tilewright-source 0.1.0

// Copyright 2026 Seraphina Studios
// SPDX-License-Identifier: MIT

//! Confined access to the files of a source tree, for tools that edit it.
//!
//! This is the loader's own path confinement: paths are root-relative and
//! `/`-separated, with no empty, `.` or `..` component and no `\`, `:` or
//! NUL in any component, and nothing beneath the root is followed if it is
//! a symbolic link. On Unix every component is opened through its parent's
//! directory handle without following links, and writes, renames and
//! removals act on names inside a directory opened that way, so a link
//! swapped in at any moment is never followed. Elsewhere components are
//! checked by path before use, on Windows for any reparse point (symbolic
//! links, junctions and mount points among them), a best effort that a
//! local process changing the tree at the same time can race.

use crate::files::{Refused, components, platform};
use std::{
    fmt,
    io::Read as _,
    path::{Path, PathBuf},
};

/// Why a path beneath the root could not be used.
#[derive(Debug)]
pub enum AccessError {
    /// The path has an empty, `.` or `..` component, or a component with
    /// `\`, `:` or NUL; or a name has a `/`.
    Path(String),
    /// A symbolic link at this prefix of the path.
    Link(String),
    /// The entry is not a regular file or directory, as the access needs.
    Unexpected(String),
    /// The file is larger than the limit of the read.
    TooLarge(String),
    /// Another holder has the root's lock.
    Locked,
    Io(String, std::io::Error),
}

impl fmt::Display for AccessError {
    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
        match self {
            Self::Path(path) => write!(formatter, "{path}: not a relative path beneath the root"),
            Self::Link(path) => write!(formatter, "{path}: symbolic links are not followed"),
            Self::Unexpected(path) => {
                write!(formatter, "{path}: not a regular file or directory")
            }
            Self::TooLarge(path) => write!(formatter, "{path}: file is too large"),
            Self::Locked => write!(formatter, "another edit holds the source tree's lock"),
            Self::Io(path, error) => write!(formatter, "{path}: {error}"),
        }
    }
}

impl std::error::Error for AccessError {}

impl AccessError {
    /// A removal of `path` that took effect but is not finished, as
    /// [`removal_pending`](Self::removal_pending) says, for tests of what
    /// is made of one; only Windows makes them otherwise.
    #[doc(hidden)]
    #[must_use]
    pub fn pending_removal(path: &str) -> Self {
        Self::Io(
            path.to_owned(),
            std::io::Error::new(std::io::ErrorKind::ResourceBusy, RemovalPending),
        )
    }

    /// Whether this is a removal that took effect but is not finished: on
    /// Windows a removed entry stays, on some file systems, until another
    /// program closes it, and goes then.
    #[must_use]
    pub fn removal_pending(&self) -> bool {
        type Inner = dyn std::error::Error + Send + Sync;
        matches!(self, Self::Io(_, error) if error.get_ref().is_some_and(Inner::is::<RemovalPending>))
    }
}

/// A removal that took effect but stays until another program closes the
/// entry; see [`AccessError::removal_pending`].
#[derive(Debug)]
pub(crate) struct RemovalPending;

impl fmt::Display for RemovalPending {
    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
        formatter.write_str("it is removed, but stays until another program closes it")
    }
}

impl std::error::Error for RemovalPending {}

fn refused(path: &str, refused: Refused) -> AccessError {
    match refused {
        Refused::Link(prefix) => AccessError::Link(prefix),
        Refused::Missing | Refused::Unexpected => AccessError::Unexpected(path.to_owned()),
        Refused::Io(error) => AccessError::Io(path.to_owned(), error),
    }
}

/// The name of the file in the root that is its lock on Windows. The
/// command line writes no output there on any platform.
pub const LOCK_NAME: &str = ".tilewright.lock";

/// The checked components of a root-relative path; the empty path is the
/// root.
pub(crate) fn checked(path: &str) -> Result<Vec<&str>, AccessError> {
    if path.is_empty() {
        return Ok(Vec::new());
    }
    if path.len() > 4096 || path.split('/').any(|name| checked_name(name).is_err()) {
        return Err(AccessError::Path(path.to_owned()));
    }
    Ok(components(path))
}

/// Whether `name` can name an entry of a directory: not empty, `.` or `..`,
/// and without `/`, `\`, `:` or NUL, so that no platform reads it as more
/// than one component, a drive or a stream.
fn checked_name(name: &str) -> Result<&str, AccessError> {
    if name.is_empty()
        || name == "."
        || name == ".."
        || name.len() > 255
        || name.contains(['/', '\\', ':', '\0'])
    {
        Err(AccessError::Path(name.to_owned()))
    } else {
        Ok(name)
    }
}

