rigtop_core/

remote.rs

//! A [`DataSource`] for a remote host, reached by running `rigtop --agent`
//! through the system's `ssh` (OpenSSH; `ssh.exe` on Windows).
//!
//! Using the system client means `~/.ssh/config`, keys, agents and
//! `ProxyJump` all work as on the command line. Authentication must not
//! need a prompt (`BatchMode`), since there is no terminal to ask on.

use std::collections::VecDeque;
use std::io::{BufRead, BufReader};
use std::process::{Child, ChildStdin, Command, Stdio};
use std::sync::mpsc::{self, Receiver, RecvTimeoutError};
use std::sync::{Arc, Mutex};
use std::thread;
use std::time::{Duration, Instant};

use crate::agent::{AgentMessage, ClientMessage, PROTOCOL_VERSION, write_message};
use crate::snapshot::Snapshot;
use crate::source::{DataSource, ProcessSignal, SourceError};

/// Command run on the remote host by default.
pub const DEFAULT_REMOTE_COMMAND: &str = "rigtop --agent";
/// Tried when the default command is not found: `cargo install` puts rigtop
/// in `~/.cargo/bin`, which non-interactive SSH sessions often lack on
/// their `PATH`.
pub const FALLBACK_REMOTE_COMMAND: &str = "~/.cargo/bin/rigtop --agent";
/// Exit status of POSIX shells for "command not found".
const NOT_FOUND_STATUS: i32 = 127;

/// Minimum time between connection attempts after a failure.
const RECONNECT_DELAY: Duration = Duration::from_secs(5);
/// How long to wait for the first snapshot of a new connection (SSH
/// handshake plus the agent's warm-up sample).
const CONNECT_TIMEOUT: Duration = Duration::from_secs(20);
/// How long to wait for the agent to answer a signal request.
const SIGNAL_TIMEOUT: Duration = Duration::from_secs(10);
/// stderr lines kept to explain failures.
const STDERR_LINES: usize = 20;

/// How to start the agent: a program and its arguments.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct RemoteTarget {
    /// Short name for messages (e.g. the SSH destination).
    pub name: String,
    /// Program to run (normally `ssh`).
    pub program: String,
    /// Its arguments.
    pub args: Vec<String>,
    /// Arguments to use instead when the remote command is not found (for
    /// the default command: [`FALLBACK_REMOTE_COMMAND`]).
    pub fallback_args: Option<Vec<String>>,
}

impl RemoteTarget {
    /// Runs `remote_command` on `destination` (anything `ssh` accepts:
    /// a `~/.ssh/config` host, `user@host`, ...).
    pub fn ssh(destination: &str, remote_command: &str) -> Self {
        let options = [
            // No password or host key prompts: fail instead of hanging.
            "BatchMode=yes",
            "ConnectTimeout=10",
            // Notice dead connections within ~45 s.
            "ServerAliveInterval=15",
            "ServerAliveCountMax=3",
        ];
        let mut args = vec!["-T".to_owned()];
        for option in options {
            args.push("-o".to_owned());
            args.push(option.to_owned());
        }
        // `--` keeps a destination starting with "-" from being an option.
        args.push("--".to_owned());
        args.push(destination.to_owned());
        args.push(remote_command.to_owned());
        // Only the default command falls back; one set by the user is
        // taken as it is.
        let fallback_args = (remote_command == DEFAULT_REMOTE_COMMAND).then(|| {
            let mut fallback = args.clone();
            *fallback.last_mut().expect("command pushed above") =
                FALLBACK_REMOTE_COMMAND.to_owned();
            fallback
        });
        Self {
            name: destination.to_owned(),
            program: "ssh".to_owned(),
            args,
            fallback_args,
        }
    }
}

/// What the reader thread passes on.
enum Incoming {
    Message(AgentMessage),
    /// stdout closed: the agent or ssh exited.
    Closed,
}

