = Some(owned);
            raw
        } else {
            let st = self.st.lock().unwrap();
            let raw = st.handles.get(&fh).ok_or(EBADF)?.as_raw_fd();
            drop(st);
            opened_fresh = None;
            raw
        };
        // Suppress unused-var lint when fh != MAX.
        let _ = &opened_fresh;
        // Spike 22 validated: PROT_READ|PROT_WRITE host backing
        // works for both R and RW DAX. We always map RW on host
        // and let the guest-side stage-2 protection enforce R-only
        // semantics. Apple's HVF requires writable host backing
        // regardless of guest-side flags.
        let _ = prot; // recorded for hv_vm_map elsewhere
        let host_prot = libc::PROT_READ | libc::PROT_WRITE;
        let ptr = unsafe {
            libc::mmap(
                std::ptr::null_mut(),
                len as usize,
                host_prot,
                libc::MAP_SHARED,
                raw,
                foffset as libc::off_t,
            )
        };
        if ptr == libc::MAP_FAILED {
            return Err(errno_now());
        }
        let mut st = self.st.lock().unwrap();
        st.dax_mmaps.insert(
            ptr as usize,
            Mmap {
                ptr: ptr as *mut u8,
                len: len as usize,
            },
        );
        Ok(ptr as *mut u8)
    }

    fn dax_unmap(&self, _nodeid: u64, host_va: *mut u8, _len: u64) -> Result<(), Errno> {
        let mut st = self.st.lock().unwrap();
        let m = st.dax_mmaps.remove(&(host_va as usize)).ok_or(EINVAL)?;
        let rc = unsafe { libc::munmap(m.ptr as *mut _, m.len) };
        if rc != 0 {
            return Err(errno_now());
        }
        Ok(())
    }
}

/// Background thread body: blocks in kevent(), dispatches NOTE_*
/// events to the notifier.
fn run_watcher(inner: Arc<WatcherInner>) {
    // Register a USER event so Drop can wake us.
    let wakeup = libc::kevent {
        ident: 0,
        filter: libc::EVFILT_USER,
        flags: libc::EV_ADD | libc::EV_CLEAR,
        fflags: 0,
        data: 0,
        udata: std::ptr::null_mut(),
    };
    let mut w = wakeup;
    unsafe {
        libc::kevent(
            inner.kq,
            &mut w as *mut _,
            1,
            std::ptr::null_mut(),
            0,
            std::ptr::null(),
        );
    }

    let mut events: [libc::kevent; 16] = unsafe { std::mem::zeroed() };
    loop {
        if inner.stop.load(Ordering::Acquire) {
            break;
        }
        let n = unsafe {
            libc::kevent(
                inner.kq,
                std::ptr::null(),
                0,
                events.as_mut_ptr(),
                events.len() as libc::c_int,
                std::ptr::null(),
            )
        };
        if n < 0 {
            let err = std::io::Error::last_os_error();
            if err.raw_os_error() == Some(libc::EINTR) {
                continue;
            }
            eprintln!("[posix-fs watcher] kevent failed: {err}; thread exiting");
            return;
        }
        if inner.stop.load(Ordering::Acquire) {
            break;
        }
        for ev in events.iter().take(n as usize) {
            if ev.filter == libc::EVFILT_USER {
                continue;
            }
            // Look up the watched entry for the kqueue ident.
            let fd = ev.ident as libc::c_int;
            let entry = inner
                .watched
                .lock()
                .unwrap()
                .get(&fd)
                .map(|e| (e.nodeid, e.parent_nodeid, e.name.clone()));
            let Some((nodeid, parent_nodeid, name)) = entry else { continue };
            if let Some(n) = inner.notifier.lock().unwrap().as_ref() {
                // Three-step invalidation:
                //   (off=0, len=-1) — invalidates DATA pages
                //   (off=0, len=0)  — invalidates ATTRS (size, mtime,
                //                      etc.). FUSE protocol uses this
                //                      (0,0) sentinel specifically.
                //   INVAL_ENTRY     — drops the dentry cache so a
                //                      re-open under the same name
                //                      hits the NEW inode after an
                //                      atomic rename.
                // All three are needed to make atomic-rename feel
                // synchronous to the guest's userspace.
                n.invalidate_inode(nodeid, 0, -1);
                n.invalidate_inode(nodeid, 0, 0);
                n.invalidate_entry(parent_nodeid, &name);
            }
            // If the file was deleted or renamed, drop the watch so
            // we don't keep an orphan fd.
            if ev.fflags & (libc::NOTE_DELETE | libc::NOTE_RENAME) != 0 {
                inner.watched.lock().unwrap().remove(&fd);
            }
        }
    }
}

