sozu_lib/protocol/tcp_preread/

shell.rs

//! I/O shell wiring the sans-io [`SniPrereadCore`] into the TCP session
//! lifecycle (issue #1279).
//!
//! [`SniPreread`] is the socket-owning half of the split described in the
//! module doc one level up: it accumulates bytes from the raw frontend
//! socket into the session's own [`Checkout`] (NEVER `consume()`d while
//! undecided -- byte-for-byte replay to the backend is the whole point),
//! re-feeds the full accumulated window to [`SniPrereadCore::handle_input`]
//! on every read, and -- once routed -- stays parked in this state through
//! backend connect. This mirrors `RelayProxyProtocol`, not `Pipe`:
//! `Pipe::set_back_socket` does not re-arm inherited buffer writes, so a
//! premature switch to `Pipe` before the backend socket exists would strand
//! the connect handshake.
//!
//! This module intentionally emits its OWN `tcp.sni_preread.*` decision
//! metrics (routed / `rejected.<reason>`) and log lines, but does **not** own
//! the `tcp.sni_preread.active` gauge lifecycle or the eventual state
//! transition out of `SniPreread` -- both live in `lib/src/tcp.rs` (gauge:
//! +1 on entry in `TcpSession::new_sni_preread`, -1 in
//! `TcpSession::upgrade_sni_preread` on the upgrade exit or in
//! `TcpSession::close`'s `StateMarker::SniPreread` arm on the
//! reject/teardown exits; transition: `TcpSession::upgrade_sni_preread`),
//! which alone has the `TcpProxy`/`TcpListener` context (per-cluster
//! `proxy_protocol`, buffers, listener) needed to pick the right next state
//! and guarantee the gauge is adjusted exactly once per session, on exactly
//! one of its three exits (reject / upgrade / teardown).

use std::{net::SocketAddr, time::Instant};

use mio::{Token, net::TcpStream};
use rusty_ulid::Ulid;

use super::{AlpnMatcher, Input, Output, PrereadConfig, SniPrereadCore};
use crate::{
    Readiness, SessionMetrics, SessionResult,
    metrics::names,
    pool::Checkout,
    socket::{SocketHandler, SocketResult},
    sozu_command::{ready::Ready, state::ClusterId},
};

/// Per-session prefix for log lines emitted with a [`SniPreread`] in scope.
/// Renders the canonical `TCP-SNI\tSession(...)\t >>>` envelope, reusing
/// the `TCP-SNI` tag already established by the core's own `log_context!`
/// (`tcp_preread/mod.rs`) so operators can grep core decisions and shell
/// orchestration together.
macro_rules! log_context {
    ($self:expr) => {{
        let (open, reset, grey, gray, white) = sozu_command::logging::ansi_palette();
        format!(
            "{open}TCP-SNI{reset}\t{grey}Session{reset}({gray}frontend{reset}={white}{frontend}{reset})\t >>>",
            open = open,
            reset = reset,
            grey = grey,
            gray = gray,
            white = white,
            frontend = $self.frontend_token.0,
        )
    }};
}

/// Captured exactly once, when [`Output::Routed`] first fires. The core's
/// own `decided` latch guarantees the SAME terminal verdict replays on every
/// subsequent `handle_input` call, so [`SniPreread::outcome`] never changes
/// once `Some`.
#[derive(Debug, Clone)]
pub struct RoutedOutcome {
    pub cluster: ClusterId,
    pub content_offset: usize,
    pub proxy_source: Option<SocketAddr>,
    pub sni: String,
    pub alpn: Vec<Vec<u8>>,
    /// The matched route's configured pattern (trie key) -- `*.example.com`
    /// for a wildcard route, not the client's concrete SNI. Together with
    /// `matched_alpn` this is the matched FRONTEND's identity, which
    /// `TcpSession::upgrade_sni_preread` uses to rebuild the per-frontend
    /// access-log tags key (`sni_tags_key` in `lib/src/tcp.rs`).
    pub matched_sni_pattern: String,
    /// Clone of the winning route entry's [`AlpnMatcher`].
    pub matched_alpn: AlpnMatcher,
}

