simdutf8-cli 0.1.6

// SPDX-License-Identifier: Apache-2.0
// SPDX-FileCopyrightText: 2025,2026 ndaal Gesellschaft für Sicherheit in der Informationstechnik mbH & Co KG, Cologne
// SPDX-FileCopyrightText: Author: Pierre Gronau <Pierre.Gronau@ndaal.eu>

//! Hardened file access for untrusted path arguments.
//!
//! A CLI that opens arbitrary user-supplied paths must defend against a number
//! of classic attacks. This module centralises that policy so the rest of the
//! program never touches the filesystem directly:
//!
//! * **Path traversal** — when a base directory is configured, a `..` component
//!   is rejected lexically *and* the fully resolved path is verified to stay
//!   inside the base directory.
//! * **Symlink escapes / TOCTOU** — paths are [canonicalized](std::fs::canonicalize)
//!   (resolving symlinks and `..`) before the containment check, and symlinks
//!   can be denied outright. The opened file handle's own metadata is then
//!   re-checked (`fstat` on the descriptor) so the type/size decision is made on
//!   the object we actually opened, not on a name that may have been swapped.
//! * **Non-regular files** — directories, devices, FIFOs and sockets are
//!   rejected; only regular files are accepted.
//! * **Resource exhaustion** — reads are hard-capped at a configurable byte
//!   limit, independent of the size the filesystem metadata claims.
//!
//! The entry points are [`PathPolicy::open`] / [`PathPolicy::read`] for files on
//! disk, [`read_capped`] for arbitrary readers such as standard input,
//! [`write_in_dir`] for capability-scoped writes (via `cap-std`), and the
//! lexical [`safe_join`] primitive for confining attacker-influenced relative
//! paths to a base directory.

use std::fs::File;
use std::io::{Read, Write};
use std::path::{Component, Path, PathBuf};

use cap_std::ambient_authority;
use cap_std::fs::Dir;

/// Default upper bound on the number of bytes read from a single input (64 MiB).
pub const DEFAULT_MAX_FILE_SIZE: u64 = 64 * 1024 * 1024;

/// Errors that can occur while securely resolving and opening an input path.
#[derive(Debug, thiserror::Error)]
pub enum PathSecurityError {
    /// The supplied path was empty.
    #[error("input path is empty")]
    EmptyPath,

    /// The supplied path contained an interior NUL byte.
    #[error("input path contains an interior NUL byte")]
    InteriorNul,

    /// A `..` component was present while a base directory was configured.
    #[error("input path contains a '..' component, which is not allowed with --base-dir: {}", .0.display())]
    ParentTraversal(PathBuf),

    /// The resolved path lay outside the configured base directory.
    #[error("resolved path {} escapes the permitted base directory {}", .path.display(), .base.display())]
    OutsideBase {
        /// The fully resolved (canonical) path that was rejected.
        path: PathBuf,
        /// The configured base directory (canonical).
        base: PathBuf,
    },

    /// A symbolic link was encountered while symlinks were disallowed.
    #[error("symbolic links are not permitted: {}", .0.display())]
    SymlinkDenied(PathBuf),

    /// The path resolved to something other than a regular file.
    #[error("not a regular file: {}", .0.display())]
    NotRegularFile(PathBuf),

    /// The input exceeded the configured byte limit.
    #[error("input is too large: {size} bytes exceeds the {limit} byte limit")]
    TooLarge {
        /// Observed size in bytes (at least `limit + 1` for streamed inputs).
        size: u64,
        /// The configured limit in bytes.
        limit: u64,
    },

    /// An underlying I/O error occurred while accessing the path.
    #[error("failed to access {}: {source}", .path.display())]
    Io {
        /// The path that was being accessed.
        path: PathBuf,
        /// The underlying I/O error.
        #[source]
        source: std::io::Error,
    },
}

/// A successfully opened, validated input file.
#[derive(Debug)]
pub struct OpenedFile {
    /// The open file handle (already confirmed to be a regular file).
    pub file: File,
    /// The fully resolved (canonical) path of the file.
    pub path: PathBuf,
    /// The size in bytes as reported by the file handle's metadata.
    pub size: u64,
}

