term_session_server/

session_server.rs

use std::collections::HashMap;
use std::sync::Arc;

use muxio_core::rpc::rpc_internals::RpcStreamEvent;
use muxio_rpc_service::prebuffered::RpcMethodPrebuffered;
use muxio_rpc_service_caller::prebuffered::RpcCallPrebuffered;
use muxio_rpc_service_endpoint::{RpcServiceEndpointInterface, StreamResponder};
use muxio_tokio_rpc_ipc_server::{RpcIpcConnectionContextHandle, RpcIpcServer, RpcIpcServerEvent};
use portable_pty::PtySize;
use tokio::sync::{Mutex, Notify, mpsc, oneshot};

use term_session_muxio_service_definitions::{
    ChannelName, CloseSession, ListSessions, OnPtyResized, ResizePty, STREAM_INPUT_METHOD_ID,
    SUBSCRIBE_OUTPUT_METHOD_ID, Spawn, WriteInput,
};
use term_wm_pty_engine::PtyStatus;

use crate::session::Session;

/// Default terminal columns when no client constrains the PTY size.
const FALLBACK_COLS: u16 = 80;
/// Default terminal rows when no client constrains the PTY size.
const FALLBACK_ROWS: u16 = 24;
/// Hardcoded singleton session ID (this server manages one PTY at a time).
const SESSION_ID: u64 = 1;
/// Bounded input channel capacity — memory safety against extreme input bursts.
const INPUT_CHANNEL_CAPACITY: usize = 128;

/// Grace period to let the transport flush end-of-stream frames after the
/// session exits, before the server process terminates.
const SESSION_EXIT_FLUSH_GRACE: std::time::Duration = std::time::Duration::from_millis(100);

/// How often the output polling task wakes to re-check the session's exit
/// status, as a fallback for a missed or raced PTY EOF notification.
const SESSION_EXIT_POLL_INTERVAL: std::time::Duration = std::time::Duration::from_millis(100);

pub struct SessionServerConfig {
    pub channel: ChannelName,
    pub cmd: Vec<String>,
}

#[derive(Clone)]
struct ClientEntry {
    caller: Option<RpcIpcConnectionContextHandle>,
    cols: u16,
    rows: u16,
}

struct SubscriberEntry {
    conn_id: usize,
    respond: StreamResponder,
}

struct ServerState {
    session: Option<Session>,
    clients: HashMap<usize, ClientEntry>,
    subscribers: Vec<SubscriberEntry>,
    notify: Arc<Notify>,
}

impl ServerState {
    fn new(notify: Arc<Notify>) -> Self {
        Self {
            session: None,
            clients: HashMap::new(),
            subscribers: Vec::new(),
            notify,
        }
    }

    /// Replace the current session and attach the Notify callback
    /// so the background polling task is woken on PTY output.
    fn set_session(&mut self, mut session: Session) {
        let n = self.notify.clone();
        session.set_status_callback(Some(Box::new(move |status| {
            if matches!(status, PtyStatus::Wakeup | PtyStatus::Exited) {
                n.notify_one();
            }
        })));
        self.session = Some(session);
        // Prime notify to process initial startup output generated
        // before the callback was registered.
        self.notify.notify_one();
    }

    /// Terminate and clear the active session, flushing remaining PTY buffers
    /// and stream completion markers to all active subscribers.
    fn clear_session(&mut self) {
        if let Some(mut session) = self.session.take() {
            let _ = session.pty.kill_child();
            let raw = session.read_output();
            if !raw.is_empty() {
                for sub in &self.subscribers {
                    sub.respond.respond(raw.clone(), false);
                }
            }
        }
        for sub in &self.subscribers {
            sub.respond.respond(Vec::new(), true);
        }
        self.subscribers.clear();
        self.notify.notify_one();
    }