/// One running agent process.
struct Connection {
    child: Child,
    stdin: ChildStdin,
    incoming: Receiver<Incoming>,
    /// Last lines of the process's stderr (ssh errors, "command not found").
    stderr: Arc<Mutex<VecDeque<String>>>,
    /// Whether a snapshot has arrived on this connection yet.
    established: bool,
}

impl Connection {
    fn spawn(target: &RemoteTarget, args: &[String]) -> Result<Self, SourceError> {
        let mut command = Command::new(&target.program);
        command
            .args(args)
            .stdin(Stdio::piped())
            .stdout(Stdio::piped())
            .stderr(Stdio::piped());
        #[cfg(windows)]
        {
            use std::os::windows::process::CommandExt;
            // CREATE_NO_WINDOW: a GUI must not flash a console for ssh.exe.
            command.creation_flags(0x0800_0000);
        }
        let mut child = command.spawn().map_err(|err| {
            SourceError::Disconnected(format!("cannot start {}: {err}", target.program))
        })?;
        let stdin = child.stdin.take().expect("piped stdin");
        let stdout = child.stdout.take().expect("piped stdout");
        let stderr_pipe = child.stderr.take().expect("piped stderr");

        let (tx, incoming) = mpsc::channel();
        thread::Builder::new()
            .name("rigtop-remote-reader".to_owned())
            .spawn(move || {
                for line in BufReader::new(stdout).lines() {
                    let Ok(line) = line else { break };
                    // Lines that are not protocol messages (e.g. a login
                    // banner printed by a shell profile) are skipped.
                    if let Ok(message) = serde_json::from_str::<AgentMessage>(&line)
                        && tx.send(Incoming::Message(message)).is_err()
                    {
                        return;
                    }
                }
                let _ = tx.send(Incoming::Closed);
            })
            .map_err(SourceError::Io)?;

        let stderr = Arc::new(Mutex::new(VecDeque::new()));
        let lines = Arc::clone(&stderr);
        thread::Builder::new()
            .name("rigtop-remote-stderr".to_owned())
            .spawn(move || {
                for line in BufReader::new(stderr_pipe).lines() {
                    let Ok(line) = line else { break };
                    let mut lines = lines.lock().unwrap_or_else(|e| e.into_inner());
                    if lines.len() == STDERR_LINES {
                        lines.pop_front();
                    }
                    lines.push_back(line);
                }
            })
            .map_err(SourceError::Io)?;

        Ok(Self {
            child,
            stdin,
            incoming,
            stderr,
            established: false,
        })
    }

    fn send(&mut self, message: &ClientMessage) -> Result<(), SourceError> {
        write_message(&mut self.stdin, message).map_err(SourceError::Io)
    }

    /// Explanation for a lost connection, from the process's stderr, and
    /// whether the remote command was not found.
    fn failure(&mut self) -> (String, bool) {
        // Give the stderr thread a moment to collect the final lines.
        thread::sleep(Duration::from_millis(100));
        let status = self.child.try_wait().ok().flatten();
        let lines = self.stderr.lock().unwrap_or_else(|e| e.into_inner());
        let detail = lines
            .iter()
            .rev()
            .find(|l| !l.trim().is_empty())
            .cloned()
            .unwrap_or_default();
        // ssh passes the remote exit status on; cmd.exe says "not recognized".
        let not_found = status.and_then(|s| s.code()) == Some(NOT_FOUND_STATUS)
            || detail.contains("command not found")
            || detail.contains("not recognized");
        let reason = match (detail.is_empty(), status) {
            (false, _) => detail,
            (true, Some(status)) => format!("connection closed ({status})"),
            (true, None) => "connection closed".to_owned(),
        };
        (reason, not_found)
    }
}

impl Drop for Connection {
    fn drop(&mut self) {
        // Closing stdin makes the agent exit; killing ssh ends the session
        // right away instead of waiting for it.
        let _ = self.child.kill();
        let _ = self.child.wait();
    }
}

