//! Proxy one GET to a peer. //! //! REST and MCP both call [`get`]. The allowlist is the peer's documented //! read routes; `ping` and `peers` are excluded so this cannot chain or //! become an open proxy. Caller credentials are never copied: the transport //! attaches the peer token itself. //! //! Callers may name a route with or without the `/api/v1` prefix; validation //! works on the route alone (`/systemd/units`) and forwards the canonical peer //! path (`/api/v1/systemd/units`), which is what a peer serves. use std::time::Duration; use tokio::time::Instant; use axum::http::{HeaderMap, HeaderName, HeaderValue, StatusCode}; use bytes::Bytes; use futures::stream::{BoxStream, Stream, StreamExt}; use tokio_util::sync::CancellationToken; use super::error::{peer_error_to_app, PeerError}; use super::registry::ProxyOpts; use super::transport::{PeerBody, PeerRequest, PeerResponse, PeerTransport}; use super::types::{PeerConfig, PeerName}; use crate::error::AppError; use crate::sources::systemd::validate_unit_name; const HOP_HEADER: &str = "x-slinuxm-hop"; const TRACEPARENT: &str = "traceparent"; /// Response header naming the peer a proxied response came from. pub(crate) const PEER_HEADER: &str = "x-slinuxm-peer"; /// Reject an inbound hop marker. Any value counts: a consolidator must not /// forward a request that has already been proxied. pub(crate) fn reject_inbound_hop(peer: &str, headers: &HeaderMap) -> Result<(), AppError> { if headers.contains_key(HOP_HEADER) { return Err(chained(peer)); } Ok(()) } fn chained(peer: &str) -> AppError { AppError::BadRequest(format!("peer {peer}: chained requests are not allowed")) } fn path_rejected(peer: &PeerName) -> AppError { AppError::BadRequest(format!("peer {peer}: path not allowed")) } fn query_rejected(peer: &str) -> AppError { AppError::BadRequest(format!("peer {peer}: unsupported query parameter")) } /// Parse a raw query string. Bad encodings are rejected before the pairs are /// handed to [`get`], which checks the allowlist again. pub(crate) fn query_from_raw( peer: &str, raw: Option<&str>, ) -> Result, AppError> { let Some(raw) = raw.filter(|s| !s.is_empty()) else { return Ok(Vec::new()); }; if raw.len() > 4096 || encoding_forbidden(raw) || raw.chars().any(|c| c.is_control()) { return Err(query_rejected(peer)); } Ok(url::form_urlencoded::parse(raw.as_bytes()) .map(|(k, v)| (k.into_owned(), v.into_owned())) .collect()) } fn header(name: &'static str, value: &str) -> (HeaderName, HeaderValue) { let name = HeaderName::from_static(name); let value = HeaderValue::from_str(value).expect("static header value"); (name, value) } pub(crate) async fn get( transport: &dyn PeerTransport, peer: &PeerConfig, path: &str, query: &[(String, String)], opts: ProxyOpts, timeout: Duration, ) -> Result { let Validated { path, query, sse } = validate(&peer.name, path, query, opts.accept_sse)?; // Marks the hop so a peer refuses chained consolidation. let mut headers = vec![header(HOP_HEADER, "1")]; if let Some(traceparent) = opts.traceparent.as_deref() { if let Ok(value) = HeaderValue::from_str(traceparent) { if !value.as_bytes().contains(&b' ') { headers.push((HeaderName::from_static(TRACEPARENT), value)); } } } let req = PeerRequest { path, query, headers, accept: sse.then(|| "text/event-stream".to_owned()), }; let result = tokio::time::timeout(timeout, transport.get(peer, req, timeout)).await; match result { Ok(Ok(resp)) => Ok(resp), Ok(Err(e)) => Err(peer_error_to_app(&peer.name, &e)), Err(_elapsed) => Err(peer_error_to_app(&peer.name, &PeerError::Timeout)), } } /// A peer response shaped for a surface: status already classified, the /// peer's content type checked, body bounded. REST and MCP both build their /// response from this so neither owns a second copy of the limits. pub(crate) struct PreparedResponse { pub status: StatusCode, pub content_type: Option, pub body: PreparedBody, } /// One table for peer statuses: the transport classifies a live response with /// it and [`prepare_response`] checks again, so neither side can drift. pub(crate) fn classify_status(status: u16) -> PeerError { match status { 404 => PeerError::NotFound, 401 | 403 => PeerError::Rejected, code => PeerError::Http(code), } } /// The contract promises pass-through only for 2xx, so anything else is /// classified here rather than relayed: a transport that forgot to map a /// redirect or the peer's own error body must not reach the caller. pub(crate) fn prepare_response( peer: &PeerName, resp: PeerResponse, max_bytes: u64, deadline: Instant, shutdown: CancellationToken, ) -> Result { let status = match StatusCode::from_u16(resp.status) { Ok(status) if status.is_success() => status, _ => return Err(peer_error_to_app(peer, &classify_status(resp.status))), }; let content_type = content_type(peer, resp.content_type.as_deref())?; let body = prepare_body(peer, resp.body, max_bytes, deadline, shutdown)?; Ok(PreparedResponse { status, content_type, body, }) } fn content_type(peer: &PeerName, value: Option<&str>) -> Result, AppError> { match value { None => Ok(None), Some(value) => HeaderValue::from_str(value) .map(Some) .map_err(|_| peer_error_to_app(peer, &PeerError::Malformed(String::new()))), } } /// Buffered bodies over the cap become `Oversize`. Streams are cut off at the /// cap, at `deadline`, or when `shutdown` fires, because a follow holds a /// connection until the peer closes it. pub(crate) fn prepare_body( peer: &PeerName, body: PeerBody, max_bytes: u64, deadline: Instant, shutdown: CancellationToken, ) -> Result { match body { PeerBody::Bytes(bytes) if bytes.len() as u64 > max_bytes => { Err(peer_error_to_app(peer, &PeerError::Oversize)) } PeerBody::Bytes(bytes) => Ok(PreparedBody::Bytes(bytes)), PeerBody::Stream(stream) => Ok(PreparedBody::Stream(Box::pin(bound_stream( stream, max_bytes, deadline, shutdown, )))), } } pub(crate) enum PreparedBody { Bytes(Bytes), Stream(BoxStream<'static, Result>), } fn bound_stream( stream: BoxStream<'static, Result>, max_bytes: u64, deadline: Instant, shutdown: CancellationToken, ) -> impl Stream> + Send { futures::stream::unfold((stream, 0u64), move |(mut stream, sent)| { let shutdown = shutdown.clone(); async move { let next = tokio::select! { biased; () = shutdown.cancelled() => return None, () = tokio::time::sleep_until(deadline) => return None, item = stream.next() => item, }; let bytes = match next? { Ok(bytes) => bytes, // Headers are already sent. Dropping the chunk keeps a // token-file path or peer URL out of the body. Err(_) => return None, }; if sent.saturating_add(bytes.len() as u64) > max_bytes { return None; } let sent = sent + bytes.len() as u64; Some((Ok(bytes), (stream, sent))) } }) } struct Validated { path: String, query: Vec<(String, String)>, sse: bool, } fn validate( peer: &PeerName, path: &str, query: &[(String, String)], accept_sse: bool, ) -> Result { let route = canonical_route(path).ok_or_else(|| path_rejected(peer))?; let query = allow_query(peer, &route, query)?; let sse = accept_sse || route_is_follow(&route, &query); Ok(Validated { // The allowlist and the query rules work on the route; the peer only // serves it below its own API prefix. path: format!("/api/v1/{route}"), query, sse, }) } fn canonical_route(path: &str) -> Option { if path.len() > 512 || encoding_forbidden(path) { return None; } let mut rel = path.trim_start_matches('/'); if let Some(rest) = rel.strip_prefix("api/v1/") { rel = rest; } if rel.is_empty() { return None; } let mut decoded = Vec::new(); let mut forward = Vec::new(); for raw in rel.split('/') { let segment = percent_encoding::percent_decode_str(raw) .decode_utf8() .ok()?; if !segment_ok(&segment) { return None; } forward.push(percent_encoding::utf8_percent_encode(&segment, SEGMENT_SET).to_string()); decoded.push(segment.into_owned()); } if !route_allowed(&decoded.join("/")) { return None; } // The forward path is re-encoded rather than the caller's text reused so // a backslash from an