/// TCP session state that owns the frontend socket and the accumulating
/// [`Checkout`] while [`SniPrereadCore`] decides a route. See the module doc
/// for the full lifecycle and the exit-accounting contract for
/// `tcp.sni_preread.active`.
pub struct SniPreread<Front: SocketHandler> {
    pub frontend: Front,
    pub frontend_token: Token,
    pub frontend_readiness: Readiness,
    pub backend_readiness: Readiness,
    pub backend: Option<TcpStream>,
    pub backend_token: Option<Token>,
    pub request_id: Ulid,
    /// The session's own frontend accumulator, growing from wire offset 0.
    /// NEVER `consume()`d while undecided.
    pub frontend_buffer: Checkout,
    /// The listener's `sni_preread_max_bytes` knob (default 16 384) clamped
    /// to `frontend_buffer.capacity()` and floored at the 5-byte TLS
    /// record-header minimum -- see `effective_sni_preread_max_bytes` in
    /// `lib/src/tcp.rs`. Captured once at construction: the buffer's
    /// capacity is fixed for its lifetime, so this never needs re-deriving.
    effective_max_bytes: usize,
    core: SniPrereadCore,
    outcome: Option<RoutedOutcome>,
    started_at: Instant,
}

impl<Front: SocketHandler> SniPreread<Front> {
    /// Instantiate a new SniPreread SessionState with:
    /// - frontend_interest: READABLE | HUP | ERROR
    /// - backend_interest: HUP | ERROR (WRITABLE armed once routed)
    pub fn new(
        frontend: Front,
        frontend_token: Token,
        request_id: Ulid,
        frontend_buffer: Checkout,
        effective_max_bytes: usize,
    ) -> Self {
        SniPreread {
            frontend,
            frontend_token,
            frontend_readiness: Readiness {
                interest: Ready::READABLE | Ready::HUP | Ready::ERROR,
                event: Ready::EMPTY,
            },
            backend_readiness: Readiness {
                interest: Ready::HUP | Ready::ERROR,
                event: Ready::EMPTY,
            },
            backend: None,
            backend_token: None,
            request_id,
            frontend_buffer,
            effective_max_bytes,
            core: SniPrereadCore::new(),
            outcome: None,
            started_at: Instant::now(),
        }
    }

    pub fn effective_max_bytes(&self) -> usize {
        self.effective_max_bytes
    }

    pub fn outcome(&self) -> Option<&RoutedOutcome> {
        self.outcome.as_ref()
    }

    pub fn is_routed(&self) -> bool {
        self.outcome.is_some()
    }

    pub fn started_at(&self) -> Instant {
        self.started_at
    }

    /// Bytes already accumulated from the frontend. Distinguishes a genuine
    /// mid-preread abort (bytes were seen) from a bare TCP health-check
    /// connect/FIN with no bytes at all (cf. `expect.rs`'s identical guard).
    pub fn has_received_bytes(&self) -> bool {
        self.frontend_buffer.available_data() > 0
    }

    pub fn front_socket(&self) -> &TcpStream {
        self.frontend.socket_ref()
    }

    pub fn back_socket_mut(&mut self) -> Option<&mut TcpStream> {
        self.backend.as_mut()
    }

    pub fn set_back_socket(&mut self, socket: TcpStream) {
        self.backend = Some(socket);
    }

    pub fn set_back_token(&mut self, token: Token) {
        self.backend_token = Some(token);
    }