/// Policy controlling how input paths are resolved and opened.
///
/// Construct with [`PathPolicy::new`] (or [`Default`]) and refine with the
/// builder-style methods.
#[derive(Clone, Debug)]
pub struct PathPolicy {
    base_dir: Option<PathBuf>,
    allow_symlinks: bool,
    max_file_size: u64,
}

impl Default for PathPolicy {
    fn default() -> Self {
        Self::new()
    }
}

impl PathPolicy {
    /// Create a policy with safe defaults: no base-directory confinement,
    /// symlinks allowed (but resolved and re-checked), and a
    /// [`DEFAULT_MAX_FILE_SIZE`] byte cap.
    #[must_use]
    pub const fn new() -> Self {
        Self {
            base_dir: None,
            allow_symlinks: true,
            max_file_size: DEFAULT_MAX_FILE_SIZE,
        }
    }

    /// Confine all inputs to `base`: resolved paths must stay within it.
    #[must_use]
    pub fn base_dir(mut self, base: impl Into<PathBuf>) -> Self {
        self.base_dir = Some(base.into());
        self
    }

    /// Allow (`true`, the default) or deny (`false`) symbolic links.
    #[must_use]
    pub const fn allow_symlinks(mut self, allow: bool) -> Self {
        self.allow_symlinks = allow;
        self
    }

    /// Set the maximum number of bytes that may be read from a single input.
    #[must_use]
    pub const fn max_file_size(mut self, limit: u64) -> Self {
        self.max_file_size = limit;
        self
    }

    /// The configured byte limit.
    #[must_use]
    pub const fn limit(&self) -> u64 {
        self.max_file_size
    }

    /// Resolve, validate and open `requested`, returning an [`OpenedFile`].
    ///
    /// # Errors
    ///
    /// Returns a [`PathSecurityError`] if the path is empty, contains an
    /// interior NUL, traverses outside the configured base directory, is a
    /// disallowed symlink, is not a regular file, exceeds the size limit, or
    /// cannot be accessed.
    pub fn open(&self, requested: &Path) -> Result<OpenedFile, PathSecurityError> {
        let canonical = self.resolve_path(requested)?;

        // Open the canonical path, then re-derive type and size from the open
        // descriptor itself (fstat) so the decision is made on the object we hold
        // open rather than on a name that could have changed (TOCTOU mitigation).
        let file = File::open(&canonical).map_err(|source| PathSecurityError::Io {
            path: canonical.clone(),
            source,
        })?;
        let meta = file.metadata().map_err(|source| PathSecurityError::Io {
            path: canonical.clone(),
            source,
        })?;
        if !meta.is_file() {
            return Err(PathSecurityError::NotRegularFile(canonical));
        }

        // Enforce the size limit (the subsequent read is hard-capped too).
        let size = meta.len();
        if size > self.max_file_size {
            return Err(PathSecurityError::TooLarge {
                size,
                limit: self.max_file_size,
            });
        }

        Ok(OpenedFile {
            file,
            path: canonical,
            size,
        })
    }

    /// Validate `requested` and resolve it to a canonical path that is confined
    /// to the configured base directory (if any). Performs no file open.
    ///
    /// # Errors
    ///
    /// Returns a [`PathSecurityError`] for an empty/NUL path, a `..` traversal,
    /// a disallowed symlink, a path escaping the base directory, or I/O failure.
    fn resolve_path(&self, requested: &Path) -> Result<PathBuf, PathSecurityError> {
        // 1. Reject an empty path outright.
        if requested.as_os_str().is_empty() {
            return Err(PathSecurityError::EmptyPath);
        }

        // 2. Reject interior NUL bytes before touching the filesystem.
        if requested.as_os_str().as_encoded_bytes().contains(&0) {
            return Err(PathSecurityError::InteriorNul);
        }

        // 3. With a base directory, reject `..` lexically as defence in depth
        //    (the canonical containment check below is the authoritative one).
        if self.base_dir.is_some()
            && requested
                .components()
                .any(|component| matches!(component, Component::ParentDir))
        {
            return Err(PathSecurityError::ParentTraversal(requested.to_path_buf()));
        }

        // 4. If symlinks are disallowed, reject a symlinked final component.
        if !self.allow_symlinks {
            let meta =
                std::fs::symlink_metadata(requested).map_err(|source| PathSecurityError::Io {
                    path: requested.to_path_buf(),
                    source,
                })?;
            if meta.file_type().is_symlink() {
                return Err(PathSecurityError::SymlinkDenied(requested.to_path_buf()));
            }
        }

