coreshift_core/

fs.rs

// This Source Code Form is subject to the terms of the Mozilla Public
// License, v. 2.0. If a copy of the MPL was not distributed with this
// file, You can obtain one at https://mozilla.org/MPL/2.0/

//! Filesystem-oriented low-level helpers.
//!
//! This module contains lightweight Linux and Android file probes and helpers
//! that are useful near the OS boundary, including path existence checks,
//! page-cache read-ahead hints, and symlink-safe read/write primitives.

use crate::CoreError;
use std::ffi::CString;
use std::fs;
use std::io::{Read, Write};
use std::os::unix::io::{AsRawFd, FromRawFd, OwnedFd, RawFd};
use std::path::Path;
use std::time::UNIX_EPOCH;

#[derive(Clone, Debug, PartialEq, Eq)]
pub struct PathFingerprint {
    pub len: u64,
    pub modified_ns: u128,
}

/// Return a fingerprint of the file metadata at the specified path.
///
/// ### Errors
/// - `EACCES`: Permission denied.
/// - `ENOENT`: The path does not exist.
pub fn path_fingerprint(path: &Path) -> Result<PathFingerprint, CoreError> {
    let metadata = std::fs::metadata(path).map_err(|err| {
        CoreError::sys(err.raw_os_error().unwrap_or(libc::EIO), "path_fingerprint")
    })?;
    let modified_ns = metadata
        .modified()
        .ok()
        .and_then(|modified| modified.duration_since(UNIX_EPOCH).ok())
        .map(|duration| duration.as_nanos())
        .unwrap_or_default();
    Ok(PathFingerprint {
        len: metadata.len(),
        modified_ns,
    })
}

/// Return a fingerprint of file metadata without following symlinks.
///
/// The parent directory is opened `O_NOFOLLOW` component-by-component and the
/// file itself is opened with `O_NOFOLLOW` via `openat`, so a symlink at the
/// final or any parent component fails with `ELOOP` rather than being followed.
/// Preferred over [`path_fingerprint`] when the caller cannot trust the path's
/// directory components.
///
/// ### Errors
/// - `EACCES`: Permission denied for a component of the path prefix.
/// - `ENOENT`: The path does not exist.
/// - `ELOOP`: The path is a symlink or passes through one.
pub fn path_fingerprint_nofollow(path: &Path) -> Result<PathFingerprint, CoreError> {
    let target = path;
    let file_name = basename(target)?;
    let dir = open_parent_nofollow(target)?;
    let fd = openat_raw(
        dir.as_raw_fd(),
        &file_name,
        libc::O_RDONLY | libc::O_NOFOLLOW | libc::O_CLOEXEC,
        0,
    )?;
    // SAFETY: `fd` from openat O_NOFOLLOW is uniquely owned.
    let metadata = unsafe { fs::File::from_raw_fd(fd) }.metadata()?;
    let modified_ns = metadata
        .modified()
        .ok()
        .and_then(|modified| modified.duration_since(UNIX_EPOCH).ok())
        .map(|duration| duration.as_nanos())
        .unwrap_or_default();
    Ok(PathFingerprint {
        len: metadata.len(),
        modified_ns,
    })
}

/// Probe whether a filesystem path is accessible and exists.
///
/// NOTE: This follows symbolic links. It uses `libc::access` with `F_OK`
/// so the check is a single syscall with no Rust allocator involvement.
/// Returns `true` if the path is accessible or visible, `false` on any error
/// (including `ENOENT`, `EACCES`, or invalid path bytes).
pub fn path_exists(path: &str) -> bool {
    match std::ffi::CString::new(path) {
        Ok(c) => unsafe { libc::access(c.as_ptr(), libc::F_OK) == 0 },
        Err(_) => false,
    }
}

/// Probe whether a path exists without following symbolic links.
///
/// Returns `true` if the path exists, including a dangling symlink.
pub fn path_lstat_exists(path: &str) -> bool {
    match std::ffi::CString::new(path) {
        Ok(c) => unsafe {
            let mut stat = std::mem::zeroed();
            libc::lstat(c.as_ptr(), &mut stat) == 0
        },
        Err(_) => false,
    }
}

/// Read a file into a string.
///
/// This stays as a small convenience helper for low-level modules that treat
/// blocking filesystem or procfs reads as an acceptable boundary cost.
/// Read the entire contents of a file into a string.
///
/// ### Errors
/// - `EACCES`: Permission denied.
/// - `ENOENT`: The path does not exist.
/// - `EIO`: Low-level I/O error.
pub fn read_to_string(path: &str) -> Result<String, CoreError> {
    std::fs::read_to_string(path)
        .map_err(|err| CoreError::sys(err.raw_os_error().unwrap_or(libc::EIO), "read_to_string"))
}

/// Advise the kernel to begin reading file data into the page cache.
///
/// This is an advisory hint only. It can help warm likely-needed file ranges,
/// but the kernel may ignore the request, perform only part of it, or return
/// before the data is fully resident in memory.
///
/// The `offset` and `len` identify the byte range to prefetch for `fd`.
/// Success means the kernel accepted the request, not that subsequent reads
/// are guaranteed to be cache hits.
/// Advise the kernel to begin reading file data into the page cache.
///
/// ### Errors
/// - `EBADF`: The file descriptor is invalid.
/// - `EINVAL`: The offset or length is invalid.
pub fn readahead(fd: impl AsRawFd, offset: u64, len: usize) -> Result<(), CoreError> {
    readahead_raw(fd.as_raw_fd(), offset, len)
}