/// Who may use a file or directory, as publishing keeps it: where there are
/// no modes, a read-only file has none of the write bits.
///
/// On Unix: the permission bits, the owner and the group, and on Linux the
/// mandatory access control labels (`SELinux`, Smack), which publishing
/// requires to be the same on every entry it compares. Access control that
/// publishing could not keep is refused when the access is read: on Linux
/// an access control list. On macOS and Windows access control lists are
/// neither read nor kept.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub struct Access {
    /// The permission bits, with the set-user, set-group and sticky bits.
    pub mode: u32,
    /// The owning user.
    pub user: u32,
    /// The owning group.
    pub group: u32,
    /// A digest of the mandatory access control labels (Linux), all zero
    /// where there are none.
    pub control: [u8; 32],
}

/// Who may reach content through a file or directory: for a file who may
/// read it, for a directory who may search it. Publishing requires the
/// entries it compares to have one audience.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub struct Audience {
    /// The classes that may reach, as read bits.
    reach: u32,
    /// The owner, where not everyone may reach.
    user: Option<u32>,
    /// The group, where only it may reach.
    group: Option<u32>,
    /// The labels.
    control: [u8; 32],
}

impl Access {
    /// Whether the group may read and others may not, so that the group
    /// says who may.
    #[must_use]
    pub const fn group_reads_alone(&self) -> bool {
        self.mode & 0o040 != 0 && self.mode & 0o004 == 0
    }

    /// Whether the classes that may reach content through an entry with
    /// this access, a directory if `directory`, nest: its owner may
    /// whenever its group may, and its group whenever others may. Where
    /// they do not, as for mode `0604`, which denies the group what it
    /// lets others read, who may reach depends on who the owner and group
    /// are, which [`audience`](Self::audience) does not say.
    #[must_use]
    pub const fn nested(&self, directory: bool) -> bool {
        let bits = if directory {
            self.mode & 0o111
        } else {
            (self.mode & 0o444) >> 2
        };
        let (owner, group, other) = (bits & 0o100 != 0, bits & 0o010 != 0, bits & 0o001 != 0);
        (owner || !group) && (group || !other)
    }

    /// Who may reach content through an entry with this access, a
    /// directory if `directory`.
    #[must_use]
    pub fn audience(&self, directory: bool) -> Audience {
        let reach = if directory {
            (self.mode & 0o111) << 2
        } else {
            self.mode & 0o444
        };
        Audience {
            reach,
            user: (reach != 0o444).then_some(self.user),
            group: (reach & 0o040 != 0 && reach & 0o004 == 0).then_some(self.group),
            control: self.control,
        }
    }

    /// The access a new file in a directory with this access is given: the
    /// directory's mode without the execute bits, its group and its labels,
    /// and as its owner the user that makes it, this process's effective
    /// user (0 where there are no users).
    #[must_use]
    pub fn new_file(&self) -> Self {
        #[cfg(unix)]
        let user = rustix::process::geteuid().as_raw();
        #[cfg(not(unix))]
        let user = 0;
        Self {
            mode: self.mode & 0o666,
            user,
            ..*self
        }
    }
}

/// A source tree root. The root itself may be a link; nothing beneath it is
/// followed.
pub struct SourceRoot {
    path: PathBuf,
    root: platform::Root,
}

impl SourceRoot {
    /// Opens the root. On Unix the directory is opened here, and every
    /// later access goes through that handle, so moving or replacing what
    /// the path names afterwards does not move the root; elsewhere the path
    /// is resolved to an absolute one here.
    #[must_use]
    pub fn new(root: &Path) -> Self {
        Self {
            path: root.to_path_buf(),
            root: platform::Root::new(root),
        }
    }

    /// A second handle on the same root directory.
    #[must_use]
    pub fn try_clone(&self) -> Self {
        Self {
            path: self.path.clone(),
            root: self.root.try_clone(),
        }
    }

    /// The path supplied when this root was opened, for display only.
    /// It may be relative, moved or replaced; using it to reopen the tree
    /// does not preserve this root's identity. Access files through this
    /// root instead.
    #[must_use]
    pub fn path(&self) -> &Path {
        &self.path
    }

    pub(crate) fn platform(&self) -> platform::Root {
        self.root.try_clone()
    }

    /// What is at `path`, not following it, if anything: as the file
    /// system finds it by that name, which on one that ignores case, or as
    /// on Windows drops trailing dots, can be an entry named otherwise.
    /// A path through a file names nothing; a link on the way is an error.
    pub fn exists(&self, path: &str) -> Result<bool, AccessError> {
        // The empty path is the root.
        if path.is_empty() {
            return self.directory(path, false).map(|root| root.is_some());
        }
        match self.root.kind(&checked(path)?) {
            Ok(crate::files::Found::Missing) | Err(Refused::Missing | Refused::Unexpected) => {
                Ok(false)
            }
            Ok(_) => Ok(true),
            // The entry itself is a link.
            Err(Refused::Link(prefix)) if prefix == path => Ok(true),
            Err(other) => Err(refused(path, other)),
        }
    }