        // 5. Canonicalize: resolves `.`/`..` and every symlink, yielding an
        //    absolute path to the real object on disk.
        let canonical = requested
            .canonicalize()
            .map_err(|source| PathSecurityError::Io {
                path: requested.to_path_buf(),
                source,
            })?;

        // 6. Containment: the resolved path must live inside the resolved base.
        if let Some(base) = &self.base_dir {
            let canonical_base = base
                .canonicalize()
                .map_err(|source| PathSecurityError::Io {
                    path: base.clone(),
                    source,
                })?;
            if !canonical.starts_with(&canonical_base) {
                return Err(PathSecurityError::OutsideBase {
                    path: canonical,
                    base: canonical_base,
                });
            }
        }

        Ok(canonical)
    }

    /// Open `requested` and read its contents, hard-capped at the byte limit.
    ///
    /// # Errors
    ///
    /// As for [`PathPolicy::open`], plus [`PathSecurityError::TooLarge`] if the
    /// file streams more than the configured limit.
    pub fn read(&self, requested: &Path) -> Result<Vec<u8>, PathSecurityError> {
        let OpenedFile { file, path, .. } = self.open(requested)?;
        read_capped(file, self.max_file_size).map_err(move |error| match error {
            // Replace the placeholder stream path with the real file path.
            PathSecurityError::Io { source, .. } => PathSecurityError::Io { path, source },
            other => other,
        })
    }
}

/// Read at most `limit` bytes from `reader`, erroring if more are available.
///
/// This bounds memory use for streamed inputs (e.g. standard input) where no
/// size is known in advance.
///
/// # Errors
///
/// Returns [`PathSecurityError::TooLarge`] if `reader` yields more than `limit`
/// bytes, or [`PathSecurityError::Io`] on an underlying read error.
pub fn read_capped<R: Read>(reader: R, limit: u64) -> Result<Vec<u8>, PathSecurityError> {
    // Read at most `limit + 1` bytes: the extra byte lets us detect overflow
    // without trusting any externally reported size.
    let mut limited = reader.take(limit.saturating_add(1));
    let mut buf = Vec::new();
    limited
        .read_to_end(&mut buf)
        .map_err(|source| PathSecurityError::Io {
            path: PathBuf::from("<stream>"),
            source,
        })?;

    let len = u64::try_from(buf.len()).unwrap_or(u64::MAX);
    if len > limit {
        return Err(PathSecurityError::TooLarge { size: len, limit });
    }
    Ok(buf)
}

/// Write `bytes` to a file named `name` inside the directory `dir`, using a
/// capability-scoped [`cap_std::fs::Dir`] handle (see `skills/rust-path-security.md`).
///
/// `name` must be a single relative file name; `cap-std` rejects any `..`,
/// absolute path, or symlink that would escape `dir` at the syscall layer, so a
/// crafted report file name cannot redirect the write outside `dir`. This is the
/// one place ambient authority crosses into the program for writes.
///
/// # Errors
///
/// Returns [`PathSecurityError::Io`] if `dir` cannot be opened or the file
/// cannot be created/written.
pub fn write_in_dir(dir: &Path, name: &str, bytes: &[u8]) -> Result<(), PathSecurityError> {
    let handle = Dir::open_ambient_dir(dir, ambient_authority()).map_err(|source| {
        PathSecurityError::Io {
            path: dir.to_path_buf(),
            source,
        }
    })?;
    let mut file = handle
        .create(name)
        .map_err(|source| PathSecurityError::Io {
            path: dir.join(name),
            source,
        })?;
    file.write_all(bytes)
        .map_err(|source| PathSecurityError::Io {
            path: dir.join(name),
            source,
        })?;
    Ok(())
}

