koh 0.3.2

koh — a resilient peer-to-peer remote shell: mosh, rewritten in Rust over iroh
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
//! # koh-pty — PTY allocation, shell spawn, resize, reaping
//!
//! The server side's plumbing to the real shell. Allocates a pseudo-terminal, spawns the
//! user's login shell under it, pumps the child's output to an async channel from a dedicated
//! blocking thread (portable-pty's reader is blocking-only), forwards input bytes to the
//! child via a second dedicated thread (so a slow child never blocks a tokio worker), and
//! propagates window-size changes (which `ioctl(TIOCSWINSZ)` turns into `SIGWINCH`).

use std::io::{self, Read, Write};
use std::sync::mpsc::{sync_channel, SyncSender, TrySendError};

use portable_pty::{
    native_pty_system, ChildKiller, CommandBuilder, ExitStatus, MasterPty, PtySize,
};
use tokio::sync::mpsc;

/// Size of each output chunk read from the PTY master.
const READ_CHUNK: usize = 8192;
/// Bound on the output channel (chunks). Backpressure here naturally slows the reader thread.
const OUTPUT_CHANNEL_DEPTH: usize = 512;
/// Bound on the input channel (chunks) feeding the writer thread. Generous, because under normal
/// interactive use the child drains its input promptly; a full queue means the child has stopped
/// reading (flow-controlled or hung), which [`Pty::write_input`] surfaces rather than blocking on.
const WRITE_CHANNEL_DEPTH: usize = 1024;

/// Resolve the session shell when the caller didn't pass `--shell`. Prefers `$SHELL`; otherwise a
/// platform default. portable-pty's `new_default_prog` falls back to `/bin/sh`, which does **not**
/// exist on Android (it's `/system/bin/sh`) — so a `koh serve` with no `--shell` would fail to spawn
/// a session there (and the Bevy Android app, which has no `$SHELL`, would hit the same). The logic
/// lives in the pure [`resolve_shell`] so it is unit-testable without touching the process env.
fn default_shell() -> String {
    resolve_shell(std::env::var_os("SHELL"))
}

/// Remove koh's operational env vars — notably the `$KOH_PASSPHRASE` second factor — from a
/// command's environment before it spawns the session shell (L-4 / KOH-15). `CommandBuilder::new`
/// seeds the full parent environment, so we strip *every* inherited `KOH_*` key by prefix (rather
/// than a hand-maintained list that silently misses future vars). Pulled out of [`Pty::spawn`] so
/// it is unit-testable without allocating a real PTY.
fn scrub_koh_env(cmd: &mut CommandBuilder) {
    for (key, _) in std::env::vars_os() {
        if key.to_string_lossy().starts_with("KOH_") {
            cmd.env_remove(&key);
        }
    }
    // Belt-and-suspenders for the security-critical one, regardless of the snapshot above.
    cmd.env_remove("KOH_PASSPHRASE");
}

fn resolve_shell(shell_env: Option<std::ffi::OsString>) -> String {
    if let Some(sh) = shell_env {
        if !sh.is_empty() {
            return sh.to_string_lossy().into_owned();
        }
    }
    if cfg!(target_os = "android") {
        "/system/bin/sh".to_string()
    } else {
        "/bin/sh".to_string()
    }
}

/// Typed errors from PTY allocation, shell spawn, and resize (mirrors the
/// `transport-iroh::SetupError` pattern so callers can match on the failure stage).
///
/// `portable-pty` surfaces its failures as `anyhow::Error`; we fold those into `io::Error`
/// (via [`io::Error::other`]) so every variant carries one concrete payload. Only the reader
/// thread's `Builder::spawn` is natively an `io::Error`, so it is the single `#[from]` source.
/// Binaries keep `anyhow` internally — their `?`/`.context()` absorb `PtyError` via anyhow's
/// blanket `From<E: Error + Send + Sync>`.
#[derive(Debug, thiserror::Error)]
pub enum PtyError {
    /// Allocating the pseudo-terminal pair (`openpty`) failed.
    #[error("opening pty: {0}")]
    OpenPty(#[source] io::Error),
    /// Spawning the shell under the slave side (`spawn_command`) failed.
    #[error("spawning shell: {0}")]
    Spawn(#[source] io::Error),
    /// Wiring up the master read/write pumps failed: cloning the reader, taking the writer, or
    /// starting the blocking reader thread (`Builder::spawn`, the native `io::Error` source).
    #[error("starting pty reader: {0}")]
    Reader(#[from] io::Error),
    /// Propagating a window-size change to the kernel (`master.resize`) failed.
    #[error("resizing pty: {0}")]
    Resize(#[source] io::Error),
}