    /// Reads a regular file of at most `limit` bytes. `Ok(None)` means the
    /// file, or a directory on its way, is missing.
    pub fn read(&self, path: &str, limit: u64) -> Result<Option<Vec<u8>>, AccessError> {
        let parts = checked(path)?;
        if parts.is_empty() {
            return Err(AccessError::Unexpected(path.to_owned()));
        }
        let (file, length) = match self.root.file(&parts) {
            Ok(opened) => opened,
            Err(Refused::Missing) => return Ok(None),
            Err(other) => return Err(refused(path, other)),
        };
        if length > limit {
            return Err(AccessError::TooLarge(path.to_owned()));
        }
        // One byte past the limit shows a file that grew after the check.
        let mut bytes = Vec::new();
        file.take(limit.saturating_add(1))
            .read_to_end(&mut bytes)
            .map_err(|error| AccessError::Io(path.to_owned(), error))?;
        if u64::try_from(bytes.len()).unwrap_or(u64::MAX) > limit {
            return Err(AccessError::TooLarge(path.to_owned()));
        }
        Ok(Some(bytes))
    }

    /// Opens the directory `path` (the empty path is the root), creating it
    /// and its missing parents when `create`. `Ok(None)` means it is missing
    /// and was not created.
    pub fn directory(
        &self,
        path: &str,
        create: bool,
    ) -> Result<Option<SourceDirectory>, AccessError> {
        match self.root.directory(&checked(path)?, create) {
            Ok(directory) => Ok(Some(SourceDirectory {
                directory,
                path: path.to_owned(),
            })),
            Err(Refused::Missing) => Ok(None),
            Err(other) => Err(refused(path, other)),
        }
    }

    /// Makes the directory `path` in its existing parent. Returns whether
    /// this call made it: `false` when something by that name was already
    /// there, which the caller then did not create.
    pub fn make_directory(&self, path: &str) -> Result<bool, AccessError> {
        self.root
            .make_directory(&checked(path)?)
            .map_err(|error| refused(path, error))
    }

    /// The UTF-8 names of the entries of the directory `path`, links
    /// included, and how many entries it has, those whose names are not
    /// UTF-8 included; `None` when it is missing. A directory with more
    /// than `limit` entries is an error, [`AccessError::TooLarge`].
    pub fn names(
        &self,
        path: &str,
        limit: usize,
    ) -> Result<Option<(std::collections::BTreeSet<String>, usize)>, AccessError> {
        let listing = match self.root.list(&checked(path)?) {
            Ok(listing) => listing,
            Err(Refused::Missing) => return Ok(None),
            Err(other) => return Err(refused(path, other)),
        };
        let mut names = std::collections::BTreeSet::new();
        let mut count = 0;
        for entry in listing {
            if count >= limit {
                return Err(AccessError::TooLarge(path.to_owned()));
            }
            count += 1;
            let entry = entry.map_err(|error| AccessError::Io(path.to_owned(), error))?;
            names.extend(entry.name);
        }
        Ok(Some((names, count)))
    }

    /// Takes the root's exclusive lock, held until the returned value is
    /// dropped. It fails at once with [`AccessError::Locked`] if another
    /// holder, in this process or another, has it. On Unix it is an
    /// advisory `flock` of the root directory; on Windows the file
    /// [`LOCK_NAME`] in the root, created afresh, shared with no one and
    /// deleted when the lock is released or its process stops, so a file by
    /// that name that no one holds is an error. Elsewhere there is no lock
    /// and this always succeeds.
    pub fn lock(&self) -> Result<SourceLock, AccessError> {
        match self.root.lock() {
            Ok(Some(lock)) => Ok(SourceLock(lock)),
            Ok(None) => Err(AccessError::Locked),
            Err(other) => Err(refused(".", other)),
        }
    }
}

/// The mounts of this process, read once to tell the file systems of
/// several directories ([`SourceDirectory::unchecked_file_system`]): on
/// Linux from `/proc/self/mountinfo`, bounded; elsewhere nothing.
pub struct Mounts(platform::Mounts);

impl Mounts {
    /// Reads the mounts. Where they cannot be read, every directory's file
    /// system is one that cannot be told.
    #[must_use]
    pub fn read() -> Self {
        Self(platform::Mounts::read())
    }
}

/// The root's lock; see [`SourceRoot::lock`].
pub struct SourceLock(#[allow(dead_code)] platform::Lock);