// ── fadvise ──────────────────────────────────────────────────────────────────

pub const FADV_NORMAL: i32 = libc::POSIX_FADV_NORMAL;
pub const FADV_RANDOM: i32 = libc::POSIX_FADV_RANDOM;
pub const FADV_SEQUENTIAL: i32 = libc::POSIX_FADV_SEQUENTIAL;
pub const FADV_WILLNEED: i32 = libc::POSIX_FADV_WILLNEED;
pub const FADV_DONTNEED: i32 = libc::POSIX_FADV_DONTNEED;
pub const FADV_NOREUSE: i32 = libc::POSIX_FADV_NOREUSE;

/// Advise the kernel on the expected access pattern for a file range.
///
/// `offset` and `len` define the byte range; `len = 0` means "to end of file".
/// `advice` is one of the `FADV_*` constants.
///
/// Unlike most syscalls, `posix_fadvise` returns the error code directly
/// rather than setting `errno`.
///
/// ### Errors
/// - `EBADF`: invalid file descriptor.
/// - `EINVAL`: invalid advice value or unsupported `len`.
/// - `ESPIPE`: the fd refers to a pipe.
pub fn fadvise(fd: impl AsRawFd, offset: u64, len: usize, advice: i32) -> Result<(), CoreError> {
    let ret = unsafe {
        libc::posix_fadvise(
            fd.as_raw_fd(),
            offset as libc::off_t,
            len as libc::off_t,
            advice,
        )
    };
    if ret == 0 {
        Ok(())
    } else {
        Err(CoreError::sys(ret, "posix_fadvise"))
    }
}

/// Map a file range, advise the kernel that it will be needed, then unmap it.
///
/// `offset` must be page-aligned. This low-level primitive rejects unaligned
/// offsets with `EINVAL` instead of silently widening the requested range.
/// Map a file range and advise the kernel with `MADV_WILLNEED`.
///
/// ### Errors
/// - `EBADF`: The file descriptor is invalid.
/// - `EINVAL`: The offset is not page-aligned or the range is invalid.
/// - `ENOMEM`: Insufficient kernel memory.
pub fn mmap_madvise(
    fd: impl AsRawFd,
    offset: u64,
    len: usize,
    touch: bool,
) -> Result<(), CoreError> {
    mmap_madvise_raw(fd.as_raw_fd(), offset, len, touch)
}

#[cfg(any(target_os = "linux", target_os = "android"))]
fn mmap_madvise_raw(
    fd: libc::c_int,
    offset: u64,
    len: usize,
    touch: bool,
) -> Result<(), CoreError> {
    if len == 0 {
        return Ok(());
    }

    let page_size = unsafe { libc::sysconf(libc::_SC_PAGESIZE) };
    if page_size <= 0 {
        return Err(CoreError::sys(libc::EINVAL, "sysconf(_SC_PAGESIZE)"));
    }
    let page_size = page_size as u64;
    if offset % page_size != 0 || offset > libc::off_t::MAX as u64 {
        return Err(CoreError::sys(libc::EINVAL, "mmap"));
    }

    // Clamp the range to the file's actual size. Mapping beyond EOF is legal,
    // but touching a page with no backing data raises SIGBUS, so a `touch`
    // request that extends past EOF would crash the process.
    let file_len = {
        let mut st: libc::stat = unsafe { std::mem::zeroed() };
        if unsafe { libc::fstat(fd, &mut st) } == -1 {
            let code = std::io::Error::last_os_error().raw_os_error().unwrap_or(0);
            return Err(CoreError::sys(code, "fstat"));
        }
        st.st_size.max(0) as u64
    };
    let len = len.min(file_len.saturating_sub(offset) as usize);
    if len == 0 {
        return Ok(());
    }

    let ptr = unsafe {
        libc::mmap(
            std::ptr::null_mut(),
            len,
            libc::PROT_READ,
            libc::MAP_PRIVATE,
            fd,
            offset as libc::off_t,
        )
    };
    if ptr == libc::MAP_FAILED {
        let code = std::io::Error::last_os_error().raw_os_error().unwrap_or(0);
        return Err(CoreError::sys(code, "mmap"));
    }

    let result = if unsafe { libc::madvise(ptr, len, libc::MADV_WILLNEED) } == -1 {
        let code = std::io::Error::last_os_error().raw_os_error().unwrap_or(0);
        Err(CoreError::sys(code, "madvise"))
    } else {
        if touch {
            // Touching a page with no backing data (a file truncated after the
            // initial fstat clamp) raises SIGBUS and would kill the daemon. A
            // SIGBUS handler is not an option here (process-wide signal state
            // is unsafe in a multithreaded root daemon), so re-validate the
            // current file size on every page and stop as soon as the file
            // shrank below the page about to be touched. This narrows the
            // crash window from "any truncate during the whole loop" to a
            // single fstat→read race on the boundary page.
            let page_size = page_size as usize;
            let mut pos = 0usize;
            while pos < len {
                let mut st: libc::stat = unsafe { std::mem::zeroed() };
                if unsafe { libc::fstat(fd, &mut st) } == -1 {
                    break;
                }
                let file_len = st.st_size.max(0) as u64;
                let page_start = offset.saturating_add(pos as u64);
                if page_start >= file_len {
                    break;
                }
                unsafe {
                    std::ptr::read_volatile((ptr as *const u8).add(pos));
                }
                pos = pos.saturating_add(page_size);
            }
        }
        Ok(())
    };

