sozu_lib/

server.rs

//! event loop management
use std::{
    cell::RefCell,
    collections::{HashMap, HashSet, VecDeque, hash_map::Entry},
    hash::{DefaultHasher, Hash, Hasher},
    io::Error as IoError,
    net::{IpAddr, Ipv4Addr, Ipv6Addr, SocketAddr},
    os::unix::io::{AsRawFd, FromRawFd, RawFd},
    rc::{Rc, Weak},
    str::FromStr,
    sync::LazyLock,
    time::{Duration, Instant},
};

use mio::{
    Events, Interest, Poll, Token,
    net::{TcpListener as MioTcpListener, TcpStream, UdpSocket as MioUdpSocket},
};
use slab::Slab;
use sozu_command::{
    channel::{Channel, ChannelError},
    config::MetricDetailLevel,
    listener_key::ListenerKey,
    logging,
    proto::command::{
        ActivateListener, AddBackend, CertificatesWithFingerprints, ClosedSessions, Cluster,
        ClusterHashes, ClusterInformations, DeactivateListener, Event, EventKind,
        HttpListenerConfig, HttpsListenerConfig, InitialState, ListenerType,
        LoadBalancingAlgorithms, LoadMetric, MetricDetail, MetricsConfiguration,
        ProtectedHealthCheckPhase, RemoveBackend, Request, ResponseContent, ResponseStatus,
        ServerConfig, TcpListenerConfig as CommandTcpListener,
        UdpListenerConfig as CommandUdpListener, UpdateHttpListenerConfig,
        UpdateHttpsListenerConfig, UpdateTcpListenerConfig, UpdateUdpListenerConfig, WorkerRequest,
        WorkerResponse, request::RequestType, response_content::ContentType,
    },
    ready::Ready,
    scm_socket::{Listeners, ScmSocket, ScmSocketError},
    state::{ClusterId, ConfigState, StateError},
};

use crate::metrics::names;
use crate::{
    AcceptError, Protocol, ProxyConfiguration, ProxySession, SessionIsToBeClosed, SessionTarget,
    backends::{Backend, BackendMap},
    features::FEATURES,
    health_check::HealthChecker,
    http, https,
    load_balancing::ShuffleSharding,
    metrics::METRICS,
    pool::Pool,
    protected_health_check::ProtectedWorker,
    tcp,
    timer::Timer,
    udp,
};

/// Number of bounded buckets for the per-source connect-rate counter.
///
/// `incr!` requires a `&'static str`, so per-IP labelling would either need
/// runtime `Box::leak` per unique source (unbounded under SYN flood — direct
/// OWASP A05 / NIST SP 800-92 cardinality-blow-up risk) or a fixed bucket
/// table. We pick the bucket table: 256 static labels precomputed at startup,
/// each masked subnet hashes into one of them.
///
/// Bucket-noise vs per-IP fidelity is a deliberate trade. Operators wanting
/// per-IP attribution should pair these counters with structured access logs
/// or a downstream rate-limiter; the metric here is for "is some /24 spamming
/// us right now?", not "which IP exactly". 256 buckets keep the memory + UDP
/// statsd cost flat regardless of attacker effort.
pub const PER_SOURCE_BUCKETS: usize = 256;

/// Pre-leaked `&'static str` table for per-source bucket counters.
/// `incr!` requires `&'static str`; we leak once at first access (LazyLock)
/// for `PER_SOURCE_BUCKETS` keys, totalling ~10 KB heap. The leak is bounded
/// by `PER_SOURCE_BUCKETS` and never grows with traffic.
static PER_SOURCE_BUCKET_KEYS: LazyLock<[&'static str; PER_SOURCE_BUCKETS]> = LazyLock::new(|| {
    let mut keys: [&'static str; PER_SOURCE_BUCKETS] = [""; PER_SOURCE_BUCKETS];
    for (i, slot) in keys.iter_mut().enumerate() {
        // e.g. "client.connect.per_source.bucket_042"
        let owned = format!("client.connect.per_source.bucket_{i:03}");
        *slot = Box::leak(owned.into_boxed_str());
    }
    keys
});

/// Mask an IP address to its bounded prefix (/24 for IPv4, /48 for IPv6) and
/// hash it into one of `PER_SOURCE_BUCKETS` slots. The hash is `DefaultHasher`,
/// which is deterministic within a process but salted across runs — fine for
/// telemetry, not suitable for cross-host correlation.
fn per_source_bucket(peer: &SocketAddr) -> &'static str {
    let mut hasher = DefaultHasher::new();
    match peer.ip() {
        IpAddr::V4(v4) => {
            let octets = v4.octets();
            // /24 mask: keep first three octets, zero the host portion.
            let masked = Ipv4Addr::new(octets[0], octets[1], octets[2], 0);
            masked.hash(&mut hasher);
        }
        IpAddr::V6(v6) => {
            let octets = v6.octets();
            // /48 mask: keep first 6 bytes, zero the rest.
            let mut masked_octets = [0u8; 16];
            masked_octets[..6].copy_from_slice(&octets[..6]);
            Ipv6Addr::from(masked_octets).hash(&mut hasher);
        }
    }
    let idx = (hasher.finish() as usize) % PER_SOURCE_BUCKETS;
    PER_SOURCE_BUCKET_KEYS[idx]
}

/// Period between two `accept_queue.saturated_seconds` ticks. The counter is
/// incremented once per period while [`SessionManager::can_accept`] is `false`,
/// distinguishing "queue spent N seconds at max" from "queue briefly hit max"
/// — the binary `accept_queue.backpressure` gauge collapses that duration.
const ACCEPT_SATURATION_TICK: Duration = Duration::from_secs(1);

pub type ProxyChannel = Channel<WorkerResponse, WorkerRequest>;

thread_local! {
  pub static QUEUE: RefCell<VecDeque<WorkerResponse>> = const { RefCell::new(VecDeque::new()) };
}

thread_local! {
  pub static TIMER: RefCell<Timer<Token>> = RefCell::new(Timer::default());
}

pub fn push_queue(message: WorkerResponse) {
    QUEUE.with(|queue| {
        (*queue.borrow_mut()).push_back(message);
    });
}

pub fn push_event(event: Event) {
    QUEUE.with(|queue| {
        (*queue.borrow_mut()).push_back(WorkerResponse {
            id: "EVENT".to_string(),
            message: String::new(),
            status: ResponseStatus::Processing.into(),
            content: Some(ContentType::Event(event).into()),
        });
    });
}

/// Whether `request` removes a frontend, of any protocol.
fn is_frontend_removal(request: &Request) -> bool {
    matches!(
        request.request_type,
        Some(
            RequestType::RemoveHttpFrontend(_)
                | RequestType::RemoveHttpFrontendWithHeaderPredicates(_)
                | RequestType::RemoveHttpsFrontend(_)
                | RequestType::RemoveHttpsFrontendWithHeaderPredicates(_)
                | RequestType::RemoveTcpFrontend(_)
                | RequestType::RemoveUdpFrontend(_)
        )
    )
}

/// Build the `WorkerMetricDetailStatus` content payload returned in
/// every successful `SetMetricDetail` worker response. The master
/// collects these across the fan-out and assembles them into
/// `MetricDetailStatus.workers[<worker_id>]` so the TUI sees each
/// worker's actual aggregator state instead of the master's view.
fn worker_metric_detail_status_content(
    configured: MetricDetailLevel,
    effective: MetricDetailLevel,
    previous_effective: MetricDetailLevel,
    active_lease_count: u32,
) -> ResponseContent {
    use sozu_command::proto::command::WorkerMetricDetailStatus;
    ContentType::WorkerMetricDetailStatus(WorkerMetricDetailStatus {
        configured: MetricDetail::from(configured) as i32,
        effective: MetricDetail::from(effective) as i32,
        previous_effective: MetricDetail::from(previous_effective) as i32,
        active_lease_count,
    })
    .into()
}

/// Build a `METRIC_DETAIL_CHANGED` event carrying the worker-local
/// transition payload (previous/effective levels + transition kind).
/// `client_id` is `Some(_)` for explicit apply/clear, `None` for the
/// polled janitor's bulk expiry. The master folds this Event into the
/// audit log alongside operator-initiated transitions emitted at the
/// dispatch site in `bin/src/command/requests.rs::worker_request`.
fn push_metric_detail_transition(
    previous: MetricDetailLevel,
    effective: MetricDetailLevel,
    transition_kind: &'static str,
    client_id: Option<String>,
) {
    use sozu_command::proto::command::MetricDetailTransition;
    // No-op when nothing actually changed. Defence-in-depth — every
    // caller already gates on `previous != effective`, but
    // double-checking here means future call sites can't accidentally
    // emit a "ghost" transition.
    if previous == effective {
        return;
    }
    push_event(Event {
        kind: EventKind::MetricDetailChanged as i32,
        cluster_id: None,
        backend_id: None,
        address: None,
        metric_detail: Some(MetricDetailTransition {
            previous_effective: MetricDetail::from(previous) as i32,
            effective: MetricDetail::from(effective) as i32,
            transition_kind: transition_kind.to_owned(),
            client_id,
        }),
    });
}

/// Polled lease-expiry janitor: `SetMetricDetail` leases self-expire after
/// their TTL so a crashed `sozu top` cannot permanently elevate metrics
/// cardinality. Called once per event-loop iteration from [`Server::run`],
/// which wakes at least once per `poll_timeout` (1 s) even with no traffic
/// and no command, and from [`Server::notify`]. `lease_tick_due` gates the
/// HashMap walk to once per `LEASE_TICK_INTERVAL` (5 s), so an abandoned
/// lease is retired within TTL + 5 s + one poll timeout (#1831).
fn tick_metric_detail_leases(now: Instant) {
    // Capture (previous, effective) before releasing the borrow so we can
    // emit an Event afterwards. Holding `METRICS.borrow_mut` across
    // `push_event` would re-enter the same thread-local from inside
    // `QUEUE.with` (safe but conceptually noisy); the two-step split keeps
    // the borrow scopes minimal. Single-threaded worker, so `borrow_mut` is
    // safe here.
    let lease_tick_transition = METRICS.with(|metrics| {
        let mut m = metrics.borrow_mut();
        if !m.lease_tick_due(now) {
            return None;
        }
        let previous = m.lease_tick(now)?;
        let effective = m.detail_effective();
        Some((previous, effective))
    });
    if let Some((previous, effective)) = lease_tick_transition {
        // The janitor retired one or more leases AND the effective level
        // moved. Surface the worker-local transition as an Event so the
        // master folds it into the audit log. `client_id` is `None` because
        // the janitor may have retired multiple leases at once.
        push_metric_detail_transition(previous, effective, "lease_tick_expired", None);
    }
}

#[derive(Copy, Clone, Debug, PartialEq, Eq, PartialOrd, Ord, Hash)]
pub struct ListenToken(pub usize);
#[derive(Copy, Clone, Debug, PartialEq, Eq, PartialOrd, Ord, Hash)]
pub struct SessionToken(pub usize);

impl From<usize> for ListenToken {
    fn from(val: usize) -> ListenToken {
        ListenToken(val)
    }
}

impl From<ListenToken> for usize {
    fn from(val: ListenToken) -> usize {
        val.0
    }
}

impl From<usize> for SessionToken {
    fn from(val: usize) -> SessionToken {
        SessionToken(val)
    }
}

impl From<SessionToken> for usize {
    fn from(val: SessionToken) -> usize {
        val.0
    }
}

/// Mask `address` down to the network named by `prefix` bits.
///
/// `prefix` is validated at config-load time
/// (`ConfigError::InvalidSubnetPrefix` in `command/src/config.rs`), but
/// this runs on the per-request hot path and does not assume that
/// happened: a prefix at or above the address width returns the address
/// unmasked.
///
/// Both ends are handled by an explicit branch rather than by the shift.
/// `u32::MAX << 32` is an overflowing shift — it panics under
/// `debug_assertions` and silently wraps to a no-op shift in release —
/// so a full-width prefix must never reach the shift. A zero prefix is
/// branched from the other end for the same reason.
fn mask_ipv4(address: Ipv4Addr, prefix: u32) -> Ipv4Addr {
    if prefix >= u32::BITS {
        return address;
    }
    if prefix == 0 {
        return Ipv4Addr::UNSPECIFIED;
    }
    // Only a strictly-interior prefix reaches the shift, which is what
    // makes `u32::BITS - prefix` a legal shift amount in 1..32.
    debug_assert!(
        prefix > 0 && prefix < u32::BITS,
        "both shift-overflowing ends must have returned before the shift"
    );
    let mask = u32::MAX << (u32::BITS - prefix);
    let masked = Ipv4Addr::from(u32::from(address) & mask);
    // Pair assertion, positive then negative space: the result keeps
    // every masked-in bit of the input, and holds none of the host bits
    // the prefix is supposed to erase.
    debug_assert_eq!(
        u32::from(masked),
        u32::from(address) & mask,
        "masking must keep exactly the network bits of the source address"
    );
    debug_assert_eq!(
        u32::from(masked) & !mask,
        0,
        "masking must clear every host bit below the prefix"
    );
    masked
}

/// IPv6 twin of [`mask_ipv4`]. The same two shift-overflowing ends are
/// branched explicitly: `u128::MAX << 128` is as ill-defined as its
/// 32-bit counterpart.
fn mask_ipv6(address: Ipv6Addr, prefix: u32) -> Ipv6Addr {
    if prefix >= u128::BITS {
        return address;
    }
    if prefix == 0 {
        return Ipv6Addr::UNSPECIFIED;
    }
    debug_assert!(
        prefix > 0 && prefix < u128::BITS,
        "both shift-overflowing ends must have returned before the shift"
    );
    let mask = u128::MAX << (u128::BITS - prefix);
    let masked = Ipv6Addr::from(u128::from(address) & mask);
    debug_assert_eq!(
        u128::from(masked),
        u128::from(address) & mask,
        "masking must keep exactly the network bits of the source address"
    );
    debug_assert_eq!(
        u128::from(masked) & !mask,
        0,
        "masking must clear every host bit below the prefix"
    );
    masked
}

/// Derive the per-(cluster, source-SUBNET) counter key for `address`.
///
/// An IPv4-mapped IPv6 address is canonicalised to its IPv4 form and
/// charged to `ipv4_prefix`. This matters: a dual-stack `[::]` listener
/// hands every IPv4 client back as `::ffff:a.b.c.d`, and masking those
/// with an IPv6 prefix would collapse the entire IPv4 internet into the
/// single `::/56` bucket — the subnet cap would then fire on unrelated
/// clients the moment it was switched on. The per-IP counter is NOT
/// canonicalised: it keeps charging the address exactly as it arrived,
/// which is part of what leaves its behaviour untouched by this feature.
pub fn subnet_key(address: &IpAddr, ipv4_prefix: u32, ipv6_prefix: u32) -> IpAddr {
    let canonical_v4 = match address {
        IpAddr::V4(v4) => Some(*v4),
        IpAddr::V6(v6) => v6.to_ipv4_mapped(),
    };
    let key = match (canonical_v4, address) {
        (Some(v4), _) => IpAddr::V4(mask_ipv4(v4, ipv4_prefix)),
        (None, IpAddr::V6(v6)) => IpAddr::V6(mask_ipv6(*v6, ipv6_prefix)),
        // Unreachable: `canonical_v4` is `Some` for every `IpAddr::V4`.
        (None, IpAddr::V4(v4)) => IpAddr::V4(mask_ipv4(*v4, ipv4_prefix)),
    };
    // Pair assertion: the key stays in the family it will be compared
    // against — a mapped source yields a V4 key (positive space), and a
    // genuine IPv6 source never yields one (negative space).
    debug_assert_eq!(
        key.is_ipv4(),
        canonical_v4.is_some(),
        "subnet key family must follow the canonicalised source family"
    );
    debug_assert_eq!(
        key.is_ipv6(),
        canonical_v4.is_none(),
        "a genuine IPv6 source must never produce an IPv4 subnet key"
    );
    key
}

pub struct SessionManager {
    pub max_connections: usize,
    pub nb_connections: usize,
    pub can_accept: bool,
    pub slab: Slab<Rc<RefCell<dyn ProxySession>>>,
    /// Default per-(cluster, source-IP) connection limit. `0` disables
    /// the feature; cluster-level overrides take precedence at check
    /// time.
    pub max_connections_per_ip: u64,
    /// Default `Retry-After` header value (seconds) for HTTP 429
    /// responses emitted on per-(cluster, source-IP) limit hit. `0`
    /// omits the header.
    pub retry_after: u32,
    /// Active **frontend connections** per `(cluster_id, source_ip)`.
    /// Each frontend session contributes AT MOST 1 to the count for any
    /// given `(cluster, ip)` pair, regardless of how many streams it
    /// multiplexes to that cluster (an H2 connection serving 100
    /// streams to cluster X from IP 1.2.3.4 still counts as 1). The
    /// counter is incremented the first time a session's
    /// `Router::plan_connect` resolves to a fresh `(cluster, ip)` pair, and
    /// decremented when the session closes. Empty when the feature is
    /// unused.
    ///
    /// ── Why nested maps instead of `HashMap<(String, IpAddr), usize>` ──
    ///
    /// The per-request hot path (`cluster_ip_at_limit`, called from
    /// `mux/router::plan_connect` for every cluster-resolving request) used
    /// to allocate a `String` to build the compound key on every
    /// lookup. Splitting the storage so the outer key is a [`ClusterId`]
    /// lets the lookup take `&str` — `HashMap::get(cluster_id)` on a
    /// `HashMap<ClusterId, _>` accepts a borrow via the `Borrow<str>`
    /// impl. The hot path no longer allocates, and `track_cluster_ip`
    /// keys the maps with the caller's reference-counted handle, so no
    /// key is ever a copy of the id. Memory footprint is unchanged
    /// in steady state — entries are still reaped to zero on session
    /// close.
    connections_per_cluster_ip: HashMap<ClusterId, HashMap<IpAddr, usize>>,
    /// Reverse index: per-token map of `cluster_id` → set of source IPs
    /// already counted against `connections_per_cluster_ip`. Used to
    /// make `track_cluster_ip` idempotent within a session (so H2
    /// streams to the same cluster from the same client only consume
    /// one slot in the limit) and to drain a session's contributions on
    /// close. Same nesting rationale as above.
    cluster_ip_tracks: HashMap<Token, HashMap<ClusterId, HashSet<IpAddr>>>,
    /// Default per-(cluster, source-SUBNET) connection limit. `0`
    /// disables the subnet limiter, which is the default; cluster-level
    /// overrides take precedence at check time.
    ///
    /// This is a SECOND, INDEPENDENT counter standing beside
    /// `max_connections_per_ip`, not a replacement for it. The per-IP
    /// cap serves the API-gateway case — one abusive client, one
    /// address. The per-subnet cap serves the denial-of-service case,
    /// where an attacker holding a routed prefix (an ordinary
    /// residential IPv6 `/64`) defeats a per-address counter outright
    /// by taking a fresh address for every connection
    /// (sozu-proxy/sozu#1270). Both gates are consulted and BOTH must
    /// admit, so "10 per IP AND 100 per /64" is expressible.
    pub max_connections_per_subnet: u64,
    /// IPv4 prefix length, in bits, defining a "subnet" for the counter
    /// above. Boot-time only: changing a mask would re-key every live
    /// counter, so there is no runtime setter for it. 32 masks nothing.
    pub subnet_ipv4_prefix: u32,
    /// IPv6 prefix length, in bits. Boot-time only, as above. 128 masks
    /// nothing.
    pub subnet_ipv6_prefix: u32,
    /// Active frontend connections per `(cluster_id, masked_source_ip)`.
    /// Exact mirror of `connections_per_cluster_ip`, keyed on the subnet
    /// key from `subnet_key` instead of the raw address; the same
    /// nesting rationale applies verbatim.
    ///
    /// Stays EMPTY while the subnet limiter is disabled: unlike the
    /// per-IP index, which is populated unconditionally, this one is
    /// only written when a positive limit resolves for the cluster (see
    /// `SessionManager::track_cluster_subnet`). That is what makes the
    /// default genuinely free rather than merely inert — no second
    /// allocation, no second hash, no second entry per session for an
    /// operator who never turns the feature on.
    connections_per_cluster_subnet: HashMap<ClusterId, HashMap<IpAddr, usize>>,
    /// Reverse index for `connections_per_cluster_subnet`, mirroring
    /// `cluster_ip_tracks`. Also empty while the limiter is disabled.
    cluster_subnet_tracks: HashMap<Token, HashMap<ClusterId, HashSet<IpAddr>>>,
    /// Default backend connection attempt budget of a request, the first
    /// attempt included: `ServerConfig::max_connection_attempts`, changed at
    /// runtime by `SetMaxConnectionAttempts`. A cluster's own
    /// `max_connection_attempts` takes precedence
    /// ([`Self::effective_max_connection_attempts`]). Validated, so between
    /// 1 and 255.
    pub max_connection_attempts: u32,
}

impl SessionManager {
    pub fn new(
        slab: Slab<Rc<RefCell<dyn ProxySession>>>,
        max_connections: usize,
        max_connections_per_ip: u64,
        retry_after: u32,
        max_connections_per_subnet: u64,
        subnet_ipv4_prefix: u32,
        subnet_ipv6_prefix: u32,
    ) -> Rc<RefCell<Self>> {
        Rc::new(RefCell::new(SessionManager {
            max_connections,
            nb_connections: 0,
            can_accept: true,
            slab,
            max_connections_per_ip,
            retry_after,
            connections_per_cluster_ip: HashMap::new(),
            cluster_ip_tracks: HashMap::new(),
            max_connections_per_subnet,
            subnet_ipv4_prefix,
            subnet_ipv6_prefix,
            connections_per_cluster_subnet: HashMap::new(),
            cluster_subnet_tracks: HashMap::new(),
            max_connection_attempts: sozu_command::config::DEFAULT_MAX_CONNECTION_ATTEMPTS,
        }))
    }

    /// Resolve the backend connection attempt budget of a request routed to
    /// a cluster whose own setting is `override_value`: the cluster's value
    /// when it sets one, the global default otherwise (sozu-proxy/sozu#1800).
    pub fn effective_max_connection_attempts(&self, override_value: Option<u32>) -> u32 {
        override_value.unwrap_or(self.max_connection_attempts)
    }

    /// Resolve the effective per-(cluster, source-IP) limit. `override_value`
    /// is the cluster-level setting from the proto `Cluster` message:
    /// `None` inherits the global default, `Some(0)` is explicit
    /// "unlimited", `Some(n > 0)` overrides.
    pub fn effective_max_connections_per_ip(&self, override_value: Option<u64>) -> u64 {
        override_value.unwrap_or(self.max_connections_per_ip)
    }

    /// Resolve the effective `Retry-After` header value. `Some(0)` (or
    /// the global default of 0) signals "omit the header" — caller
    /// must skip emission rather than render `Retry-After: 0`.
    pub fn effective_retry_after(&self, override_value: Option<u32>) -> u32 {
        override_value.unwrap_or(self.retry_after)
    }

    /// Resolve the effective per-(cluster, source-SUBNET) limit. Same
    /// three-state contract as `effective_max_connections_per_ip`:
    /// `None` inherits the global default, `Some(0)` is explicit
    /// "unlimited", `Some(n > 0)` overrides.
    pub fn effective_max_connections_per_subnet(&self, override_value: Option<u64>) -> u64 {
        override_value.unwrap_or(self.max_connections_per_subnet)
    }

    /// Derive the subnet counter key for `ip` under this manager's
    /// configured prefixes. Allocation-free — `IpAddr` is `Copy`.
    pub fn subnet_key(&self, ip: &IpAddr) -> IpAddr {
        subnet_key(ip, self.subnet_ipv4_prefix, self.subnet_ipv6_prefix)
    }

    /// Returns `true` when admitting `token` to one more connection for
    /// `(cluster, subnet_of(ip))` would exceed the resolved subnet
    /// limit. Exact mirror of `cluster_ip_at_limit` on the subnet
    /// counter: `0` is unlimited, and a token already holding a slot
    /// for this `(cluster, subnet)` is never at the limit.
    ///
    /// The subnet key is derived only AFTER the `limit == 0` early
    /// return, so a deployment that never enables this feature pays no
    /// masking work on the hot path at all.
    pub fn cluster_subnet_at_limit(
        &self,
        token: Token,
        cluster_id: &str,
        ip: &IpAddr,
        override_value: Option<u64>,
    ) -> bool {
        let limit = self.effective_max_connections_per_subnet(override_value);
        if limit == 0 {
            return false;
        }
        debug_assert!(
            limit > 0,
            "limit==0 (unlimited) must have returned before reaching the bounded check"
        );
        let subnet = self.subnet_key(ip);
        let already_tracked = self
            .cluster_subnet_tracks
            .get(&token)
            .and_then(|by_cluster| by_cluster.get(cluster_id))
            .is_some_and(|subnets| subnets.contains(&subnet));
        if already_tracked {
            debug_assert!(
                self.connections_per_cluster_subnet
                    .get(cluster_id)
                    .and_then(|by_subnet| by_subnet.get(&subnet))
                    .is_some_and(|c| *c > 0),
                "a tracked (token, cluster, subnet) slot must have a positive forward count"
            );
            return false;
        }
        self.connections_per_cluster_subnet
            .get(cluster_id)
            .and_then(|by_subnet| by_subnet.get(&subnet))
            .is_some_and(|c| (*c as u64) >= limit)
    }

    /// THE gate. Returns `true` when EITHER the per-IP cap or the
    /// per-subnet cap would be exceeded by admitting `token` to one more
    /// connection for `cluster_id` from `ip`.
    ///
    /// Both enforcement points — the H1+H2 mux router and raw TCP — call
    /// this rather than the two halves, so the conjunction "admit only
    /// if BOTH caps allow it" is stated once and cannot drift between
    /// protocols. The per-IP half is evaluated first and is byte-for-byte
    /// the pre-existing check; the subnet half is the single line this
    /// feature adds to the admission path.
    pub fn cluster_connection_at_limit(
        &self,
        token: Token,
        cluster_id: &str,
        ip: &IpAddr,
        ip_override: Option<u64>,
        subnet_override: Option<u64>,
    ) -> bool {
        self.cluster_ip_at_limit(token, cluster_id, ip, ip_override)
            || self.cluster_subnet_at_limit(token, cluster_id, ip, subnet_override)
    }

    /// Returns `true` when admitting `token` to one more connection for
    /// `(cluster, ip)` would exceed the resolved limit. `0` is treated
    /// as unlimited. A token that already holds a slot for this
    /// `(cluster, ip)` is NEVER at the limit — H2 sessions multiplex
    /// many streams to the same cluster on a single connection, and
    /// the limit governs distinct frontend connections, not streams.
    ///
    /// Hot-path: called for every cluster-resolving request from
    /// `mux/router::plan_connect`. The nested-map storage lets both lookups
    /// borrow `cluster_id` and `ip`; no per-call allocation runs here
    /// in steady state.
    pub fn cluster_ip_at_limit(
        &self,
        token: Token,
        cluster_id: &str,
        ip: &IpAddr,
        override_value: Option<u64>,
    ) -> bool {
        let limit = self.effective_max_connections_per_ip(override_value);
        if limit == 0 {
            return false;
        }
        // Pure query: the limit==0 branch already returned, so any work below
        // runs only with a positive cap.
        debug_assert!(
            limit > 0,
            "limit==0 (unlimited) must have returned before reaching the bounded check"
        );
        let already_tracked = self
            .cluster_ip_tracks
            .get(&token)
            .and_then(|by_cluster| by_cluster.get(cluster_id))
            .is_some_and(|ips| ips.contains(ip));
        if already_tracked {
            // Reverse-index/forward-count coherence: if this token already
            // holds a slot for (cluster, ip), the forward count must be > 0
            // (decrement-to-zero reaps the inner entry in untrack_all).
            debug_assert!(
                self.connections_per_cluster_ip
                    .get(cluster_id)
                    .and_then(|by_ip| by_ip.get(ip))
                    .is_some_and(|c| *c > 0),
                "a tracked (token, cluster, ip) slot must have a positive forward count"
            );
            return false;
        }
        self.connections_per_cluster_ip
            .get(cluster_id)
            .and_then(|by_ip| by_ip.get(ip))
            .is_some_and(|c| (*c as u64) >= limit)
    }