/// Snapshots from `rigtop --agent` on another machine.
pub struct RemoteSource {
    target: RemoteTarget,
    connection: Option<Connection>,
    last_attempt: Option<Instant>,
    interval: Duration,
    process_limit: Option<usize>,
    /// Snapshots received while waiting for something else (signal results).
    pending: VecDeque<Snapshot>,
    /// Why the last connection failed, repeated while waiting to retry.
    last_error: Option<String>,
    /// Whether to start the agent with the fallback command.
    use_fallback: bool,
    /// The command was not found and the fallback should be tried at once.
    retry_now: bool,
}

impl RemoteSource {
    /// Creates a source; the connection is made on the first poll.
    pub fn new(target: RemoteTarget) -> Self {
        Self {
            target,
            connection: None,
            last_attempt: None,
            interval: Duration::from_secs(1),
            process_limit: None,
            pending: VecDeque::new(),
            last_error: None,
            use_fallback: false,
            retry_now: false,
        }
    }

    /// The target this source connects to.
    pub fn target(&self) -> &RemoteTarget {
        &self.target
    }

    /// Returns the connection, (re)connecting if needed and allowed.
    fn connection(&mut self) -> Result<&mut Connection, SourceError> {
        if self.connection.is_none() {
            if let Some(at) = self.last_attempt {
                let wait = RECONNECT_DELAY.saturating_sub(at.elapsed());
                if !wait.is_zero() {
                    let reason = self.last_error.as_deref().unwrap_or("not connected");
                    return Err(SourceError::Disconnected(format!(
                        "{}: {reason} (retrying in {}s)",
                        self.target.name,
                        wait.as_secs() + 1
                    )));
                }
            }
            self.last_attempt = Some(Instant::now());
            let args = match (&self.target.fallback_args, self.use_fallback) {
                (Some(fallback), true) => fallback,
                _ => &self.target.args,
            };
            let mut connection = match Connection::spawn(&self.target, args) {
                Ok(connection) => connection,
                Err(err) => {
                    self.last_error = Some(match &err {
                        SourceError::Disconnected(reason) => reason.clone(),
                        other => other.to_string(),
                    });
                    return Err(err);
                }
            };
            // Agents start with their own default; send our settings.
            let sent = connection
                .send(&ClientMessage::SetInterval {
                    ms: self.interval.as_millis() as u64,
                })
                .and_then(|()| {
                    connection.send(&ClientMessage::SetProcessLimit {
                        limit: self.process_limit,
                    })
                });
            self.connection = Some(connection);
            if sent.is_err() {
                // ssh already exited (bad host, auth failure): its stderr
                // says why, which beats "broken pipe".
                return Err(self.disconnect(None));
            }
        }
        Ok(self.connection.as_mut().expect("connected above"))
    }

    /// Drops the connection and returns an error explaining why.
    fn disconnect(&mut self, reason: Option<String>) -> SourceError {
        let (mut reason, not_found) = match (reason, self.connection.as_mut()) {
            (Some(reason), _) => (reason, false),
            (None, Some(connection)) => connection.failure(),
            (None, None) => ("connection closed".to_owned(), false),
        };
        self.connection = None;
        if not_found && self.target.fallback_args.is_some() {
            if self.use_fallback {
                // Neither worked; start over with the PATH next time, in
                // case rigtop gets installed there.
                self.use_fallback = false;
                reason = format!("rigtop is neither on the PATH nor in ~/.cargo/bin ({reason})");
            } else {
                self.use_fallback = true;
                self.retry_now = true;
                self.last_attempt = None;
            }
        }
        let error = SourceError::Disconnected(format!("{}: {reason}", self.target.name));
        self.last_error = Some(reason);
        error
    }

    /// Waits up to `timeout` for the next message.
    fn next_message(&mut self, timeout: Duration) -> Result<Option<AgentMessage>, SourceError> {
        let connection = self.connection()?;
        match connection.incoming.recv_timeout(timeout) {
            Ok(Incoming::Message(message)) => Ok(Some(message)),
            Ok(Incoming::Closed) | Err(RecvTimeoutError::Disconnected) => {
                Err(self.disconnect(None))
            }
            Err(RecvTimeoutError::Timeout) => Ok(None),
        }
    }