// Unused-import suppression — std re-exports are conditional on macOS file_type extensions.
use std::os::unix::fs::FileTypeExt;
#[allow(unused_imports)]
use std::convert::TryFrom;
#[allow(dead_code)]
const _: () = {
    let _ = OsString::new;
    let _ = ENOSPC;
    let _ = ENOTDIR;
    let _ = EACCES;
    let _ = EIO;
};

#[cfg(test)]
mod tests {
    use super::*;

    fn tmpdir(name: &str) -> PathBuf {
        let pid = unsafe { libc::getpid() };
        let p = std::env::temp_dir().join(format!("posixfs-{pid}-{name}"));
        let _ = std::fs::remove_dir_all(&p);
        std::fs::create_dir_all(&p).unwrap();
        p
    }

    #[test]
    fn lookup_and_read_real_file() {
        let dir = tmpdir("t1");
        std::fs::write(dir.join("hello.txt"), b"hi from posix").unwrap();
        let fs = PosixFs::new(&dir).unwrap();
        let e = fs.lookup(FUSE_ROOT_ID, OsStr::new("hello.txt")).unwrap();
        assert!(e.attr.size == 13);
        let fh = fs.open(e.nodeid, libc::O_RDONLY as u32).unwrap();
        let buf = fs.read(e.nodeid, fh, 0, 64).unwrap();
        assert_eq!(buf, b"hi from posix");
        fs.release(e.nodeid, fh).unwrap();
    }

    #[test]
    fn readdir_lists_real_entries_with_types() {
        let dir = tmpdir("t2");
        std::fs::write(dir.join("a.txt"), b"a").unwrap();
        std::fs::create_dir_all(dir.join("sub")).unwrap();
        let fs = PosixFs::new(&dir).unwrap();
        let dh = fs.opendir(FUSE_ROOT_ID, 0).unwrap();
        let entries = fs.readdir(FUSE_ROOT_ID, dh, 0, 4096).unwrap();
        let by_name: std::collections::HashMap<&[u8], u32> =
            entries.iter().map(|e| (e.name.as_slice(), e.typ)).collect();
        assert_eq!(by_name[&b"a.txt"[..]], DT_REG);
        assert_eq!(by_name[&b"sub"[..]], DT_DIR);
        fs.releasedir(FUSE_ROOT_ID, dh).unwrap();
    }

    #[test]
    fn lookup_rejects_dotdot() {
        let dir = tmpdir("t3");
        let fs = PosixFs::new(&dir).unwrap();
        let err = fs.lookup(FUSE_ROOT_ID, OsStr::new("..")).unwrap_err();
        assert_eq!(err, EINVAL);
    }

    #[test]
    fn lookup_rejects_slash_in_name() {
        let dir = tmpdir("t4");
        let fs = PosixFs::new(&dir).unwrap();
        let err = fs.lookup(FUSE_ROOT_ID, OsStr::new("a/b")).unwrap_err();
        assert_eq!(err, EINVAL);
    }

    #[test]
    fn lookup_rejects_embedded_nul_in_name() {
        let dir = tmpdir("t4n");
        let fs = PosixFs::new(&dir).unwrap();
        let err = fs.lookup(FUSE_ROOT_ID, OsStr::from_bytes(b"foo\0bar")).unwrap_err();
        assert_eq!(err, EINVAL);
    }

    #[test]
    fn lookup_blocks_external_symlink_by_default() {
        // Mount root contains a symlink that points OUTSIDE the root.
        // LOOKUP must refuse with EACCES under the default (hostile-
        // tenant-safe) policy.
        let dir = tmpdir("t-sym-default");
        let outside = tmpdir("t-sym-outside");
        std::fs::write(outside.join("secret.txt"), b"do not leak").unwrap();
        std::os::unix::fs::symlink(&outside, dir.join("escape")).unwrap();

        let fs = PosixFs::new(&dir).unwrap();
        let err = fs.lookup(FUSE_ROOT_ID, OsStr::new("escape")).unwrap_err();
        assert_eq!(err, EACCES, "external symlink must be denied");
    }

    #[test]
    fn lookup_allows_internal_symlink_by_default() {
        // Symlinks that resolve INSIDE the mount root are fine even
        // under the strict default.
        let dir = tmpdir("t-sym-internal");
        std::fs::create_dir_all(dir.join("real")).unwrap();
        std::fs::write(dir.join("real/data.txt"), b"in-tree data").unwrap();
        std::os::unix::fs::symlink("real/data.txt", dir.join("link.txt")).unwrap();

        let fs = PosixFs::new(&dir).unwrap();
        // Direct lookup of the link must succeed.
        let e = fs.lookup(FUSE_ROOT_ID, OsStr::new("link.txt")).unwrap();
        assert!(e.attr.size > 0);
    }