/// Lexically join a relative, attacker-influenced `candidate` onto `base`,
/// returning `None` if the candidate would escape `base`.
///
/// This is a *string-level* primitive (see `skills/rust-path-security.md`): it
/// performs no filesystem access and does not resolve symlinks — it is the
/// auditable companion to the syscall-level confinement enforced by
/// [`PathPolicy`] / [`cap_std`]. A candidate is rejected when it:
///
/// * contains an interior NUL byte;
/// * is absolute or carries a path prefix (e.g. a Windows drive / UNC); or
/// * contains a `..` that pops above `base` (balanced `a/../b` is fine).
///
/// On success the returned path is guaranteed to start with `base` and to
/// contain no `..` component.
#[must_use]
pub fn safe_join(base: &Path, candidate: &str) -> Option<PathBuf> {
    if candidate.as_bytes().contains(&0) {
        return None;
    }

    let mut stack: Vec<std::ffi::OsString> = Vec::new();
    for component in Path::new(candidate).components() {
        match component {
            Component::CurDir => {},
            Component::Normal(segment) => stack.push(segment.to_os_string()),
            Component::ParentDir => {
                // A `..` that cannot be balanced by a previously pushed segment
                // would escape `base`.
                stack.pop()?;
            },
            Component::RootDir | Component::Prefix(_) => return None,
        }
    }

    let mut resolved = base.to_path_buf();
    for segment in &stack {
        resolved.push(segment);
    }

    // Belt-and-braces post-conditions the fuzz target also asserts.
    if !resolved.starts_with(base) {
        return None;
    }
    if resolved
        .components()
        .any(|component| matches!(component, Component::ParentDir))
    {
        return None;
    }
    Some(resolved)
}

#[cfg(test)]
mod tests {
    use std::io::Write;

    use super::*;

    fn write_temp(dir: &Path, name: &str, bytes: &[u8]) -> PathBuf {
        let path = dir.join(name);
        let mut f = File::create(&path).expect("create temp file");
        f.write_all(bytes).expect("write temp file");
        path
    }

    #[test]
    fn opens_and_reads_a_regular_file() {
        let dir = tempfile::tempdir().unwrap();
        let path = write_temp(dir.path(), "hello.txt", b"hello");

        let policy = PathPolicy::new();
        let opened = policy.open(&path).expect("open should succeed");
        assert_eq!(opened.size, 5);
        assert!(opened.path.is_absolute());

        let bytes = policy.read(&path).expect("read should succeed");
        assert_eq!(bytes, b"hello");
    }

    #[test]
    fn empty_path_is_rejected() {
        let policy = PathPolicy::new();
        let err = policy.open(Path::new("")).unwrap_err();
        assert!(matches!(err, PathSecurityError::EmptyPath));
    }

    #[test]
    fn missing_file_is_io_error() {
        let dir = tempfile::tempdir().unwrap();
        let policy = PathPolicy::new();
        let err = policy.open(&dir.path().join("nope")).unwrap_err();
        assert!(matches!(err, PathSecurityError::Io { .. }), "got: {err:?}");
    }

    #[test]
    fn directory_is_not_a_regular_file() {
        let dir = tempfile::tempdir().unwrap();
        let policy = PathPolicy::new();
        let err = policy.open(dir.path()).unwrap_err();
        assert!(
            matches!(err, PathSecurityError::NotRegularFile(_)),
            "got: {err:?}"
        );
    }

    #[test]
    fn oversize_file_is_rejected() {
        let dir = tempfile::tempdir().unwrap();
        let path = write_temp(dir.path(), "big.bin", b"0123456789");
        let policy = PathPolicy::new().max_file_size(4);
        let err = policy.open(&path).unwrap_err();
        assert!(
            matches!(err, PathSecurityError::TooLarge { limit: 4, .. }),
            "got: {err:?}"
        );
    }

    #[test]
    fn read_is_capped() {
        let dir = tempfile::tempdir().unwrap();
        let path = write_temp(dir.path(), "data.bin", b"0123456789");
        let policy = PathPolicy::new().max_file_size(4);
        let err = policy.read(&path).unwrap_err();
        assert!(matches!(err, PathSecurityError::TooLarge { .. }));
    }

    #[test]
    fn read_capped_accepts_within_limit() {
        let bytes = read_capped(&b"hello"[..], 10).unwrap();
        assert_eq!(bytes, b"hello");
    }

    #[test]
    fn read_capped_rejects_over_limit() {
        let err = read_capped(&b"hello"[..], 3).unwrap_err();
        assert!(matches!(err, PathSecurityError::TooLarge { limit: 3, .. }));
    }

    #[test]
    fn base_dir_allows_contained_file() {
        let dir = tempfile::tempdir().unwrap();
        let path = write_temp(dir.path(), "inside.txt", b"ok");
        let policy = PathPolicy::new().base_dir(dir.path());
        let opened = policy.open(&path).expect("contained file should open");
        assert!(opened.path.starts_with(dir.path().canonicalize().unwrap()));
    }