    /// Account `token`'s active connection against `(cluster, ip)`.
    /// Idempotent within a token: a second call for the same
    /// `(cluster, ip)` is a no-op so H2 retries / multi-stream opens
    /// to the same cluster do not double-count.
    ///
    /// Copies no cluster id: `entry(cluster_id.clone())` only bumps the
    /// [`ClusterId`] count. A first `(token, cluster)` observation
    /// allocates the new slot's IP set; later IPs under it reuse the
    /// set, and a repeat call allocates nothing.
    ///
    /// ── A FAILED backend dial keeps the slot, and that is deliberate ──
    ///
    /// `Router::plan_connect` (`lib/src/protocol/mux/router.rs`) claims
    /// here, after the `Self::cluster_ip_at_limit` gate but BEFORE
    /// backend selection and the dial. Six fallible exits follow the
    /// claim and none of them releases. Since sozu-proxy/sozu#1519 they
    /// sit in two functions rather than one: `Router::plan_connect`
    /// keeps the two pool-reuse refusals, and `Mux::dial_backend` with the
    /// `Mux::attach_dialed` it calls (`lib/src/protocol/mux/mod.rs`) own the
    /// four on the fresh-dial
    /// path — backend selection, buffer-pool exhaustion, the new
    /// connection refusing the stream, and the `L7Proxy::register_socket`
    /// rollback, which restores its gauges and slab entry but not the
    /// slot. The split moved no ordering: the claim still precedes all
    /// six. `Self::untrack_all_cluster_ip` is the only release, and
    /// session teardown is its only caller — the
    /// `L7Proxy::remove_session` impls on `HttpProxy` / `HttpsProxy` for
    /// the mux, `TcpSession::close` for raw TCP.
    /// An unpaired claim reads like an oversight; it is not one, and
    /// sozu-proxy/sozu#1521 is where the reasoning was measured rather
    /// than assumed. Do not pair a release with those exits as part of a
    /// refactor — it would LOOSEN a connection limit under exactly the
    /// conditions a client can induce.
    ///
    /// What makes the unpaired claim cheap on H1 is that a failed dial
    /// ENDS the frontend connection, so the slot returns with it:
    ///
    /// - `Mux::ready_inner` joins the plan and the dial into a single
    ///   `Result` with `and_then`, so every post-claim exit from either
    ///   function reaches the same `BackendConnectionError` arm of its
    ///   pending-link loop, which answers the stream from the template
    ///   registry instead of retrying. No arm there returns
    ///   `SessionResult::Close` of its own.
    /// - Every built-in template in
    ///   `lib/src/protocol/kawa_h1/answers.rs` carries
    ///   `Connection: close`, so `set_default_answer_with_retry_after`
    ///   (`lib/src/protocol/mux/answers.rs`) clears
    ///   `HttpContext::keep_alive_frontend`. On H1 that one bit is what
    ///   closes the connection; H2 does not read it.
    /// - H1: once the answer flushes, the response-complete branch of
    ///   `ConnectionH1::writable` (`lib/src/protocol/mux/h1.rs`) takes
    ///   `ConnectionH1::defer_close_for_tls_flush` instead of the
    ///   keep-alive reset, so the session closes after that one flush.
    /// - H2: the answer ends its own stream only (RFC 9113 §8.2.2:
    ///   `Connection` is not an H2 field, sozu-proxy/sozu#1740). The
    ///   connection stays open, so the slot is held until the client
    ///   closes it or it times out.
    /// - TCP never reaches an answer at all: any error out of
    ///   `TcpSession::connect_to_backend` (`lib/src/tcp.rs`) becomes
    ///   `SessionResult::Close` in the same event-loop pass.
    /// - A registered, `Normal` backend that is down may not fail on the
    ///   dial at all. The dial — `Mux::dial_backend`
    ///   (`lib/src/protocol/mux/mod.rs`) for HTTP, `Backend::try_connect`
    ///   (`lib/src/backends.rs`) for TCP — is non-blocking: a refusal the
    ///   kernel reports synchronously surfaces as
    ///   `BackendError::ConnectionFailures`, which the mux re-queues for
    ///   another backend, while an `EINPROGRESS` connect returns `Ok` and
    ///   its refusal (or its `connect_timeout`) arrives later as a backend
    ///   failure, so the stream retries through `EndStreamAction::Reconnect`.
    ///   The re-track is the no-op above, so one slot — not one per
    ///   attempt — is held for at most `max_connection_attempts` attempts
    ///   (sozu-proxy/sozu#1800).
    ///
    /// The window IS the connection's whole lifetime in two cases: any
    /// H2 connection, and an H1 connection answered by an operator
    /// template that omits `Connection: close` (the opt-out
    /// `doc/rate-limit-design.md` records for the 429 template). Either
    /// holds its one slot per (connection, cluster) until the client goes
    /// away, as a connection whose dial succeeded does.
    pub fn track_cluster_ip(&mut self, token: Token, cluster_id: ClusterId, ip: IpAddr) {
        // Snapshot the forward count for this (cluster, ip) before the insert
        // so we can pair-assert the delta. Ungated `let`: read only inside the
        // debug_assert! below → optimised out in release (no E0425).
        let count_before = self
            .connections_per_cluster_ip
            .get(&cluster_id)
            .and_then(|by_ip| by_ip.get(&ip))
            .copied()
            .unwrap_or(0);
        let inserted = self
            .cluster_ip_tracks
            .entry(token)
            .or_default()
            .entry(cluster_id.clone())
            .or_default()
            .insert(ip);
        if inserted {
            *self
                .connections_per_cluster_ip
                .entry(cluster_id.clone())
                .or_default()
                .entry(ip)
                .or_insert(0) += 1;
        }
        // Postconditions: the reverse index now records this (token, cluster,
        // ip), and the forward count advanced by exactly `inserted as usize`
        // (idempotent: a repeat call for the same triple is a no-op on both).
        debug_assert!(
            self.cluster_ip_tracks
                .get(&token)
                .and_then(|by_cluster| by_cluster.get(&cluster_id))
                .is_some_and(|ips| ips.contains(&ip)),
            "track must leave the (token, cluster, ip) recorded in the reverse index"
        );
        debug_assert_eq!(
            self.connections_per_cluster_ip
                .get(&cluster_id)
                .and_then(|by_ip| by_ip.get(&ip))
                .copied()
                .unwrap_or(0),
            count_before + inserted as usize,
            "forward count must advance by exactly 1 on first track, 0 on a repeat"
        );
        #[cfg(debug_assertions)]
        self.check_invariants();
    }

    /// Account `token`'s active connection against
    /// `(cluster, subnet_of(ip))`. Mirror of `track_cluster_ip` on the
    /// subnet counter, with one deliberate difference: it is a NO-OP
    /// while the resolved subnet limit is `0`.
    ///
    /// That conditional is what keeps the default genuinely free. The
    /// per-IP index is populated unconditionally, so an operator who
    /// never enables the subnet cap would otherwise pay a second hash,
    /// a second allocation and a second per-session entry forever for a
    /// feature they do not use. The cost is that enabling the cap at
    /// runtime does not retroactively count sessions already in flight;
    /// the counter converges as those sessions close. The per-IP path
    /// has the same shape of gap after `clear_cluster_ip_tracking`.
    pub fn track_cluster_subnet(
        &mut self,
        token: Token,
        cluster_id: ClusterId,
        ip: IpAddr,
        override_value: Option<u64>,
    ) {
        if self.effective_max_connections_per_subnet(override_value) == 0 {
            // No assertion that nothing is recorded for this token: that
            // is NOT an invariant. A cluster override can drop to
            // `Some(0)` after a session was tracked under a positive
            // limit — `AddCluster` is an upsert and clears no counters,
            // and only the global `SetMaxConnectionsPerSubnet(0)` drains
            // the bookkeeping. The stale slot is harmless (a `0` limit
            // never refuses) and is released normally on session close.
            return;
        }
        let subnet = self.subnet_key(&ip);
        // Snapshot before the insert so the delta can be pair-asserted.
        // Ungated `let`: read only inside the debug_assert! below, so it
        // is optimised out in release (no E0425).
        let count_before = self
            .connections_per_cluster_subnet
            .get(&cluster_id)
            .and_then(|by_subnet| by_subnet.get(&subnet))
            .copied()
            .unwrap_or(0);
        let inserted = self
            .cluster_subnet_tracks
            .entry(token)
            .or_default()
            .entry(cluster_id.clone())
            .or_default()
            .insert(subnet);
        if inserted {
            *self
                .connections_per_cluster_subnet
                .entry(cluster_id.clone())
                .or_default()
                .entry(subnet)
                .or_insert(0) += 1;
        }
        debug_assert!(
            self.cluster_subnet_tracks
                .get(&token)
                .and_then(|by_cluster| by_cluster.get(&cluster_id))
                .is_some_and(|subnets| subnets.contains(&subnet)),
            "track must leave the (token, cluster, subnet) recorded in the reverse index"
        );
        debug_assert_eq!(
            self.connections_per_cluster_subnet
                .get(&cluster_id)
                .and_then(|by_subnet| by_subnet.get(&subnet))
                .copied()
                .unwrap_or(0),
            count_before + inserted as usize,
            "forward count must advance by exactly 1 on first track, 0 on a repeat"
        );
        #[cfg(debug_assertions)]
        self.check_invariants();
    }

    /// Account `token`'s active connection against BOTH counters. The
    /// companion of `cluster_connection_at_limit`: the two enforcement
    /// points take the gate and the slot through one call each, so the
    /// per-IP and per-subnet bookkeeping can never fall out of step
    /// between the mux and raw TCP.
    pub fn track_cluster_connection(
        &mut self,
        token: Token,
        cluster_id: ClusterId,
        ip: IpAddr,
        subnet_override: Option<u64>,
    ) {
        self.track_cluster_ip(token, cluster_id.clone(), ip);
        self.track_cluster_subnet(token, cluster_id, ip, subnet_override);
    }

    /// Drain every slot held by `token` — per-IP AND per-subnet — and
    /// apply the matching decrements. Called on session teardown only —
    /// there is no per-stream untrack because the limit is
    /// per-connection, not per-stream. Removes empty inner maps so the
    /// outer `connections_per_cluster_ip` does not retain
    /// `(cluster_id, empty_map)` orphans across cluster lifetimes.
    ///
    /// The name is historical: this is the SINGLE release point every
    /// protocol already calls on close (`HttpSession`, `HttpsSession`,
    /// `TcpSession`), so the subnet counter is drained from here rather
    /// than from a second call each of those three sites would have to
    /// remember. A missed release leaks a slot until worker restart,
    /// which is exactly the failure a second entry point would invite.
    pub fn untrack_all_cluster_ip(&mut self, token: Token) {
        self.untrack_all_cluster_subnet(token);
        let Some(by_cluster) = self.cluster_ip_tracks.remove(&token) else {
            return;
        };
        // The reverse index for this token was just drained by `remove`; no
        // other code path re-inserts it within this call.
        debug_assert!(
            !self.cluster_ip_tracks.contains_key(&token),
            "untrack_all must evict the token from the reverse index"
        );
        for (cluster_id, ips) in by_cluster {
            let Entry::Occupied(mut outer) = self.connections_per_cluster_ip.entry(cluster_id)
            else {
                continue;
            };
            for ip in ips {
                if let Entry::Occupied(mut inner) = outer.get_mut().entry(ip) {
                    let count = inner.get_mut();
                    *count = count.saturating_sub(1);
                    if *count == 0 {
                        inner.remove();
                    }
                }
            }
            if outer.get().is_empty() {
                outer.remove();
            }
        }
        // No orphan bookkeeping survives: the forward map must not retain a
        // cluster with an empty inner ip-map, nor an ip whose count is zero.
        debug_assert!(
            self.connections_per_cluster_ip
                .values()
                .all(|by_ip| !by_ip.is_empty() && by_ip.values().all(|&c| c > 0)),
            "untrack_all must not leave empty inner maps or zero-count ips behind"
        );
        #[cfg(debug_assertions)]
        self.check_invariants();
    }

    /// Drain every `(cluster, subnet)` slot held by `token`. Mirror of
    /// the per-IP drain in `untrack_all_cluster_ip`, which is its only
    /// caller. A no-op when the subnet limiter was never enabled, since
    /// nothing was tracked in the first place.
    fn untrack_all_cluster_subnet(&mut self, token: Token) {
        let Some(by_cluster) = self.cluster_subnet_tracks.remove(&token) else {
            return;
        };
        debug_assert!(
            !self.cluster_subnet_tracks.contains_key(&token),
            "untrack_all must evict the token from the subnet reverse index"
        );
        for (cluster_id, subnets) in by_cluster {
            let Entry::Occupied(mut outer) = self.connections_per_cluster_subnet.entry(cluster_id)
            else {
                continue;
            };
            for subnet in subnets {
                if let Entry::Occupied(mut inner) = outer.get_mut().entry(subnet) {
                    let count = inner.get_mut();
                    *count = count.saturating_sub(1);
                    if *count == 0 {
                        inner.remove();
                    }
                }
            }
            if outer.get().is_empty() {
                outer.remove();
            }
        }
        debug_assert!(
            self.connections_per_cluster_subnet
                .values()
                .all(|by_subnet| !by_subnet.is_empty() && by_subnet.values().all(|&c| c > 0)),
            "untrack_all must not leave empty inner maps or zero-count subnets behind"
        );
    }

    /// Whether both halves of the subnet accounting are empty.
    ///
    /// Exists so a test can assert that a DISABLED subnet limiter leaves
    /// no bookkeeping behind at all — "inert" and "free" are different
    /// claims, and only the second one is checkable from outside.
    pub fn subnet_tracking_is_empty(&self) -> bool {
        self.cluster_subnet_tracks.is_empty() && self.connections_per_cluster_subnet.is_empty()
    }

    /// Wipe every per-(cluster, source-SUBNET) accounting bucket.
    /// Called by the runtime `SetMaxConnectionsPerSubnet(0)` path.
    /// Deliberately does NOT touch the per-IP buckets: the two limiters
    /// are independent, and disabling one must not silently reset the
    /// other's live counters.
    pub fn clear_cluster_subnet_tracking(&mut self) {
        self.cluster_subnet_tracks.clear();
        self.connections_per_cluster_subnet.clear();
        debug_assert!(
            self.cluster_subnet_tracks.is_empty() && self.connections_per_cluster_subnet.is_empty(),
            "clear must wipe both the subnet reverse index and the subnet forward count map"
        );
        #[cfg(debug_assertions)]
        self.check_invariants();
    }

    /// Wipe every per-(cluster, source-IP) accounting bucket. Called by
    /// the runtime `SetMaxConnectionsPerIp(0)` path so disabling the
    /// feature does not leave dead bookkeeping behind that a future
    /// re-enable would consult.
    pub fn clear_cluster_ip_tracking(&mut self) {
        self.cluster_ip_tracks.clear();
        self.connections_per_cluster_ip.clear();
        // Both halves of the per-(cluster, ip) accounting are now empty; a
        // future re-enable starts from a clean slate.
        debug_assert!(
            self.cluster_ip_tracks.is_empty() && self.connections_per_cluster_ip.is_empty(),
            "clear must wipe both the reverse index and the forward count map"
        );
        #[cfg(debug_assertions)]
        self.check_invariants();
    }

    /// The slab is considered at capacity if it contains more sessions than twice max_connections
    pub fn at_capacity(&self) -> bool {
        self.slab.len() >= self.accept_slab_threshold()
    }

    /// The slab fill level at which `at_capacity` flips to true and the
    /// accept queue is flushed. Reported as `slab.accept_threshold` so the
    /// per-iteration `slab.accept_threshold_percent` gauge in the run loop
    /// can chart proximity to this gate, distinct from raw slab usage.
    ///
    /// The constant `10 + 2 * max_connections` is the historical pre-knob
    /// budget; configured slab capacity is
    /// `10 + slab_entries_per_connection * max_connections` (see
    /// `command/src/config.rs`) and can be larger, so `slab.usage_percent`
    /// (against `slab.capacity()`) and `slab.accept_threshold_percent`
    /// (against this gate) are emitted as independent gauges.
    pub fn accept_slab_threshold(&self) -> usize {
        let threshold = 10 + 2 * self.max_connections;
        // The gate must leave headroom above the connection cap so listener /
        // system slots (Channel, Metrics, Timer, listeners) are never starved
        // by frontend connections alone.
        debug_assert!(
            threshold > self.max_connections,
            "accept gate must sit strictly above max_connections to reserve system slots"
        );
        threshold
    }

    /// Check the number of connections against max_connections, and the slab capacity.
    /// Returns false if limits are reached.
    pub fn check_limits(&mut self) -> bool {
        // Live-count invariant: the accounted connection count never exceeds
        // the configured cap (incr() enforces this, decr() never underflows).
        debug_assert!(
            self.nb_connections <= self.max_connections,
            "nb_connections must never exceed max_connections"
        );
        if self.nb_connections >= self.max_connections {
            error!("max number of session connection reached, flushing the accept queue");
            gauge!(names::accept_queue::BACKPRESSURE, 1);
            self.can_accept = false;
            // A negative result must have closed the accept gate.
            debug_assert!(
                !self.can_accept,
                "refusing at the cap must clear can_accept"
            );
            return false;
        }

        if self.at_capacity() {
            error!("not enough memory to accept another session, flushing the accept queue");
            error!(
                "nb_connections: {}, max_connections: {}",
                self.nb_connections, self.max_connections
            );
            gauge!(names::accept_queue::BACKPRESSURE, 1);
            self.can_accept = false;

            debug_assert!(
                !self.can_accept,
                "refusing at slab capacity must clear can_accept"
            );
            return false;
        }

        // A positive result means there is room under both gates.
        debug_assert!(
            self.nb_connections < self.max_connections && !self.at_capacity(),
            "check_limits returned room while a gate was actually saturated"
        );
        true
    }

    pub fn to_session(token: Token) -> SessionToken {
        SessionToken(token.0)
    }

    pub fn incr(&mut self) {
        let before = self.nb_connections;
        self.nb_connections += 1;
        assert!(self.nb_connections <= self.max_connections);
        // The counter advances by exactly one per accepted session.
        debug_assert_eq!(
            self.nb_connections,
            before + 1,
            "incr must raise nb_connections by exactly one"
        );
        // `client.connections_max` and `client.connections_percent` are
        // emitted from the run loop alongside `process.uptime_seconds` /
        // `server.live` so all proxy gauges advance in lock-step. Keeping
        // `client.connections` per-event preserves the high-resolution
        // signal scrapers expect.
        gauge!(names::client::CONNECTIONS, self.nb_connections);
    }

    /// Decrements the number of sessions, start accepting new connections
    /// if the capacity limit of 90% has not been reached.
    pub fn decr(&mut self) {
        assert!(self.nb_connections != 0);
        let before = self.nb_connections;
        self.nb_connections -= 1;
        // Mirror of incr: exactly one connection is released, no underflow.
        debug_assert_eq!(
            self.nb_connections,
            before - 1,
            "decr must lower nb_connections by exactly one"
        );
        gauge!(names::client::CONNECTIONS, self.nb_connections);

        // do not be ready to accept right away, wait until we get back to 10% capacity
        if !self.can_accept && self.nb_connections < self.max_connections * 90 / 100 {
            debug!(
                "nb_connections = {}, max_connections = {}, starting to accept again",
                self.nb_connections, self.max_connections
            );
            gauge!(names::accept_queue::BACKPRESSURE, 0);
            self.can_accept = true;
        }
    }

    /// Full cross-field invariant sweep for the session manager. Called as a
    /// `debug_assert!`-guarded postcondition from the mutating cluster-IP
    /// tracking methods. Asserts logic-bug conditions only — every clause
    /// holds on any well-formed manager regardless of traffic.
    #[cfg(debug_assertions)]
    fn check_invariants(&self) {
        // 1. Live connection count never exceeds the configured cap.
        debug_assert!(
            self.nb_connections <= self.max_connections,
            "nb_connections {} exceeds max_connections {}",
            self.nb_connections,
            self.max_connections
        );
        // 2. The forward count map never retains an empty inner ip-map or a
        //    zero count (both are reaped on the last untrack).
        debug_assert!(
            self.connections_per_cluster_ip
                .values()
                .all(|by_ip| !by_ip.is_empty() && by_ip.values().all(|&c| c > 0)),
            "connections_per_cluster_ip holds an empty inner map or a zero count"
        );
        // 3. Reverse-index → forward-count coherence: every (token, cluster,
        //    ip) recorded in the reverse index must have a positive forward
        //    count. The forward count is the sum of contributing tokens, so a
        //    tracked slot can never point at a missing/zero count.
        debug_assert!(
            self.cluster_ip_tracks.values().all(|by_cluster| {
                by_cluster.iter().all(|(cluster_id, ips)| {
                    ips.iter().all(|ip| {
                        self.connections_per_cluster_ip
                            .get(cluster_id)
                            .and_then(|by_ip| by_ip.get(ip))
                            .is_some_and(|&c| c > 0)
                    })
                })
            }),
            "a tracked (token, cluster, ip) slot has no positive forward count"
        );
        // 4. The reverse index never retains an empty inner structure.
        debug_assert!(
            self.cluster_ip_tracks.values().all(|by_cluster| {
                !by_cluster.is_empty() && by_cluster.values().all(|ips| !ips.is_empty())
            }),
            "cluster_ip_tracks retains an empty per-token or per-cluster entry"
        );
        // 5-7. The same three structural clauses, on the subnet mirror.
        debug_assert!(
            self.connections_per_cluster_subnet
                .values()
                .all(|by_subnet| !by_subnet.is_empty() && by_subnet.values().all(|&c| c > 0)),
            "connections_per_cluster_subnet holds an empty inner map or a zero count"
        );
        debug_assert!(
            self.cluster_subnet_tracks.values().all(|by_cluster| {
                by_cluster.iter().all(|(cluster_id, subnets)| {
                    subnets.iter().all(|subnet| {
                        self.connections_per_cluster_subnet
                            .get(cluster_id)
                            .and_then(|by_subnet| by_subnet.get(subnet))
                            .is_some_and(|&c| c > 0)
                    })
                })
            }),
            "a tracked (token, cluster, subnet) slot has no positive forward count"
        );
        debug_assert!(
            self.cluster_subnet_tracks.values().all(|by_cluster| {
                !by_cluster.is_empty() && by_cluster.values().all(|subnets| !subnets.is_empty())
            }),
            "cluster_subnet_tracks retains an empty per-token or per-cluster entry"
        );
    }
}

#[derive(thiserror::Error, Debug)]
pub enum ServerError {
    #[error("could not create event loop with MIO poll: {0}")]
    CreatePoll(IoError),
    #[error("could not clone the MIO registry: {0}")]
    CloneRegistry(IoError),
    #[error("could not register the channel: {0}")]
    RegisterChannel(IoError),
    #[error("{msg}:{scm_err}")]
    ScmSocket {
        msg: String,
        scm_err: ScmSocketError,
    },
    #[error("could not read the initial message of the main process: {0}")]
    ReadInitialMessage(ChannelError),
    #[error("could not answer the initial message of the main process: {0}")]
    WriteInitialAnswer(ChannelError),
}

/// Where the blocking handshake before `Server::run` stands.
#[derive(Debug, Clone, PartialEq, Eq)]
enum BootstrapStage {
    /// Nothing received: a direct `Status` or the capability query.
    Start,
    /// The capability is confirmed: a `Prepare`, or the closing `Status`.
    Ready,
    /// A version is staged: its `Commit` or `Abort`.
    Prepared { cluster: String, version: u64 },
    /// A version is applied: its `Release` or `Abort`.
    Committed { cluster: String, version: u64 },
}

/// What a worker answers to the one request the handshake expects next, and
/// the stage an exact answer leads to.
struct BootstrapStep {
    phase: ProtectedHealthCheckPhase,
    cluster: String,
    version: u64,
    next: BootstrapStage,
}

impl BootstrapStage {
    /// The step `request` is, when this stage accepts it. `Status` is not a
    /// step: `Server::await_initial_status` ends the handshake with it.
    fn step(&self, request: &RequestType) -> Option<BootstrapStep> {
        use ProtectedHealthCheckPhase as Phase;
        match request {
            RequestType::QueryProtectedHealthCheckSupport(_) if *self == Self::Start => {
                Some(BootstrapStep {
                    phase: Phase::ProtectedHealthCheckCapability,
                    cluster: String::new(),
                    version: 0,
                    next: Self::Ready,
                })
            }
            RequestType::ProtectedHealthCheckOrder(order) => {
                let phase = Phase::try_from(order.phase).ok()?;
                let (cluster, version) = (order.cluster_id.to_owned(), order.version);
                let outstanding = match self {
                    Self::Prepared {
                        cluster: held,
                        version: staged,
                    }
                    | Self::Committed {
                        cluster: held,
                        version: staged,
                    } => *held == cluster && *staged == version,
                    _ => false,
                };
                let next = match (self, phase) {
                    (Self::Ready, Phase::ProtectedHealthCheckPrepare) => Self::Prepared {
                        cluster: cluster.to_owned(),
                        version,
                    },
                    (Self::Prepared { .. }, Phase::ProtectedHealthCheckCommit) if outstanding => {
                        Self::Committed {
                            cluster: cluster.to_owned(),
                            version,
                        }
                    }
                    (Self::Committed { .. }, Phase::ProtectedHealthCheckRelease) if outstanding => {
                        Self::Ready
                    }
                    (
                        Self::Prepared { .. } | Self::Committed { .. },
                        Phase::ProtectedHealthCheckAbort,
                    ) if outstanding => Self::Ready,
                    _ => return None,
                };
                Some(BootstrapStep {
                    phase,
                    cluster,
                    version,
                    next,
                })
            }
            _ => None,
        }
    }

    /// Whether the headerless `Status` may end the handshake: nothing is
    /// prepared, committed or left to release.
    fn accepts_status(&self) -> bool {
        matches!(self, Self::Start | Self::Ready)
    }
}

impl BootstrapStep {
    /// Whether `response` is the positive typed reply to this very step.
    fn is_answered_by(&self, response: &WorkerResponse) -> bool {
        response.status == ResponseStatus::Ok as i32
            && matches!(
                response
                    .content
                    .as_ref()
                    .and_then(|content| content.content_type.as_ref()),
                Some(ContentType::ProtectedHealthCheckReply(reply))
                    if reply.phase == self.phase as i32
                        && reply.cluster_id == self.cluster
                        && reply.version == self.version
            )
    }
}

/// `Server` handles the event loop, the listeners, the sessions and
/// communication with the configuration channel.
///
/// A listener wraps a listen socket, the associated proxying protocols
/// (HTTP, HTTPS and TCP) and the routing configuration for clusters.
/// Listeners handle creating sessions from accepted sockets.
///
/// A session manages a "front" socket for a connected client, and all
/// of the associated data (back socket, protocol state machine, buffers,
/// metrics...).
///
/// `Server` gets configuration updates from the channel (domIN/path routes,
/// backend server address...).
///
/// Listeners and sessions are all stored in a slab structure to index them
/// by a [Token], they all have to implement the [ProxySession] trait.
pub struct Server {
    accept_queue_timeout: Duration,
    /// Tuple layout: `(socket, listen token, protocol, accept time, peer
    /// address)`. The peer is the address `accept(2)` returned with the
    /// socket: it attributes the `client.connect.per_source.*` counter at
    /// accept time and is handed to `create_session`, so no one has to read it
    /// back with `getpeername(2)` (sozu-proxy/sozu#1586). Unlike that lookup,
    /// which fails once the peer has reset, `accept(2)` always reports it.
    accept_queue: VecDeque<(TcpStream, ListenToken, Protocol, Instant, SocketAddr)>,
    /// When the accept queue saturates and `check_limits` refuses, evict the
    /// oldest non-listener sessions to make room. Default off — see
    /// `command::config::DEFAULT_EVICT_ON_QUEUE_FULL` for the rationale.
    evict_on_queue_full: bool,
    accept_ready: HashSet<ListenToken>,
    backends: Rc<RefCell<BackendMap>>,
    base_sessions_count: usize,
    channel: ProxyChannel,
    config_state: ConfigState,
    current_poll_errors: i32,
    health_checker: HealthChecker,
    http: Rc<RefCell<http::HttpProxy>>,
    https: Rc<RefCell<https::HttpsProxy>>,
    last_sessions_len: usize,
    last_shutting_down_message: Option<Instant>,
    last_zombie_check: Instant,
    loop_start: Instant,
    /// Wall-clock anchor for the `process.uptime_seconds` gauge. Captured once
    /// in [`Server::new`]; never reset on hot upgrades (the new worker that
    /// inherits FDs is a fresh process and starts its own counter).
    started_at: Instant,
    /// Last time the 1Hz `accept_queue.saturated_seconds` ticker fired. The
    /// counter is incremented once per [`ACCEPT_SATURATION_TICK`] while
    /// `SessionManager::can_accept` is `false`, so dashboards can plot the
    /// time spent saturated rather than just whether saturation occurred.
    last_saturation_tick: Instant,
    max_poll_errors: i32, // TODO: make this configurable? this defaults to 10000 for now
    /// Shared reference to the buffer pool the protocol stacks check buffers
    /// in and out of. Held here so the run loop can sample
    /// `buffer.in_use` / `buffer.capacity` / `buffer.usage_percent` once per
    /// iteration without requiring each protocol module to expose its own
    /// snapshot.
    pool: Rc<RefCell<Pool>>,
    pub poll: Poll,
    /// Protected health-check headers staged, applied or kept for an abort.
    protected_health_check: ProtectedWorker,
    poll_timeout: Option<Duration>, // TODO: make this configurable? this defaults to 1000 milliseconds for now
    scm_listeners: Option<Listeners>,
    scm: ScmSocket,
    sessions: Rc<RefCell<SessionManager>>,
    should_poll_at: Option<Instant>,
    shutting_down: Option<String>,
    /// Sessions a `CloseSessions` request ordered to close that are still
    /// finishing their in-flight HTTP work, keyed by frontend token.
    ///
    /// The `Weak` is what makes the token safe to hold: a session that closes
    /// on its own frees its slab key, which a new session can take, and
    /// `Server::drive_closing_sessions` only acts on a token whose slab entry
    /// is still this very session.
    closing_sessions: Vec<(Token, Weak<RefCell<dyn ProxySession>>)>,
    /// Sessions whose last `ready` pass yielded at its iteration budget while
    /// forwarding bytes (`ProxySession::yielded`), keyed by frontend token.
    /// `Server::resume_yielded_sessions` runs them again after the next poll;
    /// the `Weak` and the slab check guard against a session that closed and
    /// whose key was reused in between, as for `closing_sessions`.
    yielded_sessions: Vec<(Token, Weak<RefCell<dyn ProxySession>>)>,
    tcp: Rc<RefCell<tcp::TcpProxy>>,
    udp: Rc<RefCell<udp::UdpProxy>>,
    zombie_check_interval: Duration,
}

impl Server {
    pub fn try_new_from_config(
        worker_to_main_channel: ProxyChannel,
        worker_to_main_scm: ScmSocket,
        config: ServerConfig,
        initial_state: InitialState,
        expects_initial_status: bool,
    ) -> Result<Self, ServerError> {
        let event_loop = Poll::new().map_err(ServerError::CreatePoll)?;
        // Commit the operator-configured Basic-auth credential cap (or
        // keep the built-in default) once per worker process. The
        // `OnceLock` rejects any later attempt to change the value, so
        // calling here — before any L7 listener has accepted a request
        // — guarantees the cap is in force the first time `mux::auth`
        // runs.
        if let Some(cap) = config.basic_auth_max_credential_bytes {
            crate::protocol::mux::auth::set_max_decoded_credential_bytes(cap as usize);
        }
        // Same set-once-per-worker-boot pattern for the splice kernel-pipe
        // capacity. The setter no-ops on `0` so an explicit zero in config
        // does not collapse the pipe to PAGE_SIZE; the kernel still applies
        // page-rounding and `/proc/sys/fs/pipe-max-size` clamping at
        // SplicePipe::new time. Cfg-gated because the splice module only
        // exists on Linux + `splice` feature.
        #[cfg(all(target_os = "linux", feature = "splice"))]
        if let Some(cap) = config.splice_pipe_capacity_bytes {
            crate::splice::set_pipe_capacity(cap as usize);
        }
        let pool = Rc::new(RefCell::new(Pool::with_capacity(
            config.min_buffers as usize,
            config.max_buffers as usize,
            config.buffer_size as usize,
        )));
        let backends = Rc::new(RefCell::new(BackendMap::new()));