    /// Handles messages that are not what the caller waits for. Returns an
    /// error for fatal ones (protocol mismatch).
    fn absorb(&mut self, message: AgentMessage) -> Result<(), SourceError> {
        match message {
            AgentMessage::Hello { protocol, version } if protocol != PROTOCOL_VERSION => {
                let reason = format!(
                    "agent speaks protocol {protocol} (rigtop {version}), this client speaks {PROTOCOL_VERSION}; \
                     install the same rigtop version on both sides"
                );
                // Do not retry quickly against an incompatible agent.
                self.last_attempt = Some(Instant::now());
                Err(self.disconnect(Some(reason)))
            }
            AgentMessage::Hello { .. } | AgentMessage::SignalResult { .. } => Ok(()),
            AgentMessage::Snapshot { snapshot } => {
                if let Some(connection) = self.connection.as_mut() {
                    connection.established = true;
                }
                self.last_error = None;
                self.pending.push_back(*snapshot);
                Ok(())
            }
            AgentMessage::Error { message } => Err(SourceError::Remote(message)),
        }
    }
}

impl RemoteSource {
    /// One attempt at [`DataSource::poll`].
    fn poll_once(&mut self) -> Result<Snapshot, SourceError> {
        // A new connection needs time for SSH and the agent's warm-up.
        let established = self.connection.as_ref().is_some_and(|c| c.established);
        let timeout = if established {
            self.interval * 3 + Duration::from_secs(5)
        } else {
            CONNECT_TIMEOUT
        };
        let deadline = Instant::now() + timeout;
        loop {
            // Drain everything already received; keep only the newest snapshot.
            while let Some(message) = self.next_message(Duration::ZERO)? {
                self.absorb(message)?;
            }
            if let Some(newest) = self.pending.pop_back() {
                self.pending.clear();
                return Ok(newest);
            }
            let remaining = deadline.saturating_duration_since(Instant::now());
            if remaining.is_zero() {
                return Err(self.disconnect(Some("no data from the agent (timed out)".to_owned())));
            }
            if let Some(message) = self.next_message(remaining)? {
                self.absorb(message)?;
            }
        }
    }
}

impl DataSource for RemoteSource {
    /// Returns the newest snapshot from the agent, waiting for one if none
    /// arrived since the previous poll.
    fn poll(&mut self) -> Result<Snapshot, SourceError> {
        match self.poll_once() {
            // Not found on the PATH: try ~/.cargo/bin right away.
            Err(_) if std::mem::take(&mut self.retry_now) => self.poll_once(),
            result => result,
        }
    }

    fn signal_process(&mut self, pid: u32, signal: ProcessSignal) -> Result<(), SourceError> {
        if self
            .connection()?
            .send(&ClientMessage::Signal { pid, signal })
            .is_err()
        {
            return Err(self.disconnect(None));
        }
        let deadline = Instant::now() + SIGNAL_TIMEOUT;
        loop {
            let remaining = deadline.saturating_duration_since(Instant::now());
            if remaining.is_zero() {
                return Err(SourceError::Remote(format!(
                    "no answer to the signal for {pid}"
                )));
            }
            match self.next_message(remaining)? {
                Some(AgentMessage::SignalResult {
                    pid: answered,
                    error,
                    ..
                }) if answered == pid => {
                    return match error {
                        None => Ok(()),
                        Some(error) => Err(SourceError::Remote(error)),
                    };
                }
                Some(other) => self.absorb(other)?,
                None => {}
            }
        }
    }

    fn set_interval(&mut self, interval: Duration) {
        self.interval = interval;
        if let Some(connection) = self.connection.as_mut() {
            let _ = connection.send(&ClientMessage::SetInterval {
                ms: interval.as_millis() as u64,
            });
        }
    }

    fn set_process_limit(&mut self, limit: Option<usize>) {
        self.process_limit = limit;
        if let Some(connection) = self.connection.as_mut() {
            let _ = connection.send(&ClientMessage::SetProcessLimit { limit });
        }
    }
}