/// A running shell behind a PTY.
///
/// Construct with [`Pty::spawn`], which also returns the receiver of the child's output.
/// Hold the `Pty` for the life of the session: dropping it drops `writer_tx`, which lets the
/// writer thread finish and drop the PTY's write handle — and `portable-pty` writes an EOT
/// (Ctrl-D) on that drop, so the child sees EOF on its stdin.
pub struct Pty {
    master: Box<dyn MasterPty + Send>,
    /// Bounded sender to the dedicated writer thread (which owns the blocking `Box<dyn Write>`).
    /// Shared by both input producers (keystrokes + host query replies), so writes stay FIFO.
    writer_tx: SyncSender<Vec<u8>>,
    child: Box<dyn portable_pty::Child + Send + Sync>,
    killer: Box<dyn ChildKiller + Send + Sync>,
    /// Join handles for the reader/writer pump threads, kept so a graceful [`Pty::shutdown`] can
    /// join them rather than leaking detached threads. `None` only after `shutdown` takes them.
    reader_handle: Option<std::thread::JoinHandle<()>>,
    writer_handle: Option<std::thread::JoinHandle<()>>,
}

impl Pty {
    /// Allocate a PTY of `rows`×`cols`, spawn `shell` (or the user's default login shell when
    /// `None`) with `TERM` set, and start streaming its output.
    ///
    /// Returns the [`Pty`] handle plus an async receiver of raw output chunks. The reader runs
    /// on a dedicated OS thread; when the child closes the PTY the channel ends.
    pub fn spawn(
        rows: u16,
        cols: u16,
        shell: Option<&str>,
        term: &str,
    ) -> Result<(Self, mpsc::Receiver<Vec<u8>>), PtyError> {
        let pty_system = native_pty_system();
        let pair = pty_system
            .openpty(PtySize {
                rows,
                cols,
                pixel_width: 0,
                pixel_height: 0,
            })
            .map_err(|e| PtyError::OpenPty(io::Error::other(e)))?;

        let mut cmd = match shell {
            Some(prog) => CommandBuilder::new(prog),
            None => CommandBuilder::new(default_shell()),
        };
        // A real terminal type so curses apps behave; the env is otherwise inherited.
        cmd.env("TERM", term);
        // Scrub koh's operational env from the child (L-4). Most important: `$KOH_PASSPHRASE` — the
        // second-factor secret — must NOT reach the spawned shell, or any authorized user could
        // `echo $KOH_PASSPHRASE` to recover it.
        scrub_koh_env(&mut cmd);

        let child = pair
            .slave
            .spawn_command(cmd)
            .map_err(|e| PtyError::Spawn(io::Error::other(e)))?;
        let killer = child.clone_killer();
        // The slave fd is now owned by the child; drop our handle so EOF propagates correctly.
        drop(pair.slave);

        let mut reader = pair
            .master
            .try_clone_reader()
            .map_err(|e| PtyError::Reader(io::Error::other(e)))?;
        let mut writer = pair
            .master
            .take_writer()
            .map_err(|e| PtyError::Reader(io::Error::other(e)))?;

        let (tx, rx) = mpsc::channel::<Vec<u8>>(OUTPUT_CHANNEL_DEPTH);
        let reader_handle = std::thread::Builder::new()
            .name("koh-pty-reader".into())
            .spawn(move || {
                let mut buf = [0u8; READ_CHUNK];
                loop {
                    match reader.read(&mut buf) {
                        Ok(0) => break, // EOF: slave closed (EIO is mapped to 0 on unix)
                        // `Read::read` guarantees `n <= buf.len()`, so `get(..n)` is always
                        // `Some`; the `else` is a panic-free fallback that can't actually run.
                        Ok(n) => {
                            let Some(chunk) = buf.get(..n) else { break };
                            if tx.blocking_send(chunk.to_vec()).is_err() {
                                break; // receiver dropped: session over
                            }
                        }
                        Err(e) => {
                            tracing::debug!(error = %e, "pty reader stopping");
                            break;
                        }
                    }
                }
            })?;

        // Dedicated writer thread: it owns the blocking `Box<dyn Write>` and drains the bounded
        // input channel, so `write_input` never blocks a tokio worker. `recv()` yields every
        // buffered chunk before it observes the senders being dropped, so pending writes flush
        // before the writer is dropped (and `portable-pty` then writes the EOT that EOFs the
        // child). The thread exits as soon as the last sender (held in `Pty`) drops.
        let (writer_tx, writer_rx) = sync_channel::<Vec<u8>>(WRITE_CHANNEL_DEPTH);
        let writer_handle = std::thread::Builder::new()
            .name("koh-pty-writer".into())
            .spawn(move || {
                while let Ok(chunk) = writer_rx.recv() {
                    if writer
                        .write_all(&chunk)
                        .and_then(|()| writer.flush())
                        .is_err()
                    {
                        break; // master closed / child gone
                    }
                }
                // `writer` drops here -> portable-pty sends EOT -> child sees EOF on stdin.
            })?;