        // Note: slab_capacity uses 4x multiplier (up from 2x) to account for H2
        // multiplexing where each session can have multiple backend connections.
        // Newer `optional` proto fields fall through to the
        // command-lib defaults when an older worker manager omits them.
        let sessions: Rc<RefCell<SessionManager>> = SessionManager::new(
            Slab::with_capacity(config.slab_capacity() as usize),
            config.max_connections as usize,
            config
                .max_connections_per_ip
                .unwrap_or(sozu_command::config::DEFAULT_MAX_CONNECTIONS_PER_IP),
            config
                .retry_after
                .unwrap_or(sozu_command::config::DEFAULT_RETRY_AFTER),
            config
                .max_connections_per_subnet
                .unwrap_or(sozu_command::config::DEFAULT_MAX_CONNECTIONS_PER_SUBNET),
            config
                .subnet_ipv4_prefix
                .unwrap_or(sozu_command::config::DEFAULT_SUBNET_IPV4_PREFIX),
            config
                .subnet_ipv6_prefix
                .unwrap_or(sozu_command::config::DEFAULT_SUBNET_IPV6_PREFIX),
        );
        // Newer field, so an older main may omit it. The main validates it at
        // config load; a value that still breaks the bounds keeps the default
        // rather than disabling retries or overflowing the attempt counters.
        sessions.borrow_mut().max_connection_attempts = match config.max_connection_attempts {
            Some(attempts)
                if sozu_command::config::validate_max_connection_attempts(attempts).is_ok() =>
            {
                attempts
            }
            Some(attempts) => {
                error!(
                    "max_connection_attempts = {} is out of 1..=255, keeping {}",
                    attempts,
                    sozu_command::config::DEFAULT_MAX_CONNECTION_ATTEMPTS
                );
                sozu_command::config::DEFAULT_MAX_CONNECTION_ATTEMPTS
            }
            None => sozu_command::config::DEFAULT_MAX_CONNECTION_ATTEMPTS,
        };
        {
            let mut s = sessions.borrow_mut();
            let entry = s.slab.vacant_entry();
            trace!("taking token {:?} for channel", SessionToken(entry.key()));
            entry.insert(Rc::new(RefCell::new(ListenSession {
                protocol: Protocol::Channel,
            })));
        }
        {
            let mut s = sessions.borrow_mut();
            let entry = s.slab.vacant_entry();
            trace!("taking token {:?} for metrics", SessionToken(entry.key()));
            entry.insert(Rc::new(RefCell::new(ListenSession {
                protocol: Protocol::Timer,
            })));
        }
        {
            let mut s = sessions.borrow_mut();
            let entry = s.slab.vacant_entry();
            trace!("taking token {:?} for metrics", SessionToken(entry.key()));
            entry.insert(Rc::new(RefCell::new(ListenSession {
                protocol: Protocol::Metrics,
            })));
        }

        Server::new(
            event_loop,
            worker_to_main_channel,
            worker_to_main_scm,
            sessions,
            pool,
            backends,
            None,
            None,
            None,
            config,
            Some(initial_state),
            expects_initial_status,
        )
    }

    #[allow(clippy::too_many_arguments)]
    pub fn new(
        poll: Poll,
        mut channel: ProxyChannel,
        scm: ScmSocket,
        sessions: Rc<RefCell<SessionManager>>,
        pool: Rc<RefCell<Pool>>,
        backends: Rc<RefCell<BackendMap>>,
        http: Option<http::HttpProxy>,
        https: Option<https::HttpsProxy>,
        tcp: Option<tcp::TcpProxy>,
        server_config: ServerConfig,
        initial_state: Option<InitialState>,
        expects_initial_status: bool,
    ) -> Result<Self, ServerError> {
        FEATURES.with(|_features| {
            // initializing feature flags
        });

        poll.registry()
            .register(
                &mut channel,
                Token(0),
                Interest::READABLE | Interest::WRITABLE,
            )
            .map_err(ServerError::RegisterChannel)?;

        METRICS.with(|metrics| {
            if let Some(sock) = (*metrics.borrow_mut()).socket_mut() {
                poll.registry()
                    .register(sock, Token(2), Interest::WRITABLE)
                    .expect("should register the metrics socket");
            }
        });

        let base_sessions_count = sessions.borrow().slab.len();

        let http = Rc::new(RefCell::new(match http {
            Some(http) => http,
            None => {
                let registry = poll
                    .registry()
                    .try_clone()
                    .map_err(ServerError::CloneRegistry)?;

                http::HttpProxy::new(registry, sessions.clone(), pool.clone(), backends.clone())
            }
        }));

        let https = Rc::new(RefCell::new(match https {
            Some(https) => https,
            None => {
                let registry = poll
                    .registry()
                    .try_clone()
                    .map_err(ServerError::CloneRegistry)?;

                https::HttpsProxy::new(registry, sessions.clone(), pool.clone(), backends.clone())
            }
        }));

        let tcp = Rc::new(RefCell::new(match tcp {
            Some(tcp) => tcp,
            None => {
                let registry = poll
                    .registry()
                    .try_clone()
                    .map_err(ServerError::CloneRegistry)?;

                tcp::TcpProxy::new(registry, sessions.clone(), pool.clone(), backends.clone())
            }
        }));

        // UDP proxy is constructed internally (no constructor parameter) so the
        // public `Server::new` / `try_new_from_config` signatures stay
        // unchanged — bin/ is not forced to change for construction.
        let udp = Rc::new(RefCell::new({
            let registry = poll
                .registry()
                .try_clone()
                .map_err(ServerError::CloneRegistry)?;

            udp::UdpProxy::new(
                registry,
                sessions.clone(),
                pool.clone(),
                backends.clone(),
                server_config.max_connections as usize,
                server_config.buffer_size as usize,
            )
        }));

        let mut server = Server {
            accept_queue_timeout: Duration::from_secs(u64::from(
                server_config.accept_queue_timeout,
            )),
            accept_queue: VecDeque::new(),
            evict_on_queue_full: server_config.evict_on_queue_full.unwrap_or(false),
            accept_ready: HashSet::new(),
            backends,
            base_sessions_count,
            channel,
            config_state: ConfigState::new(),
            current_poll_errors: 0,
            health_checker: HealthChecker::new(),
            http,
            https,
            last_sessions_len: 0, // to be reset on server run
            last_shutting_down_message: None,
            last_zombie_check: Instant::now(), // to be reset on server run
            loop_start: Instant::now(),        // to be reset on server run
            started_at: Instant::now(),        // captured once, never reset
            last_saturation_tick: Instant::now(), // 1Hz saturation ticker anchor
            max_poll_errors: 10000,            // TODO: make it configurable?
            pool,
            poll_timeout: Some(Duration::from_millis(1000)), // TODO: make it configurable?
            poll,
            scm_listeners: None,
            scm,
            sessions,
            should_poll_at: None,
            shutting_down: None,
            closing_sessions: Vec::new(),
            yielded_sessions: Vec::new(),
            protected_health_check: ProtectedWorker::default(),
            tcp,
            udp,
            zombie_check_interval: Duration::from_secs(u64::from(
                server_config.zombie_check_interval,
            )),
        };

        // Take the retiring worker's listening sockets BEFORE applying the
        // initial state, because that state is what adopts them.
        //
        // `fork_main_into_worker` (`bin/src/worker.rs`) sends the descriptors
        // over the SCM socket right after the fork — before the main process
        // sends the initial `Status` request — and
        // `ConfigState::produce_initial_state` carries one `ActivateListener`
        // per active listener. Receiving them afterwards left `scm_listeners`
        // empty for the whole initial state, so every `ActivateListener` took
        // the `server_bind` / `udp_bind` branch (which succeeds only because
        // of `SO_REUSEPORT`), the listener came up `active`, and the inherited
        // descriptor was then dropped unused by the next, short-circuiting
        // `activate()` — discarding the accept backlog (TCP/HTTP/HTTPS) and
        // the receive buffer (UDP) the hand-off exists to preserve (sozu#1342).
        //
        // Blocking here is the same blocking read as before, only earlier: the
        // sole production caller of `fork_main_into_worker`,
        // `CommandServer::launch_new_worker`, always passes `Some(listeners)`,
        // so a fresh start receives an empty manifest immediately and every
        // listener binds its own socket exactly as it always did.
        info!("will try to receive listeners");
        server
            .scm
            .set_blocking(true)
            .map_err(|scm_err| ServerError::ScmSocket {
                msg: "Could not set the scm socket to blocking".to_string(),
                scm_err,
            })?;
        let listeners =
            server
                .scm
                .receive_listeners()
                .map_err(|scm_err| ServerError::ScmSocket {
                    msg: "could not receive listeners from the scm socket".to_string(),
                    scm_err,
                })?;
        server
            .scm
            .set_blocking(false)
            .map_err(|scm_err| ServerError::ScmSocket {
                msg: "Could not set the scm socket to unblocking".to_string(),
                scm_err,
            })?;
        info!("received listeners: {:?}", listeners);
        server.scm_listeners = Some(listeners);

        // initialize the worker with the state we got from a file
        if let Some(state) = initial_state {
            for request in state.requests {
                trace!("generating initial config request: {:#?}", request);
                server.notify_proxys(request);
            }

            // do not send back answers to the initialization messages
            QUEUE.with(|queue| {
                (*queue.borrow_mut()).clear();
            });
        }

        server.report_unclaimed_inherited_sockets();

        if expects_initial_status {
            // the main process sends a Status message, so we can notify it
            // when the initial state is loaded; a protected health check
            // bootstrap may come first, see `Server::await_initial_status`
            server.block_channel();
            server.await_initial_status()?;
            server.unblock_channel();
        }

        Ok(server)
    }

    /// Answer the main process on the blocking channel until its `Status`
    /// ends the handshake, before `Server::run` can start any probe.
    ///
    /// A worker that is sent protected health checks first gets the capability
    /// query, then any number of prepare, commit and release cycles (or an
    /// abort), each answered by `ProtectedWorker::handle` straight on the
    /// channel. `BootstrapStage` admits one request at a time and moves on
    /// only when the handler returned the exact typed reply; anything else
    /// gets a static refusal and the loop keeps waiting. There is no timer
    /// here: the loop ends with a valid `Status` or a channel error, and the
    /// main process owns the deadline.
    fn await_initial_status(&mut self) -> Result<(), ServerError> {
        let mut stage = BootstrapStage::Start;
        loop {
            let WorkerRequest { id, mut content } = self
                .channel
                .read_message()
                .map_err(ServerError::ReadInitialMessage)?;
            // the request is never logged: a prepare order carries clear values
            debug!("got initial message {}", id);

            let Some(request_type) = content.request_type.as_mut() else {
                self.answer_initial(worker_response_error(id, REFUSED_INITIAL_REQUEST))?;
                continue;
            };

            if matches!(request_type, RequestType::Status(_)) {
                if !stage.accepts_status() {
                    self.answer_initial(worker_response_error(id, REFUSED_INITIAL_REQUEST))?;
                    continue;
                }
                self.answer_initial(WorkerResponse::ok(id))?;
                return Ok(());
            }

            let Some(step) = stage.step(request_type) else {
                self.answer_initial(worker_response_error(id, REFUSED_INITIAL_REQUEST))?;
                continue;
            };
            let response = self.protected_health_check.handle(
                &id,
                request_type,
                &mut self.config_state,
                &mut self.backends.borrow_mut(),
            );
            match response {
                Some(response) if step.is_answered_by(&response) => {
                    stage = step.next;
                    self.answer_initial(response)?;
                }
                // a refusal of the handler is forwarded as it is
                Some(response) if response.status != ResponseStatus::Ok as i32 => {
                    self.answer_initial(response)?;
                }
                _ => {
                    self.answer_initial(worker_response_error(id, REFUSED_INITIAL_REQUEST))?;
                }
            }
        }
    }

    fn answer_initial(&mut self, response: WorkerResponse) -> Result<(), ServerError> {
        self.channel
            .write_message(&response)
            .map_err(ServerError::WriteInitialAnswer)
    }

    /// The server runs in a loop until a shutdown is ordered
    pub fn run(&mut self) {
        let mut events = Events::with_capacity(1024); // TODO: make event capacity configurable?
        // The session owning each event's token when `epoll_wait` returned,
        // by event position; see `Server::ready_if_still_owned`. Reused
        // across turns so the loop allocates it once.
        let mut event_owners: Vec<Option<Rc<RefCell<dyn ProxySession>>>> =
            Vec::with_capacity(events.capacity());
        self.last_sessions_len = self.sessions.borrow().slab.len();

        self.last_zombie_check = Instant::now();
        self.loop_start = Instant::now();

        loop {
            self.check_for_poll_errors();

            let timeout = self.reset_loop_time_and_get_timeout();

            match self.poll.poll(&mut events, timeout) {
                Ok(_) => self.current_poll_errors = 0,
                Err(error) => {
                    error!("Error while polling events: {:?}", error);
                    self.current_poll_errors += 1;
                    continue;
                }
            }

            // Taken before this turn's events are dispatched, so a session
            // that yields while serving them waits for the next poll instead
            // of running another budget before any other descriptor is read.
            let yielded = std::mem::take(&mut self.yielded_sessions);

            let after_epoll = Instant::now();
            time!(
                names::event_loop::EPOLL_TIME,
                (after_epoll - self.loop_start).as_millis()
            );
            self.loop_start = after_epoll;

            // Before `send_queue`, so an expiry event leaves on this turn.
            tick_metric_detail_leases(after_epoll);

            self.send_queue();

            self.snapshot_event_owners(events.iter().map(|event| event.token()), &mut event_owners);
            for (event, owner) in events.iter().zip(event_owners.drain(..)) {
                match event.token() {
                    // this is the command channel
                    Token(0) => {
                        // An error is a close too: the main process resets
                        // the channel when it closes its end with our
                        // answer still unread. The channel is registered
                        // edge-triggered, so no later event would come and
                        // nothing could drive the worker any more.
                        if event.is_error() {
                            error!(
                                "error on command channel, the main process can no longer reach this worker"
                            );
                            return;
                        }
                        if event.is_read_closed() || event.is_write_closed() {
                            error!("command channel was closed");
                            return;
                        }
                        let ready = Ready::from(event);
                        self.channel.handle_events(ready);

                        // loop here because iterations has borrow issues
                        loop {
                            QUEUE.with(|queue| {
                                if !(*queue.borrow()).is_empty() {
                                    self.channel.interest.insert(Ready::WRITABLE);
                                }
                            });

                            //trace!("WORKER[{}] channel readiness={:?}, interest={:?}, queue={} elements",
                            //  line!(), self.channel.readiness, self.channel.interest, self.queue.len());
                            if self.channel.readiness() == Ready::EMPTY {
                                break;
                            }

                            // exit the big loop if the message is HardStop
                            if self.read_channel_messages_and_notify() {
                                return;
                            }

                            QUEUE.with(|queue| {
                                if !(*queue.borrow()).is_empty() {
                                    self.channel.interest.insert(Ready::WRITABLE);
                                }
                            });

                            self.send_queue();
                        }
                    }
                    // timer tick
                    Token(1) => {
                        while let Some(t) = TIMER.with(|timer| timer.borrow_mut().poll()) {
                            self.timeout(t);
                        }
                    }
                    // metrics socket is writable
                    Token(2) => METRICS.with(|metrics| {
                        (*metrics.borrow_mut()).writable();
                    }),
                    // ListenToken: 1 listener <=> 1 token
                    // ProtocolToken (HTTP/HTTPS/TCP): 1 connection <=> 1 token
                    token if self.health_checker.owns_token(token) => {
                        self.health_checker.ready(token);
                    }
                    token if self.udp.borrow().health_owns_token(token) => {
                        self.udp.borrow_mut().health_ready(token);
                    }
                    token => self.ready_if_still_owned(token, Ready::from(event), owner),
                }
            }
            self.resume_yielded_sessions(yielded);

            if let Some(t) = self.should_poll_at.as_ref()
                && *t <= Instant::now()
            {
                while let Some(t) = TIMER.with(|timer| timer.borrow_mut().poll()) {
                    //info!("polled for timeout: {:?}", t);
                    self.timeout(t);
                }
            }
            self.handle_remaining_readiness();
            self.create_sessions();

            self.should_poll_at = TIMER.with(|timer| timer.borrow().next_poll_date());

            self.zombie_check();
            self.health_checker
                .poll(&self.backends, self.poll.registry());
            // Drive the UDP endpoint health prober (TCP-probe + hysteresis +
            // fail-open). Non-blocking; no-op when no UDP cluster has health
            // configured.
            self.udp.borrow_mut().health_poll();

            // Frontend session gauges. `client.connections` keeps the
            // per-event signal from `SessionManager::incr/decr`; the rest
            // follow the once-per-iteration batching contract this run loop
            // uses for `process.uptime_seconds` / `server.live`.
            //
            // `slab.usage_percent` charts pure slab utilisation against
            // `slab.capacity()`. `slab.accept_threshold_percent` charts how
            // close the slab is to the `at_capacity()` accept gate
            // (`10 + 2 * max_connections`, see
            // `SessionManager::accept_slab_threshold`). Configured slab
            // capacity can be larger than that gate (`slab_entries_per_connection`
            // > 2), so the two gauges are kept independent on purpose.
            {
                let sessions = self.sessions.borrow();
                let nb_connections = sessions.nb_connections;
                let max_connections = sessions.max_connections;
                let slab_len = sessions.slab.len();
                let slab_capacity = sessions.slab.capacity();
                let accept_threshold = sessions.accept_slab_threshold();

                gauge!(names::client::CONNECTIONS, nb_connections);
                gauge!(names::client::CONNECTIONS_MAX, max_connections);
                if let Some(percent) = (nb_connections * 100).checked_div(max_connections) {
                    gauge!(names::client::CONNECTIONS_PERCENT, percent);
                }

                gauge!(names::slab::ENTRIES, slab_len);
                gauge!(names::slab::CAPACITY, slab_capacity);
                if let Some(percent) = (slab_len * 100).checked_div(slab_capacity) {
                    gauge!(names::slab::USAGE_PERCENT, percent);
                }
                if let Some(percent) = (slab_len * 100).checked_div(accept_threshold) {
                    gauge!(names::slab::ACCEPT_THRESHOLD_PERCENT, percent);
                }
            }
            // Buffer pool gauges. `buffer.in_use` replaces the older
            // `buffer.number` (renamed in `lib/src/pool.rs` for naming
            // consistency with the surrounding `buffer.*` keys).
            // `buffer.usage_percent` is computed against the configured
            // `buffer.capacity` so dashboards can chart pool pressure.
            {
                let pool = self.pool.borrow();
                let used = pool.inner.used();
                let capacity = pool.inner.capacity();
                gauge!(names::buffer::IN_USE, used);
                gauge!(names::buffer::CAPACITY, capacity);
                if let Some(percent) = (used * 100).checked_div(capacity) {
                    gauge!(names::buffer::USAGE_PERCENT, percent);
                }
            }
            // 1Hz tick for `accept_queue.saturated_seconds`. Increments once
            // per `ACCEPT_SATURATION_TICK` while `SessionManager::can_accept`
            // is `false`. Distinguishes "queue spent N seconds at max" from
            // "queue briefly hit max" — the binary `accept_queue.backpressure`
            // gauge collapses that duration. Sampled here rather than via a
            // dedicated mio timer because the run loop ticks at least once
            // per `poll_timeout` (1s by default), which is granular enough.
            let now = Instant::now();
            if now.duration_since(self.last_saturation_tick) >= ACCEPT_SATURATION_TICK {
                if !self.sessions.borrow().can_accept {
                    incr!(names::accept_queue::SATURATED_SECONDS);
                }
                self.last_saturation_tick = now;
            }
            // Process / runtime gauges sampled once per loop iteration. Same
            // batch as `client.connections` so dashboards see them update in
            // lock-step.
            gauge!(
                names::process::UPTIME_SECONDS,
                self.started_at.elapsed().as_secs() as usize
            );
            // `server.live` flips to 0 once a graceful shutdown is requested,
            // matching Envoy's `server.live` semantics. L4 health checks
            // (HAProxy / cloud LBs) can poll this gauge to drain a worker
            // before the OS-level termination signal lands.
            gauge!(
                names::server::LIVE,
                if self.shutting_down.is_some() { 0 } else { 1 }
            );
            METRICS.with(|metrics| {
                (*metrics.borrow_mut()).send_data();
            });

            if self.shutting_down.is_some() && self.shut_down_sessions() {
                return;
            }
            if !self.closing_sessions.is_empty() {
                self.drive_closing_sessions();
            }
        }
    }

    fn check_for_poll_errors(&mut self) {
        if self.current_poll_errors >= self.max_poll_errors {
            error!(
                "Something is going very wrong. Last {} poll() calls failed, crashing..",
                self.current_poll_errors
            );
            panic!(
                "poll() calls failed {} times in a row",
                self.current_poll_errors
            );
        }
    }

    fn reset_loop_time_and_get_timeout(&mut self) -> Option<Duration> {
        let now = Instant::now();
        time!(
            names::event_loop::EVENT_LOOP_TIME,
            (now - self.loop_start).as_millis()
        );

        let mut timeout = match self.should_poll_at.as_ref() {
            None => self.poll_timeout,
            Some(i) => {
                if *i <= now {
                    self.poll_timeout
                } else {
                    let dur = *i - now;
                    match self.poll_timeout {
                        None => Some(dur),
                        Some(t) => {
                            if t < dur {
                                Some(t)
                            } else {
                                Some(dur)
                            }
                        }
                    }
                }
            }
        };

        // A session ordered to close is re-asked every tick, during a soft stop
        // or after a `CloseSessions` request, because nothing else wakes the
        // loop when its in-flight work completes.
        if self.shutting_down.is_some() || !self.closing_sessions.is_empty() {
            let shutdown_tick = Duration::from_millis(100);
            timeout = match timeout {
                None => Some(shutdown_tick),
                Some(current) => Some(current.min(shutdown_tick)),
            };
        }

        // A yielded session has work no socket event will announce: poll
        // without blocking, so the other descriptors get their turn and the
        // yielded session is resumed right after.
        if !self.yielded_sessions.is_empty() {
            timeout = Some(Duration::ZERO);
        }

        self.loop_start = now;
        timeout
    }

    /// Returns true if hardstop
    fn read_channel_messages_and_notify(&mut self) -> bool {
        if !self.channel.readiness().is_readable() {
            return false;
        }

        if let Err(e) = self.channel.readable() {
            error!("error reading from channel: {:?}", e);
        }

        loop {
            let request = self.channel.read_message();
            debug!("Received request {:?}", request);
            match request {
                Ok(request) => match request.content.request_type {
                    Some(RequestType::HardStop(_)) => {
                        let req_id = request.id.clone();
                        self.notify(request);
                        if let Err(e) = self.channel.write_message(&WorkerResponse::ok(req_id)) {
                            error!("Could not send ok response to the main process: {}", e);
                        }
                        if let Err(e) = self.channel.run() {
                            error!("Error while running the server channel: {}", e);
                        }
                        return true;
                    }
                    Some(RequestType::SoftStop(_)) => {
                        self.shutting_down = Some(request.id.clone());
                        self.last_sessions_len = self.sessions.borrow().slab.len();
                        self.notify(request);
                    }
                    Some(RequestType::ReturnListenSockets(_)) => {
                        info!("received ReturnListenSockets order");
                        match self.return_listen_sockets() {
                            Ok(_) => push_queue(WorkerResponse::ok(request.id)),
                            Err(error) => push_queue(worker_response_error(
                                request.id,
                                format!("Could not send listeners on scm socket: {error:?}"),
                            )),
                        }
                    }
                    _ => self.notify(request),
                },
                // Not an error per se, occurs when there is nothing to read
                Err(_) => {
                    // if the message was too large, we grow the buffer and retry to read if possible
                    if (self.channel.interest & self.channel.readiness).is_readable() {
                        if let Err(e) = self.channel.readable() {
                            error!("error reading from channel: {:?}", e);
                        }
                        continue;
                    }
                    break;
                }
            }
        }
        false
    }

    /// Scans all sessions that have been inactive for longer than the configured interval
    fn zombie_check(&mut self) {
        let now = Instant::now();
        if now - self.last_zombie_check < self.zombie_check_interval {
            return;
        }
        info!("zombie check");
        // `now` is sampled this iteration and we only get here past the
        // interval gate, so the check timestamp advances monotonically.
        debug_assert!(
            now >= self.last_zombie_check,
            "zombie-check timestamp must never move backwards"
        );
        self.last_zombie_check = now;

        let mut zombie_tokens = HashSet::new();

        // find the zombie sessions
        for (_index, session) in self
            .sessions
            .borrow_mut()
            .slab
            .iter_mut()
            .filter(|(_, c)| now - c.borrow().last_event() > self.zombie_check_interval)
        {
            let session_token = session.borrow().frontend_token();
            if !zombie_tokens.contains(&session_token) {
                session.borrow().print_session();
                zombie_tokens.insert(session_token);
            }
        }

        // Listen/system sessions report `Instant::now()` as their last event,
        // so `now - last_event` is ~0 and never exceeds the interval: a
        // listener can never be collected as a zombie. Assert the set is free
        // of listen protocols before we reap it.
        debug_assert!(
            !self.sessions.borrow().slab.iter().any(|(_, session)| {
                let s = session.borrow();
                zombie_tokens.contains(&s.frontend_token())
                    && matches!(
                        s.protocol(),
                        Protocol::HTTPListen | Protocol::HTTPSListen | Protocol::TCPListen
                    )
            }),
            "zombie reaping must never target a listener session"
        );

        let zombie_count = zombie_tokens.len() as i64;
        count!(names::misc::ZOMBIES, zombie_count);

        let remaining_count = self.shut_down_sessions_by_frontend_tokens(zombie_tokens);
        info!(
            "removing {} zombies ({} remaining entries after close)",
            zombie_count, remaining_count
        );
    }

    /// Calls close on targeted sessions, yields the number of entries in the slab
    /// that were not properly removed
    fn shut_down_sessions_by_frontend_tokens(&self, tokens: HashSet<Token>) -> usize {
        if tokens.is_empty() {
            return 0;
        }

        // close the sessions associated with the tokens
        for token in &tokens {
            if self.sessions.borrow().slab.contains(token.0) {
                let slab_before = self.sessions.borrow().slab.len();
                let session = { self.sessions.borrow_mut().slab.remove(token.0) };
                session.borrow_mut().close();
                self.sessions.borrow_mut().decr();
                // The removed token is truly gone afterwards. The slab may shrink
                // by MORE than one: `close()` also frees the session's backend
                // slab slot(s) (the multi-token pattern), so assert it shrank by
                // at least the frontend slot we just removed, not exactly one.
                debug_assert!(
                    !self.sessions.borrow().slab.contains(token.0),
                    "removed token must be absent from the slab"
                );
                debug_assert!(
                    self.sessions.borrow().slab.len() < slab_before,
                    "removing a present session must free at least its own slab slot"
                );
            }
        }

        // find the entries of closed sessions in the session manager (they should not be there)
        let mut dangling_entries = HashSet::new();
        for (entry_key, session) in &self.sessions.borrow().slab {
            if tokens.contains(&session.borrow().frontend_token()) {
                dangling_entries.insert(entry_key);
            }
        }