    #[test]
    fn lookup_allows_external_symlink_when_opted_in() {
        // `new_unchecked` disables the check; absolute symlinks resolve.
        let dir = tmpdir("t-sym-opt-in");
        let outside = tmpdir("t-sym-target");
        std::fs::write(outside.join("ok.txt"), b"opt-in data").unwrap();
        std::os::unix::fs::symlink(outside.join("ok.txt"), dir.join("link")).unwrap();

        let fs = PosixFs::new_unchecked(&dir).unwrap();
        let e = fs.lookup(FUSE_ROOT_ID, OsStr::new("link")).unwrap();
        assert!(e.attr.size > 0);
    }

    #[test]
    fn dax_map_then_unmap_round_trip() {
        let dir = tmpdir("t5");
        // 32 KiB file with known pattern.
        let path = dir.join("data.bin");
        let mut data = vec![0u8; 32 * 1024];
        for (i, b) in data.iter_mut().enumerate() {
            *b = (i % 251) as u8;
        }
        std::fs::write(&path, &data).unwrap();

        let fs = PosixFs::new(&dir).unwrap();
        let e = fs.lookup(FUSE_ROOT_ID, OsStr::new("data.bin")).unwrap();
        let fh = fs.open(e.nodeid, libc::O_RDWR as u32).unwrap();

        // mmap the whole file via dax_map.
        let host_va = fs.dax_map(e.nodeid, fh, 0, 32 * 1024, 0).unwrap();
        assert!(!host_va.is_null());
        // Verify mmap contents match the file we wrote.
        let host_slice = unsafe { std::slice::from_raw_parts(host_va, 32 * 1024) };
        assert_eq!(host_slice, &data[..]);

        // Unmap — must succeed and clear internal tracking.
        fs.dax_unmap(e.nodeid, host_va, 32 * 1024).unwrap();
        // Calling dax_unmap on a host_va we don't know about must error.
        assert_eq!(fs.dax_unmap(e.nodeid, host_va, 32 * 1024).unwrap_err(), EINVAL);
        fs.release(e.nodeid, fh).unwrap();
    }

    #[test]
    fn read_eof_returns_empty() {
        let dir = tmpdir("t6");
        std::fs::write(dir.join("x"), b"short").unwrap();
        let fs = PosixFs::new(&dir).unwrap();
        let e = fs.lookup(FUSE_ROOT_ID, OsStr::new("x")).unwrap();
        let fh = fs.open(e.nodeid, libc::O_RDONLY as u32).unwrap();
        let eof = fs.read(e.nodeid, fh, 100, 10).unwrap();
        assert!(eof.is_empty());
        fs.release(e.nodeid, fh).unwrap();
    }

    #[test]
    fn open_directory_returns_eisdir() {
        let dir = tmpdir("t7");
        std::fs::create_dir_all(dir.join("sub")).unwrap();
        let fs = PosixFs::new(&dir).unwrap();
        let e = fs.lookup(FUSE_ROOT_ID, OsStr::new("sub")).unwrap();
        let err = fs.open(e.nodeid, libc::O_RDONLY as u32).unwrap_err();
        assert_eq!(err, EISDIR);
    }

    // === Symlink / Readlink / Link =====================================

    #[test]
    fn symlink_create_inside_mount() {
        let dir = tmpdir("t-sym-create");
        let fs = PosixFs::new(&dir).unwrap();
        let e = fs
            .symlink(FUSE_ROOT_ID, OsStr::new("link"), OsStr::new("target"))
            .unwrap();
        // Symlink mode bits.
        assert_eq!(e.attr.mode & S_IFMT, S_IFLNK);
        // readlink returns the bytes we stored, verbatim, no NUL.
        let target = fs.readlink(e.nodeid).unwrap();
        assert_eq!(target, b"target");
        // The host file actually is a symlink.
        let md = std::fs::symlink_metadata(dir.join("link")).unwrap();
        assert!(md.file_type().is_symlink());
    }

