Implement a second HTTP server within compute_ctl (#10574)

The compute_ctl HTTP server has the following purposes:

- Allow management via the control plane
- Provide an endpoint for scaping metrics
- Provide APIs for compute internal clients
  - Neon Postgres extension for installing remote extensions
  - local_proxy for installing extensions and adding grants

The first two purposes require the HTTP server to be available outside
the compute.

The Neon threat model is a bad actor within our internal network. We
need to reduce the surface area of attack. By exposing unnecessary
unauthenticated HTTP endpoints to the internal network, we increase the
surface area of attack. For endpoints described in the third bullet
point, we can just run an extra HTTP server, which is only bound to the
loopback interface since all consumers of those endpoints are within the
compute.
This commit is contained in:
Tristan Partin
2025-02-11 12:02:22 -06:00
committed by GitHub
parent f7b2293317
commit da9c101939
16 changed files with 310 additions and 167 deletions

View File

@@ -48,6 +48,7 @@ use anyhow::{Context, Result};
use chrono::Utc;
use clap::Parser;
use compute_tools::disk_quota::set_disk_quota;
use compute_tools::http::server::Server;
use compute_tools::lsn_lease::launch_lsn_lease_bg_task_for_static;
use signal_hook::consts::{SIGQUIT, SIGTERM};
use signal_hook::{consts::SIGINT, iterator::Signals};
@@ -62,7 +63,6 @@ use compute_tools::compute::{
};
use compute_tools::configurator::launch_configurator;
use compute_tools::extension_server::get_pg_version_string;
use compute_tools::http::launch_http_server;
use compute_tools::logger::*;
use compute_tools::monitor::launch_monitor;
use compute_tools::params::*;
@@ -108,8 +108,20 @@ struct Cli {
#[arg(short = 'r', long, value_parser = parse_remote_ext_config)]
pub remote_ext_config: Option<String>,
#[arg(long, default_value_t = 3080)]
pub http_port: u16,
/// The port to bind the external listening HTTP server to. Clients running
/// outside the compute will talk to the compute through this port. Keep
/// the previous name for this argument around for a smoother release
/// with the control plane.
///
/// TODO: Remove the alias after the control plane release which teaches the
/// control plane about the renamed argument.
#[arg(long, alias = "http-port", default_value_t = 3080)]
pub external_http_port: u16,
/// The port to bind the internal listening HTTP server to. Clients like
/// the neon extension (for installing remote extensions) and local_proxy.
#[arg(long)]
pub internal_http_port: Option<u16>,
#[arg(short = 'D', long, value_name = "DATADIR")]
pub pgdata: String,
@@ -340,7 +352,8 @@ fn wait_spec(
pgdata: cli.pgdata.clone(),
pgbin: cli.pgbin.clone(),
pgversion: get_pg_version_string(&cli.pgbin),
http_port: cli.http_port,
external_http_port: cli.external_http_port,
internal_http_port: cli.internal_http_port.unwrap_or(cli.external_http_port + 1),
live_config_allowed,
state: Mutex::new(new_state),
state_changed: Condvar::new(),
@@ -358,9 +371,13 @@ fn wait_spec(
compute.prewarm_postgres()?;
}
// Launch http service first, so that we can serve control-plane requests
// while configuration is still in progress.
let _http_handle = launch_http_server(cli.http_port, &compute);
// Launch the external HTTP server first, so that we can serve control plane
// requests while configuration is still in progress.
Server::External(cli.external_http_port).launch(&compute);
// The internal HTTP server could be launched later, but there isn't much
// sense in waiting.
Server::Internal(cli.internal_http_port.unwrap_or(cli.external_http_port + 1)).launch(&compute);
if !spec_set {
// No spec provided, hang waiting for it.

