//! Who is on the other end of the socket, and what they may ask for.
//!
//! The daemon runs as root, so every method it exposes runs as root. Until
//! the desktop app existed that was academic — the socket was `0700 root:root`
//! and only `sudo hound` could reach it. But a GUI runs as the logged-in user
//! and has to reach the daemon without a password prompt on every status poll,
//! so the socket is now group-readable by `hound`.
//!
//! That group is a real trust grant and it is worth being exact about how far
//! it goes. `quarantine.add` moves any path into the vault, and
//! `quarantine.restore` writes a file back out as root — between them that is
//! arbitrary file replacement on the system, which is root. Group members
//! therefore get the read side of the API and nothing that writes:
//! administrative methods still require a peer with uid 0.
//!
//! Scanning is the exception worth explaining. It reads as root, so an
//! unprivileged caller could otherwise use it as an oracle for files it cannot
//! open. So a non-root caller's scan is checked against what *that caller*
//! could read, by forking a child, dropping it to the peer's uid, gid and
//! supplementary groups, and letting the kernel answer with `access(2)` —
//! which honours ACLs and mount options, unlike anything we could reconstruct
//! from a mode bitfield.

use anyhow::{bail, Result};
use std::ffi::CString;
use std::os::unix::io::AsRawFd;
use std::os::unix::net::UnixStream;
use std::path::Path;

/// The credentials the kernel attaches to a connection. Unforgeable: these
/// come from `SO_PEERCRED`, not from anything the client sent us.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct Peer {
    pub pid: i32,
    pub uid: u32,
    pub gid: u32,
}

impl Peer {
    pub fn is_root(&self) -> bool {
        self.uid == 0
    }

    /// May this peer use the methods that write?
    ///
    /// Root, or whoever the daemon is running as. The second clause is not a
    /// loophole: a user running `houndd` under their own account can already
    /// kill it, attach a debugger to it, or edit its settings file directly,
    /// so refusing them the API buys nothing and breaks every developer and
    /// test run. In the shipped configuration the daemon is uid 0, and this
    /// collapses to exactly "the peer is root".
    pub fn may_administer(&self) -> bool {
        // SAFETY: getuid takes no arguments and cannot fail.
        self.is_root() || self.uid == unsafe { libc::getuid() }
    }

    pub fn from_stream(stream: &UnixStream) -> Result<Self> {
        let mut cred = libc::ucred {
            pid: 0,
            uid: u32::MAX,
            gid: u32::MAX,
        };
        let mut len = std::mem::size_of::<libc::ucred>() as libc::socklen_t;
        // SAFETY: cred and len are correctly sized for SO_PEERCRED, and the fd
        // is owned by the borrowed stream for the duration of the call.
        let rc = unsafe {
            libc::getsockopt(
                stream.as_raw_fd(),
                libc::SOL_SOCKET,
                libc::SO_PEERCRED,
                (&mut cred as *mut libc::ucred).cast(),
                &mut len,
            )
        };
        if rc != 0 {
            bail!(
                "reading peer credentials: {}",
                std::io::Error::last_os_error()
            );
        }
        Ok(Peer {
            pid: cred.pid,
            uid: cred.uid,
            gid: cred.gid,
        })
    }
}

/// What a method costs the caller in trust.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum Access {
    /// Reports the daemon's own state. Anyone who can open the socket.
    Read,
    /// Reads a caller-supplied path as root; allowed once we have confirmed
    /// the caller could have read it themselves.
    ReadsPath,
    /// Changes the system, or the daemon's configuration. Root only.
    Admin,
}

pub fn access_for(method: &str) -> Access {
    match method {
        "status"
        | "settings.get"
        | "events.list"
        | "quarantine.list"
        | "realtime.status"
        | "rootkit.scan"
        | "persistence.scan"
        // Asking whether a newer version exists changes nothing, and reveals
        // nothing the website does not already say.
        | "release.check"
        // Reports on this daemon's own blind spots. Reveals nothing about
        // the machine that its own operator cannot already see.
        | "selfcheck" => Access::Read,

        "scan" | "supply.sweep" => Access::ReadsPath,

        // Everything that writes: definitions updates, settings, the event
        // log, and every quarantine mutation.
        _ => Access::Admin,
    }
}