    #[test]
    fn symlink_create_external_target_stored_verbatim() {
        // POSIX symlink(2) doesn't resolve the target. A guest creating
        // a symlink whose target is `/etc/passwd` simply gets a symlink
        // whose CONTENTS say `/etc/passwd`. The host never opens that
        // file as part of symlink(); only readlink returns the bytes.
        // This is the safe behaviour: read-side LOOKUP's external-
        // symlink check rejects any subsequent `lookup("escape")`.
        let dir = tmpdir("t-sym-external");
        let fs = PosixFs::new(&dir).unwrap();
        let e = fs
            .symlink(FUSE_ROOT_ID, OsStr::new("escape"), OsStr::new("/etc/passwd"))
            .unwrap();
        let target = fs.readlink(e.nodeid).unwrap();
        assert_eq!(target, b"/etc/passwd");
        // Sanity: the host symlink really points there.
        let read = std::fs::read_link(dir.join("escape")).unwrap();
        assert_eq!(read.as_os_str(), OsStr::new("/etc/passwd"));
        // And LOOKUP (Opaque default) refuses to traverse it.
        let err = fs.lookup(FUSE_ROOT_ID, OsStr::new("escape")).unwrap_err();
        assert_eq!(err, EACCES);
    }

    #[test]
    fn symlink_create_blocked_by_deny() {
        let dir = tmpdir("t-sym-deny");
        let fs = PosixFs::new_with_symlinks(&dir, SymlinkPolicy::Deny).unwrap();
        let err = fs
            .symlink(FUSE_ROOT_ID, OsStr::new("link"), OsStr::new("target"))
            .unwrap_err();
        assert_eq!(err, EPERM);
        // And no host symlink was created.
        assert!(std::fs::symlink_metadata(dir.join("link")).is_err());
    }

    #[test]
    fn link_create_inside_mount() {
        let dir = tmpdir("t-link");
        std::fs::write(dir.join("orig"), b"hello").unwrap();
        let fs = PosixFs::new(&dir).unwrap();
        let src = fs.lookup(FUSE_ROOT_ID, OsStr::new("orig")).unwrap();
        let new_entry = fs
            .link(src.nodeid, FUSE_ROOT_ID, OsStr::new("alias"))
            .unwrap();
        // nlink on the new entry should be 2 — two names pointing at
        // the same underlying inode.
        assert_eq!(new_entry.attr.nlink, 2);
        // Reading via either name gives the same bytes.
        let fh = fs.open(new_entry.nodeid, libc::O_RDONLY as u32).unwrap();
        let buf = fs.read(new_entry.nodeid, fh, 0, 64).unwrap();
        assert_eq!(buf, b"hello");
        fs.release(new_entry.nodeid, fh).unwrap();
    }

    #[test]
    fn link_blocked_by_deny() {
        let dir = tmpdir("t-link-deny");
        std::fs::write(dir.join("orig"), b"hi").unwrap();
        let fs = PosixFs::new_with_symlinks(&dir, SymlinkPolicy::Deny).unwrap();
        let src = fs.lookup(FUSE_ROOT_ID, OsStr::new("orig")).unwrap();
        let err = fs
            .link(src.nodeid, FUSE_ROOT_ID, OsStr::new("alias"))
            .unwrap_err();
        assert_eq!(err, EPERM);
        // No host link was created.
        assert!(std::fs::symlink_metadata(dir.join("alias")).is_err());
    }

    #[test]
    fn lookup_with_opaque_blocks_external_symlink() {
        // Same as `lookup_blocks_external_symlink_by_default`, but
        // pinned to the new explicit Opaque policy so the test still
        // exercises the right code path once Default's value changes.
        let dir = tmpdir("t-sym-opaque-blocks");
        let outside = tmpdir("t-sym-opaque-out");
        std::fs::write(outside.join("x"), b"secret").unwrap();
        std::os::unix::fs::symlink(&outside, dir.join("escape")).unwrap();
        let fs = PosixFs::new_with_symlinks(&dir, SymlinkPolicy::Opaque).unwrap();
        let err = fs.lookup(FUSE_ROOT_ID, OsStr::new("escape")).unwrap_err();
        assert_eq!(err, EACCES);
    }

    #[test]
    fn lookup_with_follow_allows_external_symlink() {
        // Parity with the legacy `new_unchecked` / pre-0.5.5
        // `allow_external_symlinks: true` behaviour.
        let dir = tmpdir("t-sym-follow");
        let outside = tmpdir("t-sym-follow-out");
        std::fs::write(outside.join("ok.txt"), b"out-of-mount data").unwrap();
        std::os::unix::fs::symlink(outside.join("ok.txt"), dir.join("link")).unwrap();
        let fs = PosixFs::new_with_symlinks(&dir, SymlinkPolicy::Follow).unwrap();
        let e = fs.lookup(FUSE_ROOT_ID, OsStr::new("link")).unwrap();
        assert!(e.attr.size > 0);
    }