View File

@@ -82,8 +82,10 @@ pub struct ComputeNode {
/// - we push spec and it does configuration
/// - but then it is restarted without any spec again
pub live_config_allowed: bool,
/// The port that the compute's HTTP server listens on
pub http_port: u16,
/// The port that the compute's external HTTP server listens on
pub external_http_port: u16,
/// The port that the compute's internal HTTP server listens on
pub internal_http_port: u16,
/// Volatile part of the `ComputeNode`, which should be used under `Mutex`.
/// To allow HTTP API server to serving status requests, while configuration
/// is in progress, lock should be held only for short periods of time to do
@@ -631,7 +633,7 @@ impl ComputeNode {
config::write_postgres_conf(
&pgdata_path.join("postgresql.conf"),
&pspec.spec,
self.http_port,
self.internal_http_port,
)?;
// Syncing safekeepers is only safe with primary nodes: if a primary
@@ -1396,7 +1398,7 @@ impl ComputeNode {
// Write new config
let pgdata_path = Path::new(&self.pgdata);
let postgresql_conf_path = pgdata_path.join("postgresql.conf");
config::write_postgres_conf(&postgresql_conf_path, &spec, self.http_port)?;
config::write_postgres_conf(&postgresql_conf_path, &spec, self.internal_http_port)?;
let max_concurrent_connections = spec.reconfigure_concurrency;

View File

@@ -4,11 +4,9 @@ use http::{header::CONTENT_TYPE, StatusCode};
use serde::Serialize;
use tracing::error;
pub use server::launch_http_server;
mod extract;
mod routes;
mod server;
pub mod server;
/// Convenience response builder for JSON responses
struct JsonResponse;