    if unsafe { libc::munmap(ptr, len) } == -1 {
        let code = std::io::Error::last_os_error().raw_os_error().unwrap_or(0);
        return Err(CoreError::sys(code, "munmap"));
    }
    result
}

#[cfg(not(any(target_os = "linux", target_os = "android")))]
fn mmap_madvise_raw(
    _fd: libc::c_int,
    _offset: u64,
    _len: usize,
    _touch: bool,
) -> Result<(), CoreError> {
    Err(CoreError::sys(libc::ENOSYS, "mmap"))
}

#[cfg(any(target_os = "linux", target_os = "android"))]
fn readahead_raw(fd: libc::c_int, offset: u64, len: usize) -> Result<(), CoreError> {
    if offset > libc::off64_t::MAX as u64 {
        return Err(CoreError::sys(libc::EINVAL, "readahead"));
    }

    let count = len as libc::size_t;
    let offset = offset as libc::off64_t;

    loop {
        let ret = unsafe { libc::syscall(readahead_syscall_number(), fd, offset, count) };
        if ret == -1 {
            let code = std::io::Error::last_os_error().raw_os_error().unwrap_or(0);
            if code == libc::EINTR {
                continue;
            }
            return Err(CoreError::sys(code, "readahead"));
        }
        return Ok(());
    }
}

#[cfg(not(any(target_os = "linux", target_os = "android")))]
fn readahead_raw(_fd: libc::c_int, _offset: u64, _len: usize) -> Result<(), CoreError> {
    Err(CoreError::sys(libc::ENOSYS, "readahead"))
}

#[cfg(target_os = "linux")]
#[inline(always)]
const fn readahead_syscall_number() -> libc::c_long {
    libc::SYS_readahead
}

#[cfg(all(target_os = "android", target_arch = "aarch64"))]
#[inline(always)]
const fn readahead_syscall_number() -> libc::c_long {
    213
}

#[cfg(all(target_os = "android", target_arch = "arm"))]
#[inline(always)]
const fn readahead_syscall_number() -> libc::c_long {
    225
}

#[cfg(all(target_os = "android", target_arch = "x86_64"))]
#[inline(always)]
const fn readahead_syscall_number() -> libc::c_long {
    187
}

#[cfg(all(target_os = "android", target_arch = "x86"))]
#[inline(always)]
const fn readahead_syscall_number() -> libc::c_long {
    225
}

const TEMP_ATTEMPTS: usize = 32;

/// Fill `buf` with random bytes from the kernel CSPRNG (getrandom(2)).
///
/// Fails if the syscall is unavailable or interrupted too often; callers treat
/// this as a hard error so a temp name never falls back to a predictable
/// value.
fn getrandom_bytes(buf: &mut [u8]) -> Result<(), CoreError> {
    let mut off = 0;
    while off < buf.len() {
        // SAFETY: `buf` is a valid writable region; flags=0 blocks until the
        // pool is seeded, so names are never derived from a weak pool.
        let n = unsafe { libc::getrandom(buf[off..].as_mut_ptr() as *mut _, buf.len() - off, 0) };
        if n < 0 {
            if std::io::Error::last_os_error().raw_os_error() == Some(libc::EINTR) {
                continue;
            }
            return Err(std::io::Error::last_os_error().into());
        }
        off += n as usize;
    }
    Ok(())
}

/// Lowercase hex encoding of `bytes` (for temp-name suffixes).
fn hex(bytes: &[u8]) -> String {
    let mut out = String::with_capacity(bytes.len() * 2);
    for b in bytes {
        out.push_str(&format!("{b:02x}"));
    }
    out
}

fn cstr(s: &str) -> Result<CString, CoreError> {
    CString::new(s).map_err(|_| CoreError::sys(libc::EINVAL, "path:nul_byte"))
}

fn openat_raw(dirfd: RawFd, name: &str, flags: i32, mode: u32) -> Result<RawFd, CoreError> {
    let c = cstr(name)?;
    // SAFETY: `name` is a single path component (no `/`) converted to a
    // NUL-terminated string, and `dirfd` is a valid open directory fd.
    let fd = unsafe { libc::openat(dirfd, c.as_ptr(), flags, mode as libc::c_uint) };
    if fd < 0 {
        Err(std::io::Error::last_os_error().into())
    } else {
        Ok(fd)
    }
}

fn openat_dir(dirfd: RawFd, name: &str) -> Result<OwnedFd, CoreError> {
    let fd = openat_raw(
        dirfd,
        name,
        libc::O_RDONLY | libc::O_DIRECTORY | libc::O_NOFOLLOW | libc::O_CLOEXEC,
        0,
    )?;
    // SAFETY: `fd` came from openat and is uniquely owned here.
    Ok(unsafe { OwnedFd::from_raw_fd(fd) })
}