        // remove these from the session manager
        let mut dangling_entries_count = 0;
        for entry_key in dangling_entries {
            let mut sessions = self.sessions.borrow_mut();
            if sessions.slab.contains(entry_key) {
                sessions.slab.remove(entry_key);
                dangling_entries_count += 1;
            }
        }
        // Postcondition: no surviving slab entry still references any of the
        // closed frontend tokens — both the direct remove and the dangling
        // sweep together leave the slab clean of these sessions.
        debug_assert!(
            !self
                .sessions
                .borrow()
                .slab
                .iter()
                .any(|(_, session)| tokens.contains(&session.borrow().frontend_token())),
            "no slab entry may reference a closed frontend token after teardown"
        );
        dangling_entries_count
    }

    /// Order sessions to shut down, check that they are all down
    fn shut_down_sessions(&mut self) -> bool {
        let sessions_count = self.sessions.borrow().slab.len();
        let mut sessions_to_shut_down = HashSet::new();

        for (_key, session) in &self.sessions.borrow().slab {
            let mut session = session.borrow_mut();
            if session.shutting_down() {
                debug!(
                    "Server killing session from shutting_down: token={:?}, protocol={:?}",
                    session.frontend_token(),
                    session.protocol()
                );
                sessions_to_shut_down.insert(Token(session.frontend_token().0));
            }
        }
        let _ = self.shut_down_sessions_by_frontend_tokens(sessions_to_shut_down);

        let new_sessions_count = self.sessions.borrow().slab.len();

        if new_sessions_count < sessions_count {
            let now = Instant::now();
            if let Some(last) = self.last_shutting_down_message
                && (now - last) > Duration::from_secs(5)
            {
                info!(
                    "closed {} sessions, {} sessions left, base_sessions_count = {}",
                    sessions_count - new_sessions_count,
                    new_sessions_count,
                    self.base_sessions_count
                );
            }
            self.last_shutting_down_message = Some(now);
        }

        if new_sessions_count <= self.base_sessions_count {
            info!("last session stopped, shutting down!");
            if let Err(e) = self.channel.run() {
                error!("Error while running the server channel: {}", e);
            }
            // self.block_channel();
            let id = self
                .shutting_down
                .take()
                .expect("should have shut down correctly"); // panicking here makes sense actually

            debug!("Responding OK to main process for request {}", id);

            let proxy_response = WorkerResponse::ok(id);
            if let Err(e) = self.channel.write_message(&proxy_response) {
                error!("Could not write response to the main process: {}", e);
            }
            if let Err(e) = self.channel.run() {
                error!("Error while running the server channel: {}", e);
            }
            return true;
        }

        if new_sessions_count < self.last_sessions_len {
            info!(
                "shutting down, {} slab elements remaining (base: {})",
                new_sessions_count - self.base_sessions_count,
                self.base_sessions_count
            );
            self.last_sessions_len = new_sessions_count;
        }

        false
    }

    /// Order every session `target` selects to close, the way a soft stop
    /// orders all of them: `ProxySession::shutting_down` closes it now when it
    /// is idle (or TCP, or WebSocket), and otherwise lets it finish its
    /// in-flight HTTP work, which `Server::drive_closing_sessions` then
    /// watches.
    fn close_sessions(&mut self, target: &SessionTarget) -> ClosedSessions {
        // One slab entry per token: a session appears under its frontend
        // token and under each backend token, so deduplicate by frontend.
        let mut targeted: HashMap<Token, Rc<RefCell<dyn ProxySession>>> = HashMap::new();
        for (_key, session) in &self.sessions.borrow().slab {
            let borrowed = session.borrow();
            if borrowed.targets(target) {
                targeted
                    .entry(borrowed.frontend_token())
                    .or_insert_with(|| session.clone());
            }
        }

        let mut to_close = HashSet::new();
        let mut draining = 0u64;
        for (token, session) in targeted {
            if session.borrow_mut().shutting_down() {
                to_close.insert(token);
            } else {
                draining += 1;
                let already_closing = self.closing_sessions.iter().any(|(_, pending)| {
                    pending.upgrade().is_some_and(|p| Rc::ptr_eq(&p, &session))
                });
                if !already_closing {
                    self.closing_sessions.push((token, Rc::downgrade(&session)));
                }
            }
        }
        let closed = to_close.len() as u64;
        self.shut_down_sessions_by_frontend_tokens(to_close);

        // Post: every session left draining is tracked for the next ticks.
        debug_assert!(
            self.closing_sessions.len() as u64 >= draining,
            "every draining session must be watched by drive_closing_sessions"
        );
        ClosedSessions { closed, draining }
    }

    /// Re-ask every session a `CloseSessions` request left draining whether it
    /// can close now, and close those that can. Run once per loop tick while
    /// any is left.
    fn drive_closing_sessions(&mut self) {
        let pending = std::mem::take(&mut self.closing_sessions);
        let mut to_close = HashSet::new();
        for (token, weak) in pending {
            let Some(session) = weak.upgrade() else {
                continue;
            };
            // The session may have closed on its own and its key been reused.
            let still_there = self
                .sessions
                .borrow()
                .slab
                .get(token.0)
                .is_some_and(|entry| Rc::ptr_eq(entry, &session));
            if !still_there {
                continue;
            }
            if session.borrow_mut().shutting_down() {
                to_close.insert(token);
            } else {
                self.closing_sessions.push((token, weak));
            }
        }
        // Post: a session is either closed now or still watched, never both.
        debug_assert!(
            self.closing_sessions
                .iter()
                .all(|(token, _)| !to_close.contains(token)),
            "a session closed this tick must not stay watched"
        );
        self.shut_down_sessions_by_frontend_tokens(to_close);
    }

    /// Queue `session` for `Server::resume_yielded_sessions` when its last
    /// `ready` pass yielded (`ProxySession::yielded`), once.
    fn queue_if_yielded(&mut self, session: &Rc<RefCell<dyn ProxySession>>) {
        if !session.borrow_mut().yielded() {
            return;
        }
        let queued = self
            .yielded_sessions
            .iter()
            .any(|(_, weak)| weak.upgrade().is_some_and(|q| Rc::ptr_eq(&q, session)));
        if !queued {
            let token = session.borrow().frontend_token();
            self.yielded_sessions.push((token, Rc::downgrade(session)));
        }
    }

    /// Run `ready` again for the sessions that yielded before this turn's
    /// poll (`pending`, taken from `yielded_sessions` right after it), with
    /// the readiness they kept: no new event is merged in. A session that
    /// closed since, its key possibly taken by another session, is skipped,
    /// as `Server::drive_closing_sessions` skips one. A session that yields
    /// again is queued for the next turn (sozu-proxy/sozu#1906).
    fn resume_yielded_sessions(&mut self, pending: Vec<(Token, Weak<RefCell<dyn ProxySession>>)>) {
        for (token, weak) in pending {
            let Some(session) = weak.upgrade() else {
                continue;
            };
            let still_there = self
                .sessions
                .borrow()
                .slab
                .get(token.0)
                .is_some_and(|entry| Rc::ptr_eq(entry, &session));
            if !still_there {
                trace!(
                    "PROXY\t{:?} dropping the resumption of a yielded session that closed since",
                    token
                );
                continue;
            }
            if session.borrow_mut().ready(session.clone()) {
                debug!(
                    "Server killing session from a resumed ready: token={:?}, protocol={:?}",
                    token,
                    session.borrow().protocol()
                );
                self.kill_session(session);
            } else {
                self.queue_if_yielded(&session);
            }
        }
    }

    fn kill_session(&self, session: Rc<RefCell<dyn ProxySession>>) {
        let token = session.borrow().frontend_token();
        let _ = self.shut_down_sessions_by_frontend_tokens(HashSet::from([token]));
    }

    fn send_queue(&mut self) {
        if self.channel.readiness.is_writable() {
            QUEUE.with(|q| {
                let mut queue = q.borrow_mut();
                loop {
                    if let Some(resp) = queue.pop_front() {
                        debug!("Sending response {:?}", resp);
                        if let Err(e) = self.channel.write_message(&resp) {
                            error!("Could not write message {} on the channel: {}", resp, e);
                            queue.push_front(resp);
                        }
                    }

                    if self.channel.back_buf.available_data() > 0
                        && let Err(e) = self.channel.writable()
                    {
                        error!("error writing to channel: {:?}", e);
                    }

                    if !self.channel.readiness.is_writable() {
                        break;
                    }

                    if self.channel.back_buf.available_data() == 0 && queue.is_empty() {
                        break;
                    }
                }
            });
        }
    }

    fn notify(&mut self, message: WorkerRequest) {
        // Run the lease janitor here too, not only from the run loop: the
        // same `now` reading is threaded into `lease_apply` below
        // (`SetMetricDetail` arm) so the whole lease lifecycle for this
        // `notify` call is anchored to one clock read, and a query never
        // reports a lease the janitor is already due to retire.
        let now = std::time::Instant::now();
        tick_metric_detail_leases(now);
        match &message.content.request_type {
            Some(RequestType::ConfigureMetrics(configuration)) => {
                match MetricsConfiguration::try_from(*configuration) {
                    Ok(metrics_config) => {
                        METRICS.with(|metrics| {
                            (*metrics.borrow_mut()).configure(&metrics_config);
                            push_queue(WorkerResponse::ok(message.id));
                        });
                    }
                    Err(e) => {
                        error!("Error configuring metrics: {}", e);
                        push_queue(WorkerResponse::error(message.id, e));
                    }
                }
                return;
            }
            Some(RequestType::QueryMetrics(query_metrics_options)) => {
                METRICS.with(|metrics| {
                    match (*metrics.borrow_mut()).query(query_metrics_options) {
                        Ok(c) => push_queue(WorkerResponse::ok_with_content(message.id, c)),
                        Err(e) => {
                            error!("Error querying metrics: {}", e);
                            push_queue(WorkerResponse::error(message.id, e))
                        }
                    }
                });
                return;
            }
            // Runtime cardinality lease verb — apply, renew, or clear a lease
            // on this worker's `Aggregator`. The lease bumps `effective` to
            // `max(configured, max(active leases))`; expiry runs on the polled
            // janitor, `tick_metric_detail_leases`. Master-side aggregation
            // into `MetricDetailStatus` lands in a follow-up; for now the
            // worker acks with a bare OK so the existing `worker_request`
            // fan-out path can collect.
            Some(RequestType::SetMetricDetail(req)) => {
                // Master populates the peer binding from the connecting
                // `ClientSession` before fan-out (`bin/src/command/
                // requests.rs::worker_request`). A pre-binding caller or a
                // platform without `SO_PEERCRED` yields `PeerBinding::default()`
                // — clears against that lease are accepted from anyone, per
                // the proto contract on `SetMetricDetail.peer_pid`.
                let presented_binding = crate::metrics::PeerBinding {
                    pid: req.peer_pid,
                    // Master sends Crockford-base32 ULIDs (`Ulid::to_string`);
                    // accept those, with a fallback hex parse for callers that
                    // happen to send `0x…` form. A failed parse degrades to
                    // `None` — the lease store treats that as "binding
                    // unknown" per the proto contract.
                    session_ulid: req.peer_session_ulid.as_deref().and_then(|s| {
                        rusty_ulid::Ulid::from_str(s)
                            .map(u128::from)
                            .ok()
                            .or_else(|| u128::from_str_radix(s.trim_start_matches("0x"), 16).ok())
                    }),
                };
                if req.clear.unwrap_or(false) {
                    // Defense-in-depth: the master pre-validates `client_id`
                    // length at the dispatch site, but worker IPC is not
                    // master-only — fuzz harnesses, serial_test-flagged
                    // integration tests, and future internal callers can
                    // issue an oversized clear directly. Mirror the apply
                    // path's `ClientIdTooLong` arm so an unbounded HashMap
                    // lookup is never driven by an operator-supplied string
                    // here either. The reason string echoes the byte length
                    // but not the operator bytes themselves (symmetric with
                    // the audit-column-smuggling guard on the apply path).
                    if req.client_id.len() > crate::metrics::LEASE_CLIENT_ID_MAX_BYTES {
                        let msg = format!(
                            "SetMetricDetail: clear client_id length {} exceeds {} bytes",
                            req.client_id.len(),
                            crate::metrics::LEASE_CLIENT_ID_MAX_BYTES,
                        );
                        error!("{}", msg);
                        push_queue(WorkerResponse::error(message.id.clone(), msg));
                        return;
                    }
                    // Capture transition fields + post-clear snapshot
                    // before releasing the borrow so we can emit an
                    // Event after AND build the WorkerMetricDetailStatus
                    // payload that the master folds into
                    // `MetricDetailStatus.workers[<worker_id>]`. Without
                    // this payload the master used its own view as a
                    // stand-in for the worker's per-aggregator state.
                    let (outcome, effective_after, configured_after, lease_count_after) = METRICS
                        .with(|metrics| {
                            let mut m = metrics.borrow_mut();
                            let outcome = m.lease_clear(&req.client_id, presented_binding);
                            (
                                outcome,
                                m.detail_effective(),
                                m.detail_configured(),
                                m.lease_count(),
                            )
                        });
                    match outcome {
                        crate::metrics::LeaseClearOutcome::Cleared { previous_effective } => {
                            push_metric_detail_transition(
                                previous_effective,
                                effective_after,
                                "lease_clear",
                                Some(req.client_id.clone()),
                            );
                            push_queue(WorkerResponse::ok_with_content(
                                message.id.clone(),
                                worker_metric_detail_status_content(
                                    configured_after,
                                    effective_after,
                                    previous_effective,
                                    lease_count_after,
                                ),
                            ));
                        }
                        crate::metrics::LeaseClearOutcome::NotFound => {
                            // Silent no-op: no lease existed for that
                            // id. The worker's state is unchanged so
                            // previous_effective == effective.
                            push_queue(WorkerResponse::ok_with_content(
                                message.id.clone(),
                                worker_metric_detail_status_content(
                                    configured_after,
                                    effective_after,
                                    effective_after,
                                    lease_count_after,
                                ),
                            ));
                        }
                        crate::metrics::LeaseClearOutcome::Unauthorized => {
                            // Do NOT echo `req.client_id` here: the operator-
                            // supplied bytes flow back through the master's
                            // worker→reason aggregation into the audit line's
                            // `reason=` column, which is sanitised for control
                            // bytes only. The dedicated `lease_id=` audit
                            // column already carries the operator string
                            // through `sanitize_for_audit_kv`, so re-embedding
                            // it here would let a value containing `,` or `=`
                            // forge a sibling KV pair against SIEM consumers
                            // that split on `, key=value`.
                            let msg = "SetMetricDetail: clear refused (peer \
                                 binding does not match the apply-time owner)"
                                .to_owned();
                            error!("{}", msg);
                            push_queue(WorkerResponse::error(message.id.clone(), msg));
                        }
                    }
                    return;
                }
                let detail_proto = match req.detail {
                    Some(d) => d,
                    None => {
                        // Operator-supplied `client_id` is intentionally
                        // omitted from the reason string: the dedicated
                        // `lease_id=` audit column carries it through the
                        // strict KV sanitiser. See the matching comment on
                        // the `Unauthorized` arm above for the column-
                        // smuggling rationale.
                        let msg = "SetMetricDetail without `detail` and without `clear`".to_owned();
                        error!("{}", msg);
                        push_queue(WorkerResponse::error(message.id.clone(), msg));
                        return;
                    }
                };
                let detail_enum = match MetricDetail::try_from(detail_proto) {
                    Ok(d) => d,
                    Err(e) => {
                        let msg =
                            format!("SetMetricDetail: invalid MetricDetail variant {detail_proto}");
                        error!("{}: {}", msg, e);
                        push_queue(WorkerResponse::error(message.id.clone(), msg));
                        return;
                    }
                };
                let level = MetricDetailLevel::from(detail_enum);
                // Bound the worst case BEFORE we touch the aggregator: the
                // proto contract on `SetMetricDetail.ttl_seconds` says the
                // worker rejects values larger than `LEASE_TTL_MAX` so a
                // stuck operator-side renewer (or a buggy third-party client)
                // cannot lock the worker into elevated cardinality. The
                // `Aggregator::lease_apply` clamp is still in place as a
                // defence-in-depth net for code paths that bypass this
                // dispatch (proto fuzzing, future internal callers).
                if let Some(t) = req.ttl_seconds
                    && u64::from(t) > crate::metrics::LEASE_TTL_MAX.as_secs()
                {
                    let msg = format!(
                        "SetMetricDetail: ttl_seconds={t} exceeds LEASE_TTL_MAX={}",
                        crate::metrics::LEASE_TTL_MAX.as_secs()
                    );
                    error!("{}", msg);
                    push_queue(WorkerResponse::error(message.id.clone(), msg));
                    return;
                }
                let ttl_seconds = req.ttl_seconds.filter(|&t| t > 0).unwrap_or_else(|| {
                    // The default fits in a u32 by construction
                    // (LEASE_TTL_DEFAULT = 60 s); the lossy `as u32` cast
                    // is replaced with a checked conversion so any
                    // future tweak past `u32::MAX` seconds (≈ 136 years)
                    // can't silently truncate. Falls through to 60 s on
                    // the theoretical overflow path.
                    u32::try_from(crate::metrics::LEASE_TTL_DEFAULT.as_secs()).unwrap_or(60)
                });
                let ttl = std::time::Duration::from_secs(ttl_seconds.into());
                let (outcome, configured_after, lease_count_after) = METRICS.with(|metrics| {
                    let mut m = metrics.borrow_mut();
                    let outcome =
                        m.lease_apply(req.client_id.clone(), level, ttl, presented_binding, now);
                    (outcome, m.detail_configured(), m.lease_count())
                });
                match outcome {
                    crate::metrics::LeaseApplyOutcome::Applied {
                        previous_effective,
                        new_effective,
                    } => {
                        push_metric_detail_transition(
                            previous_effective,
                            new_effective,
                            "lease_apply",
                            Some(req.client_id.clone()),
                        );
                        push_queue(WorkerResponse::ok_with_content(
                            message.id.clone(),
                            worker_metric_detail_status_content(
                                configured_after,
                                new_effective,
                                previous_effective,
                                lease_count_after,
                            ),
                        ));
                    }
                    crate::metrics::LeaseApplyOutcome::ClientIdTooLong => {
                        let msg = format!(
                            "SetMetricDetail: client_id length {} exceeds {} bytes",
                            req.client_id.len(),
                            crate::metrics::LEASE_CLIENT_ID_MAX_BYTES,
                        );
                        error!("{}", msg);
                        push_queue(WorkerResponse::error(message.id.clone(), msg));
                    }
                    crate::metrics::LeaseApplyOutcome::TableFull => {
                        // Same audit-column-smuggling guard as the
                        // `Unauthorized` and missing-detail arms: the
                        // operator-supplied `client_id` is already rendered
                        // safely through the strict KV sanitiser in the
                        // audit envelope's `lease_id=` column, so we keep
                        // it out of the reason string.
                        let msg = format!(
                            "SetMetricDetail: lease table at capacity ({} entries); reject new \
                             apply — operators must retry after an active lease expires or is \
                             cleared",
                            crate::metrics::LEASE_TABLE_CAP,
                        );
                        error!("{}", msg);
                        push_queue(WorkerResponse::error(message.id.clone(), msg));
                    }
                    crate::metrics::LeaseApplyOutcome::TtlOutOfRange => {
                        // Unreachable in the normal flow: the dispatch-time
                        // gate above already rejected ttl > LEASE_TTL_MAX.
                        // Surface explicitly so any future bypass (proto
                        // fuzzing, internal callers) fails loud rather
                        // than silently capping the lessor's intent.
                        let msg = format!(
                            "SetMetricDetail: ttl exceeds LEASE_TTL_MAX={} (internal contract \
                             violation: dispatch gate should have rejected)",
                            crate::metrics::LEASE_TTL_MAX.as_secs(),
                        );
                        error!("{}", msg);
                        push_queue(WorkerResponse::error(message.id.clone(), msg));
                    }
                    crate::metrics::LeaseApplyOutcome::Unauthorized => {
                        // A renewal arrived against an existing lease whose
                        // apply-time peer binding does not match the
                        // presented one. The `client_id` is intentionally
                        // omitted from the error string — the audit-log
                        // row already carries it in the dedicated
                        // `lease_id` column, and echoing it here would
                        // route operator-controlled bytes through the
                        // freeform reason field.
                        let msg = "SetMetricDetail: renewal refused (peer binding does not \
                                   match the apply-time owner)"
                            .to_owned();
                        error!("{}", msg);
                        push_queue(WorkerResponse::error(message.id.clone(), msg));
                    }
                }
                return;
            }
            Some(RequestType::Logging(logging_filter)) => {
                info!(
                    "{} changing logging filter to {}",
                    message.id, logging_filter
                );
                // there should not be any errors as it was already parsed by the main process
                let (directives, _errors) = logging::parse_logging_spec(logging_filter);
                logging::LOGGER.with(|logger| {
                    logger.borrow_mut().set_directives(directives);
                });
                push_queue(WorkerResponse::ok(message.id));
                return;
            }
            Some(RequestType::QueryClustersHashes(_)) => {
                push_queue(WorkerResponse::ok_with_content(
                    message.id.clone(),
                    ContentType::ClusterHashes(ClusterHashes {
                        map: self
                            .config_state
                            .hash_state()
                            .into_iter()
                            .map(|(cluster_id, hash)| (cluster_id.to_string(), hash))
                            .collect(),
                    })
                    .into(),
                ));
                return;
            }
            Some(RequestType::QueryClusterById(cluster_id)) => {
                push_queue(WorkerResponse::ok_with_content(
                    message.id.clone(),
                    ContentType::Clusters(ClusterInformations {
                        vec: self
                            .config_state
                            .cluster_state(cluster_id)
                            .map_or(vec![], |ci| vec![ci]),
                    })
                    .into(),
                ));
            }
            Some(RequestType::SetMaxConnectionsPerIp(limit)) => {
                let mut sessions = self.sessions.borrow_mut();
                let previous = sessions.max_connections_per_ip;
                sessions.max_connections_per_ip = *limit;
                // Disabling the feature on the fly should not leave
                // stale `(cluster, ip)` entries behind: drain the
                // bookkeeping so a re-enable starts from a clean slate
                // and `cluster_ip_at_limit` does not consult dead state.
                if *limit == 0 {
                    sessions.clear_cluster_ip_tracking();
                }
                info!(
                    "{} updated global max_connections_per_ip from {} to {}",
                    message.id, previous, limit
                );
                push_queue(WorkerResponse::ok(message.id));
                return;
            }
            Some(RequestType::QueryMaxConnectionsPerIp(_)) => {
                let limit = self.sessions.borrow().max_connections_per_ip;
                push_queue(WorkerResponse::ok_with_content(
                    message.id,
                    ContentType::MaxConnectionsPerIpLimit(
                        sozu_command::proto::command::MaxConnectionsPerIpLimit { limit },
                    )
                    .into(),
                ));
                return;
            }
            Some(RequestType::SetMaxConnectionsPerSubnet(limit)) => {
                let mut sessions = self.sessions.borrow_mut();
                let previous = sessions.max_connections_per_subnet;
                sessions.max_connections_per_subnet = *limit;
                // Disabling the subnet limiter on the fly should not
                // leave stale `(cluster, subnet)` entries behind: drain
                // its bookkeeping so a re-enable starts from a clean
                // slate. The per-IP buckets are deliberately untouched —
                // the two limiters are independent.
                if *limit == 0 {
                    sessions.clear_cluster_subnet_tracking();
                }
                info!(
                    "{} updated global max_connections_per_subnet from {} to {}",
                    message.id, previous, limit
                );
                push_queue(WorkerResponse::ok(message.id));
                return;
            }
            Some(RequestType::SetMaxConnectionAttempts(attempts)) => {
                if let Err(reason) =
                    sozu_command::config::validate_max_connection_attempts(*attempts)
                {
                    push_queue(WorkerResponse::error(message.id, reason));
                    return;
                }
                let mut sessions = self.sessions.borrow_mut();
                let previous = sessions.max_connection_attempts;
                sessions.max_connection_attempts = *attempts;
                info!(
                    "{} updated global max_connection_attempts from {} to {}",
                    message.id, previous, attempts
                );
                push_queue(WorkerResponse::ok(message.id));
                return;
            }
            Some(RequestType::CloseSessions(request)) => {
                let target = SessionTarget::from(request);
                let closed = self.close_sessions(&target);
                info!(
                    "{} CloseSessions on {:?}: {} sessions closed, {} draining",
                    message.id, target, closed.closed, closed.draining
                );
                push_queue(WorkerResponse::ok_with_content(
                    message.id,
                    ContentType::ClosedSessions(closed).into(),
                ));
                return;
            }
            Some(RequestType::QueryMaxConnectionsPerSubnet(_)) => {
                let sessions = self.sessions.borrow();
                let limit = sessions.max_connections_per_subnet;
                // The prefixes ride along because they are boot-time
                // only: this reply is the sole runtime surface that can
                // tell an operator what the live cap is a cap ON.
                let ipv4_prefix = sessions.subnet_ipv4_prefix;
                let ipv6_prefix = sessions.subnet_ipv6_prefix;
                drop(sessions);
                push_queue(WorkerResponse::ok_with_content(
                    message.id,
                    ContentType::MaxConnectionsPerSubnetLimit(
                        sozu_command::proto::command::MaxConnectionsPerSubnetLimit {
                            limit,
                            ipv4_prefix,
                            ipv6_prefix,
                        },
                    )
                    .into(),
                ));
                return;
            }
            Some(RequestType::QueryClustersByDomain(domain)) => {
                let cluster_ids = self
                    .config_state
                    .get_cluster_ids_by_domain(domain.hostname.clone(), domain.path.clone());
                let vec = cluster_ids
                    .iter()
                    .filter_map(|cluster_id| self.config_state.cluster_state(cluster_id))
                    .collect();

                push_queue(WorkerResponse::ok_with_content(
                    message.id.clone(),
                    ContentType::Clusters(ClusterInformations { vec }).into(),
                ));
                return;
            }
            // A filter carrying a fingerprint is answered from this worker's
            // own `ConfigState` rather than from its resolver. Note that
            // `get_certificates` tests its domain arm FIRST, so a request
            // carrying BOTH filters is answered by exact SAN equality on the
            // domain and the fingerprint is ignored — the one path where
            // `--domain` is not the trie lookup `query_certificate_for_domain`
            // and `bin`'s `certificates_serving_domain` both perform
            // (sozu#1383). `QueryCertificatesFilters` documents that the two
            // filters do not compound; `doc/configure_cli.md` says to pass one
            // at a time.
            Some(RequestType::QueryCertificatesFromWorkers(filters))
                if filters.fingerprint.is_some() =>
            {
                let certs = self.config_state.get_certificates(filters.clone());
                let response = if !certs.is_empty() {
                    WorkerResponse::ok_with_content(
                        message.id.clone(),
                        ContentType::CertificatesWithFingerprints(CertificatesWithFingerprints {
                            certs,
                        })
                        .into(),
                    )
                } else {
                    worker_response_error(
                        message.id.clone(),
                        "Could not find certificate for this fingerprint",
                    )
                };
                push_queue(response);
                return;
            }
            // if all certificates are queried, or filtered by domain name,
            // the request will be handled by the https proxy
            _other_request => {}
        }
        self.notify_proxys(message);
    }

    pub fn notify_proxys(&mut self, mut request: WorkerRequest) {
        // The public protected verb carries the clear values to the main
        // process only. A worker never applies it as one step: the main
        // process drives the prepare, commit, abort and release orders below.
        if matches!(
            request.content.request_type,
            Some(RequestType::SetProtectedHealthCheck(_))
        ) {
            push_queue(worker_response_error(
                request.id,
                "SetProtectedHealthCheck is not a worker order",
            ));
            return;
        }
        if let Some(request_type) = request.content.request_type.as_mut()
            && let Some(response) = self.protected_health_check.handle(
                &request.id,
                request_type,
                &mut self.config_state,
                &mut self.backends.borrow_mut(),
            )
        {
            push_queue(response);
            return;
        }

        // A health check `port` and a `unix_path` backend never share a
        // cluster. The state refuses the pair, but a refusal other than
        // `ClusterNotFound` does not stop the worker below, so ask first and
        // leave the state, the `BackendMap` and the `HealthChecker` as they
        // are.
        if let Some(reason) = request
            .content
            .request_type
            .as_ref()
            .and_then(|request_type| {
                self.config_state
                    .health_check_port_unix_conflict(request_type)
            })
        {
            push_queue(worker_response_error(request.id, reason));
            return;
        }

        // Read before the dispatch below: `ConfigState::remove_cluster`
        // (`command/src/state.rs`) removes these objects from the state, and
        // the `RemoveCluster` arm removes them from the proxies.
        let cluster_objects = match &request.content.request_type {
            Some(RequestType::RemoveCluster(cluster_id)) => {
                self.config_state.cluster_removal_requests(cluster_id)
            }
            _ => Vec::new(),
        };