        Ok((
            Self {
                master: pair.master,
                writer_tx,
                child,
                killer,
                reader_handle: Some(reader_handle),
                writer_handle: Some(writer_handle),
            },
            rx,
        ))
    }

    /// Gracefully tear down the session and join both I/O pump threads (rather than leaking them
    /// as detached threads). Consumes the `Pty`. It first kills the child — so the reader's
    /// blocking `read` returns EOF — then drops the writer sender — so the writer's `recv` returns
    /// — guaranteeing both threads unblock before we join them, so this never deadlocks.
    pub fn shutdown(mut self) {
        // A failed kill is logged, not ignored: if the child somehow survives it keeps the slave
        // fd open, the reader stays blocked on read(), and the join below would hang — so a warning
        // is the breadcrumb for that (otherwise impossible-looking) stall.
        if let Err(e) = self.killer.kill() {
            tracing::warn!(error = %e, "pty kill on shutdown failed; reader join may stall");
        }
        let reader = self.reader_handle.take();
        let writer = self.writer_handle.take();
        // Dropping `self` drops `writer_tx`, which lets the writer thread observe the channel
        // close and exit; the child kill above lets the reader thread hit EOF and exit.
        drop(self);
        if let Some(h) = writer {
            let _ = h.join();
        }
        if let Some(h) = reader {
            let _ = h.join();
        }
    }

    /// Forward input bytes to the child (verbatim — keystrokes or host query replies).
    ///
    /// Takes `&self` and never blocks: it enqueues `data` onto the bounded channel feeding the
    /// writer thread. Both producers share one sender, and callers enqueue while holding the
    /// session lock, so bytes stay FIFO (a DSR reply can't overtake the keystroke that triggered
    /// it). Returns [`io::ErrorKind::BrokenPipe`] if the writer thread is gone, and
    /// [`io::ErrorKind::WouldBlock`] if the queue is full — the defined over-limit policy: surface
    /// backpressure rather than block a tokio worker or silently drop input (a full 1024-deep
    /// queue means the child has stopped reading, i.e. the session is effectively dead).
    pub fn write_input(&self, data: &[u8]) -> io::Result<()> {
        match self.writer_tx.try_send(data.to_vec()) {
            Ok(()) => Ok(()),
            Err(TrySendError::Full(_)) => Err(io::Error::new(
                io::ErrorKind::WouldBlock,
                "pty writer queue full (child not draining its input)",
            )),
            Err(TrySendError::Disconnected(_)) => Err(io::Error::from(io::ErrorKind::BrokenPipe)),
        }
    }

    /// Propagate a window-size change; the kernel raises `SIGWINCH` in the child.
    pub fn resize(&self, rows: u16, cols: u16) -> Result<(), PtyError> {
        self.master
            .resize(PtySize {
                rows,
                cols,
                pixel_width: 0,
                pixel_height: 0,
            })
            .map_err(|e| PtyError::Resize(io::Error::other(e)))
    }

    /// Non-blocking check for child exit.
    pub fn try_wait(&mut self) -> std::io::Result<Option<ExitStatus>> {
        self.child.try_wait()
    }

    /// Block until the child exits, returning its status.
    pub fn wait(&mut self) -> std::io::Result<ExitStatus> {
        self.child.wait()
    }

    /// A standalone killer that can be moved to another thread/task to terminate the child.
    pub fn killer(&self) -> Box<dyn ChildKiller + Send + Sync> {
        self.killer.clone_killer()
    }

    /// Terminate the child (SIGHUP via the portable-pty killer).
    pub fn kill(&mut self) -> std::io::Result<()> {
        self.killer.kill()
    }

    /// Force-kill the child with SIGKILL (which cannot be trapped). portable-pty's cloned killer
    /// only sends SIGHUP, so a child that ignores SIGHUP (e.g. `trap '' HUP`) would otherwise keep
    /// the PTY slave fd open and wedge the reader thread on a blocking `read()` forever — leaking a
    /// thread + fds per session and defeating the reaper (KOH-10). Best effort: a child that has
    /// already exited yields `ESRCH`, which is fine. Off-unix this is a no-op (the killer's own
    /// terminate applies).
    pub fn kill_hard(&self) {
        #[cfg(unix)]
        if let Some(pid) = self.process_id() {
            use nix::sys::signal::{kill, Signal};
            use nix::unistd::Pid;
            if let Ok(pid) = i32::try_from(pid) {
                let _ = kill(Pid::from_raw(pid), Signal::SIGKILL);
            }
        }
    }