    /// Read available bytes and feed them to the core. Emits the
    /// `tcp.sni_preread.routed` / `tcp.sni_preread.rejected.<reason>`
    /// metrics and log lines itself; the caller (`TcpSession::readable` in
    /// `lib/src/tcp.rs`) only needs to react to the returned
    /// [`SessionResult`] and, on a fresh route, sync `TcpSession::cluster_id`
    /// from [`Self::outcome`].
    pub fn readable(
        &mut self,
        metrics: &mut SessionMetrics,
        cfg: &PrereadConfig<'_>,
    ) -> SessionResult {
        if self.outcome.is_some() {
            // Already decided -- and deliberately not reading: bytes the
            // client sends past the routed window must stay in the kernel
            // socket buffer for the post-upgrade state to consume. Defensive
            // only since sozu-proxy/sozu#1373: the `Output::Routed` arm of
            // `handle_output` drops frontend-readable INTEREST at the same
            // moment it latches the outcome, so `TcpSession::ready_inner`
            // stops selecting this dispatch entirely. Returning `Continue`
            // without consuming the (still latched, deliberately preserved)
            // READABLE event is what spun the session to
            // `MAX_LOOP_ITERATIONS` while the backend was still connecting.
            return SessionResult::Continue;
        }

        // Enforce `effective_max_bytes` as a HARD read bound, not merely an
        // advisory the core applies on its NeedMore path. The frontend
        // `Checkout` is sized to the pool buffer (`buffer_size`, 16 393
        // bytes by default), far larger than a tight
        // `sni_preread_max_bytes`; draining all of it would read an
        // oversized-but-COMPLETE ClientHello in full, which the core then
        // routes (it caps only would-be-NeedMore windows -- see
        // `mod.rs::need_more_or_too_large`). Capping the read makes an
        // over-cap hello reach the core INCOMPLETE with `buf.len() ==
        // max_bytes`, the exact `TooLarge` reject condition. A hello that
        // fits within the cap still routes, and any coalesced bytes past
        // the cap stay in the kernel socket buffer for the `Pipe` to
        // replay byte-for-byte.
        let cap_remaining = self
            .effective_max_bytes
            .saturating_sub(self.frontend_buffer.available_data());
        let space_before = self.frontend_buffer.available_space();
        let read_len = space_before.min(cap_remaining);
        let (sz, socket_result) = self
            .frontend
            .socket_read(&mut self.frontend_buffer.space()[..read_len]);
        // The socket can only write into the free space it was handed.
        debug_assert!(
            sz <= read_len,
            "socket_read cannot return more bytes than the capped space it was handed"
        );

        if sz > 0 {
            let data_before = self.frontend_buffer.available_data();
            self.frontend_buffer.fill(sz);
            debug_assert_eq!(
                self.frontend_buffer.available_data(),
                data_before + sz,
                "fill must expose exactly the bytes just read"
            );

            count!(names::backend::BYTES_IN, sz as i64);
            metrics.bin += sz;

            if socket_result == SocketResult::Error {
                return self.front_gone(cfg);
            }
            if socket_result == SocketResult::WouldBlock {
                self.frontend_readiness.event.remove(Ready::READABLE);
            }

            let output = self.core.handle_input(
                cfg,
                Input::Bytes {
                    buf: self.frontend_buffer.data(),
                    now: Instant::now(),
                },
            );
            return self.handle_output(output);
        }

        match socket_result {
            SocketResult::Error | SocketResult::Closed => self.front_gone(cfg),
            SocketResult::WouldBlock => {
                self.frontend_readiness.event.remove(Ready::READABLE);
                SessionResult::Continue
            }
            SocketResult::Continue => SessionResult::Continue,
        }
    }

    /// Feed the preread deadline firing. Always terminal: a fired deadline
    /// resolves to `Reject(Fragmented)` (or replays an already-latched
    /// verdict in the vanishingly unlikely race where routing and the
    /// deadline land in the same tick -- the core's `decided` latch makes
    /// that safe either way).
    pub fn on_timeout(&mut self, cfg: &PrereadConfig<'_>) {
        let output = self.core.handle_input(
            cfg,
            Input::Timeout {
                now: Instant::now(),
            },
        );
        // The frontend timeout always closes the session regardless; this
        // call exists purely for its metric/log side effect.
        let _ = self.handle_output(output);
    }