/// Resolve `path` from the root down to a directory fd, refusing to traverse
/// Walk `path` from the root down to a directory fd, refusing to traverse any
/// symlink component (ELOOP) or non-directory (ENOTDIR). When `create_mode` is
/// `Some(mode)`, missing components are created via `mkdirat` relative to the
/// previously held fd — never by re-resolving a path, so an attacker link can
/// never make root create a directory at an attacker-chosen location (N5).
fn walk_dir(path: &Path, create_mode: Option<u32>) -> Result<OwnedFd, CoreError> {
    let abs = if path.is_absolute() {
        path.to_path_buf()
    } else {
        std::env::current_dir()?.join(path)
    };
    let mut dir = openat_dir(libc::AT_FDCWD, "/")?;
    for comp in abs.components() {
        use std::path::Component;
        match comp {
            Component::RootDir | Component::CurDir => {}
            Component::Normal(name) => {
                let name = name
                    .to_str()
                    .ok_or_else(|| CoreError::sys(libc::EINVAL, "path:non_utf8"))?;
                let next = match openat_dir(dir.as_raw_fd(), name) {
                    Ok(fd) => fd,
                    Err(e) if e.raw_os_error() == Some(libc::ENOENT) && create_mode.is_some() => {
                        let c = cstr(name)?;
                        // SAFETY: mkdirat creates a single directory entry
                        // relative to the held fd; no pathname resolution of
                        // attacker components.
                        if unsafe {
                            libc::mkdirat(
                                dir.as_raw_fd(),
                                c.as_ptr(),
                                create_mode.unwrap() as libc::mode_t,
                            )
                        } != 0
                        {
                            // Raced with another creator; if it now exists as a
                            // real directory, use it. A symlink raced into place
                            // still fails O_NOFOLLOW below.
                            if std::io::Error::last_os_error().raw_os_error() == Some(libc::EEXIST)
                            {
                                openat_dir(dir.as_raw_fd(), name)?
                            } else {
                                return Err(std::io::Error::last_os_error().into());
                            }
                        } else {
                            openat_dir(dir.as_raw_fd(), name)?
                        }
                    }
                    Err(e) => return Err(e),
                };
                dir = next;
            }
            Component::ParentDir => {
                return Err(CoreError::sys(libc::EINVAL, "path:parent_component"));
            }
            Component::Prefix(_) => unreachable!("non-Windows path"),
        }
    }
    Ok(dir)
}

/// Resolve `path` from the root down to a directory fd, refusing to traverse
/// any symlink component (ELOOP) or non-directory (ENOTDIR).
fn open_dir_nofollow(path: &Path) -> Result<OwnedFd, CoreError> {
    walk_dir(path, None)
}

fn fstat(fd: RawFd) -> Result<libc::stat, CoreError> {
    let mut st: libc::stat = unsafe { std::mem::zeroed() };
    // SAFETY: `fd` is valid and `st` remains alive for the call.
    if unsafe { libc::fstat(fd, &mut st) } != 0 {
        Err(std::io::Error::last_os_error().into())
    } else {
        Ok(st)
    }
}

fn unlink_name(dirfd: RawFd, name: &str) -> Result<(), CoreError> {
    let c = cstr(name)?;
    // SAFETY: unlinkat without AT_SYMLINK_FOLLOW removes the directory entry
    // itself and never follows a final symlink.
    if unsafe { libc::unlinkat(dirfd, c.as_ptr(), 0) } != 0 {
        Err(std::io::Error::last_os_error().into())
    } else {
        Ok(())
    }
}

fn basename(target: &Path) -> Result<String, CoreError> {
    target
        .file_name()
        .map(|n| n.to_string_lossy().into_owned())
        .ok_or_else(|| CoreError::sys(libc::EINVAL, "path:no_file_name"))
}

fn parent_dir(target: &Path) -> &Path {
    target
        .parent()
        .filter(|p| !p.as_os_str().is_empty())
        .unwrap_or_else(|| Path::new("."))
}

/// Open the parent directory of `target` fd-anchored (O_NOFOLLOW walk) and
/// verify it is owned by the current effective uid.
///
/// Requiring ownership on every operation closes the directory-swap vector:
/// an attacker who renames the state dir and substitutes their own can never
/// make the daemon read from or write into the substituted directory, because
/// the substituted dir fails the ownership check regardless of the path.
fn open_parent_nofollow(target: &Path) -> Result<OwnedFd, CoreError> {
    let dir = open_dir_nofollow(parent_dir(target))?;
    let st = fstat(dir.as_raw_fd())?;
    let euid = unsafe { libc::geteuid() };
    if st.st_uid != euid {
        return Err(CoreError::sys(libc::EACCES, "parent_dir_owner"));
    }
    Ok(dir)
}

struct TmpGuard {
    dirfd: RawFd,
    name: String,
}

impl Drop for TmpGuard {
    fn drop(&mut self) {
        let _ = unlink_name(self.dirfd, &self.name);
    }
}