    /// Constrain the PTY to the smallest geometry across all connected clients.
    /// This guarantees the virtual buffer never exceeds any attached monitor.
    fn recalculate_pty_size(&mut self) {
        let Some(session) = self.session.as_mut() else {
            return;
        };
        if self.clients.is_empty() {
            return;
        }
        let min_cols = self
            .clients
            .values()
            .map(|c| c.cols)
            .filter(|&c| c != u16::MAX)
            .min()
            .unwrap_or(FALLBACK_COLS);
        let min_rows = self
            .clients
            .values()
            .map(|c| c.rows)
            .filter(|&r| r != u16::MAX)
            .min()
            .unwrap_or(FALLBACK_ROWS);
        let size = PtySize {
            rows: min_rows,
            cols: min_cols,
            pixel_width: 0,
            pixel_height: 0,
        };
        let _ = session.pty.resize(size);
        session.cols = min_cols;
        session.rows = min_rows;
    }

    /// Broadcast geometry to all clients via detached async tasks.
    /// Call AFTER releasing the ServerState lock.
    fn notify_clients(clients: &[ClientEntry], cols: u16, rows: u16) {
        for client in clients {
            let Some(caller) = client.caller.clone() else {
                continue;
            };
            tokio::spawn(async move {
                if let Err(e) = OnPtyResized::call(&caller, (cols, rows)).await {
                    tracing::debug!(error = ?e, "Failed to deliver OnPtyResized notification");
                }
            });
        }
    }
}

type SharedState = Arc<Mutex<ServerState>>;

/// Run the session server. Returns the PTY child's exit code on success.
pub async fn run_server(
    config: SessionServerConfig,
) -> Result<i32, Box<dyn std::error::Error + Send + Sync>> {
    let socket_name = config.channel.to_string();
    let notify = Arc::new(Notify::new());
    let state: SharedState = Arc::new(Mutex::new(ServerState::new(notify.clone())));

    {
        let mut st = state.lock().await;
        let cmd = if config.cmd.is_empty() {
            None
        } else {
            Some(config.cmd.clone())
        };
        let session = Session::spawn(
            SESSION_ID,
            cmd,
            FALLBACK_COLS,
            FALLBACK_ROWS,
            Some(&config.channel),
        )?;
        st.set_session(session);
    }

    let channel_id = config.channel.clone();

    let (event_tx, mut event_rx) = mpsc::unbounded_channel();
    let server = RpcIpcServer::new(Some(event_tx));
    let endpoint = server.endpoint();

    // Register Spawn
    let st = Arc::clone(&state);
    let ch = channel_id.clone();
    endpoint
        .register_prebuffered(Spawn::METHOD_ID, move |payload, ctx| {
            let state = Arc::clone(&st);
            let ch = ch.clone();
            async move {
                let mut guard = state.lock().await;
                let (cmd, cols, rows) = Spawn::decode_request(&payload)
                    .map_err(|e| Box::new(e) as Box<dyn std::error::Error + Send + Sync>)?;
                let entry = guard
                    .clients
                    .entry(ctx.conn_id)
                    .or_insert_with(|| ClientEntry {
                        caller: None,
                        cols,
                        rows,
                    });
                entry.cols = cols;
                entry.rows = rows;