    // === Host → Linux errno translation ================================

    /// rmdir on a non-empty directory must surface as Linux ENOTEMPTY
    /// (39), not the macOS host value of 66 (which the guest's libc
    /// decodes as EREMOTE — the bug an integrator hit before this fix).
    #[test]
    fn rmdir_nonempty_returns_linux_enotempty() {
        let dir = tmpdir("t-rmdir-nonempty");
        std::fs::create_dir_all(dir.join("sub")).unwrap();
        std::fs::write(dir.join("sub/inside"), b"x").unwrap();
        let fs = PosixFs::new(&dir).unwrap();
        let err = fs.rmdir(FUSE_ROOT_ID, OsStr::new("sub")).unwrap_err();
        // Linux ENOTEMPTY == 39; Errno is the negative form.
        assert_eq!(err, -39, "rmdir non-empty must surface as Linux ENOTEMPTY (39), got {err}");
    }

    /// rmdir on an empty directory still succeeds (sanity that we
    /// didn't break the happy path while translating errnos).
    #[test]
    fn rmdir_empty_succeeds() {
        let dir = tmpdir("t-rmdir-empty");
        std::fs::create_dir_all(dir.join("sub")).unwrap();
        let fs = PosixFs::new(&dir).unwrap();
        fs.rmdir(FUSE_ROOT_ID, OsStr::new("sub")).unwrap();
        assert!(!dir.join("sub").exists());
    }

    /// Table-driven check that the translator emits the Linux numbers
    /// for the divergent errnos we care about. On non-macOS hosts the
    /// translator is the identity, so the assertions are written
    /// against the per-host expectation.
    #[test]
    fn host_to_linux_errno_table() {
        // (macOS host value, Linux wire value) — only meaningful on macOS.
        // On Linux the host value == Linux value already, so we just
        // sanity-check identity for a handful of POSIX-shared codes.
        #[cfg(target_os = "macos")]
        {
            // EAGAIN/EWOULDBLOCK
            assert_eq!(host_to_linux_errno(35), 11);
            // EDEADLK
            assert_eq!(host_to_linux_errno(11), 35);
            // ENAMETOOLONG
            assert_eq!(host_to_linux_errno(63), 36);
            // ENOSYS
            assert_eq!(host_to_linux_errno(78), 38);
            // ENOTEMPTY  ← the regression
            assert_eq!(host_to_linux_errno(66), 39);
            // ELOOP
            assert_eq!(host_to_linux_errno(62), 40);
            // EOVERFLOW
            assert_eq!(host_to_linux_errno(84), 75);
            // ENOTSOCK
            assert_eq!(host_to_linux_errno(38), 88);
            // EOPNOTSUPP / ENOTSUP (== 102 on macOS, 95 on Linux)
            assert_eq!(host_to_linux_errno(102), 95);
            // ETIMEDOUT
            assert_eq!(host_to_linux_errno(60), 110);
            // ECONNREFUSED
            assert_eq!(host_to_linux_errno(61), 111);
            // EINPROGRESS  (connect() pending — the one called out in the spec)
            assert_eq!(host_to_linux_errno(36), 115);
            // ESTALE
            assert_eq!(host_to_linux_errno(70), 116);
            // ECANCELED
            assert_eq!(host_to_linux_errno(89), 125);
            // EOWNERDEAD
            assert_eq!(host_to_linux_errno(105), 130);
            // ENOTRECOVERABLE
            assert_eq!(host_to_linux_errno(104), 131);
        }

        // POSIX-identical 1..=34 must always pass through.
        assert_eq!(host_to_linux_errno(libc::EPERM), 1);
        assert_eq!(host_to_linux_errno(libc::ENOENT), 2);
        assert_eq!(host_to_linux_errno(libc::EINTR), 4);
        assert_eq!(host_to_linux_errno(libc::EIO), 5);
        assert_eq!(host_to_linux_errno(libc::EINVAL), 22);
        assert_eq!(host_to_linux_errno(libc::ERANGE), 34);
    }

    // === Integration test stub (in-VM) ================================

    /// Boot a python image, virtio-fs mount a tmpdir, exec
    /// `ln -s a b && readlink b`, assert. This is the end-to-end test
    /// that proves npm install works inside guests — but kicking off a
    /// real VM requires the kernel patch series + the rest of the
    /// supermachine harness. Skipped for now; tracked as a follow-up
    /// once the read/write symlink path lands in the in-tree kernel.
    #[test]
    #[ignore = "requires a booted VM; turn this into a tsi_loopback-style integration test"]
    fn symlink_inside_guest_via_python_image() {
        // TODO(0.5.5+): build a python image, mount a host tmpdir, run
        // `python -c "import os; os.symlink('a','b'); print(os.readlink('b'))"`
        // via Vm::exec, assert stdout == "a\n".
    }
}