    /// Feed a frontend close observed before a routing decision (e.g. HUP
    /// racing ahead of `readable`'s own 0-byte detection, dispatched from
    /// `TcpSession::front_hup`). Mirrors `readable`'s
    /// `SocketResult::Closed` handling: silent when no bytes were ever
    /// received (the bare TCP health-check pattern), metered otherwise.
    pub fn on_front_closed(&mut self, cfg: &PrereadConfig<'_>) {
        let _ = self.front_gone(cfg);
    }

    /// Shared "the frontend is gone" handling for both the read-path
    /// (`Ok(0)` / socket error) and the HUP-path (`on_front_closed`).
    fn front_gone(&mut self, cfg: &PrereadConfig<'_>) -> SessionResult {
        if self.has_received_bytes() {
            let output = self.core.handle_input(cfg, Input::FrontClosed);
            self.handle_output(output)
        } else {
            // Bare TCP health-check pattern (SYN/ACK/FIN, no bytes ever
            // sent): silent, no rejection metric -- cf. `expect.rs`'s
            // identical guard.
            trace!(
                "{} front socket closed with 0 bytes during SNI preread",
                log_context!(self)
            );
            self.frontend_readiness.reset();
            SessionResult::Close
        }
    }

    fn handle_output(&mut self, output: Output) -> SessionResult {
        match output {
            Output::NeedMore { .. } => SessionResult::Continue,
            Output::Routed {
                cluster,
                content_offset,
                proxy_source,
                sni,
                alpn,
                matched_sni_pattern,
                matched_alpn,
            } => {
                debug_assert!(
                    self.outcome.is_none(),
                    "a route must be captured at most once per SniPreread lifetime"
                );
                let alpn_bytes = alpn
                    .iter()
                    .map(Vec::len)
                    .fold(0usize, usize::saturating_add);
                info!(
                    "{} SNI preread routed (cluster_id_bytes={}, sni_bytes={}, alpn_count={}, alpn_bytes={})",
                    log_context!(self),
                    cluster.len(),
                    sni.len(),
                    alpn.len(),
                    alpn_bytes,
                );
                incr!(names::tcp::sni_preread::ROUTED);
                // Arm backend-writable interest now so the FIRST backend
                // connect-writable event (whenever `connect_to_backend`
                // completes) immediately dispatches into `back_writable`
                // without waiting for a second, unrelated readiness event --
                // mirrors `RelayProxyProtocol::readable`'s identical arm
                // once its own header has been parsed.
                self.backend_readiness.interest.insert(Ready::WRITABLE);
                // ...and drop frontend-readable interest in the same breath,
                // completing the mirror: `relay.rs` drops READABLE on the line
                // ABOVE the arm above, and copying only half of that pair is
                // what left this state spinning. From here on
                // `readable` deliberately reads nothing (the
                // coalesced tail must stay in the kernel buffer for the
                // post-upgrade state to replay byte-for-byte), so a frontend
                // READABLE bit that stays SELECTED is a dispatch that can
                // never make progress. `tcp_socket_read` returns
                // `SocketResult::Continue` -- not `WouldBlock` -- when it
                // fills the capped slice it was handed, which is exactly what
                // a client coalescing its ClientHello with payload produces,
                // so the event below is latched and nothing will ever clear
                // it. Left selected, `front_interest = interest & event` in
                // `TcpSession::ready_inner` stays readable on every pass and
                // the session spins to `MAX_LOOP_ITERATIONS` with `bin` stuck
                // at the cap and `bout == 0` (sozu-proxy/sozu#1373).
                //
                // Mask the INTEREST, never the event: `upgrade_sni_preread`
                // hands `frontend_readiness.event` to the post-upgrade `Pipe`
                // (`restore_readiness_events`) or `SendProxyProtocol`
                // (`into_pipe`), both of which re-insert READABLE interest,
                // and under edge-triggered epoll that carried-over bit is the
                // only wake-up the coalesced tail will ever get. Clearing the
                // event here would trade the spin for a hang. HUP and ERROR
                // stay armed, so a frontend that goes away mid-connect is
                // still noticed.
                self.frontend_readiness.interest.remove(Ready::READABLE);
                self.outcome = Some(RoutedOutcome {
                    cluster,
                    content_offset,
                    proxy_source,
                    sni,
                    alpn,
                    matched_sni_pattern,
                    matched_alpn,
                });
                SessionResult::Continue
            }
            Output::Reject(reason) => {
                debug!("{} SNI preread rejected: {:?}", log_context!(self), reason);
                incr!(names::tcp::sni_preread::rejected_name(reason));
                self.frontend_readiness.reset();
                self.backend_readiness.reset();
                SessionResult::Close
            }
        }
    }