                // If a session already exists and hasn't exited, reuse it.
                if guard.session.as_ref().is_some_and(|s| !s.exited) {
                    guard.recalculate_pty_size();
                    let (ncols, nrows) = guard
                        .session
                        .as_ref()
                        .map(|s| (s.cols, s.rows))
                        .unwrap_or((FALLBACK_COLS, FALLBACK_ROWS));
                    let targets: Vec<ClientEntry> = guard.clients.values().cloned().collect();
                    let id = guard.session.as_ref().map(|s| s.id).unwrap_or(SESSION_ID);
                    let cols = guard.session.as_ref().map(|s| s.cols).unwrap_or(cols);
                    let rows = guard.session.as_ref().map(|s| s.rows).unwrap_or(rows);
                    drop(guard);
                    ServerState::notify_clients(&targets, ncols, nrows);
                    return Spawn::encode_response((id, cols, rows))
                        .map_err(|e| Box::new(e) as Box<dyn std::error::Error + Send + Sync>);
                }
                let id = SESSION_ID;
                let session = Session::spawn(id, cmd, cols, rows, Some(&ch))?;
                guard.set_session(session);
                // Enforce global geometric constraints on the newly instantiated PTY
                guard.recalculate_pty_size();
                let (ncols, nrows) = guard
                    .session
                    .as_ref()
                    .map(|s| (s.cols, s.rows))
                    .unwrap_or((FALLBACK_COLS, FALLBACK_ROWS));
                let targets: Vec<ClientEntry> = guard.clients.values().cloned().collect();
                let session = guard.session.as_ref().unwrap();
                let (sid, scol, srow) = (session.id, session.cols, session.rows);
                drop(guard);
                ServerState::notify_clients(&targets, ncols, nrows);
                Spawn::encode_response((sid, scol, srow))
                    .map_err(|e| Box::new(e) as Box<dyn std::error::Error + Send + Sync>)
            }
        })
        .await
        .map_err(|e| format!("register Spawn: {e:?}"))?;

    // Register ResizePty
    let st = Arc::clone(&state);
    endpoint
        .register_prebuffered(ResizePty::METHOD_ID, move |payload, ctx| {
            let state = Arc::clone(&st);
            async move {
                let (_id, cols, rows) = ResizePty::decode_request(&payload)
                    .map_err(|e| Box::new(e) as Box<dyn std::error::Error + Send + Sync>)?;
                let mut guard = state.lock().await;
                if let Some(client) = guard.clients.get_mut(&ctx.conn_id) {
                    client.cols = cols;
                    client.rows = rows;
                }
                guard.recalculate_pty_size();
                let (ncols, nrows) = guard
                    .session
                    .as_ref()
                    .map(|s| (s.cols, s.rows))
                    .unwrap_or((cols, rows));
                let targets: Vec<ClientEntry> = guard.clients.values().cloned().collect();
                drop(guard);
                ServerState::notify_clients(&targets, ncols, nrows);
                ResizePty::encode_response((ncols, nrows))
                    .map_err(|e| Box::new(e) as Box<dyn std::error::Error + Send + Sync>)
            }
        })
        .await
        .map_err(|e| format!("register ResizePty: {e:?}"))?;

    // Register CloseSession
    let st = Arc::clone(&state);
    endpoint
        .register_prebuffered(CloseSession::METHOD_ID, move |payload, _ctx| {
            let state = Arc::clone(&st);
            async move {
                let _id = CloseSession::decode_request(&payload)
                    .map_err(|e| Box::new(e) as Box<dyn std::error::Error + Send + Sync>)?;
                let mut guard = state.lock().await;
                guard.clear_session();
                CloseSession::encode_response(())
                    .map_err(|e| Box::new(e) as Box<dyn std::error::Error + Send + Sync>)
            }
        })
        .await
        .map_err(|e| format!("register CloseSession: {e:?}"))?;

    // Register ListSessions
    let st = Arc::clone(&state);
    endpoint
        .register_prebuffered(ListSessions::METHOD_ID, move |_payload, _ctx| {
            let state = Arc::clone(&st);
            async move {
                let guard = state.lock().await;
                let sessions = match &guard.session {
                    Some(s) => vec![(s.id, String::new(), s.exited)],
                    None => vec![],
                };
                ListSessions::encode_response(sessions)
                    .map_err(|e| Box::new(e) as Box<dyn std::error::Error + Send + Sync>)
            }
        })
        .await
        .map_err(|e| format!("register ListSessions: {e:?}"))?;