/// Can `peer` read `path` in its own right? Answered by the kernel, in a child
/// process that has actually become the peer.
///
/// Errs on the side of refusal: a fork failure, a child that dies on a signal,
/// or any unexpected exit status is a no.
pub fn peer_can_read(peer: &Peer, path: &Path) -> bool {
    if peer.is_root() {
        return true;
    }
    let Ok(c_path) = CString::new(path.as_os_str().as_encoded_bytes()) else {
        return false; // an interior NUL is not a path we will scan
    };

    // Impersonating someone requires being root. The packaged daemon is, but
    // a developer running `houndd` from a build tree is not — and there the
    // only peer it can answer for is itself.
    // SAFETY: geteuid takes no arguments and cannot fail.
    if unsafe { libc::geteuid() } != 0 {
        // SAFETY: as above.
        if unsafe { libc::getuid() } != peer.uid {
            return false;
        }
        // SAFETY: c_path is NUL-terminated and outlives the call.
        return unsafe { libc::access(c_path.as_ptr(), libc::R_OK) } == 0;
    }

    let groups = supplementary_groups(peer.uid, peer.gid);

    // SAFETY: between fork and _exit the child calls only async-signal-safe
    // functions — setgroups, setgid, setuid, access, _exit. It allocates
    // nothing and unwinds nothing.
    let pid = unsafe { libc::fork() };
    if pid < 0 {
        return false;
    }
    if pid == 0 {
        unsafe {
            // Order matters: setgroups and setgid must precede setuid, or the
            // child no longer has the privilege to make those calls.
            if libc::setgroups(groups.len() as _, groups.as_ptr()) != 0 {
                libc::_exit(101);
            }
            if libc::setgid(peer.gid) != 0 {
                libc::_exit(102);
            }
            if libc::setuid(peer.uid) != 0 {
                libc::_exit(103);
            }
            // Belt and braces: if the drop somehow left us root, refuse.
            if libc::getuid() == 0 || libc::geteuid() == 0 {
                libc::_exit(104);
            }
            let ok = libc::access(c_path.as_ptr(), libc::R_OK) == 0;
            libc::_exit(if ok { 0 } else { 1 });
        }
    }

    let mut status = 0;
    // SAFETY: pid is a child we just created; status is a valid out-param.
    if unsafe { libc::waitpid(pid, &mut status, 0) } != pid {
        return false;
    }
    if !libc::WIFEXITED(status) {
        return false;
    }
    match libc::WEXITSTATUS(status) {
        0 => true,
        1 => false,
        // The child could not become the peer, so the question was never
        // asked. Refuse — but say so, because a daemon that has lost
        // CAP_SETUID will refuse every non-root scan and the reason should
        // not have to be guessed from a permission error.
        code => {
            eprintln!(
                "peer check: could not drop to uid {} (stage {code}) — refusing to \
                 vouch for a read this daemon cannot verify",
                peer.uid
            );
            false
        }
    }
}

/// The peer's supplementary groups, so the check honours group-readable files.
/// Falls back to the primary gid alone if the user cannot be resolved.
fn supplementary_groups(uid: u32, gid: u32) -> Vec<libc::gid_t> {
    let Some(name) = username_for(uid) else {
        return vec![gid];
    };
    let Ok(c_name) = CString::new(name) else {
        return vec![gid];
    };
    let mut ngroups: libc::c_int = 32;
    let mut groups: Vec<libc::gid_t> = vec![0; ngroups as usize];
    // SAFETY: c_name outlives the call; groups has ngroups capacity. A -1
    // return rewrites ngroups with the count actually needed, so we retry once.
    let rc = unsafe { libc::getgrouplist(c_name.as_ptr(), gid, groups.as_mut_ptr(), &mut ngroups) };
    if rc < 0 {
        groups = vec![0; ngroups.max(1) as usize];
        // SAFETY: as above, now with the size the kernel asked for.
        if unsafe { libc::getgrouplist(c_name.as_ptr(), gid, groups.as_mut_ptr(), &mut ngroups) } < 0
        {
            return vec![gid];
        }
    }
    groups.truncate(ngroups.max(0) as usize);
    if groups.is_empty() {
        groups.push(gid);
    }
    groups
}