    /// The child's process id, if known.
    pub fn process_id(&self) -> Option<u32> {
        self.child.process_id()
    }
}

impl Drop for Pty {
    fn drop(&mut self) {
        // A `Pty` dropped without an explicit [`Pty::shutdown`] (e.g. the Err path in
        // `session::teardown`, or any future caller) must still guarantee the child dies, so the
        // detached reader thread can't block forever on a still-open slave fd (KOH-10). SIGHUP
        // first (a well-behaved shell exits cleanly), then SIGKILL so a SIGHUP-immune child also
        // dies → the reader hits EOF and the pump threads exit. `writer_tx` drops with the struct,
        // EOFing the child's stdin. We deliberately do NOT join the threads here (that could block
        // the dropping thread, possibly a tokio worker); SIGKILL makes them exit promptly on their
        // own, and `shutdown` remains the path that joins.
        let _ = self.killer.kill();
        self.kill_hard();
    }
}

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

    #[test]
    fn resolve_shell_prefers_env_then_platform_default() {
        use std::ffi::OsString;
        // An explicit non-empty `$SHELL` wins.
        assert_eq!(
            resolve_shell(Some(OsString::from("/usr/bin/fish"))),
            "/usr/bin/fish"
        );
        // Empty `$SHELL` behaves like unset → a concrete absolute platform default.
        let empty = resolve_shell(Some(OsString::new()));
        let unset = resolve_shell(None);
        assert_eq!(empty, unset, "empty SHELL falls through like unset");
        assert!(
            unset.starts_with('/') && !unset.is_empty(),
            "an absolute fallback path"
        );
        // On Android the default must be the shell that actually exists (NOT /bin/sh).
        if cfg!(target_os = "android") {
            assert_eq!(unset, "/system/bin/sh");
        } else {
            assert_eq!(unset, "/bin/sh");
        }
    }

    #[test]
    fn scrub_removes_koh_passphrase_even_when_inherited() {
        // L-4: even when the parent process has $KOH_PASSPHRASE set, the spawned shell's env must
        // not — otherwise any authorized user could `echo $KOH_PASSPHRASE`. CommandBuilder::new
        // seeds the full parent env, so this proves env_remove strips an *inherited* secret.
        std::env::set_var("KOH_PASSPHRASE", "topsecret-unit");
        std::env::set_var("KOH_DNS", "1.1.1.1");
        let mut cmd = CommandBuilder::new("/bin/sh");
        assert!(
            cmd.get_env("KOH_PASSPHRASE").is_some(),
            "the builder seeds the parent env, so the var is present before scrubbing"
        );
        scrub_koh_env(&mut cmd);
        assert!(
            cmd.get_env("KOH_PASSPHRASE").is_none(),
            "the second-factor passphrase must be scrubbed from the child env"
        );
        assert!(
            cmd.get_env("KOH_DNS").is_none(),
            "operational KOH_* vars are scrubbed too"
        );
        std::env::remove_var("KOH_PASSPHRASE");
        std::env::remove_var("KOH_DNS");
    }

    #[test]
    #[allow(
        clippy::items_after_statements,
        reason = "`_assert_typed` is a deliberate compile-time signature assertion kept beside the runtime checks it documents"
    )]
    fn pty_error_variants_are_constructible_and_reachable() {
        let mk = || io::Error::other("boom");
        // Each stage variant is constructible and renders a non-empty message.
        for e in [
            PtyError::OpenPty(mk()),
            PtyError::Spawn(mk()),
            PtyError::Reader(mk()),
            PtyError::Resize(mk()),
        ] {
            assert!(!e.to_string().is_empty(), "variant must Display");
        }
        // The `#[from] io::Error` source (the reader-thread spawn path) yields `Reader`.
        let from_io: PtyError = mk().into();
        assert!(matches!(from_io, PtyError::Reader(_)));
        // A binary's `?`/`.context()` absorbs PtyError via anyhow's blanket `From` — the
        // typed error stays internal to the lib but composes with anyhow at the edges.
        let absorbed: anyhow::Error = PtyError::OpenPty(mk()).into();
        assert!(absorbed.to_string().contains("opening pty"));
        // The public spawn signature now carries the typed error.
        fn _assert_typed(r: Result<(), PtyError>) -> Result<(), PtyError> {
            r
        }
    }

    // The real-PTY / real-shell tests (spawn + stream + teardown) live in `tests/pty.rs` — a
    // dedicated integration-test binary — so they don't contend with the ~100 inline tests in
    // this crate's parallel test binary (which starved the PTY reader thread under load).
}