    // Register WriteInput
    let st = Arc::clone(&state);
    endpoint
        .register_prebuffered(WriteInput::METHOD_ID, move |payload, _ctx| {
            let state = Arc::clone(&st);
            async move {
                let (id, data) = WriteInput::decode_request(&payload)
                    .map_err(|e| Box::new(e) as Box<dyn std::error::Error + Send + Sync>)?;
                let writer = {
                    let guard = state.lock().await;
                    guard
                        .session
                        .as_ref()
                        .filter(|s| s.id == id)
                        .map(|s| s.pty.writer_handle())
                };
                // PTY writes are blocking I/O (kernel input buffer); offload
                // to the blocking pool so a full buffer never stalls an async
                // worker or holds the state lock.
                if let Some(writer) = writer {
                    let _ = tokio::task::spawn_blocking(move || writer.write_bytes(&data)).await;
                }
                WriteInput::encode_response(())
                    .map_err(|e| Box::new(e) as Box<dyn std::error::Error + Send + Sync>)
            }
        })
        .await
        .map_err(|e| format!("register WriteInput: {e:?}"))?;

    // Register StreamInput (streaming handler for PTY input)
    // The channel persists across client disconnects so reconnecting
    // clients can still send input — we drop it only when the server
    // shuts down.
    // Bounded to 128 items + try_send provides memory safety when the
    // PTY write task falls behind under extreme input bursts.  Dropped
    // chunks may fragment the PTY byte stream (multi-byte sequences);
    // client-side coalescing (in term-session-client) prevents most
    // over-production, but the bound is the last line of defense.
    let (input_tx, mut input_rx) = mpsc::channel::<Vec<u8>>(INPUT_CHANNEL_CAPACITY);
    endpoint
        .register_stream_handler(STREAM_INPUT_METHOD_ID, move |event, _responder, _ctx| {
            if let RpcStreamEvent::PayloadChunk { bytes, .. } = event
                && let Err(e) = input_tx.try_send(bytes)
            {
                tracing::warn!(error = %e, "server input buffer full; dropping input chunk");
            }
            // Intentionally ignore End/Error — the channel stays alive.
        })
        .await
        .map_err(|e| format!("register stream handler STREAM_INPUT: {e:?}"))?;

    // Background task: write received input bytes to the PTY session
    let input_st = Arc::clone(&state);
    tokio::spawn(async move {
        while let Some(data) = input_rx.recv().await {
            let writer = {
                let guard = input_st.lock().await;
                guard.session.as_ref().map(|s| s.pty.writer_handle())
            };
            // PTY writes are blocking I/O (kernel input buffer); offload to
            // the blocking pool so a full buffer never stalls an async worker
            // or holds the state lock. Awaiting per chunk preserves order.
            if let Some(writer) = writer {
                let _ = tokio::task::spawn_blocking(move || writer.write_bytes(&data)).await;
            }
        }
    });

    // Register SubscribeOutput
    let st = Arc::clone(&state);
    endpoint
        .register_stream_handler(SUBSCRIBE_OUTPUT_METHOD_ID, move |event, respond, ctx| {
            let is_new = matches!(&event, RpcStreamEvent::Header { .. });
            if is_new {
                let st = Arc::clone(&st);
                tokio::spawn(async move {
                    let mut guard = st.lock().await;

                    // Drain accumulated PTY output and capture the raw bytes
                    // so they can be sent to the new subscriber (not just the snapshot).
                    let early = guard.session.as_mut().and_then(|s| {
                        let data = s.read_output();
                        if data.is_empty() { None } else { Some(data) }
                    });
                    let snapshot = guard.session.as_mut().map(|s| s.generate_snapshot());
                    guard.subscribers.push(SubscriberEntry {
                        conn_id: ctx.conn_id,
                        respond: respond.clone(),
                    });

                    // Wake the polling loop — the session may have pending
                    // output or exit state that needs processing.
                    guard.notify.notify_one();
                    let is_dead = guard.session.is_none();
                    drop(guard);
                    if let Some(data) = snapshot
                        && !data.is_empty()
                    {
                        respond.respond(data, false);
                    }
                    if let Some(data) = early {
                        respond.respond(data, false);
                    }
                    if is_dead {
                        respond.respond(Vec::new(), true);
                    }
                });
            }
        })
        .await
        .map_err(|e| format!("register SubscribeOutput: {e:?}"))?;