        // Snapshot a compensating request before dispatch: removal requests
        // carry only an identity selector, while the stored frontend carries
        // the metadata an exact re-add needs.
        let rollback_candidate = request
            .content
            .request_type
            .as_ref()
            .and_then(|request_type| self.config_state.rollback_for(request_type));
        let mut rollback = None;
        if let Err(e) = self.config_state.dispatch(&request.content) {
            if self
                .config_state
                .removed_with_its_cluster(&request.content, &e)
            {
                // The object went with its cluster (`RemoveCluster` cascades),
                // for instance when a main process that predates the cascade
                // follows its `RemoveCluster` with the removal of each object.
                // There is nothing left to remove: answer ok rather than fail
                // on a missing route.
                debug!("{} already removed with its cluster: {}", request.id, e);
                push_queue(WorkerResponse::ok(request.id));
                return;
            }
            if is_frontend_removal(&request.content)
                && matches!(e, StateError::NotFound { .. } | StateError::NoChange)
            {
                // The state holds no frontend of the named cluster under
                // this identity: `NotFound` when it holds none of that
                // cluster at all (or the route key is another cluster's),
                // `NoChange` when the TCP or UDP frontends it holds for that
                // cluster do not include this one. The proxies look a
                // frontend up by address, hostname and path, or by listener,
                // whatever cluster owns it: applying the removal to them
                // could evict another cluster's route or listener mapping.
                // Refuse it without touching them.
                push_queue(WorkerResponse::error(
                    request.id,
                    format!("Could not execute order on config state: {e}"),
                ));
                return;
            }
            if matches!(e, StateError::ClusterNotFound { .. }) {
                // A backend belongs to its cluster
                // (`ConfigState::add_backend`, `command/src/state.rs`): the
                // state refused one of a cluster it does not hold, never
                // added or already removed. Refuse it without adding it to
                // the `BackendMap`, where it would outlive its cluster.
                push_queue(WorkerResponse::error(
                    request.id,
                    format!("Could not execute order on config state: {e}"),
                ));
                return;
            }
            error!("Could not execute order on config state: {}", e);
        } else {
            rollback = rollback_candidate;
        }

        let req_id = request.id.clone();

        match request.content.request_type {
            Some(RequestType::AddCluster(ref cluster)) => {
                // Mirror the master-side ConfigState::add_cluster check so
                // off-channel paths (TOML reload, SaveState/LoadState, direct
                // API) that smuggle a malformed `cluster.health_check` cannot
                // arm the worker's BackendList::set_health_check_config with
                // a CRLF/NUL/C0 URI or zero thresholds. The SetHealthCheck
                // handler below already runs the same validation; this is the
                // AddCluster mirror.
                if let Some(hc) = cluster.health_check.as_ref()
                    && let Err(reason) = sozu_command::config::validate_health_check_config(hc)
                {
                    push_queue(worker_response_error(req_id, reason));
                    return;
                }
                // Same mirror for the shuffle-sharding knobs (#524): a worker
                // never arms a shard the master would have refused.
                if let Err(reason) = sozu_command::config::validate_shuffle_sharding(
                    cluster.shard_percent,
                    cluster.shard_min_backends,
                    cluster.shard_mode,
                ) {
                    push_queue(worker_response_error(req_id, reason));
                    return;
                }
                if let Some(Err(reason)) = cluster
                    .max_connection_attempts
                    .map(sozu_command::config::validate_max_connection_attempts)
                {
                    push_queue(worker_response_error(req_id, reason));
                    return;
                }
                self.protected_health_check
                    .forget_cluster(&cluster.cluster_id);
                self.add_cluster(cluster);
                // An upsert without a health check clears it like
                // `RemoveHealthCheck`: `add_cluster` already reset the
                // backends, drop the probes still in flight too (#1811).
                if cluster.health_check.is_none() {
                    self.health_checker.remove_cluster(&cluster.cluster_id);
                }
                // Re-arm the metric drain tombstone in case this cluster id
                // was previously removed — without this the drain would
                // continue dropping every emission for the resurrected
                // cluster. Idempotent on a fresh id.
                METRICS.with(|metrics| {
                    (*metrics.borrow_mut()).add_cluster(&cluster.cluster_id);
                });
                //not returning because the message must still be handled by each proxy
            }
            Some(RequestType::RemoveCluster(ref cluster_id)) => {
                self.protected_health_check.forget_cluster(cluster_id);
                self.remove_cluster_objects(&req_id, cluster_id, cluster_objects);
                self.remove_health_check_state(cluster_id);
                METRICS.with(|metrics| {
                    (*metrics.borrow_mut()).remove_cluster(cluster_id);
                });
                //not returning because the message must still be handled by each proxy
            }
            Some(RequestType::SetHealthCheck(ref set)) => {
                if let Err(reason) = sozu_command::config::validate_health_check_config(&set.config)
                {
                    push_queue(worker_response_error(req_id, reason));
                    return;
                }
                self.protected_health_check.forget_cluster(&set.cluster_id);
                self.backends
                    .borrow_mut()
                    .set_health_check_config(&set.cluster_id, Some(set.config.to_owned()));
                push_queue(WorkerResponse::ok(req_id));
                return;
            }
            Some(RequestType::RemoveHealthCheck(ref cluster_id)) => {
                self.remove_health_check_state(cluster_id);
                push_queue(WorkerResponse::ok(req_id));
                return;
            }
            Some(RequestType::AddBackend(ref backend))
            | Some(RequestType::AddUnixBackend(ref backend)) => {
                push_queue(self.add_backend(&req_id, backend));
                return;
            }
            Some(RequestType::RemoveBackend(ref remove_backend)) => {
                push_queue(self.remove_backend(&req_id, remove_backend));
                return;
            }
            _ => {}
        };

        let proxy_destinations = request.content.get_destinations();
        let mut notify_response = None;
        if proxy_destinations.to_http_proxy {
            notify_response = Some(self.http.borrow_mut().notify(request.clone()));
        }
        if proxy_destinations.to_https_proxy {
            let http_proxy_response = self.https.borrow_mut().notify(request.clone());
            if http_proxy_response.is_failure() || notify_response.is_none() {
                notify_response = Some(http_proxy_response);
            }
        }
        if proxy_destinations.to_tcp_proxy {
            let tcp_proxy_response = self.tcp.borrow_mut().notify(request.clone());
            if tcp_proxy_response.is_failure() || notify_response.is_none() {
                notify_response = Some(tcp_proxy_response);
            }
        }
        if proxy_destinations.to_udp_proxy {
            let udp_proxy_response = self.udp.borrow_mut().notify(request.clone());
            if udp_proxy_response.is_failure() || notify_response.is_none() {
                notify_response = Some(udp_proxy_response);
            }
        }
        if let Some(response) = notify_response {
            self.finish_proxy_order(&mut rollback, response);
        }

        match request.content.request_type {
            // special case for adding listeners, because we need to register a listener
            Some(RequestType::AddHttpListener(listener)) => {
                let response = self.notify_add_http_listener(&req_id, listener);
                self.finish_proxy_order(&mut rollback, response);
            }
            Some(
                RequestType::AddHttpsListener(listener)
                | RequestType::AddHttpsListenerWithClientAuth(listener),
            ) => {
                let response = self.notify_add_https_listener(&req_id, listener);
                self.finish_proxy_order(&mut rollback, response);
            }
            Some(RequestType::AddTcpListener(listener)) => {
                let response = self.notify_add_tcp_listener(&req_id, listener);
                self.finish_proxy_order(&mut rollback, response);
            }
            Some(RequestType::AddUdpListener(listener)) => {
                let response = self.notify_add_udp_listener(&req_id, listener);
                self.finish_proxy_order(&mut rollback, response);
            }
            Some(RequestType::UpdateHttpListener(patch)) => {
                let response = self.notify_update_http_listener(&req_id, patch);
                self.finish_proxy_order(&mut rollback, response);
            }
            Some(
                RequestType::UpdateHttpsListener(patch)
                | RequestType::UpdateHttpsListenerWithClientAuth(patch),
            ) => {
                let response = self.notify_update_https_listener(&req_id, patch);
                self.finish_proxy_order(&mut rollback, response);
            }
            Some(RequestType::UpdateTcpListener(patch)) => {
                let response = self.notify_update_tcp_listener(&req_id, patch);
                self.finish_proxy_order(&mut rollback, response);
            }
            Some(RequestType::UpdateUdpListener(patch)) => {
                let response = self.notify_update_udp_listener(&req_id, patch);
                self.finish_proxy_order(&mut rollback, response);
            }
            Some(RequestType::RemoveListener(ref remove)) => {
                debug!("{} remove {:?} listener {:?}", req_id, remove.proxy, remove);
                // We only remove a listener that was previously added, so the
                // base count is at least 1 — the subtraction cannot underflow.
                debug_assert!(
                    self.base_sessions_count > 0,
                    "removing a listener with base_sessions_count == 0 would underflow"
                );
                self.base_sessions_count -= 1;
                let listener_type = ListenerType::try_from(remove.proxy);
                // `RemoveListener` is the sole release point of the slab slot
                // `AddListener` reserved (see `reserve_listen_token`): no
                // proxy's `remove_listener` touches the session slab, so
                // without this the slot leaks one slab key per add/remove
                // cycle — and `base_sessions_count`, decremented just above,
                // drifts below the number of reserved slots. Read the token
                // BEFORE the proxy drops the listener that holds it.
                let address = remove.listener_key();
                let listen_token = match listener_type {
                    Ok(ListenerType::Http) => self.http.borrow().listener_token(&address),
                    Ok(ListenerType::Https) => self.https.borrow().listener_token(&address),
                    Ok(ListenerType::Tcp) => self.tcp.borrow().listener_token(&address),
                    Ok(ListenerType::Udp) => self.udp.borrow().listener_token(&address),
                    Err(_) => None,
                };
                let response = match listener_type {
                    Ok(ListenerType::Http) => self.http.borrow_mut().notify(request),
                    Ok(ListenerType::Https) => self.https.borrow_mut().notify(request),
                    Ok(ListenerType::Tcp) => self.tcp.borrow_mut().notify(request),
                    Ok(ListenerType::Udp) => self.udp.borrow_mut().notify(request),
                    Err(_) => WorkerResponse::error(req_id, "Wrong variant ListenerType"),
                };
                if let Some(token) = listen_token {
                    self.sessions.borrow_mut().slab.try_remove(token.0);
                    // Same rule as the deactivate arms: a listen token with no
                    // slab slot must not stay queued for a deferred accept.
                    self.accept_ready.remove(&ListenToken(token.0));
                    info!("released listen token {:?}", token);
                }
                self.finish_proxy_order(&mut rollback, response);
            }
            Some(RequestType::ActivateListener(ref activate)) => {
                let response = self.notify_activate_listener(&req_id, activate);
                self.finish_proxy_order(&mut rollback, response);
            }
            Some(RequestType::DeactivateListener(ref deactivate)) => {
                let response = self.notify_deactivate_listener(&req_id, deactivate);
                self.finish_proxy_order(&mut rollback, response);
            }
            _other_request => {}
        };
    }

    /// Queue a proxy response and restore the worker's persisted view when a
    /// single-plane runtime order was refused after `ConfigState::dispatch`.
    fn finish_proxy_order(&mut self, rollback: &mut Option<Request>, response: WorkerResponse) {
        if response.is_failure()
            && let Some(rollback) = rollback.take()
        {
            debug_assert!(
                rollback.request_type.is_some(),
                "a ConfigState rollback must carry one request"
            );
            if let Err(error) = self.config_state.dispatch(&rollback) {
                error!(
                    "Could not roll back worker config state after a proxy refused an order: {}",
                    error
                );
            }
            debug_assert!(
                rollback.request_type.is_some(),
                "dispatching a rollback must not consume its request type"
            );
        }

        push_queue(response);
    }

    fn add_cluster(&mut self, cluster: &Cluster) {
        let mut backends = self.backends.borrow_mut();
        backends.set_load_balancing_policy_for_cluster(
            &cluster.cluster_id,
            LoadBalancingAlgorithms::try_from(cluster.load_balancing).unwrap_or_default(),
            cluster
                .load_metric
                .and_then(|n| LoadMetric::try_from(n).ok()),
        );
        backends.set_shuffle_sharding_for_cluster(
            &cluster.cluster_id,
            ShuffleSharding::from_cluster(
                cluster.shard_percent,
                cluster.shard_min_backends,
                cluster.shard_mode,
            ),
        );
        backends.set_health_check_config(&cluster.cluster_id, cluster.health_check.to_owned());
        backends.set_cluster_http2(&cluster.cluster_id, cluster.http2.unwrap_or(false));
    }

    fn add_backend(&mut self, req_id: &str, add_backend: &AddBackend) -> WorkerResponse {
        if let Some(Err(reason)) = add_backend
            .unix_path
            .as_deref()
            .map(sozu_command::config::parse_backend_unix_path)
        {
            return WorkerResponse::error(req_id, format!("invalid unix_path: {reason}"));
        }
        let new_backend = Backend::new(
            &add_backend.backend_id,
            add_backend.address.into(),
            add_backend.sticky_id.clone(),
            add_backend.load_balancing_parameters,
            add_backend.backup,
        )
        .with_unix_path(add_backend.unix_path.clone());
        self.backends
            .borrow_mut()
            .add_backend(&add_backend.cluster_id, new_backend);

        WorkerResponse::ok(req_id)
    }

    /// Remove from the proxies and the `BackendMap` the frontends and backends
    /// of a cluster being removed, `requests` being the removal orders
    /// `ConfigState::cluster_removal_requests` (`command/src/state.rs`)
    /// returned for it. Each object goes through the removal its own order
    /// takes, so the router, the tags, the TCP and UDP listener to cluster
    /// mappings and the per-backend metrics are cleaned the same way.
    ///
    /// A session already established keeps the backend connection it holds
    /// and drains on it; only new requests and connections stop routing. A
    /// failure is logged and does not stop the others: the `RemoveCluster`
    /// answer stays the single response the main process waits for.
    fn remove_cluster_objects(&mut self, req_id: &str, cluster_id: &str, requests: Vec<Request>) {
        for request in requests {
            let result = match request.request_type {
                Some(
                    RequestType::RemoveHttpFrontend(front)
                    | RequestType::RemoveHttpFrontendWithHeaderPredicates(front),
                ) => self.http.borrow_mut().remove_http_frontend(front),
                Some(
                    RequestType::RemoveHttpsFrontend(front)
                    | RequestType::RemoveHttpsFrontendWithHeaderPredicates(front),
                ) => self
                    .https
                    .borrow_mut()
                    .remove_https_frontend(front)
                    .map(|_| ()),
                Some(RequestType::RemoveTcpFrontend(front)) => {
                    self.tcp.borrow_mut().remove_tcp_front(front)
                }
                Some(RequestType::RemoveUdpFrontend(front)) => {
                    self.udp.borrow_mut().remove_udp_front(front)
                }
                Some(RequestType::RemoveBackend(backend)) => {
                    let response = self.remove_backend(req_id, &backend);
                    debug_assert_eq!(
                        response.status,
                        ResponseStatus::Ok as i32,
                        "Server::remove_backend always answers ok"
                    );
                    Ok(())
                }
                other => {
                    debug_assert!(
                        false,
                        "cluster_removal_requests only returns frontend and backend removals, got {other:?}"
                    );
                    Ok(())
                }
            };
            if let Err(error) = result {
                warn!(
                    "{} removing cluster {}: could not remove one of its objects: {}",
                    req_id, cluster_id, error
                );
            }
        }
        self.backends.borrow_mut().remove_cluster(cluster_id);
        debug_assert!(
            !self.backends.borrow().backends.contains_key(cluster_id),
            "a removed cluster must leave no backend list behind"
        );
    }

    /// Stop health checking `cluster_id`: drop its in-flight probes, then
    /// forget its configuration and reset every backend it marked DOWN, since
    /// no probe will ever mark them UP again (#1811). Dropping the probes first
    /// keeps one already in flight from marking a backend DOWN after the reset.
    fn remove_health_check_state(&mut self, cluster_id: &str) {
        self.protected_health_check.forget_cluster(cluster_id);
        self.health_checker.remove_cluster(cluster_id);
        self.backends
            .borrow_mut()
            .set_health_check_config(cluster_id, None);
    }

    fn remove_backend(&mut self, req_id: &str, backend: &RemoveBackend) -> WorkerResponse {
        let address = backend.address.into();
        // Runtime removal keys on `(backend_id, address)`, as `ConfigState`
        // does: another id at the same address stays (#1821). The metrics
        // layer is id-keyed, so one row aggregates every entry carrying the
        // id: it is dropped only once no entry of that id remains at any
        // address, and is still tidied when the backend was already gone,
        // clearing any orphan row.
        let id_still_live = {
            let mut backends = self.backends.borrow_mut();
            backends.remove_backend(&backend.cluster_id, &backend.backend_id, &address);
            backends.has_backend_id(&backend.cluster_id, &backend.backend_id)
        };
        if !id_still_live {
            METRICS.with(|metrics| {
                (*metrics.borrow_mut()).remove_backend(&backend.cluster_id, &backend.backend_id);
            });
        }

        WorkerResponse::ok(req_id)
    }

    fn notify_add_http_listener(
        &mut self,
        req_id: &str,
        listener: HttpListenerConfig,
    ) -> WorkerResponse {
        debug!("{} add http listener {:?}", req_id, listener);

        let address: SocketAddr = listener.address.into();
        let retained_frontends = self
            .config_state
            .http_fronts
            .values()
            .filter(|frontend| frontend.address == address)
            .cloned()
            .map(Into::into)
            .collect();

        if self.sessions.borrow().at_capacity() {
            return worker_response_error(req_id, "session list is full, cannot add a listener");
        }

        let mut session_manager = self.sessions.borrow_mut();
        // The vacant entry's key is free now and becomes the listener's token.
        let slab_before = session_manager.slab.len();
        debug_assert!(
            !session_manager
                .slab
                .contains(session_manager.slab.vacant_key()),
            "the next vacant slab key must be free before insertion"
        );
        let entry = session_manager.slab.vacant_entry();
        let token = Token(entry.key());

        match self.http.borrow_mut().add_listener_with_retained_frontends(
            listener,
            token,
            retained_frontends,
        ) {
            Ok(_token) => {
                entry.insert(Rc::new(RefCell::new(ListenSession {
                    protocol: Protocol::HTTPListen,
                })));
                // The listener session occupies exactly the token's slab key,
                // and the slab grew by exactly one slot.
                debug_assert!(
                    session_manager.slab.contains(token.0),
                    "listener insert must occupy the token's slab key"
                );
                debug_assert_eq!(
                    session_manager.slab.len(),
                    slab_before + 1,
                    "adding a listener must occupy exactly one slab slot"
                );
                self.base_sessions_count += 1;
                WorkerResponse::ok(req_id)
            }
            Err(e) => worker_response_error(req_id, format!("Could not add HTTP listener: {e}")),
        }
    }

    fn notify_add_https_listener(
        &mut self,
        req_id: &str,
        listener: HttpsListenerConfig,
    ) -> WorkerResponse {
        debug!("{} add https listener {:?}", req_id, listener);

        let wire_address = listener.address;
        let address: SocketAddr = wire_address.into();
        let retained_frontends = self
            .config_state
            .https_fronts
            .values()
            .filter(|frontend| frontend.address == address)
            .cloned()
            .map(Into::into)
            .collect();
        let retained_certificates = self
            .config_state
            .certificate_add_requests_for_address(address);

        if self.sessions.borrow().at_capacity() {
            return worker_response_error(req_id, "session list is full, cannot add a listener");
        }

        let mut session_manager = self.sessions.borrow_mut();
        let slab_before = session_manager.slab.len();
        debug_assert!(
            !session_manager
                .slab
                .contains(session_manager.slab.vacant_key()),
            "the next vacant slab key must be free before insertion"
        );
        let entry = session_manager.slab.vacant_entry();
        let token = Token(entry.key());

        match self.https.borrow_mut().add_listener_with_retained_state(
            listener.clone(),
            token,
            retained_frontends,
            retained_certificates,
        ) {
            Ok(_token) => {
                entry.insert(Rc::new(RefCell::new(ListenSession {
                    protocol: Protocol::HTTPSListen,
                })));
                debug_assert!(
                    session_manager.slab.contains(token.0),
                    "listener insert must occupy the token's slab key"
                );
                debug_assert_eq!(
                    session_manager.slab.len(),
                    slab_before + 1,
                    "adding a listener must occupy exactly one slab slot"
                );
                self.base_sessions_count += 1;
                WorkerResponse::ok(req_id)
            }
            Err(e) => worker_response_error(req_id, format!("Could not add HTTPS listener: {e}")),
        }
    }

    fn notify_add_tcp_listener(
        &mut self,
        req_id: &str,
        listener: CommandTcpListener,
    ) -> WorkerResponse {
        debug!("{} add tcp listener {:?}", req_id, listener);

        let address: SocketAddr = listener.address.into();
        let retained_frontends = self
            .config_state
            .tcp_fronts
            .values()
            .flatten()
            .filter(|frontend| frontend.address == address)
            .cloned()
            .map(Into::into)
            .collect();

        if self.sessions.borrow().at_capacity() {
            return worker_response_error(req_id, "session list is full, cannot add a listener");
        }

        let mut session_manager = self.sessions.borrow_mut();
        let slab_before = session_manager.slab.len();
        debug_assert!(
            !session_manager
                .slab
                .contains(session_manager.slab.vacant_key()),
            "the next vacant slab key must be free before insertion"
        );
        let entry = session_manager.slab.vacant_entry();
        let token = Token(entry.key());

        match self.tcp.borrow_mut().add_listener_with_retained_frontends(
            listener,
            token,
            retained_frontends,
        ) {
            Ok(_token) => {
                entry.insert(Rc::new(RefCell::new(ListenSession {
                    protocol: Protocol::TCPListen,
                })));
                debug_assert!(
                    session_manager.slab.contains(token.0),
                    "listener insert must occupy the token's slab key"
                );
                debug_assert_eq!(
                    session_manager.slab.len(),
                    slab_before + 1,
                    "adding a listener must occupy exactly one slab slot"
                );
                self.base_sessions_count += 1;
                WorkerResponse::ok(req_id)
            }
            Err(e) => worker_response_error(req_id, format!("Could not add TCP listener: {e}")),
        }
    }

    fn notify_add_udp_listener(
        &mut self,
        req_id: &str,
        listener: CommandUdpListener,
    ) -> WorkerResponse {
        debug!("{} add udp listener {:?}", req_id, listener);

        let address: SocketAddr = listener.address.into();
        let retained_frontends = self
            .config_state
            .udp_fronts
            .values()
            .flatten()
            .filter(|frontend| frontend.address == address)
            .cloned()
            .map(Into::into)
            .collect();

        if self.sessions.borrow().at_capacity() {
            return worker_response_error(req_id, "session list is full, cannot add a listener");
        }

        let mut session_manager = self.sessions.borrow_mut();
        let entry = session_manager.slab.vacant_entry();
        let token = Token(entry.key());

        match self.udp.borrow_mut().add_listener_with_retained_frontends(
            listener,
            token,
            retained_frontends,
        ) {
            Ok(_token) => {
                entry.insert(Rc::new(RefCell::new(ListenSession {
                    protocol: Protocol::UDPListen,
                })));
                self.base_sessions_count += 1;
                WorkerResponse::ok(req_id)
            }
            Err(e) => worker_response_error(req_id, format!("Could not add UDP listener: {e}")),
        }
    }

    fn notify_update_udp_listener(
        &mut self,
        req_id: &str,
        patch: UpdateUdpListenerConfig,
    ) -> WorkerResponse {
        debug!("{} update udp listener {:?}", req_id, patch.address);
        match self.udp.borrow_mut().update_listener(patch) {
            Ok(()) => WorkerResponse::ok(req_id),
            Err(e) => worker_response_error(req_id, format!("Could not update UDP listener: {e}")),
        }
    }

    fn notify_update_http_listener(
        &mut self,
        req_id: &str,
        patch: UpdateHttpListenerConfig,
    ) -> WorkerResponse {
        debug!("{} update http listener {:?}", req_id, patch.address);
        match self.http.borrow_mut().update_listener(patch) {
            Ok(()) => WorkerResponse::ok(req_id),
            Err(e) => worker_response_error(req_id, format!("Could not update HTTP listener: {e}")),
        }
    }

    fn notify_update_https_listener(
        &mut self,
        req_id: &str,
        patch: UpdateHttpsListenerConfig,
    ) -> WorkerResponse {
        debug!("{} update https listener {:?}", req_id, patch.address);
        match self.https.borrow_mut().update_listener(patch) {
            Ok(()) => WorkerResponse::ok(req_id),
            Err(e) => {
                worker_response_error(req_id, format!("Could not update HTTPS listener: {e}"))
            }
        }
    }

    fn notify_update_tcp_listener(
        &mut self,
        req_id: &str,
        patch: UpdateTcpListenerConfig,
    ) -> WorkerResponse {
        debug!("{} update tcp listener {:?}", req_id, patch.address);
        match self.tcp.borrow_mut().update_listener(patch) {
            Ok(()) => WorkerResponse::ok(req_id),
            Err(e) => worker_response_error(req_id, format!("Could not update TCP listener: {e}")),
        }
    }

    /// Re-arm a deactivated listener's slab slot with the inert
    /// [`ListenSession`] placeholder that `AddListener` first installed there.
    ///
    /// DESIGN (2026-09-18) — a listener owns exactly one slab slot for its
    /// whole `AddListener` -> `RemoveListener` lifetime, activated or not.
    /// The alternative was to stop retaining the token across a deactivate and
    /// allocate a fresh slab key on every activate; it was rejected because
    /// the token is the *identity* of a listener in all four proxies — their
    /// `listeners` maps (plus UDP's `managers`, `cluster_for_listener` and
    /// `listener_sessions`) are keyed by it, `give_back_listener` hands the
    /// same token back and `activate()` returns it again — so a fresh key
    /// would mean rekeying five maps across four proxies on every activate,
    /// far more machinery than the defect warrants.
    ///
    /// Reserving the slot is what makes the retained token honest, and it
    /// closes both halves of the defect at once:
    ///
    /// * the slot is never vacant, so `Server::ready` keeps dispatching the
    ///   listen token after a deactivate/activate cycle. Freeing it left the
    ///   socket re-registered under a token with no slab entry — `ready()`
    ///   silently drops the event, and the reactivated listener was deaf
    ///   while `ActivateListener` had reported success;
    /// * the slab can never hand that key to another session, so the UDP
    ///   activate path can no longer overwrite a live session with its
    ///   listener session. Mode 2 is gone by construction, not by check.
    ///
    /// `RemoveListener` is the sole release point (see `notify_proxys`). No
    /// proxy's `remove_listener` touches the session slab, so before this
    /// change the deactivate arms were the ONLY code that ever freed a
    /// listener's slot — removing a listener that had not been deactivated
    /// first already stranded its slab key (and, for UDP, the last
    /// `UdpListenerSession` reference, hence the listener's open `UdpSocket`).
    /// Reserving the slot here makes that release mandatory rather than
    /// incidental.
    fn reserve_listen_token(&mut self, token: Token, protocol: Protocol) {
        let mut sessions = self.sessions.borrow_mut();
        match sessions.slab.get_mut(token.0) {
            Some(slot) => {
                *slot = Rc::new(RefCell::new(ListenSession { protocol }));
                info!("reserved listen token {:?} for {:?}", token, protocol);
            }
            // Unreachable while the lifetime above holds: `AddListener`
            // inserted the slot and only `RemoveListener` frees it. Loud
            // rather than silent, because the listener is deaf if it happens.
            None => error!(
                "listen token {:?} ({:?}) has no slab slot to reserve; the listener would not be reachable if reactivated",
                token, protocol
            ),
        }
    }