    /// The `back_writable` dispatch point: only ever reachable once routed
    /// -- the not-yet-routed guard in
    /// `TcpSession::attempt_backend_connect_if_needed` (`lib/src/tcp.rs`,
    /// both call sites inside `ready_inner`) guarantees
    /// `connect_to_backend` (and therefore any backend-writable event)
    /// never runs before a route decision. The actual per-`ProxyProtocolConfig`
    /// dispatch lives in `TcpSession::upgrade_sni_preread` (it needs the
    /// `TcpProxy`/listener context this struct deliberately does not hold),
    /// so this just signals the transition.
    pub fn back_writable(&self) -> SessionResult {
        debug_assert!(
            self.outcome.is_some(),
            "back_writable on SniPreread must only fire after a route decision"
        );
        SessionResult::Upgrade
    }
}

#[cfg(test)]
mod tests {
    use std::time::Duration;

    use super::*;
    use crate::protocol::tcp_preread::{AlpnMatcher, RejectReason};
    use crate::router::pattern_trie::TrieNode;

    /// `names::tcp::sni_preread::rejected_name` must be a TOTAL function
    /// over every `RejectReason` variant, each mapping to a DISTINCT dotted
    /// name -- a new variant added to the core without a matching arm fails
    /// to compile (see the `match` in `metrics/names.rs`), and this test
    /// additionally guards against two variants silently sharing one name.
    #[test]
    fn every_reject_reason_has_a_distinct_metric_name() {
        let reasons = [
            RejectReason::NotTls,
            RejectReason::MalformedRecord,
            RejectReason::MalformedHandshake,
            RejectReason::Fragmented,
            RejectReason::TooLarge,
            RejectReason::NoSni,
            RejectReason::EchOuterAbsent,
            RejectReason::SniUnmatched,
            RejectReason::AlpnUnmatched,
            RejectReason::ProxyHeaderInvalid,
            RejectReason::FrontClosed,
        ];
        let mut names_seen = std::collections::HashSet::new();
        for reason in reasons {
            let name = names::tcp::sni_preread::rejected_name(reason);
            assert!(
                name.starts_with("tcp.sni_preread.rejected."),
                "unexpected metric name shape for {reason:?}: {name}"
            );
            assert!(
                names_seen.insert(name),
                "two RejectReason variants mapped to the same metric name: {name}"
            );
        }
    }

    fn cfg(routes: &TrieNode<Vec<(AlpnMatcher, ClusterId)>>) -> PrereadConfig<'_> {
        PrereadConfig {
            routes,
            inbound_proxy: false,
            max_bytes: 16 * 1024,
            timeout: Duration::from_secs(3),
            accept_wildcard: true,
        }
    }