/// A directory beneath the root. Every name is a single entry of it.
pub struct SourceDirectory {
    directory: platform::Directory,
    path: String,
}

impl SourceDirectory {
    fn entry(&self, name: &str) -> String {
        crate::files::join(&self.path, name)
    }

    /// Writes a new file `name` with `bytes` and syncs it. Whatever was at
    /// `name` is removed first, so a link or hard link there is never
    /// written through. With `access`, the file is its owner's alone until
    /// written, then takes that mode and group, and on Linux no access
    /// control list it would get from the directory; it must have the
    /// labels of `access`, which is checked before anything is written.
    /// Where there are no modes, it is read-only if `access` has no write
    /// bits. Without `access`, it has the platform's default.
    pub fn write(
        &self,
        name: &str,
        access: Option<&Access>,
        bytes: &[u8],
    ) -> Result<(), AccessError> {
        self.directory
            .write(checked_name(name)?, access, bytes)
            .map_err(|error| refused(&self.entry(name), error))
    }

    /// Who may use the regular file `name`; `None` when there is no regular
    /// file there. A file whose access publishing could not keep is
    /// refused: on Linux one with an access control list, and on Windows a
    /// read-only or encrypted one. macOS and
    /// Windows access control lists are neither read nor kept.
    pub fn access(&self, name: &str) -> Result<Option<Access>, AccessError> {
        self.directory
            .access(checked_name(name)?)
            .map_err(|error| refused(&self.entry(name), error))
    }

    /// Who may use this directory. On Linux, a directory with an access
    /// control list, its own or one it gives new files, is refused.
    pub fn own_access(&self) -> Result<Access, AccessError> {
        self.directory
            .own_access()
            .map_err(|error| refused(&self.path, error))
    }

    /// Gives this directory the mode and group of `access`, and checks that
    /// it has the labels of `access`.
    pub fn set_own_access(&self, access: &Access) -> Result<(), AccessError> {
        self.directory
            .set_own_access(access)
            .map_err(|error| refused(&self.path, error))
    }

    /// Checks, before either is made, that a file made here as `name` would
    /// get the access control beside the mode bits that one made as
    /// `staged` would: on Linux with `SELinux`, the label its policy gives a
    /// new file of that name, which may depend on the name and is not given
    /// again on renaming. Elsewhere there is nothing to check.
    pub fn check_final_name(&self, staged: &str, name: &str) -> Result<(), AccessError> {
        self.directory
            .check_final_name(checked_name(staged)?, checked_name(name)?)
            .map_err(|error| refused(&self.entry(name), error))
    }

    /// Renames `from` to `to`, replacing `to`, atomically.
    pub fn rename(&self, from: &str, to: &str) -> Result<(), AccessError> {
        self.directory
            .rename(checked_name(from)?, checked_name(to)?)
            .map_err(|error| refused(&self.entry(to), error))
    }

    /// Removes the file `name`; `false` when there was none.
    pub fn remove_file(&self, name: &str) -> Result<bool, AccessError> {
        self.directory
            .remove_file(checked_name(name)?)
            .map_err(|error| refused(&self.entry(name), error))
    }

    /// Removes the directory `name` if it is empty; `false` when it is
    /// missing or not empty.
    pub fn remove_directory(&self, name: &str) -> Result<bool, AccessError> {
        self.directory
            .remove_directory(checked_name(name)?)
            .map_err(|error| refused(&self.entry(name), error))
    }

    /// The file system this directory is on, in words, where publishing
    /// does not read the access control lists of what is in it: on Linux
    /// any but the local ones that keep them as the attributes publishing
    /// reads (ext2, ext3, ext4, XFS, Btrfs, tmpfs, F2FS, ZFS and bcachefs),
    /// so network, host-shared (such as WSL's Windows drives or a virtual
    /// machine's shared folders), FUSE, NTFS and stacked ones (overlayfs)
    /// among them. The directory's mount is found by its ID in `mounts`,
    /// and both the type the kernel names it with and the magic number
    /// `statfs` gives must be such a file system's; one that cannot be
    /// told, as where `/proc` cannot be read or the mount's line is
    /// malformed, is not checked
    /// either, and says why. `None` where it is checked, and on other
    /// platforms, which do not read access control lists at all.
    pub fn unchecked_file_system(&self, mounts: &Mounts) -> Result<Option<String>, AccessError> {
        self.directory
            .unchecked_file_system(&mounts.0)
            .map_err(|error| refused(&self.path, error))
    }

    /// Syncs the directory, making renames and removals in it durable.
    /// Returns `false` on platforms that cannot sync a directory, where
    /// they may not be durable yet.
    pub fn sync(&self) -> Result<bool, AccessError> {
        self.directory
            .sync()
            .map_err(|error| refused(&self.path, error))
    }
}

impl fmt::Debug for SourceRoot {
    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
        formatter.write_str("SourceRoot")
    }
}