    /// Say what the initial state left unclaimed in the SCM table.
    ///
    /// A descriptor is kept whenever no listener exists at its address yet,
    /// because an `AddListener` + `ActivateListener` pair can still arrive and
    /// adopt it — `UpgradeWorkerTask` (`bin/src/command/upgrade.rs`) scatters
    /// `generate_activate_requests()` only after `Server::new` has returned, so
    /// there is no point during startup at which a leftover is provably
    /// unclaimable. Retention is therefore open-ended, and the one thing this
    /// worker can do is make it visible.
    ///
    /// An address that already has a listener is the expected shape (an initial
    /// state carrying the listeners inactive, activated straight afterwards) and
    /// is reported at `info`. An address with no listener at all is the one an
    /// operator needs to see: nothing in this worker's state mentions it, so the
    /// descriptor is the retiring worker's listening socket, still bound with
    /// `SO_REUSEPORT` and registered with no event loop, which is expected to
    /// keep taking a share of new connections that nothing accepts. (That
    /// expectation follows from `SO_REUSEPORT` load-balancing semantics; nothing
    /// here measures it.) That is reported at `warn`.
    fn report_unclaimed_inherited_sockets(&self) {
        let Some(scm_listeners) = self.scm_listeners.as_ref() else {
            return;
        };
        let leftovers: [(&str, &Vec<(ListenerKey, RawFd)>); 4] = [
            ("HTTP", &scm_listeners.http),
            ("HTTPS", &scm_listeners.tls),
            ("TCP", &scm_listeners.tcp),
            ("UDP", &scm_listeners.udp),
        ];
        for (label, table) in leftovers {
            for (address, fd) in table {
                let has_listener = match label {
                    "HTTP" => self.http.borrow().listener_token(address).is_some(),
                    "HTTPS" => self.https.borrow().listener_token(address).is_some(),
                    "TCP" => self.tcp.borrow().listener_token(address).is_some(),
                    _ => self.udp.borrow().listener_token(address).is_some(),
                };
                if has_listener {
                    info!(
                        "inherited {} listening socket for {} (fd {}) is held for an \
                         ActivateListener this worker has not processed yet",
                        label, address, fd
                    );
                } else {
                    warn!(
                        "inherited {} listening socket for {} (fd {}) is held with no listener \
                         at that address: until an AddListener plus ActivateListener for {} \
                         adopts it, it is expected to keep taking a share of new connections \
                         that nothing accepts (inferred from SO_REUSEPORT semantics, not \
                         measured here)",
                        label, address, fd, address
                    );
                }
            }
        }
    }

    fn notify_activate_listener(
        &mut self,
        req_id: &str,
        activate: &ActivateListener,
    ) -> WorkerResponse {
        use crate::InheritedSocketFate;

        debug!(
            "{} activate {:?} listener {:?}",
            req_id, activate.proxy, activate
        );

        let address = activate.listener_key();

        match ListenerType::try_from(activate.proxy) {
            Ok(ListenerType::Http) => {
                let listener = match self.http.borrow().inherited_socket_fate(&address) {
                    InheritedSocketFate::Adopted => self
                        .scm_listeners
                        .as_mut()
                        .and_then(|listeners| listeners.get_http(&address))
                        // SAFETY: `fd` was just received from the supervisor via SCM_RIGHTS
                        // (see `command/src/scm_socket.rs`) and is not owned elsewhere — the
                        // `Listeners` table removes it on `get_http`. Ownership transfers to
                        // the mio wrapper, whose `Drop` closes the descriptor.
                        .map(|fd| unsafe { MioTcpListener::from_raw_fd(fd) }),
                    InheritedSocketFate::Refused => {
                        discard_inherited_socket::<MioTcpListener>(
                            self.scm_listeners
                                .as_mut()
                                .and_then(|listeners| listeners.get_http(&address)),
                            &address,
                            "HTTP",
                        );
                        None
                    }
                    // Left in the SCM table on purpose: a later `AddListener` +
                    // `ActivateListener` for this address can still adopt it.
                    InheritedSocketFate::Unclaimed => None,
                };

                let activated_token = self.http.borrow_mut().activate_listener(&address, listener);
                match activated_token {
                    Ok(token) => {
                        self.accept(ListenToken(token.0), Protocol::HTTPListen);
                        WorkerResponse::ok(req_id)
                    }
                    Err(activate_error) => worker_response_error(
                        req_id,
                        format!("Could not activate HTTP listener: {activate_error}"),
                    ),
                }
            }
            Ok(ListenerType::Https) => {
                let listener = match self.https.borrow().inherited_socket_fate(&address) {
                    InheritedSocketFate::Adopted => self
                        .scm_listeners
                        .as_mut()
                        .and_then(|listeners| listeners.get_https(&address))
                        // SAFETY: `fd` was just received from the supervisor via SCM_RIGHTS
                        // (see `command/src/scm_socket.rs`) and is not owned elsewhere — the
                        // `Listeners` table removes it on `get_https`. Ownership transfers to
                        // the mio wrapper, whose `Drop` closes the descriptor.
                        .map(|fd| unsafe { MioTcpListener::from_raw_fd(fd) }),
                    InheritedSocketFate::Refused => {
                        discard_inherited_socket::<MioTcpListener>(
                            self.scm_listeners
                                .as_mut()
                                .and_then(|listeners| listeners.get_https(&address)),
                            &address,
                            "HTTPS",
                        );
                        None
                    }
                    // Left in the SCM table on purpose: a later `AddListener` +
                    // `ActivateListener` for this address can still adopt it.
                    InheritedSocketFate::Unclaimed => None,
                };

                let activated_token = self
                    .https
                    .borrow_mut()
                    .activate_listener(&address, listener);
                match activated_token {
                    Ok(token) => {
                        self.accept(ListenToken(token.0), Protocol::HTTPSListen);
                        WorkerResponse::ok(req_id)
                    }
                    Err(activate_error) => worker_response_error(
                        req_id,
                        format!("Could not activate HTTPS listener: {activate_error}"),
                    ),
                }
            }
            Ok(ListenerType::Tcp) => {
                let listener = match self.tcp.borrow().inherited_socket_fate(&address) {
                    InheritedSocketFate::Adopted => self
                        .scm_listeners
                        .as_mut()
                        .and_then(|listeners| listeners.get_tcp(&address))
                        // SAFETY: `fd` was just received from the supervisor via SCM_RIGHTS
                        // (see `command/src/scm_socket.rs`) and is not owned elsewhere — the
                        // `Listeners` table removes it on `get_tcp`. Ownership transfers to
                        // the mio wrapper, whose `Drop` closes the descriptor.
                        .map(|fd| unsafe { MioTcpListener::from_raw_fd(fd) }),
                    InheritedSocketFate::Refused => {
                        discard_inherited_socket::<MioTcpListener>(
                            self.scm_listeners
                                .as_mut()
                                .and_then(|listeners| listeners.get_tcp(&address)),
                            &address,
                            "TCP",
                        );
                        None
                    }
                    // Left in the SCM table on purpose: a later `AddListener` +
                    // `ActivateListener` for this address can still adopt it.
                    InheritedSocketFate::Unclaimed => None,
                };

                let listener_token = self.tcp.borrow_mut().activate_listener(&address, listener);
                match listener_token {
                    Ok(token) => {
                        self.accept(ListenToken(token.0), Protocol::TCPListen);
                        WorkerResponse::ok(req_id)
                    }
                    Err(activate_error) => worker_response_error(
                        req_id,
                        format!("Could not activate TCP listener: {activate_error}"),
                    ),
                }
            }
            Ok(ListenerType::Udp) => {
                let socket = match self.udp.borrow().inherited_socket_fate(&address) {
                    InheritedSocketFate::Adopted => self
                        .scm_listeners
                        .as_mut()
                        .and_then(|listeners| listeners.get_udp(&address))
                        // SAFETY: `fd` was just received from the supervisor via
                        // SCM_RIGHTS (see `command/src/scm_socket.rs`) and is not
                        // owned elsewhere — `Listeners::get_udp` removes it from
                        // the table. Ownership transfers to the mio `UdpSocket`
                        // wrapper, whose `Drop` closes the descriptor. `O_NONBLOCK`
                        // + `SO_REUSE*` are file-description flags preserved across
                        // SCM + exec.
                        .map(|fd| unsafe { MioUdpSocket::from_raw_fd(fd) }),
                    InheritedSocketFate::Refused => {
                        discard_inherited_socket::<MioUdpSocket>(
                            self.scm_listeners
                                .as_mut()
                                .and_then(|listeners| listeners.get_udp(&address)),
                            &address,
                            "UDP",
                        );
                        None
                    }
                    // Left in the SCM table on purpose: a later `AddListener` +
                    // `ActivateListener` for this address can still adopt it.
                    InheritedSocketFate::Unclaimed => None,
                };

                let activated_token = self.udp.borrow_mut().activate_listener(&address, socket);
                match activated_token {
                    Ok(token) => {
                        // UDP never uses accept()/create_session: replace the
                        // `ListenSession` placeholder at the listener token with
                        // the real `UdpListenerSession` so the READABLE
                        // registration drives `Server::ready`'s generic path
                        // into `UdpListenerSession::update_readiness`.
                        //
                        // Install it once per ACTIVATION, not once per request.
                        // `Ok(token)` does not mean this call activated
                        // anything: `UdpListener::activate` short-circuits on
                        // its own `active` flag and answers `Ok(self.token)` for
                        // a listener that is already up. Nothing upstream
                        // filters that repeat out either —
                        // `ConfigState::activate_listener` sets `active = true`
                        // and answers `Ok(())` however often it is asked, so the
                        // main process forwards a second
                        // `sozu listener udp activate`, and
                        // `ConfigState::generate_requests` re-emits
                        // `ActivateListener` for every active listener on each
                        // state replay.
                        //
                        // Rebuilding on a repeat would therefore overwrite a
                        // LIVE session: the proxy keeps the shared `UdpManager`
                        // and its flow table, but the replacement session starts
                        // with empty `upstream_sockets` / `upstream_to_flow` /
                        // `flow_to_upstream`, so every in-flight flow stops
                        // forwarding and its upstream slab slot can never be
                        // released (`on_close_flow` reaches it only through
                        // `flow_to_upstream`). `close()` is a `ProxySession`
                        // method and not `Drop`, so dropping the displaced
                        // session runs no teardown on the way out.
                        //
                        // `has_listener_session` is the precise "already
                        // installed" test — `build_session` is its only
                        // producer, and deactivate / remove / soft+hard stop are
                        // its only consumers — paired with the reserved slab
                        // slot still being present, so a slot that vanished
                        // still falls through to the invariant-break report
                        // below. It deliberately tracks the INSTALLED session
                        // and not the listener's `active` flag: the SCM
                        // hand-off's `give_back_listeners` clears `active` and
                        // takes the socket without removing the session, and
                        // skipping the rebuild there is right — the retained
                        // session reaches the freshly bound socket through the
                        // shared `Rc<RefCell<UdpListener>>`, so its flows
                        // survive the re-exec instead of being orphaned. A repeat then answers `ok` having touched
                        // nothing: the listener is active and its session is in
                        // the slab, which is exactly what the caller asked for,
                        // and it matches what the HTTP/HTTPS/TCP arms already
                        // answer for their own repeats.
                        let session_installed = self.udp.borrow().has_listener_session(token)
                            && self.sessions.borrow().slab.contains(token.0);
                        if !session_installed {
                            // The slot is reserved for the listener's whole
                            // lifetime (`reserve_listen_token`), so both failures
                            // below are invariant breaks; report them instead of
                            // returning `ok` for a listener whose socket is
                            // registered under a token `Server::ready` would
                            // ignore.
                            let session = match self.udp.borrow_mut().build_session(token) {
                                Some(session) => session,
                                None => {
                                    return worker_response_error(
                                        req_id,
                                        format!(
                                            "Could not build the UDP listener session for {address}"
                                        ),
                                    );
                                }
                            };
                            let mut sessions = self.sessions.borrow_mut();
                            match sessions.slab.get_mut(token.0) {
                                Some(slot) => *slot = session,
                                None => {
                                    return worker_response_error(
                                        req_id,
                                        format!(
                                            "UDP listener {address} has no session slot at {token:?}"
                                        ),
                                    );
                                }
                            }
                        }
                        WorkerResponse::ok(req_id)
                    }
                    Err(activate_error) => worker_response_error(
                        req_id,
                        format!("Could not activate UDP listener: {activate_error}"),
                    ),
                }
            }
            Err(_) => worker_response_error(req_id, "Wrong variant for ListenerType on request"),
        }
    }

    fn notify_deactivate_listener(
        &mut self,
        req_id: &str,
        deactivate: &DeactivateListener,
    ) -> WorkerResponse {
        debug!(
            "{} deactivate {:?} listener {:?}",
            req_id, deactivate.proxy, deactivate
        );

        let address = deactivate.listener_key();

        match ListenerType::try_from(deactivate.proxy) {
            Ok(ListenerType::Http) => {
                let (token, mut listener) =
                    match self.http.borrow_mut().give_back_listener(&address) {
                        Ok((token, listener)) => (token, listener),
                        Err(e) => {
                            return worker_response_error(
                                req_id,
                                format!("Couldn't deactivate HTTP listener at {address}: {e}"),
                            );
                        }
                    };

                if let Err(e) = self.poll.registry().deregister(&mut listener) {
                    error!(
                        "error deregistering HTTP listen socket({:?}): {:?}",
                        deactivate, e
                    );
                }

                // The slot stays RESERVED for the deactivated listener — see
                // `reserve_listen_token`. It carries the same inert placeholder
                // `AddListener` installed, so a later `ActivateListener` finds
                // its token still valid and no other session can take the key.
                self.reserve_listen_token(token, Protocol::HTTPListen);
                // The listen token may still be queued for a deferred accept
                // (`ready()` enqueues it whenever the listener is readable but
                // `can_accept` is false). The listener has just given its
                // socket back, so `handle_remaining_readiness` must not call
                // `accept()` on it.
                self.accept_ready.remove(&ListenToken(token.0));

                if deactivate.to_scm {
                    self.unblock_scm_socket();
                    let listeners = Listeners {
                        http: vec![(address.clone(), listener.as_raw_fd())],
                        tls: vec![],
                        tcp: vec![],
                        udp: vec![],
                    };
                    info!("sending HTTP listener: {:?}", listeners);
                    let res = self.scm.send_listeners(&listeners);

                    self.block_scm_socket();

                    info!("sent HTTP listener: {:?}", res);
                }
                WorkerResponse::ok(req_id)
            }
            Ok(ListenerType::Https) => {
                let (token, mut listener) =
                    match self.https.borrow_mut().give_back_listener(&address) {
                        Ok((token, listener)) => (token, listener),
                        Err(e) => {
                            return worker_response_error(
                                req_id,
                                format!("Couldn't deactivate HTTPS listener at {address}: {e}",),
                            );
                        }
                    };
                if let Err(e) = self.poll.registry().deregister(&mut listener) {
                    error!(
                        "error deregistering HTTPS listen socket({:?}): {:?}",
                        deactivate, e
                    );
                }
                // See the HTTP arm: the slot stays reserved, and a
                // socket-less listener must not stay queued for an accept.
                self.reserve_listen_token(token, Protocol::HTTPSListen);
                self.accept_ready.remove(&ListenToken(token.0));

                if deactivate.to_scm {
                    self.unblock_scm_socket();
                    let listeners = Listeners {
                        http: vec![],
                        tls: vec![(address.clone(), listener.as_raw_fd())],
                        tcp: vec![],
                        udp: vec![],
                    };
                    info!("sending HTTPS listener: {:?}", listeners);
                    let res = self.scm.send_listeners(&listeners);

                    self.block_scm_socket();

                    info!("sent HTTPS listener: {:?}", res);
                }
                WorkerResponse::ok(req_id)
            }
            Ok(ListenerType::Tcp) => {
                let (token, mut listener) = match self.tcp.borrow_mut().give_back_listener(&address)
                {
                    Ok((token, listener)) => (token, listener),
                    Err(e) => {
                        return worker_response_error(
                            req_id,
                            format!("Could not deactivate TCP listener at {address}: {e}"),
                        );
                    }
                };

                if let Err(e) = self.poll.registry().deregister(&mut listener) {
                    error!(
                        "error deregistering TCP listen socket({:?}): {:?}",
                        deactivate, e
                    );
                }
                // See the HTTP arm: the slot stays reserved, and a
                // socket-less listener must not stay queued for an accept.
                self.reserve_listen_token(token, Protocol::TCPListen);
                self.accept_ready.remove(&ListenToken(token.0));

                if deactivate.to_scm {
                    self.unblock_scm_socket();
                    let listeners = Listeners {
                        http: vec![],
                        tls: vec![],
                        tcp: vec![(address.clone(), listener.as_raw_fd())],
                        udp: vec![],
                    };
                    info!("sending TCP listener: {:?}", listeners);
                    let res = self.scm.send_listeners(&listeners);

                    self.block_scm_socket();

                    info!("sent TCP listener: {:?}", res);
                }
                WorkerResponse::ok(req_id)
            }
            Ok(ListenerType::Udp) => {
                let (token, mut listener) = match self.udp.borrow_mut().give_back_listener(&address)
                {
                    Ok((token, listener)) => (token, listener),
                    Err(e) => {
                        return worker_response_error(
                            req_id,
                            format!("Could not deactivate UDP listener at {address}: {e}"),
                        );
                    }
                };

                if let Err(e) = self.poll.registry().deregister(&mut listener) {
                    error!(
                        "error deregistering UDP listen socket({:?}): {:?}",
                        deactivate, e
                    );
                }
                // See the HTTP arm. For UDP this ALSO drops the live
                // `UdpListenerSession` the activate path installed here,
                // putting the slot back to the placeholder — the flows it owned
                // were already torn down by `give_back_listener`.
                self.reserve_listen_token(token, Protocol::UDPListen);
                // A UDP listen token never reaches `accept_ready` (`ready()`
                // only enqueues the three accept-driven listen protocols), but
                // the removal keeps the rule uniform across the four arms.
                self.accept_ready.remove(&ListenToken(token.0));

                if deactivate.to_scm {
                    self.unblock_scm_socket();
                    let listeners = Listeners {
                        http: vec![],
                        tls: vec![],
                        tcp: vec![],
                        udp: vec![(address.clone(), listener.as_raw_fd())],
                    };
                    info!("sending UDP listener: {:?}", listeners);
                    let res = self.scm.send_listeners(&listeners);

                    self.block_scm_socket();

                    info!("sent UDP listener: {:?}", res);
                }
                WorkerResponse::ok(req_id)
            }
            Err(_) => worker_response_error(req_id, "Wrong variant for ListenerType on request"),
        }
    }

    /// Send all socket addresses and file descriptors of all proxies, via the scm socket
    pub fn return_listen_sockets(&mut self) -> Result<(), ScmSocketError> {
        self.unblock_scm_socket();

        let mut http_listeners = self.http.borrow_mut().give_back_listeners();
        for &mut (_, ref mut sock) in http_listeners.iter_mut() {
            if let Err(e) = self.poll.registry().deregister(sock) {
                error!(
                    "error deregistering HTTP listen socket({:?}): {:?}",
                    sock, e
                );
            }
        }

        let mut https_listeners = self.https.borrow_mut().give_back_listeners();
        for &mut (_, ref mut sock) in https_listeners.iter_mut() {
            if let Err(e) = self.poll.registry().deregister(sock) {
                error!(
                    "error deregistering HTTPS listen socket({:?}): {:?}",
                    sock, e
                );
            }
        }

        let mut tcp_listeners = self.tcp.borrow_mut().give_back_listeners();
        for &mut (_, ref mut sock) in tcp_listeners.iter_mut() {
            if let Err(e) = self.poll.registry().deregister(sock) {
                error!("error deregistering TCP listen socket({:?}): {:?}", sock, e);
            }
        }

        let mut udp_listeners = self.udp.borrow_mut().give_back_listeners();
        for &mut (_, ref mut sock) in udp_listeners.iter_mut() {
            if let Err(e) = self.poll.registry().deregister(sock) {
                error!("error deregistering UDP listen socket({:?}): {:?}", sock, e);
            }
        }

        // use as_raw_fd because the listeners should be dropped after sending them
        let listeners = Listeners {
            http: http_listeners
                .iter()
                .map(|(key, listener)| (key.clone(), listener.as_raw_fd()))
                .collect(),
            tls: https_listeners
                .iter()
                .map(|(key, listener)| (key.clone(), listener.as_raw_fd()))
                .collect(),
            tcp: tcp_listeners
                .iter()
                .map(|(key, listener)| (key.clone(), listener.as_raw_fd()))
                .collect(),
            udp: udp_listeners
                .iter()
                .map(|(key, listener)| (key.clone(), listener.as_raw_fd()))
                .collect(),
        };
        // Each handed-back listener is collected exactly once: the assembled
        // fd lists mirror the give_back lists one-to-one (the maps above are
        // straight `.iter().map().collect()` with no filtering or dedup).
        debug_assert_eq!(
            listeners.http.len(),
            http_listeners.len(),
            "every HTTP listener must be collected exactly once"
        );
        debug_assert_eq!(
            listeners.tls.len(),
            https_listeners.len(),
            "every HTTPS listener must be collected exactly once"
        );
        debug_assert_eq!(
            listeners.tcp.len(),
            tcp_listeners.len(),
            "every TCP listener must be collected exactly once"
        );
        info!("sending default listeners: {:?}", listeners);
        let res = self.scm.send_listeners(&listeners);

        self.block_scm_socket();

        info!("sent default listeners: {:?}", res);
        res
    }

    fn block_scm_socket(&mut self) {
        if let Err(e) = self.scm.set_blocking(true) {
            error!("Could not block scm socket: {}", e);
        }
    }

    fn unblock_scm_socket(&mut self) {
        if let Err(e) = self.scm.set_blocking(false) {
            error!("Could not unblock scm socket: {}", e);
        }
    }

    pub fn to_session(&self, token: Token) -> SessionToken {
        SessionToken(token.0)
    }

    pub fn from_session(&self, token: SessionToken) -> Token {
        Token(token.0)
    }

    pub fn accept(&mut self, token: ListenToken, protocol: Protocol) {
        // Per-protocol counter key. Keeping the namespace static (3 keys +
        // aggregate) is a deliberate cardinality cap: per-listener-address
        // labelling would require runtime `Box::leak` because `incr!` takes
        // `&'static str`, and listener addresses can be reconfigured at
        // runtime by the control plane. Operators wanting per-listener
        // attribution should correlate with the listener-protocol breakdown
        // below.
        //
        // Non-listen protocols reach this code only on an invariant break
        // upstream (`ready()` dispatched a non-listen `Protocol` to
        // `accept()`). Log and return rather than panicking — defense in
        // depth on the accept path, which is process-fatal if it aborts.
        let (proto_key, accepted_protocol) = match protocol {
            Protocol::TCPListen => ("listener.accepted.tcp", Protocol::TCPListen),
            Protocol::HTTPListen => ("listener.accepted.http", Protocol::HTTPListen),
            Protocol::HTTPSListen => ("listener.accepted.https", Protocol::HTTPSListen),
            other => {
                warn!(
                    "accept() called with non-listen protocol {:?} on token {:?}; skipping",
                    other, token
                );
                return;
            }
        };

        // Past the guard, `accepted_protocol` is one of the three listen
        // variants — the inner dispatch's `unreachable!` arm relies on this.
        debug_assert!(
            matches!(
                accepted_protocol,
                Protocol::TCPListen | Protocol::HTTPListen | Protocol::HTTPSListen
            ),
            "accept dispatch must run with a listen protocol only"
        );

        loop {
            let result = match accepted_protocol {
                Protocol::TCPListen => self.tcp.borrow_mut().accept(token),
                Protocol::HTTPListen => self.http.borrow_mut().accept(token),
                Protocol::HTTPSListen => self.https.borrow_mut().accept(token),
                // The outer match populates `accepted_protocol` only with the
                // three listen variants and returns early otherwise — this
                // arm is structurally unreachable.
                other => unreachable!(
                    "accept dispatch reached non-listen protocol {:?} after outer guard",
                    other
                ),
            };
            match result {
                Ok((sock, peer)) => {
                    // `peer` is the address `accept(2)` returned: the counter
                    // and the session below both use it, and neither costs a
                    // `getpeername(2)`.
                    incr!(names::listener::ACCEPTED_TOTAL);
                    incr!(proto_key);
                    incr!(per_source_bucket(&peer));
                    let queue_before = self.accept_queue.len();
                    self.accept_queue.push_back((
                        sock,
                        token,
                        accepted_protocol,
                        Instant::now(),
                        peer,
                    ));
                    // One accepted socket enqueues exactly one entry.
                    debug_assert_eq!(
                        self.accept_queue.len(),
                        queue_before + 1,
                        "each accepted socket must enqueue exactly one entry"
                    );
                }
                Err(AcceptError::WouldBlock) => {
                    self.accept_ready.remove(&token);
                    break;
                }
                Err(other) => {
                    error!(
                        "error accepting {:?} sockets: {:?}",
                        accepted_protocol, other
                    );
                    self.accept_ready.remove(&token);
                    break;
                }
            }
        }

        gauge!(names::accept_queue::CONNECTIONS, self.accept_queue.len());
    }

    pub fn create_sessions(&mut self) {
        while let Some((sock, token, protocol, timestamp, peer)) = self.accept_queue.pop_back() {
            let wait_time = Instant::now() - timestamp;
            time!(names::accept_queue::WAIT_TIME, wait_time.as_millis());
            if wait_time > self.accept_queue_timeout {
                incr!(names::accept_queue::TIMEOUT);
                continue;
            }

            if !self.sessions.borrow_mut().check_limits() {
                // Every `break` below drops the socket we just popped,
                // unserved, while the remaining queued sockets wait for the
                // next pass or time out. `listener.connection_capped` counts
                // exactly that dropped socket, once, at each `break`: a socket
                // served after eviction made room is not a refusal.
                if !self.evict_on_queue_full {
                    incr!(names::listener::CONNECTION_CAPPED);
                    break;
                }

                // Skip eviction during graceful shutdown — defeats the
                // shutting_down semantics and is wasted work since the
                // worker is winding down anyway.
                if self.shutting_down.is_some() {
                    incr!(names::listener::CONNECTION_CAPPED);
                    break;
                }

                // Evict 1% of `max_connections` per iteration. Conservative
                // ratio: large enough to make meaningful progress clearing
                // the accept queue, small enough to limit collateral damage
                // to active sessions. The cap loop re-checks limits after
                // each eviction round, so multiple rounds can run if the
                // queue has many pending connections. Decoupled from
                // `slab_entries_per_connection` (which only sizes the slab,
                // not `max_connections`).
                let to_evict = (self.sessions.borrow().max_connections / 100).max(1);
                let evicted = self.evict_least_active_sessions(to_evict);
                if evicted == 0 {
                    // Informational, not an invariant break: the worker may
                    // be at boot, or every active session is a system
                    // protocol (Channel/Metrics/Timer/listeners) and is
                    // ineligible. Stay at warn so operators see it in info-
                    // level production logs.
                    warn!("evict_on_queue_full enabled but no candidate sessions to evict");
                    incr!(names::listener::CONNECTION_CAPPED);
                    break;
                }

                count!(names::sessions::EVICTED, evicted as i64);
                warn!(
                    "evicted {} least recently active sessions to make room",
                    evicted
                );

                if !self.sessions.borrow_mut().check_limits() {
                    incr!(names::listener::CONNECTION_CAPPED);
                    break;
                }
            }

            //FIXME: check the timestamp
            //TODO: create_session should return the session and
            // the server should insert it in the the SessionManager
            // The accept path only ever enqueues listen protocols, so the
            // `_ => panic!` arm below is genuinely unreachable. Assert the set
            // here so a future enqueue regression trips in debug, not prod.
            debug_assert!(
                matches!(
                    protocol,
                    Protocol::TCPListen | Protocol::HTTPListen | Protocol::HTTPSListen
                ),
                "accept queue must only hold listen protocols, got {protocol:?}"
            );
            match protocol {
                Protocol::TCPListen => {
                    let proxy = self.tcp.clone();
                    if self
                        .tcp
                        .borrow_mut()
                        .create_session(sock, peer, token, wait_time, proxy)
                        .is_err()
                    {
                        break;
                    }
                }
                Protocol::HTTPListen => {
                    let proxy = self.http.clone();
                    if self
                        .http
                        .borrow_mut()
                        .create_session(sock, peer, token, wait_time, proxy)
                        .is_err()
                    {
                        break;
                    }
                }
                Protocol::HTTPSListen => {
                    if self
                        .https
                        .borrow_mut()
                        .create_session(sock, peer, token, wait_time, self.https.clone())
                        .is_err()
                    {
                        break;
                    }
                }
                _ => panic!("should not call accept() on a HTTP, HTTPS or TCP session"),
            };
            let nb_before = self.sessions.borrow().nb_connections;
            self.sessions.borrow_mut().incr();
            // A successfully created session bumps the live count by one.
            debug_assert_eq!(
                self.sessions.borrow().nb_connections,
                nb_before + 1,
                "create_sessions must account exactly one new connection per created session"
            );
        }

        gauge!(names::accept_queue::CONNECTIONS, self.accept_queue.len());
    }

    /// Record, for each token of one `epoll_wait` result, the session its
    /// slab slot holds right now, before any of the result is dispatched.
    fn snapshot_event_owners(
        &self,
        tokens: impl Iterator<Item = Token>,
        owners: &mut Vec<Option<Rc<RefCell<dyn ProxySession>>>>,
    ) {
        debug_assert!(
            owners.is_empty(),
            "the previous turn must have consumed every event owner"
        );
        let sessions = self.sessions.borrow();
        owners.extend(tokens.map(|token| sessions.slab.get(token.0).cloned()));
    }