/// Resolve a uid to a login name via /etc/passwd. Reading the file directly
/// keeps this free of NSS, which can block on a network directory service —
/// not something a scan request should ever wait on.
fn username_for(uid: u32) -> Option<String> {
    let passwd = std::fs::read_to_string("/etc/passwd").ok()?;
    for line in passwd.lines() {
        let mut f = line.split(':');
        let name = f.next()?;
        let _pw = f.next();
        let this_uid: u32 = f.next()?.parse().ok()?;
        if this_uid == uid {
            return Some(name.to_string());
        }
    }
    None
}

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

    #[test]
    fn reads_are_open_and_writes_are_not() {
        assert_eq!(access_for("status"), Access::Read);
        assert_eq!(access_for("quarantine.list"), Access::Read);
        // Asking whether a newer version exists changes nothing and reveals
        // nothing the website does not.
        assert_eq!(access_for("release.check"), Access::Read);
        assert_eq!(access_for("scan"), Access::ReadsPath);
        assert_eq!(access_for("supply.sweep"), Access::ReadsPath);
        for admin in [
            "update",
            "settings.set",
            "events.clear",
            "quarantine.add",
            "quarantine.restore",
            "quarantine.remove",
            "realtime.set_enabled",
            // Installing a licence writes a file the daemon acts on.
            "license.install",
        ] {
            assert_eq!(access_for(admin), Access::Admin, "{admin} must require root");
        }
    }

    /// An unknown method must not fall through to the permissive arm. If
    /// someone adds a mutating method and forgets to classify it, it lands
    /// in Admin and fails closed rather than silently becoming public.
    #[test]
    fn unclassified_methods_fail_closed() {
        assert_eq!(access_for("some.future.method"), Access::Admin);
        assert_eq!(access_for(""), Access::Admin);
    }

    /// The owning user is an administrator of their own daemon; anybody else
    /// is not. Under a root daemon the two clauses are the same clause.
    #[test]
    fn the_owner_administers_and_strangers_do_not() {
        let me = unsafe { libc::getuid() };
        assert!(Peer { pid: 1, uid: 0, gid: 0 }.may_administer(), "root always");
        assert!(Peer { pid: 1, uid: me, gid: 0 }.may_administer(), "the owner");
        let stranger = if me == 0 { 1000 } else { me + 1 };
        assert!(
            !Peer { pid: 1, uid: stranger, gid: 0 }.may_administer(),
            "a different unprivileged user must not administer the daemon"
        );
    }

    #[test]
    fn root_peer_reads_anything() {
        let root = Peer {
            pid: 1,
            uid: 0,
            gid: 0,
        };
        assert!(peer_can_read(&root, Path::new("/etc/shadow")));
    }

    /// The check is only meaningful if it can say no. Running as an
    /// unprivileged user, /etc/shadow is the canonical unreadable file;
    /// running as root there is nothing to prove, so skip. This exercises
    /// the non-root branch, where the daemon can only answer for itself.
    #[test]
    fn unprivileged_peer_is_refused_a_file_it_cannot_open() {
        let uid = unsafe { libc::getuid() };
        if uid == 0 {
            return;
        }
        let gid = unsafe { libc::getgid() };
        let me = Peer {
            pid: std::process::id() as i32,
            uid,
            gid,
        };
        if Path::new("/etc/shadow").exists() {
            assert!(!peer_can_read(&me, Path::new("/etc/shadow")));
        }
        // ...and it must still say yes to something the peer owns.
        let mine = std::env::temp_dir().join(format!("hound-peer-{uid}-{}", std::process::id()));
        std::fs::write(&mine, b"x").unwrap();
        assert!(peer_can_read(&me, &mine));
        let _ = std::fs::remove_file(&mine);
    }

    /// An unprivileged daemon cannot impersonate anybody, and must refuse
    /// rather than answer a question it has no way to check.
    #[test]
    fn an_unprivileged_daemon_refuses_to_vouch_for_another_user() {
        if unsafe { libc::geteuid() } == 0 {
            return;
        }
        let someone_else = Peer {
            pid: 1,
            uid: unsafe { libc::getuid() } + 1,
            gid: 0,
        };
        let readable = std::env::temp_dir();
        assert!(!peer_can_read(&someone_else, &readable));
    }

    #[test]
    fn a_missing_file_is_not_readable() {
        let uid = unsafe { libc::getuid() };
        if uid == 0 {
            return;
        }
        let me = Peer {
            pid: 1,
            uid,
            gid: unsafe { libc::getgid() },
        };
        assert!(!peer_can_read(&me, Path::new("/nonexistent/hound/path")));
    }
}