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

    #[test]
    fn ssh_target_uses_batch_mode_and_keeps_the_destination_safe() {
        let target = RemoteTarget::ssh("zefat", DEFAULT_REMOTE_COMMAND);
        assert_eq!(target.program, "ssh");
        assert_eq!(target.name, "zefat");
        assert!(target.args.windows(2).any(|w| w == ["-o", "BatchMode=yes"]));
        let n = target.args.len();
        assert_eq!(&target.args[n - 3..], ["--", "zefat", "rigtop --agent"]);
        // The default command falls back to ~/.cargo/bin ...
        let fallback = target.fallback_args.as_ref().unwrap();
        assert_eq!(&fallback[n - 3..], ["--", "zefat", FALLBACK_REMOTE_COMMAND]);
        // ... a command the user chose does not.
        let custom = RemoteTarget::ssh("zefat", "/opt/rigtop/bin/rigtop --agent");
        assert_eq!(custom.fallback_args, None);
    }

    /// `sh -c <script>` standing in for ssh.
    #[cfg(unix)]
    fn sh_target(script: &str, fallback: Option<&str>) -> RemoteTarget {
        let args = |script: &str| vec!["-c".to_owned(), script.to_owned()];
        RemoteTarget {
            name: "fake".to_owned(),
            program: "sh".to_owned(),
            args: args(script),
            fallback_args: fallback.map(args),
        }
    }

    #[cfg(unix)]
    #[test]
    fn command_not_found_tries_the_fallback_at_once() {
        // Found nowhere: both attempts in one poll, and a clear message.
        let missing = "echo 'zsh:1: command not found: rigtop' >&2; exit 127";
        let mut source = RemoteSource::new(sh_target(missing, Some(missing)));
        let err = source.poll().unwrap_err().to_string();
        assert!(
            err.contains("neither on the PATH nor in ~/.cargo/bin"),
            "{err}"
        );
        assert!(!source.use_fallback, "the PATH comes first again next time");

        // Other failures do not switch to the fallback.
        let mut source = RemoteSource::new(sh_target(
            "echo 'Permission denied (publickey)' >&2; exit 255",
            Some(missing),
        ));
        let err = source.poll().unwrap_err().to_string();
        assert!(err.contains("Permission denied"), "{err}");
        assert!(!source.use_fallback);
    }

    #[test]
    fn missing_program_is_reported_and_retried_later() {
        let mut source = RemoteSource::new(RemoteTarget {
            name: "nowhere".to_owned(),
            program: "/nonexistent/ssh".to_owned(),
            args: Vec::new(),
            fallback_args: None,
        });
        let err = source.poll().unwrap_err().to_string();
        assert!(err.contains("cannot start"), "{err}");
        // The next attempt waits for the reconnect delay, repeating the cause.
        let err = source.poll().unwrap_err().to_string();
        assert!(
            err.contains("retrying") && err.contains("cannot start"),
            "{err}"
        );
    }

    #[cfg(unix)]
    #[test]
    fn stderr_explains_an_agent_that_exits() {
        // Stands in for ssh failing, e.g. "rigtop: command not found".
        let mut source = RemoteSource::new(RemoteTarget {
            name: "fake".to_owned(),
            program: "sh".to_owned(),
            args: vec![
                "-c".to_owned(),
                "echo 'rigtop: command not found' >&2; exit 127".to_owned(),
            ],
            fallback_args: None,
        });
        let err = source.poll().unwrap_err().to_string();
        assert!(err.contains("fake: rigtop: command not found"), "{err}");
    }

    #[cfg(unix)]
    #[test]
    fn incompatible_protocol_is_rejected() {
        let mut source = RemoteSource::new(RemoteTarget {
            name: "old".to_owned(),
            program: "sh".to_owned(),
            args: vec![
                "-c".to_owned(),
                r#"echo '{"type":"hello","protocol":999,"version":"9.9.9"}'; sleep 5"#.to_owned(),
            ],
            fallback_args: None,
        });
        let err = source.poll().unwrap_err().to_string();
        assert!(err.contains("protocol 999"), "{err}");
    }
}