    #[test]
    fn routed_runtime_log_bounds_cluster_sni_and_alpn() {
        const CLUSTER_SECRET: &str = "TCP_PREREAD_CLUSTER_SECRET_SENTINEL";
        const SNI_SECRET: &str = "TCP_PREREAD_SNI_SECRET_SENTINEL";
        const ALPN_SECRET: &str = "TCP_PREREAD_ALPN_SECRET_SENTINEL";

        let cluster = format!("{CLUSTER_SECRET}{}", "x".repeat(4096));
        let sni = format!("{SNI_SECRET}{}", "x".repeat(4096));
        let alpn = format!("{ALPN_SECRET}{}", "x".repeat(4096)).into_bytes();
        let cluster_len = cluster.len();
        let sni_len = sni.len();
        let alpn_len = alpn.len();
        let output = crate::capture_test_logs(move || {
            use std::net::{TcpListener as StdTcpListener, TcpStream as StdTcpStream};

            use mio::net::TcpStream as MioTcpStream;

            use crate::pool::Pool;

            let listener = StdTcpListener::bind("127.0.0.1:0").expect("bind test listener");
            let address = listener.local_addr().expect("listener address");
            let _client = StdTcpStream::connect(address).expect("connect test client");
            let (server, _) = listener.accept().expect("accept test server");
            server.set_nonblocking(true).expect("server nonblocking");
            let mut pool = Pool::with_capacity(1, 1, 16 * 1024);
            let frontend_buffer = pool.checkout().expect("frontend buffer");
            let mut preread = SniPreread::new(
                MioTcpStream::from_std(server),
                Token(0),
                Ulid::generate(),
                frontend_buffer,
                16 * 1024,
            );

            assert_eq!(
                preread.handle_output(Output::Routed {
                    cluster: cluster.into(),
                    content_offset: 0,
                    proxy_source: None,
                    sni: sni.clone(),
                    alpn: vec![alpn],
                    matched_sni_pattern: sni,
                    matched_alpn: AlpnMatcher::Any,
                }),
                SessionResult::Continue
            );
        });

        for secret in [CLUSTER_SECRET, SNI_SECRET, ALPN_SECRET] {
            assert!(
                !output.contains(secret),
                "SNI preread routed log leaked {secret}: {output}"
            );
        }
        for metadata in [
            format!("cluster_id_bytes={cluster_len}"),
            format!("sni_bytes={sni_len}"),
            "alpn_count=1".to_owned(),
            format!("alpn_bytes={alpn_len}"),
        ] {
            assert!(
                output.contains(&metadata),
                "SNI preread routed log omitted {metadata}: {output}"
            );
        }
        assert!(
            output.len() <= 512,
            "SNI preread routed log is not bounded: {} bytes",
            output.len()
        );
    }

    #[test]
    fn preread_config_helper_shape_is_sane() {
        // Not a behavioral test of the core (already covered exhaustively
        // in `tcp_preread/mod.rs`) -- just pins the `PrereadConfig` field
        // shape this shell constructs against, so a field rename there is
        // caught here too.
        let routes = TrieNode::root();
        let cfg = cfg(&routes);
        assert!(!cfg.inbound_proxy);
        assert_eq!(cfg.max_bytes, 16 * 1024);
        assert!(cfg.accept_wildcard);
    }

    /// Regression guard for the SNI-preread read cap (sozu-proxy/sozu#1279):
    /// `readable` must never pull more than
    /// `effective_max_bytes` into the accumulator in a single read, even
    /// though the backing buffer is far larger and the socket has far more
    /// queued. Without the cap the shell drains the whole buffer, so an
    /// oversized-but-COMPLETE ClientHello is read in full and ROUTED (the core
    /// caps only would-be-NeedMore windows) instead of rejected `TooLarge` and
    /// the connection closed.
    #[test]
    fn readable_caps_the_accumulator_at_effective_max_bytes() {
        use std::io::Write as _;
        use std::net::{TcpListener as StdTcpListener, TcpStream as StdTcpStream};

        use mio::net::TcpStream as MioTcpStream;

        use crate::pool::Pool;

        let listener = StdTcpListener::bind("127.0.0.1:0").expect("bind test listener");
        let addr = listener.local_addr().expect("listener local addr");
        let mut client = StdTcpStream::connect(addr).expect("connect test client");
        let (server, _) = listener.accept().expect("accept test server");
        server.set_nonblocking(true).expect("server nonblocking");