/// Atomically replace `path` with `content`.
///
/// The parent directory is opened `O_NOFOLLOW` component-by-component and held
/// as a directory fd; the content is written to a fresh `O_CREAT|O_EXCL`
/// temp file via `openat` and then `renameat`d over the destination. A symlink
/// anywhere in the parent path fails the open; a symlink at the destination is
/// replaced by the rename, never followed (N5).
pub fn write_atomic(path: impl AsRef<Path>, content: &[u8]) -> Result<(), CoreError> {
    let target = path.as_ref();
    let file_name = basename(target)?;
    let dir = open_parent_nofollow(target)?;
    let dirfd = dir.as_raw_fd();

    for attempt in 0..TEMP_ATTEMPTS {
        // Random suffix (getrandom) so an attacker in a shared dir cannot
        // pre-create all attempt slots and force a permanent EEXIST; the
        // attempt counter is a tiebreak for a (vanishingly rare) collision.
        let mut rand_bytes = [0u8; 16];
        getrandom_bytes(&mut rand_bytes)?;
        let tmp_name = format!(
            ".{file_name}.{}.{}.{attempt:x}",
            std::process::id(),
            hex(&rand_bytes)
        );
        match openat_raw(
            dirfd,
            &tmp_name,
            libc::O_WRONLY | libc::O_CREAT | libc::O_EXCL | libc::O_CLOEXEC,
            0o600,
        ) {
            Ok(raw) => {
                // SAFETY: `raw` came from openat with O_EXCL and is uniquely
                // owned; it becomes an owned File for the write.
                let mut file = unsafe { fs::File::from_raw_fd(raw) };
                let _guard = TmpGuard {
                    dirfd,
                    name: tmp_name.clone(),
                };
                file.write_all(content)?;
                file.sync_all()?;
                drop(file);
                let c_tmp = cstr(&tmp_name)?;
                let c_final = cstr(&file_name)?;
                // SAFETY: renameat replaces the destination entry and never
                // follows a final symlink; both names are single components
                // relative to the anchored directory fd.
                if unsafe { libc::renameat(dirfd, c_tmp.as_ptr(), dirfd, c_final.as_ptr()) } != 0 {
                    return Err(std::io::Error::last_os_error().into());
                }
                // rename succeeded; the temp entry no longer exists.
                std::mem::forget(_guard);
                return Ok(());
            }
            Err(e) if e.raw_os_error() == Some(libc::EEXIST) => continue,
            Err(e) => return Err(e),
        }
    }

    Err(CoreError::sys(libc::EEXIST, "temp:exhausted"))
}

/// Read a file to a string, refusing to follow a symlink at the final or any
/// parent component.
///
/// Fails with `ELOOP` if `path` is a symlink or passes through one. Used for
/// reading attacker-placed state files so their contents can never be injected
/// through a link.
pub fn read_nofollow(path: impl AsRef<Path>) -> Result<String, CoreError> {
    let target = path.as_ref();
    let file_name = basename(target)?;
    let dir = open_parent_nofollow(target)?;
    let fd = openat_raw(
        dir.as_raw_fd(),
        &file_name,
        libc::O_RDONLY | libc::O_NOFOLLOW | libc::O_CLOEXEC,
        0,
    )?;
    // SAFETY: `fd` is uniquely owned; it becomes an owned File.
    let mut file = unsafe { fs::File::from_raw_fd(fd) };
    let mut content = String::new();
    file.read_to_string(&mut content)?;
    Ok(content)
}

/// Open an existing-or-create append stream that refuses to follow symlinks.
///
/// Appends to a regular file at `path`. If `path` is a symlink — or sits under
/// a symlinked parent — the open fails with `ELOOP` rather than writing
/// through it.
pub fn open_append_nofollow(path: impl AsRef<Path>) -> Result<fs::File, CoreError> {
    let target = path.as_ref();
    let file_name = basename(target)?;
    let dir = open_parent_nofollow(target)?;
    let fd = openat_raw(
        dir.as_raw_fd(),
        &file_name,
        libc::O_WRONLY | libc::O_CREAT | libc::O_APPEND | libc::O_NOFOLLOW | libc::O_CLOEXEC,
        0o644,
    )?;
    // SAFETY: `fd` is uniquely owned; it becomes an owned File.
    Ok(unsafe { fs::File::from_raw_fd(fd) })
}

/// Remove a directory entry without following a final symlink, anchored to its
/// (symlink-free) parent. The entry itself is removed even if it is a symlink;
/// its target is never touched.
pub fn remove_nofollow(path: impl AsRef<Path>) -> Result<(), CoreError> {
    let target = path.as_ref();
    let file_name = basename(target)?;
    let dir = open_parent_nofollow(target)?;
    unlink_name(dir.as_raw_fd(), &file_name)
}

/// Create `dir` if needed, then verify it is a real directory with no symlink
/// component and that it is owned by the current effective uid. Returning the
/// anchored fd lets callers refuse to operate inside an attacker-pre-created
/// or attacker-redirected state directory.
pub fn ensure_state_dir(dir: impl AsRef<Path>) -> Result<OwnedFd, CoreError> {
    let dir = dir.as_ref();
    // Anchored create-then-verify: missing components are made with mkdirat
    // relative to the held fd; a symlink at any component ELOOPs rather than
    // being followed, and the opened fd below is the authority for ownership.
    let fd = walk_dir(dir, Some(0o700))?;
    let st = fstat(fd.as_raw_fd())?;
    let euid = unsafe { libc::geteuid() };
    if st.st_uid != euid {
        return Err(CoreError::sys(libc::EACCES, "state_dir_owner"));
    }
    Ok(fd)
}