    /// Dispatch `events` through `Server::ready` only if `token` still
    /// belongs to `owner`, the session `Server::snapshot_event_owners`
    /// recorded for it when `epoll_wait` returned.
    ///
    /// A session that closes while the loop walks one `epoll_wait` result
    /// frees its slab slots at once, and the slab hands the most recently
    /// freed key to the next allocation. A session dispatched later in the
    /// same result can therefore take that key for its new backend
    /// (`TcpSession::connect_to_backend` in `lib/src/tcp.rs`, or the mux's
    /// backend dial through `L7Proxy::add_session`, `lib/src/lib.rs`) while the
    /// result still holds an event, typically a HUP, from the closed
    /// session's own socket. Delivered under the reused token, that event
    /// reads as the new backend hanging up, and the new session closes
    /// (sozu-proxy/sozu#1877). Comparing with the recorded owner drops it.
    /// The recorded `Rc` stays alive until this call, so no new session can
    /// sit at its address and pass the comparison by accident.
    ///
    /// The guard compares sessions, not sockets: a session that closes one of
    /// its own backends and dials another in the same result, on a different
    /// descriptor's event, gets the freed key back under the same `Rc`, and a
    /// queued event of the old socket still reaches it.
    fn ready_if_still_owned(
        &mut self,
        token: Token,
        events: Ready,
        owner: Option<Rc<RefCell<dyn ProxySession>>>,
    ) {
        let still_owned = owner.is_some_and(|owner| {
            self.sessions
                .borrow()
                .slab
                .get(token.0)
                .is_some_and(|current| Rc::ptr_eq(current, &owner))
        });
        if !still_owned {
            trace!(
                "PROXY\t{:?} dropping events {:?} from a socket whose session closed earlier in this epoll result",
                token, events
            );
            return;
        }
        self.ready(token, events);
    }

    pub fn ready(&mut self, token: Token, events: Ready) {
        trace!("PROXY\t{:?} got events: {:?}", token, events);

        let session_token = token.0;
        if self.sessions.borrow().slab.contains(session_token) {
            //info!("sessions contains {:?}", session_token);
            let protocol = self.sessions.borrow().slab[session_token]
                .borrow()
                .protocol();
            // NOTE: `token` is NOT necessarily the session's frontend token. A
            // session is registered under BOTH its frontend and its backend slab
            // slots (the multi-token pattern: `connect_to_backend` inserts the
            // same session Rc under a second vacant key), so `ready()` is also
            // dispatched with the backend token, where `frontend_token() !=
            // session_token`. There is therefore no `token == frontend_token`
            // identity to assert here.
            //info!("protocol: {:?}", protocol);
            match protocol {
                Protocol::HTTPListen | Protocol::HTTPSListen | Protocol::TCPListen => {
                    //info!("PROTOCOL IS LISTEN");
                    if events.is_readable() {
                        self.accept_ready.insert(ListenToken(token.0));
                        if self.sessions.borrow().can_accept {
                            self.accept(ListenToken(token.0), protocol);
                        }
                        return;
                    }

                    if events.is_writable() {
                        error!(
                            "received writable for listener {:?}, this should not happen",
                            token
                        );
                        return;
                    }

                    if events.is_hup() {
                        error!("should not happen: server {:?} closed", token);
                        return;
                    }

                    unreachable!();
                }
                _ => {}
            }

            let session = self.sessions.borrow_mut().slab[session_token].clone();
            session.borrow_mut().update_readiness(token, events);
            if session.borrow_mut().ready(session.clone()) {
                debug!(
                    "Server killing session from ready: token={:?}, protocol={:?}, events={:?}",
                    token, protocol, events
                );
                self.kill_session(session);
            } else {
                self.queue_if_yielded(&session);
            }
        }
    }

    pub fn timeout(&mut self, token: Token) {
        trace!("PROXY\t{:?} got timeout", token);

        let session_token = token.0;
        if self.sessions.borrow().slab.contains(session_token) {
            let session = self.sessions.borrow_mut().slab[session_token].clone();
            if session.borrow_mut().timeout(token) {
                debug!(
                    "Server killing session from timeout: token={:?}, protocol={:?}",
                    token,
                    session.borrow().protocol()
                );
                self.kill_session(session);
            } else if session.borrow().needs_ready_pass() {
                // The timeout left work only `ready` can do, and no socket
                // event will come to trigger it: a backend whose connect timed
                // out must be counted as failed and its requests sent to
                // another backend (sozu-proxy/sozu#1800).
                if session.borrow_mut().ready(session.clone()) {
                    debug!(
                        "Server killing session from ready after timeout: token={:?}, protocol={:?}",
                        token,
                        session.borrow().protocol()
                    );
                    self.kill_session(session);
                } else {
                    self.queue_if_yielded(&session);
                }
            }
        }
    }

    pub fn handle_remaining_readiness(&mut self) {
        // try to accept again after handling all session events,
        // since we might have released a few session slots
        if self.sessions.borrow().can_accept && !self.accept_ready.is_empty() {
            while let Some(token) = self
                .accept_ready
                .iter()
                .next()
                .map(|token| ListenToken(token.0))
            {
                // A listen token queued here while `can_accept` was false can
                // outlive its slab entry. `notify_deactivate_listener` no
                // longer frees that entry — it RESERVES the slot, re-arming it
                // with the inert `ListenSession` placeholder (see
                // `reserve_listen_token`) — so after a deactivate the lookup
                // below still succeeds and it is the `accept_ready` purge in
                // those arms that keeps a socket-less listener out of
                // `accept()`. `RemoveListener` is the sole release point of the
                // slot (`notify_proxys`), and it purges the token too, so a
                // token reaching this loop with no slab entry means those
                // purges were bypassed. Indexing the slab with it panicked the
                // worker, hence the fallible `get`. Purging it HERE as well as
                // at those sites is what keeps the loop finite: `accept()`
                // below only removes the token on `WouldBlock`/error, so a
                // `continue` without the removal would hand `iter().next()` the
                // same stale token forever.
                let protocol = self
                    .sessions
                    .borrow()
                    .slab
                    .get(token.0)
                    .map(|session| session.borrow().protocol());
                let Some(protocol) = protocol else {
                    error!(
                        "accept_ready holds listen token {:?}, which no longer has a session; dropping it",
                        token
                    );
                    self.accept_ready.remove(&token);
                    continue;
                };
                self.accept(token, protocol);
                if !self.sessions.borrow().can_accept || self.accept_ready.is_empty() {
                    break;
                }
            }
        }
    }
    fn block_channel(&mut self) {
        if let Err(e) = self.channel.blocking() {
            error!("Could not block channel: {}", e);
        }
    }
    fn unblock_channel(&mut self) {
        if let Err(e) = self.channel.nonblocking() {
            error!("Could not block channel: {}", e);
        }
    }

    /// Evict the `count` least-recently-active non-listener sessions and
    /// return how many tokens were enqueued for shutdown. Used by
    /// `create_sessions` when the accept queue is saturated and the
    /// `evict_on_queue_full` knob is set.
    ///
    /// Uses `select_nth_unstable_by_key` (introselect, O(n) average) to
    /// partition the oldest `count` sessions in-place rather than a full
    /// O(n log n) sort. The candidate `Vec` is unavoidable because of
    /// `RefCell` borrow rules — the immutable borrow on `self.sessions`
    /// must drop before `shut_down_sessions_by_frontend_tokens` can take
    /// its mutable borrow.
    fn evict_least_active_sessions(&self, count: usize) -> usize {
        if count == 0 {
            return 0;
        }

        let tokens = {
            let sessions = self.sessions.borrow();
            let mut candidates: Vec<(Token, Instant)> = sessions
                .slab
                .iter()
                .filter(|(_, session)| {
                    !matches!(
                        session.borrow().protocol(),
                        Protocol::HTTPListen
                            | Protocol::HTTPSListen
                            | Protocol::TCPListen
                            | Protocol::UDPListen
                            | Protocol::Channel
                            | Protocol::Metrics
                            | Protocol::Timer
                    )
                })
                .map(|(_, session)| {
                    let s = session.borrow();
                    (s.frontend_token(), s.last_event())
                })
                .collect();

            // Early return is load-bearing: the `pivot` computation below
            // does `count.min(len) - 1`, which underflows on empty input.
            if candidates.is_empty() {
                return 0;
            }

            let pivot = count.min(candidates.len()) - 1;
            candidates.select_nth_unstable_by_key(pivot, |&(_, last_event)| last_event);

            candidates[..=pivot]
                .iter()
                .map(|&(token, _)| token)
                .collect::<HashSet<Token>>()
        };

        let evicted = tokens.len();
        self.shut_down_sessions_by_frontend_tokens(tokens);
        evicted
    }
}

/// Close an inherited listening socket this worker is not going to use, and
/// say so.
///
/// Reached only for [`crate::InheritedSocketFate::Refused`]: a listener is
/// already up at `address` on a socket of its own, so `activate()`
/// short-circuits and nothing consumes this descriptor. That is a statement
/// about now, not about the lifetime — `DeactivateListener` then
/// `ActivateListener` would bring the address back to `Adopted`, and because
/// `scm_listeners` is filled once and never refilled, closing here means such a
/// reactivation binds a fresh socket. See [`crate::InheritedSocketFate::Refused`]
/// for why that trade is taken: a queued backlog decays, whereas an unaccepted
/// socket in the address's `SO_REUSEPORT` group is expected to go on taking a
/// share of new connections for as long as it stays open. That expectation
/// follows from `SO_REUSEPORT` load-balancing semantics; nothing here measures
/// it.
///
/// The close is explicit rather than a dropped wrapper falling out of scope: it
/// is the same syscall either way, but this one is deliberate, attributable and
/// logged, which is precisely what sozu#1342 was missing.
fn discard_inherited_socket<T: FromRawFd>(
    fd: Option<RawFd>,
    address: &ListenerKey,
    protocol: &str,
) {
    let Some(fd) = fd else {
        return;
    };
    warn!(
        "closing the inherited {} listening socket for {}: this worker already has an active \
         listener there, so nothing will adopt this descriptor unless that listener is \
         deactivated and reactivated, and until then it is expected to keep taking a share of \
         new connections that nothing accepts (inferred from SO_REUSEPORT semantics, not \
         measured here)",
        protocol, address
    );
    // SAFETY: `fd` was just removed from the SCM table by `Listeners::get_*`,
    // which hands over ownership, so no other owner survives. The wrapper is
    // built only to be dropped: `Drop` is what performs the close.
    drop(unsafe { T::from_raw_fd(fd) });
}

/// The refusal of a request the initial handshake does not admit now.
const REFUSED_INITIAL_REQUEST: &str = "unexpected request before the initial status";

/// log the error together with the request id
/// create a WorkerResponse
fn worker_response_error<S: ToString, T: ToString>(request_id: S, error: T) -> WorkerResponse {
    error!(
        "error on request {}, {}",
        request_id.to_string(),
        error.to_string()
    );
    WorkerResponse::error(request_id, error)
}

pub struct ListenSession {
    pub protocol: Protocol,
}

impl ProxySession for ListenSession {
    fn last_event(&self) -> Instant {
        Instant::now()
    }

    fn print_session(&self) {}

    fn frontend_token(&self) -> Token {
        Token(0)
    }

    fn protocol(&self) -> Protocol {
        self.protocol
    }

    fn ready(&mut self, _session: Rc<RefCell<dyn ProxySession>>) -> SessionIsToBeClosed {
        false
    }

    fn shutting_down(&mut self) -> SessionIsToBeClosed {
        false
    }

    fn update_readiness(&mut self, _token: Token, _events: Ready) {}

    fn close(&mut self) {}

    fn timeout(&mut self, _token: Token) -> SessionIsToBeClosed {
        error!(
            "called ProxySession::timeout(token={:?}, time) on ListenSession {{ protocol: {:?} }}",
            _token, self.protocol
        );
        false
    }
}

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

    /// Two IPv4 addresses sharing the same /24 must hash to the same bucket;
    /// the masking logic guarantees this regardless of the host octet.
    #[test]
    fn per_source_bucket_collapses_ipv4_slash24() {
        let a: SocketAddr = "203.0.113.5:1234".parse().unwrap();
        let b: SocketAddr = "203.0.113.250:9999".parse().unwrap();
        assert_eq!(
            per_source_bucket(&a),
            per_source_bucket(&b),
            "addresses in the same /24 must land in the same bucket"
        );
    }

    /// Two IPv6 addresses sharing the same /48 must hash to the same bucket.
    #[test]
    fn per_source_bucket_collapses_ipv6_slash48() {
        let a: SocketAddr = "[2001:db8:1234::1]:443".parse().unwrap();
        let b: SocketAddr = "[2001:db8:1234:abcd::ffff]:8443".parse().unwrap();
        assert_eq!(
            per_source_bucket(&a),
            per_source_bucket(&b),
            "addresses in the same /48 must land in the same bucket"
        );
    }

    /// Every bucket label must be one of `PER_SOURCE_BUCKETS` precomputed
    /// statics — the cardinality cap is the load-bearing property.
    #[test]
    fn per_source_bucket_keys_are_bounded() {
        assert_eq!(PER_SOURCE_BUCKET_KEYS.len(), PER_SOURCE_BUCKETS);
        for (i, key) in PER_SOURCE_BUCKET_KEYS.iter().enumerate() {
            let expected = format!("client.connect.per_source.bucket_{i:03}");
            assert_eq!(*key, expected.as_str());
        }
    }

    /// A modest sample across distinct /24 prefixes should hit a healthy
    /// number of distinct buckets — guards against the hash collapsing.
    #[test]
    fn per_source_bucket_distributes_distinct_subnets() {
        let mut hits = std::collections::HashSet::new();
        for i in 0..200u8 {
            let addr: SocketAddr = format!("10.0.{i}.42:80").parse().unwrap();
            hits.insert(per_source_bucket(&addr));
        }
        // With 200 distinct /24 prefixes hashed into 256 buckets we expect
        // many distinct labels — assert a conservative lower bound that
        // tolerates birthday collisions.
        assert!(
            hits.len() >= 100,
            "expected at least 100 distinct buckets across 200 /24s, got {}",
            hits.len()
        );
    }
}

#[cfg(test)]
mod state_error_log_tests {
    use sozu_command::state::StateError;

    #[test]
    fn worker_state_dispatch_error_sink_bounds_the_reason() {
        const REASON_SECRET: &str = "WORKER_STATE_ERROR_REASON_SECRET_SENTINEL";

        let reason = format!("{REASON_SECRET}{}", "x".repeat(4096));
        let reason_len = reason.len();
        let output = crate::capture_test_logs(move || {
            let error = StateError::InvalidTcpFrontend {
                address: "127.0.0.1:443"
                    .parse()
                    .expect("test TCP frontend address must parse"),
                reason,
            };
            error!("Could not execute order on config state: {}", error);
        });

        assert!(
            !output.contains(REASON_SECRET),
            "worker state-error sink leaked the reason: {output}"
        );
        assert!(
            output.contains(&format!("reason_bytes={reason_len}")),
            "worker state-error sink omitted bounded reason metadata: {output}"
        );
        assert!(
            output.len() <= 512,
            "worker state-error sink is not bounded: {} bytes",
            output.len()
        );
    }
}

#[cfg(test)]
mod eviction_tests {
    use std::collections::HashSet;
    use std::time::{Duration, Instant};

    use mio::Token;

    /// `select_nth_unstable_by_key` partitions in O(n) so that the first
    /// `pivot + 1` entries are the `pivot + 1` smallest by key. This guards
    /// against a future refactor swapping the comparator orientation.
    #[test]
    fn select_nth_finds_oldest_sessions() {
        let now = Instant::now();
        let mut candidates = [
            (Token(1), now - Duration::from_secs(10)), // 10s old
            (Token(2), now - Duration::from_secs(50)), // 50s old (oldest)
            (Token(3), now - Duration::from_secs(5)),  // 5s old (newest)
            (Token(4), now - Duration::from_secs(30)), // 30s old
            (Token(5), now - Duration::from_secs(20)), // 20s old
        ];

        let count = 2;
        let pivot = count.min(candidates.len()) - 1;
        candidates.select_nth_unstable_by_key(pivot, |&(_, last_event)| last_event);

        let selected: HashSet<Token> = candidates[..=pivot]
            .iter()
            .map(|&(token, _)| token)
            .collect();

        assert_eq!(selected.len(), 2);
        assert!(
            selected.contains(&Token(2)),
            "should contain 50s-old session"
        );
        assert!(
            selected.contains(&Token(4)),
            "should contain 30s-old session"
        );
    }

    /// When `count` exceeds available candidates, the pivot collapses to
    /// `len - 1` so we evict everything; this test pins that behaviour
    /// against a future refactor that might silently truncate.
    #[test]
    fn select_nth_with_count_exceeding_candidates() {
        let now = Instant::now();
        let mut candidates = [(Token(1), now - Duration::from_secs(10))];

        let count = 5;
        let pivot = count.min(candidates.len()) - 1;
        candidates.select_nth_unstable_by_key(pivot, |&(_, last_event)| last_event);

        let selected: HashSet<Token> = candidates[..=pivot]
            .iter()
            .map(|&(token, _)| token)
            .collect();

        assert_eq!(selected.len(), 1);
        assert!(selected.contains(&Token(1)));
    }
}

#[cfg(test)]
mod accept_ready_tests {
    use sozu_command::{
        config::ListenerBuilder,
        proto::command::{DeactivateListener, ListenerType, SocketAddress},
    };

    use super::*;
    use crate::testing::{ServerParts, prebuild_server, provide_port};

    /// A worker holding one activated TCP listener, reachable through the very
    /// same `Server` surface the event loop drives. The listen token carries a
    /// `ListenSession` placeholder in the slab, exactly like
    /// `tcp::testing::start_tcp_worker` installs it.
    pub(super) fn server_with_tcp_listener() -> (Server, Token, SocketAddress) {
        let ServerParts {
            event_loop,
            registry,
            sessions,
            pool,
            backends,
            server_scm_socket,
            server_config,
            ..
        // `send_scm = true`: `Server::new` makes a BLOCKING
        // `receive_listeners()` call before it applies the initial state, so
        // the client side must have queued its (empty) listener set first --
        // exactly what `tcp::testing::start_tcp_worker` does.
        } = prebuild_server(16, 16384, true).expect("could not prebuild a test server");
        let (_command_channel, proxy_channel) =
            Channel::generate(1000, 10000).expect("could not generate a test channel");

        let address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let listener_config = ListenerBuilder::new_tcp(address)
            .to_tcp(None)
            .expect("could not build a TcpListenerConfig for the test");

        let listen_token = {
            let mut session_manager = sessions.borrow_mut();
            let entry = session_manager.slab.vacant_entry();
            let token = Token(entry.key());
            entry.insert(Rc::new(RefCell::new(ListenSession {
                protocol: Protocol::TCPListen,
            })));
            token
        };

        let mut tcp_proxy =
            tcp::TcpProxy::new(registry, sessions.clone(), pool.clone(), backends.clone());
        tcp_proxy
            .add_listener(listener_config, listen_token)
            .expect("could not add the test TCP listener");
        tcp_proxy
            .activate_listener(&address.into(), None)
            .expect("could not activate the test TCP listener");

        let server = Server::new(
            event_loop,
            proxy_channel,
            server_scm_socket,
            sessions,
            pool,
            backends,
            None,
            None,
            Some(tcp_proxy),
            server_config,
            None,
            false,
        )
        .expect("could not build the test server");

        // `_command_channel` is dropped here on purpose: nothing in these
        // tests reads the worker's side of the channel.
        (server, listen_token, address)
    }

    /// `ready()` queues a listen token in `accept_ready` whenever its listener
    /// is readable but the worker cannot accept (buffer-pool backpressure), and
    /// only `accept()` — which never runs while `can_accept` is false — takes
    /// it back out. Deactivating that listener hands its socket back, so
    /// without this purge the deferred accept pass called `accept()` on a
    /// socket-less listener (and, before the slot was reserved for the
    /// listener's whole lifetime, indexed a vacant slab key).
    ///
    /// To SEE THIS RED: drop `self.accept_ready.remove(&ListenToken(token.0))`
    /// from `notify_deactivate_listener`'s TCP arm — the last assertion below
    /// fails.
    #[test]
    fn deactivating_a_listener_drops_its_pending_accept_token() {
        let (mut server, listen_token, address) = server_with_tcp_listener();

        // The backpressure state: readable listener, worker at capacity.
        server.accept_ready.insert(ListenToken(listen_token.0));

        let response = server.notify_deactivate_listener(
            "test-deactivate",
            &DeactivateListener {
                interface: None,
                address,
                proxy: ListenerType::Tcp as i32,
                to_scm: false,
            },
        );
        assert_eq!(
            response.status,
            ResponseStatus::Ok as i32,
            "deactivating an activated TCP listener must succeed: {response:?}"
        );
        assert!(
            server.sessions.borrow().slab.contains(listen_token.0),
            "deactivating a listener must KEEP its slab slot reserved — see \
             `reserve_listen_token`; the accept purge below is what stops the \
             deferred pass from accepting on a socket-less listener"
        );
        assert!(
            !server.accept_ready.contains(&ListenToken(listen_token.0)),
            "a deactivated listener must not stay queued for a deferred accept"
        );

        // The pass that used to panic.
        server.handle_remaining_readiness();
    }

    /// Structural safety for the deferred-accept pass itself: whatever else
    /// ever drops a listen token's slab entry, indexing the slab with a token
    /// that is no longer there must not take the worker down. Dropping the
    /// stale token is also what keeps THIS branch finite: `accept()` only
    /// removes a token on `WouldBlock`/error, so re-entering the loop without
    /// the removal would spin on the same vacant slot forever.
    ///
    /// To SEE THIS RED: restore `self.sessions.borrow().slab[token.0]` in
    /// `handle_remaining_readiness` — this test then panics with
    /// "invalid key".
    #[test]
    fn a_stale_accept_token_is_dropped_instead_of_panicking() {
        let (mut server, listen_token, _address) = server_with_tcp_listener();

        server.sessions.borrow_mut().slab.remove(listen_token.0);
        server.accept_ready.insert(ListenToken(listen_token.0));

        server.handle_remaining_readiness();

        assert!(
            server.accept_ready.is_empty(),
            "a listen token with no session must be dropped from accept_ready"
        );
    }
}
/// One `epoll_wait` result can hold events for a session that closes while
/// the loop walks it, after another session took over one of its slab keys.
/// `Server::ready_if_still_owned` must deliver such an event to nobody
/// (sozu-proxy/sozu#1877).
#[cfg(test)]
mod event_owner_tests {
    use super::*;

    /// A session that closes on its first `ready` and frees its backend slot
    /// from `close`, as `TcpSession::close_backend` (`lib/src/tcp.rs`) does.
    struct ClosingSession {
        sessions: Rc<RefCell<SessionManager>>,
        frontend: Token,
        backend: Token,
    }

    impl ProxySession for ClosingSession {
        fn protocol(&self) -> Protocol {
            Protocol::TCP
        }
        fn ready(&mut self, _session: Rc<RefCell<dyn ProxySession>>) -> SessionIsToBeClosed {
            true
        }
        fn shutting_down(&mut self) -> SessionIsToBeClosed {
            true
        }
        fn update_readiness(&mut self, _token: Token, _events: Ready) {}
        fn close(&mut self) {
            self.sessions.borrow_mut().slab.try_remove(self.backend.0);
        }
        fn timeout(&mut self, _token: Token) -> SessionIsToBeClosed {
            false
        }
        fn last_event(&self) -> Instant {
            Instant::now()
        }
        fn print_session(&self) {}
        fn frontend_token(&self) -> Token {
            self.frontend
        }
    }

    /// A session whose first `ready` dials a backend, taking the slab's next
    /// vacant key as `TcpSession::connect_to_backend` (`lib/src/tcp.rs`) does,
    /// and that records every token it is handed readiness for.
    struct DialingSession {
        sessions: Rc<RefCell<SessionManager>>,
        frontend: Token,
        backend: Option<Token>,
        delivered: Vec<Token>,
    }

    impl ProxySession for DialingSession {
        fn protocol(&self) -> Protocol {
            Protocol::TCP
        }
        fn ready(&mut self, session: Rc<RefCell<dyn ProxySession>>) -> SessionIsToBeClosed {
            if self.backend.is_none() {
                self.backend = Some(Token(self.sessions.borrow_mut().slab.insert(session)));
            }
            false
        }
        fn shutting_down(&mut self) -> SessionIsToBeClosed {
            true
        }
        fn update_readiness(&mut self, token: Token, _events: Ready) {
            self.delivered.push(token);
        }
        fn close(&mut self) {}
        fn timeout(&mut self, _token: Token) -> SessionIsToBeClosed {
            false
        }
        fn last_event(&self) -> Instant {
            Instant::now()
        }
        fn print_session(&self) {}
        fn frontend_token(&self) -> Token {
            self.frontend
        }
    }

    /// The batch of sozu-proxy/sozu#1877: the client's FIN on the first
    /// session's frontend, the first readiness of the second session, then
    /// the HUP of the first session's backend socket, in that order.
    ///
    /// To SEE THIS RED: make `Server::ready_if_still_owned` call
    /// `self.ready(token, events)` unconditionally — the dialing session then
    /// receives the closed backend's HUP under its own backend token.
    #[test]
    fn an_event_of_a_closed_session_is_not_delivered_to_the_session_reusing_its_token() {
        let (mut server, _listen_token, _address) =
            super::accept_ready_tests::server_with_tcp_listener();
        let sessions = server.sessions.clone();

        let closing = Rc::new(RefCell::new(ClosingSession {
            sessions: sessions.clone(),
            frontend: Token(0),
            backend: Token(0),
        }));
        let dialing = Rc::new(RefCell::new(DialingSession {
            sessions: sessions.clone(),
            frontend: Token(0),
            backend: None,
            delivered: Vec::new(),
        }));
        {
            let mut sessions = sessions.borrow_mut();
            let closing_session: Rc<RefCell<dyn ProxySession>> = closing.clone();
            let frontend = Token(sessions.slab.insert(closing_session.clone()));
            let backend = Token(sessions.slab.insert(closing_session));
            let mut closing = closing.borrow_mut();
            closing.frontend = frontend;
            closing.backend = backend;
            let dialing_session: Rc<RefCell<dyn ProxySession>> = dialing.clone();
            dialing.borrow_mut().frontend = Token(sessions.slab.insert(dialing_session));
            sessions.incr();
            sessions.incr();
        }
        let closed_frontend = closing.borrow().frontend;
        let closed_backend = closing.borrow().backend;
        let dialing_frontend = dialing.borrow().frontend;

        let batch = [
            (closed_frontend, Ready::READABLE | Ready::HUP),
            (dialing_frontend, Ready::READABLE | Ready::WRITABLE),
            (
                closed_backend,
                Ready::READABLE | Ready::WRITABLE | Ready::HUP,
            ),
        ];
        let mut owners = Vec::new();
        server.snapshot_event_owners(batch.iter().map(|(token, _)| *token), &mut owners);
        for ((token, events), owner) in batch.into_iter().zip(owners.drain(..)) {
            server.ready_if_still_owned(token, events, owner);
        }

        assert_eq!(
            dialing.borrow().backend,
            Some(closed_backend),
            "precondition: the dial must reuse the key the closed session freed in this batch"
        );
        assert!(
            !sessions.borrow().slab.contains(closed_frontend.0),
            "the closed session must have left the slab"
        );
        assert_eq!(
            dialing.borrow().delivered,
            vec![dialing_frontend],
            "the closed backend's HUP must not reach the session that reused its token"
        );
    }
}
/// A session a `CloseSessions` request left draining can close on its own
/// before `Server::drive_closing_sessions` revisits it, and another session
/// then takes its slab key. The watcher must leave the new tenant alone.
#[cfg(test)]
mod closing_sessions_tests {
    use std::cell::Cell;

    use super::*;

    /// A session whose `shutting_down` answer is scripted and whose `close`
    /// is counted.
    struct ScriptedSession {
        frontend: Token,
        can_close_now: bool,
        closed: Rc<Cell<bool>>,
    }

    impl ProxySession for ScriptedSession {
        fn protocol(&self) -> Protocol {
            Protocol::HTTP
        }
        fn ready(&mut self, _session: Rc<RefCell<dyn ProxySession>>) -> SessionIsToBeClosed {
            false
        }
        fn shutting_down(&mut self) -> SessionIsToBeClosed {
            self.can_close_now
        }
        fn update_readiness(&mut self, _token: Token, _events: Ready) {}
        fn close(&mut self) {
            self.closed.set(true);
        }
        fn timeout(&mut self, _token: Token) -> SessionIsToBeClosed {
            false
        }
        fn last_event(&self) -> Instant {
            Instant::now()
        }
        fn print_session(&self) {}
        fn frontend_token(&self) -> Token {
            self.frontend
        }
    }

    fn insert_scripted(
        server: &Server,
        can_close_now: bool,
    ) -> (Token, Rc<RefCell<dyn ProxySession>>, Rc<Cell<bool>>) {
        let closed = Rc::new(Cell::new(false));
        let mut sessions = server.sessions.borrow_mut();
        let entry = sessions.slab.vacant_entry();
        let token = Token(entry.key());
        let session: Rc<RefCell<dyn ProxySession>> = Rc::new(RefCell::new(ScriptedSession {
            frontend: token,
            can_close_now,
            closed: closed.clone(),
        }));
        entry.insert(session.clone());
        sessions.incr();
        (token, session, closed)
    }

    /// The draining session (closable by now) leaves the slab (its `Rc` is
    /// still held, so its `Weak` still upgrades) and an unrelated session
    /// takes the same key. Draining must not close it.
    ///
    /// To SEE THIS RED: drop the `still_there` check from
    /// `Server::drive_closing_sessions`, so a live `Weak` alone selects the
    /// token -- the unrelated session is then closed and removed.
    #[test]
    fn a_draining_sessions_reused_slab_key_is_not_closed() {
        let (mut server, _listen_token, _address) =
            super::accept_ready_tests::server_with_tcp_listener();

        let (token, draining, draining_closed) = insert_scripted(&server, true);
        server
            .closing_sessions
            .push((token, Rc::downgrade(&draining)));

        // The draining session closed by itself; an unrelated one reuses its key.
        server.sessions.borrow_mut().slab.remove(token.0);
        let (reused, unrelated, unrelated_closed) = insert_scripted(&server, true);
        assert_eq!(reused, token, "precondition: the slab must reuse the key");
        assert!(
            !Rc::ptr_eq(&draining, &unrelated),
            "precondition: the two sessions are distinct"
        );

        server.drive_closing_sessions();

        assert!(
            server.sessions.borrow().slab.contains(token.0),
            "the unrelated session that reused the key must stay in the slab"
        );
        assert!(
            !unrelated_closed.get() && !draining_closed.get(),
            "neither session may be closed by the stale watch"
        );
        assert!(
            server.closing_sessions.is_empty(),
            "the stale watch must be dropped, not kept"
        );
    }

    /// Positive twin: the same session still in its slot, once able to close,
    /// is closed and unwatched.
    #[test]
    fn a_draining_session_still_in_its_slot_is_closed_when_ready() {
        let (mut server, _listen_token, _address) =
            super::accept_ready_tests::server_with_tcp_listener();

        let (token, session, closed) = insert_scripted(&server, true);
        server
            .closing_sessions
            .push((token, Rc::downgrade(&session)));

        server.drive_closing_sessions();

        assert!(
            closed.get(),
            "a still-owned closable session must be closed"
        );
        assert!(!server.sessions.borrow().slab.contains(token.0));
        assert!(server.closing_sessions.is_empty());
    }
}
/// A session whose `ready` pass yielded at its iteration budget is resumed
/// after the next poll by `Server::resume_yielded_sessions`, unless it closed
/// in between, possibly leaving its slab key to another session
/// (sozu-proxy/sozu#1906).
#[cfg(test)]
mod yielded_sessions_tests {
    use std::cell::Cell;

    use super::*;