        // Flood far more than the tight cap, into a buffer far larger than it.
        let flood = vec![0x16u8; 4096];
        client.write_all(&flood).expect("write flood");
        client.flush().ok();

        let mut pool = Pool::with_capacity(1, 1, 16 * 1024);
        let frontend_buffer = pool.checkout().expect("frontend buffer");
        assert!(
            frontend_buffer.available_space() > 4096,
            "buffer must dwarf the cap for this test to be meaningful"
        );

        let effective_max_bytes = 16usize;
        let mut preread = SniPreread::new(
            MioTcpStream::from_std(server),
            Token(0),
            Ulid::generate(),
            frontend_buffer,
            effective_max_bytes,
        );

        // Keep the core cap consistent with the shell cap, exactly as
        // `TcpListener::preread_config` does (both = effective_max_bytes).
        let routes = TrieNode::root();
        let cfg = PrereadConfig {
            routes: &routes,
            inbound_proxy: false,
            max_bytes: effective_max_bytes,
            timeout: Duration::from_secs(3),
            accept_wildcard: true,
        };
        let mut metrics = SessionMetrics::new(Some(Duration::ZERO));

        let result = preread.readable(&mut metrics, &cfg);
        assert!(
            preread.frontend_buffer.available_data() <= effective_max_bytes,
            "readable accumulated {} bytes, past the {}-byte cap",
            preread.frontend_buffer.available_data(),
            effective_max_bytes
        );
        // A window that reached the cap without a complete hello rejects and
        // closes -- it must never sit routed or spin needing more.
        assert_eq!(
            result,
            SessionResult::Close,
            "an over-cap preread window must reject-and-close"
        );

        drop(client);
    }

    /// Regression guard for sozu-proxy/sozu#1373: a routed SNI preread must
    /// stop advertising frontend-readable interest while it waits for the
    /// backend connect, or `TcpSession::ready_inner` spins to
    /// `MAX_LOOP_ITERATIONS` and kills a perfectly healthy session.
    ///
    /// A client that coalesces its ClientHello with the first payload bytes
    /// -- what a `--release` e2e client does every time -- fills the whole
    /// `effective_max_bytes` window in one read. `tcp_socket_read` returns
    /// `SocketResult::Continue` (not `WouldBlock`) when it fills the slice it
    /// was handed, so the frontend READABLE *event* stays latched. The route
    /// decision lands on that same read, and from then on `readable` is
    /// deliberately a no-op: the coalesced tail must stay in the kernel
    /// buffer for the post-upgrade state to replay byte-for-byte. Nothing
    /// clears the latch, so if READABLE also stays in
    /// `frontend_readiness.interest`, `front_interest = interest & event`
    /// (`lib/src/tcp.rs`, the loop's re-entry predicate, mirrored below)
    /// stays readable on every pass until the 10 000-iteration guard closes
    /// the connection with `bin` stuck at the cap and `bout == 0`.
    ///
    /// The latched *event* must SURVIVE. `upgrade_sni_preread` hands
    /// `frontend_readiness.event` to the post-upgrade `Pipe`
    /// (`restore_readiness_events`) or `SendProxyProtocol` (`into_pipe`),
    /// each of which re-inserts READABLE interest; under edge-triggered
    /// epoll that carried-over bit is the ONLY wake-up the coalesced tail
    /// will ever get. Clearing the event instead of masking the interest
    /// would trade this spin for a hang.
    ///
    /// To SEE THIS RED: delete
    /// `self.frontend_readiness.interest.remove(Ready::READABLE);` from the
    /// `Output::Routed` arm of `handle_output`.
    #[test]
    fn routed_preread_stops_re_arming_frontend_readable_while_connecting() {
        use std::io::Write as _;
        use std::net::{TcpListener as StdTcpListener, TcpStream as StdTcpStream};