#[cfg(test)]
mod tests {
    #[cfg(target_os = "linux")]
    #[test]
    fn test_readahead_syscall_number_linux_matches_libc() {
        assert_eq!(super::readahead_syscall_number(), libc::SYS_readahead);
    }

    #[cfg(target_os = "linux")]
    #[test]
    fn test_mmap_madvise_touch_past_eof_does_not_sigbus() {
        use std::os::unix::io::{AsRawFd, FromRawFd};

        // A file whose length is not page-aligned; requesting a touch range
        // that extends far past EOF must be clamped to the file size instead of
        // SIGBUS-ing the process on the out-of-file pages.
        let dir =
            std::env::temp_dir().join(format!("coreshift_mmap_past_eof_{}", std::process::id()));
        let _ = std::fs::remove_file(&dir);
        std::fs::write(&dir, b"x").unwrap();

        let f = std::fs::File::open(&dir).unwrap();
        let fd = f.as_raw_fd();
        // Reopen to hand a raw fd to the helper without double-closing.
        let dup = unsafe { libc::fcntl(fd, libc::F_DUPFD_CLOEXEC, 0) };
        assert!(dup >= 0);
        let owned = unsafe { std::fs::File::from_raw_fd(dup) };

        let page = unsafe { libc::sysconf(libc::_SC_PAGESIZE) } as u64;
        let result = super::mmap_madvise(owned.as_raw_fd(), 0, (page * 4) as usize, true);
        drop(owned);
        let _ = std::fs::remove_file(&dir);

        // Clamped to 1 byte; MADV_WILLNEED on a tiny range is a no-op that
        // still reports success on Linux.
        assert!(result.is_ok(), "touch past EOF must not SIGBUS: {result:?}");
    }

    #[cfg(all(target_os = "android", target_arch = "aarch64"))]
    #[test]
    fn test_readahead_syscall_number_android_aarch64() {
        assert_eq!(super::readahead_syscall_number(), 213);
    }

    #[cfg(all(target_os = "android", target_arch = "arm"))]
    #[test]
    fn test_readahead_syscall_number_android_arm() {
        assert_eq!(super::readahead_syscall_number(), 225);
    }

    #[cfg(all(target_os = "android", target_arch = "x86_64"))]
    #[test]
    fn test_readahead_syscall_number_android_x86_64() {
        assert_eq!(super::readahead_syscall_number(), 187);
    }

    #[cfg(all(target_os = "android", target_arch = "x86"))]
    #[test]
    fn test_readahead_syscall_number_android_x86() {
        assert_eq!(super::readahead_syscall_number(), 225);
    }
}

#[cfg(test)]
mod safe_fs_tests {

    use super::*;
    use std::os::unix::fs::symlink;
    use std::path::PathBuf;

    fn tmpdir(name: &str) -> PathBuf {
        let d = std::env::temp_dir().join(format!(
            "coreshift_safe_fs_dir_{}_{name}",
            std::process::id()
        ));
        let _ = fs::remove_dir_all(&d);
        fs::create_dir_all(&d).unwrap();
        d
    }

    #[test]
    fn test_write_atomic_creates_regular_file() {
        let dir = tmpdir("w");
        let p = dir.join("out.txt");

        write_atomic(&p, b"hello").unwrap();
        assert_eq!(fs::read_to_string(&p).unwrap(), "hello");
        assert!(fs::symlink_metadata(&p).unwrap().file_type().is_file());
    }

    #[test]
    fn test_write_atomic_replaces_existing_symlink_not_target() {
        let dir = tmpdir("s1");
        let target = dir.join("victim");
        let link = dir.join("link");

        fs::write(&target, b"precious").unwrap();
        symlink(&target, &link).unwrap();

        // Write through the symlink: the victim must stay untouched and the
        // path must become a regular file (the symlink is replaced).
        write_atomic(&link, b"new").unwrap();

        assert_eq!(fs::read_to_string(&target).unwrap(), "precious");
        assert_eq!(fs::read_to_string(&link).unwrap(), "new");
        assert!(fs::symlink_metadata(&link).unwrap().file_type().is_file());
    }

    #[test]
    fn test_read_nofollow_refuses_symlink() {
        let dir = tmpdir("r");
        let target = dir.join("victim2");
        let link = dir.join("link2");

        fs::write(&target, b"secret").unwrap();
        symlink(&target, &link).unwrap();

        assert_eq!(read_nofollow(&target).unwrap(), "secret");
        assert!(read_nofollow(&link).is_err());
    }

    #[test]
    fn test_path_fingerprint_nofollow_refuses_symlink() {
        let dir = tmpdir("fp");
        let target = dir.join("target_fp");
        let link = dir.join("link_fp");

        fs::write(&target, b"secret").unwrap();
        symlink(&target, &link).unwrap();