    // Connection event handler
    let st = Arc::clone(&state);
    tokio::spawn(async move {
        while let Some(event) = event_rx.recv().await {
            match event {
                RpcIpcServerEvent::ClientConnected(handle) => {
                    tracing::info!("Client {} connected", handle.0.conn_id);
                    let mut guard = st.lock().await;
                    let handle_clone = handle.clone();
                    guard.clients.insert(
                        handle.0.conn_id,
                        ClientEntry {
                            caller: Some(handle_clone),
                            cols: u16::MAX,
                            rows: u16::MAX,
                        },
                    );
                }
                RpcIpcServerEvent::ClientDisconnected(conn_id) => {
                    tracing::info!("Client {conn_id} disconnected");
                    let mut guard = st.lock().await;
                    guard.clients.remove(&conn_id);
                    guard.subscribers.retain(|s| s.conn_id != conn_id);
                    guard.recalculate_pty_size();
                    let (ncols, nrows) = guard
                        .session
                        .as_ref()
                        .map(|s| (s.cols, s.rows))
                        .unwrap_or((FALLBACK_COLS, FALLBACK_ROWS));
                    let targets: Vec<ClientEntry> = guard.clients.values().cloned().collect();
                    drop(guard);
                    ServerState::notify_clients(&targets, ncols, nrows);
                }
            }
        }
    });

    // Output polling via Notify — blocks until PTY produces output.
    // When the session exits, the exit code is sent back through this
    // channel so run_server can return it.
    //
    // A periodic timer also wakes the loop so a session exit is detected even
    // when the reader thread's EOF/Exited notification is missed or raced.
    let (exit_tx, mut exit_rx) = oneshot::channel::<i32>();
    let st = Arc::clone(&state);
    tokio::spawn(async move {
        loop {
            tokio::select! {
                _ = notify.notified() => {}
                _ = tokio::time::sleep(SESSION_EXIT_POLL_INTERVAL) => {}
            }
            let mut guard = st.lock().await;
            if guard.subscribers.is_empty() {
                let mut exited = false;
                if let Some(session) = guard.session.as_mut() {
                    session.sync_screen();
                    exited = session.check_exited();
                }
                if exited {
                    tracing::info!("Session exited, tearing down");
                    guard.session = None;
                }
                continue;
            }
            let (raw, exited, code) = {
                let Some(session) = guard.session.as_mut() else {
                    let _ = exit_tx.send(0);
                    break;
                };
                let raw = session.read_output();
                let exited = session.check_exited();
                let code = session.exit_code;
                (raw, exited, code)
            };
            if raw.is_empty() && !guard.subscribers.is_empty() {
                tracing::debug!(
                    "PTY output empty with {} subscribers",
                    guard.subscribers.len()
                );
            }

            // Push raw PTY output to all subscribers
            if !raw.is_empty() {
                for sub in &guard.subscribers {
                    sub.respond.respond(raw.clone(), false);
                }
            }

            // On exit: finalize all streams and clean up
            if exited {
                for sub in &guard.subscribers {
                    sub.respond.respond(Vec::new(), true);
                }
                guard.subscribers.clear();
                let _ = exit_tx.send(code.unwrap_or(0));
                tracing::info!("Session exited with code {:?}", code);
                break;
            }
        }
    });

    tracing::info!("Session server listening on channel {}", config.channel);

    // Wait for either the server to finish or the session to exit.
    let exit_code = tokio::select! {
        result = async {
            server
                .serve(&socket_name)
                .await
                .map_err(|e| format!("serve: {e:?}"))
        } => {
            result?;
            0
        }
        code = &mut exit_rx => {
            // The polling task just queued the end-of-stream frames to each
            // subscriber. The transport flushes them asynchronously, so give
            // it a short grace period before the process exits — otherwise the
            // frames are dropped with the runtime and the client never learns
            // the session ended (it hangs waiting for output).
            tokio::time::sleep(SESSION_EXIT_FLUSH_GRACE).await;
            code.unwrap_or(0)
        }
    };

    Ok(exit_code)
}