        use mio::net::TcpStream as MioTcpStream;

        use crate::pool::Pool;
        use crate::protocol::tcp_preread::parser;

        let listener = StdTcpListener::bind("127.0.0.1:0").expect("bind test listener");
        let addr = listener.local_addr().expect("listener local addr");
        let mut client = StdTcpStream::connect(addr).expect("connect test client");
        let (server, _) = listener.accept().expect("accept test server");
        server.set_nonblocking(true).expect("server nonblocking");

        // The coalescing shape of sozu-proxy/sozu#1373: one write carrying a
        // complete ClientHello immediately followed by payload, far more than
        // the preread cap, so the capped read fills its slice exactly and
        // `socket_read` reports `Continue` rather than `WouldBlock`.
        let hello = parser::build_client_hello_wire(&[parser::encode_sni_extension("example.com")]);
        let effective_max_bytes = 16 * 1024usize;
        let mut coalesced = hello.clone();
        coalesced.resize(hello.len() + 2 * effective_max_bytes, 0x17);
        client.write_all(&coalesced).expect("write coalesced hello");
        client.flush().ok();

        let mut routes = TrieNode::root();
        routes.domain_insert(
            b"example.com".to_vec(),
            vec![(AlpnMatcher::Any, "cluster-a".into())],
        );
        let cfg = PrereadConfig {
            routes: &routes,
            inbound_proxy: false,
            max_bytes: effective_max_bytes,
            timeout: Duration::from_secs(3),
            accept_wildcard: true,
        };

        let mut pool = Pool::with_capacity(1, 1, effective_max_bytes);
        let frontend_buffer = pool.checkout().expect("frontend buffer");
        let mut preread = SniPreread::new(
            MioTcpStream::from_std(server),
            Token(0),
            Ulid::generate(),
            frontend_buffer,
            effective_max_bytes,
        );
        let mut metrics = SessionMetrics::new(Some(Duration::ZERO));

        // epoll delivered one frontend-readable event; the backend is still
        // `Connecting`, so no backend readiness ever fires during this pass.
        preread.frontend_readiness.event.insert(Ready::READABLE);

        // `TcpSession::ready_inner`'s re-entry predicate, bounded far below
        // `MAX_LOOP_ITERATIONS` so the spin is a failed assertion in
        // milliseconds rather than a 10 000-iteration burn.
        const DISPATCH_BUDGET: usize = 4;
        let mut dispatches = 0usize;
        while (preread.frontend_readiness.interest & preread.frontend_readiness.event).is_readable()
        {
            assert!(
                dispatches < DISPATCH_BUDGET,
                "routed SNI preread re-armed frontend-readable {DISPATCH_BUDGET} times \
                 without consuming the event: this is the spin to MAX_LOOP_ITERATIONS \
                 (interest {:?}, event {:?}, bin {})",
                preread.frontend_readiness.interest,
                preread.frontend_readiness.event,
                metrics.bin
            );
            dispatches += 1;
            assert_eq!(
                preread.readable(&mut metrics, &cfg),
                SessionResult::Continue
            );
        }

        assert!(
            preread.is_routed(),
            "the coalesced hello must route, otherwise this test witnesses nothing"
        );
        assert_eq!(
            metrics.bin, effective_max_bytes,
            "the routing read must fill the whole cap, the coalescing shape of the bug"
        );
        assert!(
            preread.frontend_readiness.event.is_readable(),
            "the latched frontend-readable event must survive for the post-upgrade \
             Pipe/SendProxyProtocol to drain the coalesced tail: under edge-triggered \
             epoll it is the only wake-up that tail will ever get"
        );
        assert!(
            preread.backend_readiness.interest.is_writable(),
            "routing must still arm backend-writable so the connect completion upgrades"
        );

        drop(client);
    }
}