        // Fingerprinting the regular file works; a symlink must fail.
        assert!(path_fingerprint_nofollow(&target).is_ok());
        assert!(path_fingerprint_nofollow(&link).is_err());
        // A symlinked parent must fail too (component-level O_NOFOLLOW).
        let parent_link = dir.join("plink");
        symlink(&dir, &parent_link).unwrap();
        assert!(path_fingerprint_nofollow(&parent_link.join("target_fp")).is_err());
    }

    #[test]
    fn test_open_append_nofollow_refuses_symlink() {
        let dir = tmpdir("a");
        let target = dir.join("target3");
        let link = dir.join("link3");

        fs::write(&target, b"x").unwrap();
        symlink(&target, &link).unwrap();

        // Append to the regular file is fine.
        assert!(open_append_nofollow(&target).is_ok());
        // Appending through a symlink must fail.
        assert!(open_append_nofollow(&link).is_err());

        let _ = fs::remove_file(&target);
        let _ = fs::remove_file(&link);
    }

    #[test]
    fn test_write_atomic_refuses_symlinked_parent() {
        // An intermediate directory replaced by a symlink must make the
        // anchored open fail, and nothing may be written through to the
        // linked directory.
        let dir = tmpdir("parent_symlink");
        let elsewhere = tmpdir("parent_dest");
        let link = dir.join("coreshift");
        symlink(&elsewhere, &link).unwrap();

        assert!(write_atomic(link.join("payload.txt"), b"boom").is_err());
        assert!(!elsewhere.join("payload.txt").exists());
        assert!(!link.join("payload.txt").exists());
    }

    #[test]
    fn test_read_nofollow_refuses_symlinked_parent() {
        let dir = tmpdir("read_parent_symlink");
        let elsewhere = tmpdir("read_parent_dest");
        fs::write(elsewhere.join("conf"), b"injected").unwrap();
        let link = dir.join("coreshift");
        symlink(&elsewhere, &link).unwrap();

        assert!(read_nofollow(link.join("conf")).is_err());
    }

    #[test]
    fn test_open_append_nofollow_refuses_symlinked_parent() {
        let dir = tmpdir("append_parent_symlink");
        let elsewhere = tmpdir("append_parent_dest");
        let link = dir.join("coreshift");
        symlink(&elsewhere, &link).unwrap();

        assert!(open_append_nofollow(link.join("daemon.log")).is_err());
        assert!(!elsewhere.join("daemon.log").exists());
    }

    #[test]
    fn test_ensure_state_dir_refuses_symlink() {
        let dir = tmpdir("state_symlink");
        let elsewhere = tmpdir("state_dest");
        let link = dir.join("state");
        symlink(&elsewhere, &link).unwrap();

        assert!(ensure_state_dir(&link).is_err());
        // A real directory is accepted (owned by the current euid).
        let real = tmpdir("state_real");
        assert!(ensure_state_dir(&real).is_ok());
    }

    #[test]
    fn test_remove_nofollow_removes_entry_not_target() {
        let dir = tmpdir("unlink");
        let target = dir.join("victim4");
        let link = dir.join("link4");
        fs::write(&target, b"keep").unwrap();
        symlink(&target, &link).unwrap();

        remove_nofollow(&link).unwrap();
        assert!(!link.exists());
        assert_eq!(fs::read_to_string(&target).unwrap(), "keep");
    }

    #[test]
    fn test_write_atomic_requires_parent_to_exist() {
        let dir = tmpdir("missing_parent");
        let p = dir.join("nope").join("file.txt");

        assert!(write_atomic(&p, b"x").is_err());
        assert!(!p.exists());
    }

    #[test]
    fn test_ops_refuse_foreign_owned_parent() {
        // Simulates an attacker who renames the root-owned state dir and
        // substitutes their own: the real (non-symlink) substituted dir must
        // fail every operation because it is not owned by the effective uid.
        // Only exercisable as root (chown); otherwise skipped.
        if unsafe { libc::geteuid() } != 0 {
            return;
        }
        let dir = tmpdir("foreign_owner");
        let path = dir.join("f");
        let owned = cstr(&dir.to_string_lossy()).unwrap();
        // SAFETY: chown to 65534 (nobody) while running as root.
        assert_eq!(unsafe { libc::chown(owned.as_ptr(), 65534, 65534) }, 0);

        assert!(write_atomic(&path, b"x").is_err());
        assert!(read_nofollow(&path).is_err());
        assert!(open_append_nofollow(&path).is_err());
        assert!(remove_nofollow(&path).is_err());
        assert!(ensure_state_dir(&dir).is_err());
        assert!(!path.exists());
    }

    fn temp_leftovers(dir: &Path, file_name: &str) -> Vec<String> {
        let prefix = format!(".{file_name}.");
        fs::read_dir(dir)
            .map(|rd| {
                rd.filter_map(|e| e.ok())
                    .filter_map(|e| e.file_name().to_str().map(str::to_owned))
                    .filter(|n| n.starts_with(&prefix))
                    .collect()
            })
            .unwrap_or_default()
    }