    #[test]
    fn base_dir_rejects_parent_traversal() {
        let base = tempfile::tempdir().unwrap();
        let outside = tempfile::tempdir().unwrap();
        let _secret = write_temp(outside.path(), "secret.txt", b"top secret");

        let policy = PathPolicy::new().base_dir(base.path());
        // Lexical traversal out of the base directory.
        let traversal = base
            .path()
            .join("..")
            .join(outside.path().file_name().unwrap());
        let err = policy.open(&traversal.join("secret.txt")).unwrap_err();
        assert!(
            matches!(
                err,
                PathSecurityError::ParentTraversal(_) | PathSecurityError::OutsideBase { .. }
            ),
            "got: {err:?}"
        );
    }

    #[cfg(unix)]
    #[test]
    fn base_dir_rejects_symlink_escape() {
        use std::os::unix::fs::symlink;

        let base = tempfile::tempdir().unwrap();
        let outside = tempfile::tempdir().unwrap();
        let secret = write_temp(outside.path(), "secret.txt", b"top secret");

        let link = base.path().join("link.txt");
        symlink(&secret, &link).unwrap();

        // Symlinks allowed, but the resolved target escapes the base dir.
        let policy = PathPolicy::new().base_dir(base.path());
        let err = policy.open(&link).unwrap_err();
        assert!(
            matches!(err, PathSecurityError::OutsideBase { .. }),
            "got: {err:?}"
        );
    }

    #[cfg(unix)]
    #[test]
    fn symlinks_can_be_denied() {
        use std::os::unix::fs::symlink;

        let dir = tempfile::tempdir().unwrap();
        let target = write_temp(dir.path(), "target.txt", b"data");
        let link = dir.path().join("link.txt");
        symlink(&target, &link).unwrap();

        let policy = PathPolicy::new().allow_symlinks(false);
        let err = policy.open(&link).unwrap_err();
        assert!(
            matches!(err, PathSecurityError::SymlinkDenied(_)),
            "got: {err:?}"
        );
    }

    #[cfg(unix)]
    #[test]
    fn interior_nul_is_rejected() {
        use std::ffi::OsStr;
        use std::os::unix::ffi::OsStrExt;

        let policy = PathPolicy::new();
        let path = Path::new(OsStr::from_bytes(b"a\0b"));
        let err = policy.open(path).unwrap_err();
        assert!(
            matches!(err, PathSecurityError::InteriorNul),
            "got: {err:?}"
        );
    }

    // -- safe_join (lexical confinement) --------------------------------------

    const BASE: &str = "/var/lib/simdutf8-cli/data";

    fn base() -> PathBuf {
        PathBuf::from(BASE)
    }

    #[test]
    fn safe_join_accepts_well_formed_relative_paths() {
        for candidate in [
            "advisory.json",
            "2026/001/file.json",
            "./a/./b.json",
            ".hidden",
            "",
        ] {
            let resolved = safe_join(&base(), candidate)
                .unwrap_or_else(|| panic!("expected accept for {candidate:?}"));
            assert!(resolved.starts_with(base()), "{candidate:?} escaped base");
            assert!(!resolved
                .components()
                .any(|c| matches!(c, Component::ParentDir)));
        }
    }

    #[test]
    fn safe_join_accepts_balanced_parent() {
        assert_eq!(
            safe_join(&base(), "a/../b.json"),
            Some(base().join("b.json"))
        );
        assert_eq!(safe_join(&base(), "2026/.."), Some(base()));
    }

    #[test]
    fn safe_join_rejects_traversal_and_absolute_and_nul() {
        for candidate in [
            "..",
            "../etc/passwd",
            "../../../../etc/passwd",
            "2026/../../etc/passwd",
            "/etc/passwd",
            "advisory.json\0",
            "a\0b",
        ] {
            assert!(
                safe_join(&base(), candidate).is_none(),
                "expected reject for {candidate:?}"
            );
        }
    }

    #[test]
    fn safe_join_rejects_every_traversal_depth() {
        for depth in 1..=64 {
            let attack = "../".repeat(depth) + "etc/passwd";
            assert!(
                safe_join(&base(), &attack).is_none(),
                "depth {depth} should be rejected"
            );
        }
    }
}