View File

@@ -1,9 +1,11 @@
use std::{
fmt::Display,
net::{IpAddr, Ipv6Addr, SocketAddr},
sync::Arc,
time::Duration,
};
use anyhow::Result;
use axum::{
extract::Request,
middleware::{self, Next},
@@ -24,45 +26,65 @@ use super::routes::{
};
use crate::compute::ComputeNode;
async fn handle_404() -> Response {
StatusCode::NOT_FOUND.into_response()
}
const X_REQUEST_ID: &str = "x-request-id";
/// This middleware function allows compute_ctl to generate its own request ID
/// if one isn't supplied. The control plane will always send one as a UUID. The
/// neon Postgres extension on the other hand does not send one.
async fn maybe_add_request_id_header(mut request: Request, next: Next) -> Response {
let headers = request.headers_mut();
if headers.get(X_REQUEST_ID).is_none() {
headers.append(X_REQUEST_ID, Uuid::new_v4().to_string().parse().unwrap());
}
next.run(request).await
/// `compute_ctl` has two servers: internal and external. The internal server
/// binds to the loopback interface and handles communication from clients on
/// the compute. The external server is what receives communication from the
/// control plane, the metrics scraper, etc. We make the distinction because
/// certain routes in `compute_ctl` only need to be exposed to local processes
/// like Postgres via the neon extension and local_proxy.
#[derive(Clone, Copy, Debug)]
pub enum Server {
Internal(u16),
External(u16),
}
/// Run the HTTP server and wait on it forever.
async fn serve(port: u16, compute: Arc<ComputeNode>) {
let mut app = Router::new()
.route("/check_writability", post(check_writability::is_writable))
.route("/configure", post(configure::configure))
.route("/database_schema", get(database_schema::get_schema_dump))
.route("/dbs_and_roles", get(dbs_and_roles::get_catalog_objects))
.route(
"/extension_server/{*filename}",
post(extension_server::download_extension),
)
.route("/extensions", post(extensions::install_extension))
.route("/grants", post(grants::add_grant))
.route("/insights", get(insights::get_insights))
.route("/metrics", get(metrics::get_metrics))
.route("/metrics.json", get(metrics_json::get_metrics))
.route("/status", get(status::get_status))
.route("/terminate", post(terminate::terminate))
.fallback(handle_404)
.layer(
impl Display for Server {
fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
match self {
Server::Internal(_) => f.write_str("internal"),
Server::External(_) => f.write_str("external"),
}
}
}
impl From<Server> for Router<Arc<ComputeNode>> {
fn from(server: Server) -> Self {
let mut router = Router::<Arc<ComputeNode>>::new();
router = match server {
Server::Internal(_) => {
router = router
.route(
"/extension_server/{*filename}",
post(extension_server::download_extension),
)
.route("/extensions", post(extensions::install_extension))
.route("/grants", post(grants::add_grant));
// Add in any testing support
if cfg!(feature = "testing") {
use super::routes::failpoints;
router = router.route("/failpoints", post(failpoints::configure_failpoints));
}
router
}
Server::External(_) => router
.route("/check_writability", post(check_writability::is_writable))
.route("/configure", post(configure::configure))
.route("/database_schema", get(database_schema::get_schema_dump))
.route("/dbs_and_roles", get(dbs_and_roles::get_catalog_objects))
.route("/insights", get(insights::get_insights))
.route("/metrics", get(metrics::get_metrics))
.route("/metrics.json", get(metrics_json::get_metrics))
.route("/status", get(status::get_status))
.route("/terminate", post(terminate::terminate)),
};
router.fallback(Server::handle_404).method_not_allowed_fallback(Server::handle_405).layer(
ServiceBuilder::new()
// Add this middleware since we assume the request ID exists
.layer(middleware::from_fn(maybe_add_request_id_header))
@@ -102,43 +124,88 @@ async fn serve(port: u16, compute: Arc<ComputeNode>) {
)
.layer(PropagateRequestIdLayer::x_request_id()),
)
.with_state(compute);
}
}
// Add in any testing support
if cfg!(feature = "testing") {
use super::routes::failpoints;
app = app.route("/failpoints", post(failpoints::configure_failpoints))
impl Server {
async fn handle_404() -> impl IntoResponse {
StatusCode::NOT_FOUND
}
// This usually binds to both IPv4 and IPv6 on Linux, see
// https://github.com/rust-lang/rust/pull/34440 for more information
let addr = SocketAddr::new(IpAddr::from(Ipv6Addr::UNSPECIFIED), port);
let listener = match TcpListener::bind(&addr).await {
Ok(listener) => listener,
Err(e) => {
error!(
"failed to bind the compute_ctl HTTP server to port {}: {}",
port, e
);
return;
async fn handle_405() -> impl IntoResponse {
StatusCode::METHOD_NOT_ALLOWED
}
async fn listener(&self) -> Result<TcpListener> {
let addr = SocketAddr::new(self.ip(), self.port());
let listener = TcpListener::bind(&addr).await?;
Ok(listener)
}
fn ip(&self) -> IpAddr {
match self {
// TODO: Change this to Ipv6Addr::LOCALHOST when the GitHub runners
// allow binding to localhost
Server::Internal(_) => IpAddr::from(Ipv6Addr::UNSPECIFIED),
Server::External(_) => IpAddr::from(Ipv6Addr::UNSPECIFIED),
}
};
if let Ok(local_addr) = listener.local_addr() {
info!("compute_ctl HTTP server listening on {}", local_addr);
} else {
info!("compute_ctl HTTP server listening on port {}", port);
}
if let Err(e) = axum::serve(listener, app).await {
error!("compute_ctl HTTP server error: {}", e);
fn port(self) -> u16 {
match self {
Server::Internal(port) => port,
Server::External(port) => port,
}
}
async fn serve(self, compute: Arc<ComputeNode>) {
let listener = self.listener().await.unwrap_or_else(|e| {
// If we can't bind, the compute cannot operate correctly
panic!(
"failed to bind the compute_ctl {} HTTP server to {}: {}",
self,
SocketAddr::new(self.ip(), self.port()),
e
);
});
if tracing::enabled!(tracing::Level::INFO) {
let local_addr = match listener.local_addr() {
Ok(local_addr) => local_addr,
Err(_) => SocketAddr::new(self.ip(), self.port()),
};
info!(
"compute_ctl {} HTTP server listening at {}",
self, local_addr
);
}
let router = Router::from(self).with_state(compute);
if let Err(e) = axum::serve(listener, router).await {
error!("compute_ctl {} HTTP server error: {}", self, e);
}
}
pub fn launch(self, compute: &Arc<ComputeNode>) {
let state = Arc::clone(compute);
info!("Launching the {} server", self);
tokio::spawn(self.serve(state));
}
}
/// Launch HTTP server in a new task and return its `JoinHandle`.
pub fn launch_http_server(port: u16, state: &Arc<ComputeNode>) -> tokio::task::JoinHandle<()> {
let state = Arc::clone(state);
/// This middleware function allows compute_ctl to generate its own request ID
/// if one isn't supplied. The control plane will always send one as a UUID. The
/// neon Postgres extension on the other hand does not send one.
async fn maybe_add_request_id_header(mut request: Request, next: Next) -> Response {
let headers = request.headers_mut();
if headers.get(X_REQUEST_ID).is_none() {
headers.append(X_REQUEST_ID, Uuid::new_v4().to_string().parse().unwrap());
}
tokio::spawn(serve(port, state))
next.run(request).await
}