escaped unit name cannot be read as a separator by // the HTTP client. Some(forward.join("/")) } /// Characters that must stay percent-encoded in a forwarded segment. const SEGMENT_SET: &percent_encoding::AsciiSet = &percent_encoding::CONTROLS .add(b' ') .add(b'"') .add(b'#') .add(b'%') .add(b'/') .add(b'<') .add(b'>') .add(b'?') .add(b'[') .add(b'\\') .add(b']') .add(b'^') .add(b'`') .add(b'{') .add(b'|') .add(b'}'); /// Segment-level rules only; the allowlist below decides which characters a /// segment may carry for its route. fn segment_ok(segment: &str) -> bool { !segment.is_empty() && segment != "." && segment != ".." && !segment .chars() .any(|c| c.is_control() || matches!(c, '%' | '?' | '#' | '/')) } /// Encoded slash, dot-dot, NUL and CRLF, including a second layer of encoding /// (`%252f`, `%255c`). A single `%5c` is not listed: systemd escapes names /// such as `dev-disk-by\x2did.device`, and the local route accepts them, so /// the allowlist decides where a backslash is legal. fn encoding_forbidden(raw: &str) -> bool { let lower = raw.to_ascii_lowercase(); ["%2f", "%00", "%0a", "%0d", "%2e%2e", "%252f", "%255c"] .iter() .any(|needle| lower.contains(needle)) } fn route_allowed(route: &str) -> bool { matches!( route, "systemd/units" | "systemd/timers" | "journal" | "journal/follow" | "docker/ps" ) || unit_path(route) || container_path(route) } /// Unit names go through the same validation as the local route. fn unit_path(route: &str) -> bool { route .strip_prefix("systemd/units/") .is_some_and(|name| validate_unit_name(name).is_ok()) } fn container_path(route: &str) -> bool { let Some(rest) = route.strip_prefix("docker/containers/") else { return false; }; match rest.strip_suffix("/logs") { Some(name) => resource_segment(name), None => resource_segment(rest), } } fn resource_segment(name: &str) -> bool { let mut chars = name.chars(); let Some(first) = chars.next() else { return false; }; first.is_ascii_alphanumeric() && name .chars() .all(|c| c.is_ascii_alphanumeric() || matches!(c, '.' | '_' | '-' | '@' | ':')) } fn allowed_keys(route: &str) -> &'static [&'static str] { match route { "systemd/units" => &["state"], "journal" | "journal/follow" => &["unit", "priority", "lines", "since"], "docker/ps" => &["all"], logs if logs.starts_with("docker/containers/") && logs.ends_with("/logs") => { &["tail", "since", "follow"] } _ => &[], } } fn allow_query( peer: &PeerName, route: &str, query: &[(String, String)], ) -> Result, AppError> { if query.len() > 8 { return Err(query_rejected(peer.as_str())); } let allowed = allowed_keys(route); let mut seen = Vec::new(); let mut out = Vec::with_capacity(query.len()); for (key, value) in query { let known = allowed.iter().any(|name| *name == key); if !known || !query_key_ok(key) || !query_value_ok(key, value) || seen.contains(&key.as_str()) { return Err(query_rejected(peer.as_str())); } seen.push(key.as_str()); out.push((key.clone(), value.clone())); } Ok(out) } fn query_key_ok(key: &str) -> bool { !key.is_empty() && key.len() <= 32 && key .chars() .all(|c| c.is_ascii_lowercase() || c.is_ascii_digit() || c == '_') } /// Values arrive decoded and leave as encoded query pairs, so only control /// characters and a second encoding layer are unsafe here. The journal `unit` /// filter is a unit name, so it gets the local route's unit check verbatim: /// that is where a backslash (systemd's escape) is legal. fn query_value_ok(key: &str, value: &str) -> bool { if value.len() > 1024 || encoding_forbidden(value) { return false; } if key == "unit" { return validate_unit_name(value).is_ok(); } !value .chars() .any(|c| c.is_control() || c == '\\' || c == '/') && !value.contains("..") } fn route_is_follow(route: &str, query: &[(String, String)]) -> bool { if route == "journal/follow" { return true; } route.ends_with("/logs") && query .iter() .any(|(key, value)| key == "follow" && matches!(value.as_str(), "1" | "true")) } #[cfg(test)] #[path = "proxy_tests.rs"] mod tests;