    #[test]
    fn test_write_atomic_cleans_temp_on_rename_failure() {
        // A non-empty directory at the destination forces renameat to fail
        // (EISDIR/ENOTEMPTY) after the temp file has been fully written; the
        // TmpGuard must remove the temp entry.
        let dir = tmpdir("rename_fail");
        let dest = dir.join("dest");
        fs::create_dir_all(&dest).unwrap();
        fs::write(dest.join("keep"), b"x").unwrap();

        assert!(write_atomic(&dest, b"boom").is_err());
        assert_eq!(fs::read_to_string(dest.join("keep")).unwrap(), "x");
        assert!(
            temp_leftovers(&dir, "dest").is_empty(),
            "temp file must be cleaned up"
        );
    }

    #[test]
    fn test_write_atomic_leaves_no_temp_on_success() {
        let dir = tmpdir("no_temp_success");
        let p = dir.join("out.txt");

        write_atomic(&p, b"hello").unwrap();
        assert_eq!(fs::read_to_string(&p).unwrap(), "hello");
        assert!(
            temp_leftovers(&dir, "out.txt").is_empty(),
            "no temp left behind"
        );
    }

    #[test]
    fn test_write_atomic_retries_when_temp_name_exists() {
        let dir = tmpdir("temp_collision");
        let p = dir.join("out.txt");
        let pid = std::process::id();
        let collided = dir.join(format!(".out.txt.{pid}.0"));
        fs::write(&collided, b"not mine").unwrap();

        write_atomic(&p, b"hello").unwrap();
        assert_eq!(fs::read_to_string(&p).unwrap(), "hello");
        // The pre-existing colliding temp is neither clobbered nor mistaken for
        // ours; a later attempt slot is used.
        assert_eq!(fs::read_to_string(&collided).unwrap(), "not mine");
        let leftovers = temp_leftovers(&dir, "out.txt");
        assert_eq!(
            leftovers.len(),
            1,
            "only the pre-existing colliding temp remains"
        );
        assert_eq!(leftovers[0], format!(".out.txt.{pid}.0"));
    }

    #[test]
    fn test_ops_refuse_foreign_owned_writable_parent() {
        // A world-writable directory owned by a uid != euid (e.g. /tmp owned by
        // root) is a privilege-free proxy for a swapped state dir: a vulnerable
        // implementation with no ownership check would happily write into it,
        // while the fixed one must refuse every operation. Runs whenever such a
        // directory is available; skips as root (root owns /tmp).
        use std::os::unix::fs::MetadataExt;
        let euid = unsafe { libc::geteuid() };
        if euid == 0 {
            return;
        }
        let tmp = std::env::temp_dir();
        let meta = match fs::symlink_metadata(&tmp) {
            Ok(m) => m,
            Err(_) => return,
        };
        if meta.uid() == euid {
            return; // no ownership mismatch available in this environment
        }
        if meta.mode() & 0o002 == 0 && meta.mode() & 0o020 == 0 {
            return; // not writable by us; the proxy doesn't apply
        }
        let p = tmp.join(format!(
            "coreshift_fs_foreign_{}_{}",
            std::process::id(),
            "out"
        ));
        let _ = fs::remove_file(&p);
        fs::write(&p, b"probe").unwrap();

        assert!(read_nofollow(&p).is_err());
        assert!(open_append_nofollow(&p).is_err());
        assert!(write_atomic(&p, b"boom").is_err());
        assert!(remove_nofollow(&p).is_err());
        assert!(ensure_state_dir(&tmp).is_err());
        assert_eq!(fs::read_to_string(&p).unwrap(), "probe");
        let _ = fs::remove_file(&p);
    }

    #[test]
    fn test_ensure_state_dir_creates_fresh_dir() {
        let base = tmpdir("fresh_base");
        let nested = base.join("a").join("b").join("state");

        let fd = ensure_state_dir(&nested).unwrap();
        assert!(nested.is_dir());
        drop(fd);
        assert!(ensure_state_dir(&nested).is_ok());
    }

    #[test]
    fn test_open_append_nofollow_refuses_dangling_symlink() {
        let dir = tmpdir("dangling_append");
        let missing = dir.join("not_there.txt");
        let link = dir.join("linkd");
        symlink(&missing, &link).unwrap();

        assert!(open_append_nofollow(&link).is_err());
        assert!(
            !missing.exists(),
            "must not create the target through a dangling link"
        );
    }

    #[test]
    fn test_read_nofollow_refuses_dangling_symlink() {
        let dir = tmpdir("dangling_read");
        let missing = dir.join("not_there2.txt");
        let link = dir.join("linkd2");
        symlink(&missing, &link).unwrap();

        assert!(read_nofollow(&link).is_err());
        assert!(!missing.exists());
    }

    #[test]
    fn test_ops_refuse_regular_file_parent() {
        // A parent component that is a regular file must yield ENOTDIR for
        // every anchored operation — no write-through, no creation.
        let dir = tmpdir("regfile_parent");
        let f = dir.join("notadir");
        fs::write(&f, b"x").unwrap();

        assert!(write_atomic(f.join("out"), b"y").is_err());
        assert!(read_nofollow(f.join("out")).is_err());
        assert!(open_append_nofollow(f.join("out")).is_err());
        assert!(remove_nofollow(f.join("out")).is_err());
        assert!(ensure_state_dir(&f).is_err());
        assert_eq!(fs::read_to_string(&f).unwrap(), "x");
    }
}