    /// A session that yields on its first `yields` ready passes and counts
    /// its `ready` and `update_readiness` calls.
    struct YieldingSession {
        frontend: Token,
        yields: Rc<Cell<u32>>,
        yielded: bool,
        ready_calls: Rc<Cell<u32>>,
        readiness_updates: Rc<Cell<u32>>,
    }

    impl ProxySession for YieldingSession {
        fn protocol(&self) -> Protocol {
            Protocol::HTTP
        }
        fn ready(&mut self, _session: Rc<RefCell<dyn ProxySession>>) -> SessionIsToBeClosed {
            self.ready_calls.set(self.ready_calls.get() + 1);
            self.yielded = self.yields.get() > 0;
            if self.yielded {
                self.yields.set(self.yields.get() - 1);
            }
            false
        }
        fn yielded(&mut self) -> bool {
            std::mem::take(&mut self.yielded)
        }
        fn shutting_down(&mut self) -> SessionIsToBeClosed {
            true
        }
        fn update_readiness(&mut self, _token: Token, _events: Ready) {
            self.readiness_updates.set(self.readiness_updates.get() + 1);
        }
        fn close(&mut self) {}
        fn timeout(&mut self, _token: Token) -> SessionIsToBeClosed {
            false
        }
        fn last_event(&self) -> Instant {
            Instant::now()
        }
        fn print_session(&self) {}
        fn frontend_token(&self) -> Token {
            self.frontend
        }
    }

    struct Counters {
        ready_calls: Rc<Cell<u32>>,
        readiness_updates: Rc<Cell<u32>>,
    }

    fn insert_yielding(
        server: &Server,
        yields: u32,
    ) -> (Token, Rc<RefCell<dyn ProxySession>>, Counters) {
        let ready_calls = Rc::new(Cell::new(0));
        let readiness_updates = Rc::new(Cell::new(0));
        let mut sessions = server.sessions.borrow_mut();
        let entry = sessions.slab.vacant_entry();
        let token = Token(entry.key());
        let session: Rc<RefCell<dyn ProxySession>> = Rc::new(RefCell::new(YieldingSession {
            frontend: token,
            yields: Rc::new(Cell::new(yields)),
            yielded: false,
            ready_calls: ready_calls.clone(),
            readiness_updates: readiness_updates.clone(),
        }));
        entry.insert(session.clone());
        sessions.incr();
        (
            token,
            session,
            Counters {
                ready_calls,
                readiness_updates,
            },
        )
    }

    /// A yielded session still in its slot is run again with no readiness
    /// update, queued again while it keeps yielding, and dropped from the
    /// queue once it does not.
    #[test]
    fn a_yielded_session_is_resumed_without_a_new_event_until_it_stops_yielding() {
        let (mut server, _listen_token, _address) =
            super::accept_ready_tests::server_with_tcp_listener();
        let (token, _session, counters) = insert_yielding(&server, 2);

        server.ready(token, Ready::READABLE);
        assert_eq!(server.yielded_sessions.len(), 1, "the yield must be queued");
        assert_eq!(
            server.reset_loop_time_and_get_timeout(),
            Some(Duration::ZERO),
            "a queued yield must make the next poll non-blocking"
        );

        let pending = std::mem::take(&mut server.yielded_sessions);
        server.resume_yielded_sessions(pending);
        assert_eq!(counters.ready_calls.get(), 2, "the session must be resumed");
        assert_eq!(
            counters.readiness_updates.get(),
            1,
            "a resumption merges no new readiness"
        );
        assert_eq!(
            server.yielded_sessions.len(),
            1,
            "a session that yields again is queued for the next turn"
        );

        let pending = std::mem::take(&mut server.yielded_sessions);
        server.resume_yielded_sessions(pending);
        assert_eq!(counters.ready_calls.get(), 3);
        assert!(
            server.yielded_sessions.is_empty(),
            "a session that did not yield leaves the queue"
        );
        assert!(server.sessions.borrow().slab.contains(token.0));
    }

    /// The session yielded, then closed (a timeout) before its resumption,
    /// and another session took its slab key. The resumption must reach
    /// neither: no panic, no `ready` on a closed session, and none on the new
    /// tenant, which has no work queued.
    ///
    /// To SEE THIS RED: drop the `still_there` check from
    /// `Server::resume_yielded_sessions`. The closed session's `Weak` still
    /// upgrades (this test holds its `Rc`), so its `ready` runs again.
    #[test]
    fn a_yielded_session_that_closed_before_its_resumption_is_not_resumed() {
        let (mut server, _listen_token, _address) =
            super::accept_ready_tests::server_with_tcp_listener();
        let (token, yielded, yielded_counters) = insert_yielding(&server, 1);

        server.ready(token, Ready::READABLE);
        assert_eq!(server.yielded_sessions.len(), 1, "precondition: queued");

        // Closed by a timeout before the next turn; a new session reuses the key.
        server.sessions.borrow_mut().slab.remove(token.0);
        let (reused, tenant, tenant_counters) = insert_yielding(&server, 0);
        assert_eq!(reused, token, "precondition: the slab must reuse the key");
        assert!(!Rc::ptr_eq(&yielded, &tenant));

        let pending = std::mem::take(&mut server.yielded_sessions);
        server.resume_yielded_sessions(pending);

        assert_eq!(
            yielded_counters.ready_calls.get(),
            1,
            "a closed session must not be resumed"
        );
        assert_eq!(
            tenant_counters.ready_calls.get(),
            0,
            "the session that reused the key must not be run on the stale entry"
        );
        assert!(server.yielded_sessions.is_empty());
        assert!(server.sessions.borrow().slab.contains(token.0));
    }
}
/// The listener slab slot is reserved for a listener's whole `AddListener` ->
/// `RemoveListener` lifetime. These tests pin both ends of that lifetime and
/// the deactivate/reactivate cycle in between; see `reserve_listen_token` for
/// the design and the alternative it rejects.
#[cfg(test)]
mod listener_lifecycle_tests {
    use std::net::UdpSocket as StdUdpSocket;

    use sozu_command::{
        config::ListenerBuilder,
        proto::command::{
            ActivateListener, DeactivateListener, ListenerType, LoadBalancingParams,
            RemoveListener, RequestUdpFrontend, SocketAddress, filtered_metrics,
        },
    };

    use super::accept_ready_tests::server_with_tcp_listener;
    use super::*;
    use crate::{
        metrics::{METRICS, names},
        testing::{ServerParts, prebuild_server, provide_port},
    };

    /// A worker with no listener at all, driven through `Server`'s own request
    /// surface. Same shape as `accept_ready_tests::server_with_tcp_listener`,
    /// minus the pre-installed listener — `Server::new` builds its own UDP
    /// proxy, which is the one these tests exercise.
    pub(super) fn bare_server() -> Server {
        let ServerParts {
            event_loop,
            sessions,
            pool,
            backends,
            server_scm_socket,
            server_config,
            ..
        } = prebuild_server(16, 16384, true).expect("could not prebuild a test server");
        let (_command_channel, proxy_channel) =
            Channel::generate(1000, 10000).expect("could not generate a test channel");

        Server::new(
            event_loop,
            proxy_channel,
            server_scm_socket,
            sessions,
            pool,
            backends,
            None,
            None,
            None,
            server_config,
            None,
            false,
        )
        .expect("could not build the test server")
    }

    fn add_udp_listener(server: &mut Server, address: SocketAddress) -> Token {
        let listener_config = ListenerBuilder::new_udp(address)
            .to_udp(None)
            .expect("could not build a UdpListenerConfig for the test");
        let response = server.notify_add_udp_listener("test-add-udp", listener_config);
        assert_eq!(
            response.status,
            ResponseStatus::Ok as i32,
            "adding a UDP listener must succeed: {response:?}"
        );
        server
            .udp
            .borrow()
            .listener_token(&address.into())
            .expect("the added UDP listener must own a token")
    }

    fn activate(server: &mut Server, address: SocketAddress, proxy: ListenerType) {
        let response = server.notify_activate_listener(
            "test-activate",
            &ActivateListener {
                interface: None,
                address,
                proxy: proxy as i32,
                from_scm: false,
            },
        );
        assert_eq!(
            response.status,
            ResponseStatus::Ok as i32,
            "activating a {proxy:?} listener must succeed: {response:?}"
        );
    }

    fn deactivate(server: &mut Server, address: SocketAddress, proxy: ListenerType) {
        let response = server.notify_deactivate_listener(
            "test-deactivate",
            &DeactivateListener {
                interface: None,
                address,
                proxy: proxy as i32,
                to_scm: false,
            },
        );
        assert_eq!(
            response.status,
            ResponseStatus::Ok as i32,
            "deactivating a {proxy:?} listener must succeed: {response:?}"
        );
    }

    /// The protocol recorded in the slab slot at `token`, and whether that slot
    /// holds the inert `ListenSession` placeholder. `ListenSession` reports
    /// `Token(0)` as its frontend token while a real `UdpListenerSession`
    /// reports its own listen token — the only way to tell the two apart, since
    /// both answer `Protocol::UDPListen`.
    fn slot_state(server: &Server, token: Token) -> Option<(Protocol, bool)> {
        let sessions = server.sessions.borrow();
        let session = sessions.slab.get(token.0)?;
        let session = session.borrow();
        Some((session.protocol(), session.frontend_token() == Token(0)))
    }

    /// FAILURE MODE 1, the whole point of this change: a deactivated listener
    /// that is activated again must be reachable from the event loop.
    ///
    /// The proxies retain the listener's token across a deactivate
    /// (`give_back_listener` and `activate()` both hand back the same one), and
    /// `Server::ready` ignores any token the slab does not contain. Freeing the
    /// slot on deactivate therefore left `ActivateListener` re-registering the
    /// socket under a dead token: the request answered `ok` and the listener
    /// never saw another readiness event.
    ///
    /// `can_accept` is forced to `false` so `ready()` parks the token in
    /// `accept_ready` instead of immediately draining it in `accept()` —
    /// membership there is the observable proof that the event reached the
    /// listener's dispatch path at all.
    ///
    /// To SEE THIS RED: in `notify_deactivate_listener`'s TCP arm, replace
    /// `self.reserve_listen_token(token, Protocol::TCPListen);` with
    /// `self.sessions.borrow_mut().slab.remove(token.0);` — the reactivated
    /// listener's readiness event is then dropped, and the assertion fails
    /// with "a reactivated listener must still be dispatched by `ready()`".
    #[test]
    fn a_reactivated_tcp_listener_is_reachable_from_the_event_loop() {
        let (mut server, listen_token, address) = server_with_tcp_listener();

        deactivate(&mut server, address, ListenerType::Tcp);
        activate(&mut server, address, ListenerType::Tcp);

        server.sessions.borrow_mut().can_accept = false;
        server.ready(listen_token, Ready::READABLE);

        assert!(
            server.accept_ready.contains(&ListenToken(listen_token.0)),
            "a reactivated listener must still be dispatched by `ready()`"
        );
    }

    /// The UDP half of failure mode 1. UDP never accepts, so its activate path
    /// installs a real `UdpListenerSession` at the listen token instead — which
    /// it skipped outright when the slot had been freed, leaving the socket
    /// registered against the placeholder-less token and every datagram
    /// unread, while the request still answered `ok`.
    ///
    /// To SEE THIS RED: in `notify_deactivate_listener`'s UDP arm, replace
    /// `self.reserve_listen_token(token, Protocol::UDPListen);` with
    /// `self.sessions.borrow_mut().slab.remove(token.0);` — the activate arm
    /// then returns an error response ("has no session slot"), so `activate()`
    /// panics on the status assertion.
    #[test]
    fn a_reactivated_udp_listener_gets_its_session_back() {
        let mut server = bare_server();
        let address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let listen_token = add_udp_listener(&mut server, address);

        assert_eq!(
            slot_state(&server, listen_token),
            Some((Protocol::UDPListen, true)),
            "an added, not-yet-activated UDP listener holds the placeholder"
        );

        activate(&mut server, address, ListenerType::Udp);
        assert_eq!(
            slot_state(&server, listen_token),
            Some((Protocol::UDPListen, false)),
            "activating must install the real UdpListenerSession"
        );

        deactivate(&mut server, address, ListenerType::Udp);
        activate(&mut server, address, ListenerType::Udp);

        assert_eq!(
            slot_state(&server, listen_token),
            Some((Protocol::UDPListen, false)),
            "reactivating must reinstall the real UdpListenerSession"
        );
    }

    /// FAILURE MODE 2, now impossible by construction: while the slot was
    /// freed on deactivate, the slab was free to hand that key to the next
    /// session, and the UDP activate path — which found `contains` true and
    /// assigned — overwrote that live session with its listener session.
    ///
    /// The reserved slot is never vacant, so `vacant_entry()` can no longer
    /// return it. This test pins that, and ONLY that: the session created while
    /// the listener is deactivated must land on a different key and survive the
    /// reactivate. It says nothing about what the activate arm writes into the
    /// listener's OWN key, because the reactivate here follows a deactivate and
    /// so takes the full activation path. The repeat with no deactivate in
    /// between — where the arm's answer is `Ok(token)` from
    /// `UdpListener::activate`'s `active` short-circuit and the listener's own
    /// key holds a LIVE `UdpListenerSession` — is FAILURE MODE 3, pinned by
    /// `a_duplicate_udp_activate_keeps_the_live_listener_session` and the three
    /// tests beside it. Naming this one for "a live session" hid that gap: it
    /// stayed green throughout.
    ///
    /// To SEE THIS RED: in `notify_deactivate_listener`'s UDP arm, replace
    /// `self.reserve_listen_token(token, Protocol::UDPListen);` with
    /// `self.sessions.borrow_mut().slab.remove(token.0);` — the interloper
    /// then takes the listener's key and the `assert_ne!` below fails with
    /// "a deactivated listener's key must not be handed to another session".
    #[test]
    fn a_deactivated_udp_listeners_key_is_not_handed_to_another_session() {
        let mut server = bare_server();
        let address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let listen_token = add_udp_listener(&mut server, address);
        activate(&mut server, address, ListenerType::Udp);
        deactivate(&mut server, address, ListenerType::Udp);

        // Whatever the worker admits next takes the slab's next vacant key.
        // `Protocol::Channel` is just a marker no listener ever reports.
        let interloper = {
            let mut sessions = server.sessions.borrow_mut();
            let entry = sessions.slab.vacant_entry();
            let key = entry.key();
            entry.insert(Rc::new(RefCell::new(ListenSession {
                protocol: Protocol::Channel,
            })));
            Token(key)
        };
        assert_ne!(
            interloper, listen_token,
            "a deactivated listener's key must not be handed to another session"
        );

        activate(&mut server, address, ListenerType::Udp);

        assert_eq!(
            slot_state(&server, interloper),
            Some((Protocol::Channel, true)),
            "reactivating a listener must not overwrite another session's slot"
        );
    }

    /// The cluster every UDP data-path test in this module routes through.
    const UDP_FLOW_CLUSTER: &str = "udp-duplicate-activate";

    /// The session Rc currently occupying `token`. Identity (`Rc::ptr_eq`)
    /// across an operation is the direct statement that the operation did not
    /// replace the live session — `UdpListenerSession`'s per-flow maps are
    /// private to `udp.rs`, so identity plus an end-to-end datagram are what
    /// this module can observe.
    fn slab_session(server: &Server, token: Token) -> Rc<RefCell<dyn ProxySession>> {
        server
            .sessions
            .borrow()
            .slab
            .get(token.0)
            .expect("the listener must own a slab slot")
            .clone()
    }

    /// The current value of the `udp.active_flows` gauge for this test thread.
    /// `METRICS` is a `thread_local!` and libtest gives each test its own
    /// thread, so this reads only what the test itself emitted.
    fn active_flows_gauge() -> i64 {
        METRICS.with(|metrics| {
            metrics
                .borrow_mut()
                .dump_local_proxy_metrics()
                .get(names::udp::ACTIVE_FLOWS)
                .and_then(|metric| match metric.inner {
                    Some(filtered_metrics::Inner::Gauge(value)) => Some(value as i64),
                    _ => None,
                })
                .unwrap_or(0)
        })
    }

    /// A worker holding one activated UDP listener, a frontend and one backend,
    /// plus the real client and backend sockets. The UDP data path is driven for
    /// real on loopback because the orphaning these tests pin — the shell's
    /// `upstream_to_flow` / `flow_to_upstream` / `upstream_sockets` maps going
    /// empty under a live `UdpManager` — is only observable end to end.
    fn udp_worker_with_backend(
        server: &mut Server,
        address: SocketAddress,
    ) -> (StdUdpSocket, StdUdpSocket) {
        // A backend belongs to its cluster: add the cluster first.
        server.notify_proxys(WorkerRequest {
            id: "test-add-udp-cluster".to_owned(),
            content: RequestType::AddCluster(Cluster {
                cluster_id: UDP_FLOW_CLUSTER.to_owned(),
                ..Default::default()
            })
            .into(),
        });
        server.notify_proxys(WorkerRequest {
            id: "test-add-udp-front".to_owned(),
            content: RequestType::AddUdpFrontend(RequestUdpFrontend {
                cluster_id: UDP_FLOW_CLUSTER.to_owned(),
                address,
                tags: Default::default(),
            })
            .into(),
        });

        let backend_socket =
            StdUdpSocket::bind("127.0.0.1:0").expect("could not bind the test UDP backend");
        backend_socket
            .set_read_timeout(Some(Duration::from_millis(500)))
            .expect("could not set the backend read timeout");
        let backend_address = backend_socket
            .local_addr()
            .expect("the test UDP backend must have a local address");

        server.notify_proxys(WorkerRequest {
            id: "test-add-udp-backend".to_owned(),
            content: RequestType::AddBackend(AddBackend {
                cluster_id: UDP_FLOW_CLUSTER.to_owned(),
                backend_id: format!("{UDP_FLOW_CLUSTER}-0"),
                address: backend_address.into(),
                load_balancing_parameters: Some(LoadBalancingParams::default()),
                sticky_id: None,
                backup: None,
                unix_path: None,
            })
            .into(),
        });

        let client_socket =
            StdUdpSocket::bind("127.0.0.1:0").expect("could not bind the test UDP client");
        (client_socket, backend_socket)
    }

    /// Send one datagram from `client` to the listener and drive a single
    /// READABLE event through `Server::ready`, exactly as the event loop does.
    /// Returns what reached the backend within its read timeout, or `None` when
    /// nothing was forwarded. Loopback `send_to` has already queued the datagram
    /// on the listener socket by the time it returns, so one pass is enough.
    fn forward_one(
        server: &mut Server,
        listen_token: Token,
        listen_address: SocketAddress,
        client: &StdUdpSocket,
        backend: &StdUdpSocket,
        payload: &[u8],
    ) -> Option<Vec<u8>> {
        let target: std::net::SocketAddr = listen_address.into();
        client
            .send_to(payload, target)
            .expect("the test client must reach the UDP listener");
        server.ready(listen_token, Ready::READABLE);

        let mut received = [0u8; 64];
        backend
            .recv_from(&mut received)
            .ok()
            .map(|(len, _)| received[..len].to_vec())
    }

    /// FAILURE MODE 3, the residual of the listener-slot change: a SECOND
    /// `ActivateListener` for a listener that is already active must not rebuild
    /// its session.
    ///
    /// `UdpListener::activate` short-circuits on its own `active` flag and
    /// answers `Ok(self.token)` without doing any work, so `Ok(token)` does not
    /// mean this call activated anything — and the arm used to run
    /// `build_session` + `*slot = session` on that answer regardless.
    ///
    /// To SEE THIS RED: in `notify_activate_listener`'s UDP arm, drop the
    /// `if !session_installed` guard so `build_session` and the slab write run
    /// unconditionally again — the second activate then installs a fresh
    /// session and the `Rc::ptr_eq` assertion fails with "a repeated
    /// ActivateListener must not replace the live UDP listener session".
    #[test]
    fn a_duplicate_udp_activate_keeps_the_live_listener_session() {
        let mut server = bare_server();
        let address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let listen_token = add_udp_listener(&mut server, address);
        activate(&mut server, address, ListenerType::Udp);

        let installed = slab_session(&server, listen_token);

        let response = server.notify_activate_listener(
            "test-activate-twice",
            &ActivateListener {
                interface: None,
                address,
                proxy: ListenerType::Udp as i32,
                from_scm: false,
            },
        );
        // The response an already-active listener must get. `ok` and not an
        // error: `ConfigState::activate_listener` answers `Ok(())` for the same
        // repeat, `generate_activate_requests` re-emits one per active listener
        // on every state replay, and the HTTP/HTTPS/TCP arms all answer `ok`
        // for their own repeats. Erroring here would fail an ordinary replay and
        // make the worker disagree with the state the main process persisted.
        assert_eq!(
            response.status,
            ResponseStatus::Ok as i32,
            "a repeated ActivateListener on an active UDP listener must answer ok: {response:?}"
        );
        assert_eq!(
            slot_state(&server, listen_token),
            Some((Protocol::UDPListen, false)),
            "the real UdpListenerSession must still be the one in the slab"
        );
        assert!(
            Rc::ptr_eq(&installed, &slab_session(&server, listen_token)),
            "a repeated ActivateListener must not replace the live UDP listener session"
        );
    }

    /// The consequence of that rebuild, end to end. The displaced session took
    /// the whole shell side of every flow with it — `upstream_sockets`,
    /// `upstream_to_flow`, `flow_to_upstream` — while the proxy kept the shared
    /// `UdpManager` and its flow table. `close()` is a `ProxySession` method and
    /// not `Drop`, so nothing tore those flows down; they simply stopped
    /// forwarding while `udp.active_flows` went on counting them.
    ///
    /// To SEE THIS RED: in `notify_activate_listener`'s UDP arm, drop the
    /// `if !session_installed` guard — the datagram sent after the duplicate
    /// activate then never reaches the backend and the assertion fails with
    /// "an in-flight UDP flow must survive a repeated ActivateListener".
    #[test]
    fn in_flight_udp_flows_survive_a_duplicate_udp_activate() {
        let mut server = bare_server();
        let address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let listen_token = add_udp_listener(&mut server, address);
        activate(&mut server, address, ListenerType::Udp);
        let (client, backend) = udp_worker_with_backend(&mut server, address);

        assert_eq!(
            forward_one(
                &mut server,
                listen_token,
                address,
                &client,
                &backend,
                b"before"
            )
            .as_deref(),
            Some(&b"before"[..]),
            "the flow must be established before the duplicate activate"
        );
        assert_eq!(
            active_flows_gauge(),
            1,
            "one established flow must be on the gauge"
        );

        activate(&mut server, address, ListenerType::Udp);

        // The gauge counts what the manager holds, and the manager is retained
        // across the rebuild — so the gauge alone cannot tell a live flow from
        // an orphaned one. Assert both together: the flow the gauge claims must
        // still be a flow the data path can serve.
        assert_eq!(
            active_flows_gauge(),
            1,
            "a repeated ActivateListener must neither open nor evict a flow"
        );
        assert_eq!(
            forward_one(
                &mut server,
                listen_token,
                address,
                &client,
                &backend,
                b"after"
            )
            .as_deref(),
            Some(&b"after"[..]),
            "an in-flight UDP flow must survive a repeated ActivateListener"
        );
    }

    /// The resource the rebuild strands for good. Every flow holds one extra
    /// slab slot for its upstream socket, and `on_close_flow` reaches that slot
    /// only through `flow_to_upstream`. A replacement session's map is empty, so
    /// the eventual manager-driven teardown frees the flow, balances the gauge —
    /// and leaves the upstream slot allocated for the worker's whole life.
    ///
    /// To SEE THIS RED: in `notify_activate_listener`'s UDP arm, drop the
    /// `if !session_installed` guard — the deactivate then leaves the upstream
    /// slot behind and the assertion fails with "tearing the listener down must
    /// release every per-flow slab slot".
    #[test]
    fn a_duplicate_udp_activate_strands_no_upstream_slab_slot() {
        let mut server = bare_server();
        let address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let listen_token = add_udp_listener(&mut server, address);
        activate(&mut server, address, ListenerType::Udp);
        let (client, backend) = udp_worker_with_backend(&mut server, address);

        // Baseline AFTER the listener exists: its own reserved slot is part of
        // `base_sessions_count` and is not what this test measures.
        let baseline = server.sessions.borrow().slab.len();

        assert!(
            forward_one(
                &mut server,
                listen_token,
                address,
                &client,
                &backend,
                b"flow"
            )
            .is_some(),
            "the flow must be established before the duplicate activate"
        );
        assert_eq!(
            server.sessions.borrow().slab.len(),
            baseline + 1,
            "an established flow owns exactly one upstream slab slot"
        );

        activate(&mut server, address, ListenerType::Udp);
        deactivate(&mut server, address, ListenerType::Udp);

        assert_eq!(
            server.sessions.borrow().slab.len(),
            baseline,
            "tearing the listener down must release every per-flow slab slot"
        );
        assert_eq!(
            active_flows_gauge(),
            0,
            "tearing the listener down must return the active-flows gauge to zero"
        );
    }

    /// The path that reaches this arm without anyone typing a command:
    /// `load_state` replays the state the main process persisted, and
    /// `ConfigState::generate_activate_requests` emits one `ActivateListener`
    /// per ACTIVE listener. For a worker whose listener is already up, that
    /// replay is a duplicate activate by construction — the ordinary case, not
    /// an operator mistake, which is why this arm answers `ok` instead of
    /// surfacing the repeat as an error.
    ///
    /// Driven through the real `ConfigState` rather than a hand-written
    /// `ActivateListener`, so the test breaks if that emission ever changes.
    ///
    /// To SEE THIS RED: in `notify_activate_listener`'s UDP arm, drop the
    /// `if !session_installed` guard — the replay then rebuilds the session and
    /// the `Rc::ptr_eq` assertion fails with "replaying ActivateListener must
    /// leave the live UDP listener session in place".
    #[test]
    fn replaying_activate_listener_is_a_clean_no_op_for_an_active_udp_listener() {
        let mut server = bare_server();
        let address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let listener_config = ListenerBuilder::new_udp(address)
            .to_udp(None)
            .expect("could not build a UdpListenerConfig for the test");

        // Through `notify_proxys` so the worker's own `ConfigState` records the
        // listener and its activation, exactly as a live worker's does.
        server.notify_proxys(WorkerRequest {
            id: "test-add-udp".to_owned(),
            content: RequestType::AddUdpListener(listener_config).into(),
        });
        server.notify_proxys(WorkerRequest {
            id: "test-activate".to_owned(),
            content: RequestType::ActivateListener(ActivateListener {
                interface: None,
                address,
                proxy: ListenerType::Udp as i32,
                from_scm: false,
            })
            .into(),
        });
        let listen_token = server
            .udp
            .borrow()
            .listener_token(&address.into())
            .expect("the added UDP listener must own a token");
        let installed = slab_session(&server, listen_token);

        let replayed: Vec<ActivateListener> = server
            .config_state
            .generate_activate_requests()
            .into_iter()
            .filter_map(|request| match request.request_type {
                Some(RequestType::ActivateListener(activate)) => Some(activate),
                _ => None,
            })
            .collect();
        assert_eq!(
            replayed.len(),
            1,
            "an active UDP listener must be replayed exactly once: {replayed:?}"
        );

        for activate in &replayed {
            let response = server.notify_activate_listener("test-replay", activate);
            assert_eq!(
                response.status,
                ResponseStatus::Ok as i32,
                "a replayed ActivateListener must answer ok: {response:?}"
            );
        }

        assert_eq!(
            slot_state(&server, listen_token),
            Some((Protocol::UDPListen, false)),
            "the replay must leave the real UdpListenerSession in the slab"
        );
        assert!(
            Rc::ptr_eq(&installed, &slab_session(&server, listen_token)),
            "replaying ActivateListener must leave the live UDP listener session in place"
        );
    }

    /// The other end of the lifetime. Reserving the slot across a deactivate is
    /// only sound if something releases it: no proxy's `remove_listener`
    /// touches the session slab, so `RemoveListener` is the sole release point.
    /// Without it each add/remove cycle would strand one slab key — and
    /// `base_sessions_count`, decremented on every `RemoveListener`, would
    /// drift below the number of occupied slots, so a soft stop would wait for
    /// sessions that do not exist.
    ///
    /// Three full cycles, because a leak is only visible as growth, and every
    /// other one removes the listener while it is still ACTIVATED: that path
    /// drops a live `UdpListenerSession` out of the slab — the Rc that, left
    /// there, also pinned the listener's `UdpSocket` open and kept `ready()`
    /// dispatching to a removed listener.
    ///
    /// To SEE THIS RED: drop the `if let Some(token) = listen_token` block from
    /// `notify_proxys`'s `RemoveListener` arm — the slab then grows by one slot
    /// per cycle and the assertion fails with "cycle 0: removing a listener
    /// must release its reserved slot".
    #[test]
    fn removing_a_listener_releases_its_reserved_slab_slot() {
        let mut server = bare_server();
        let baseline = server.sessions.borrow().slab.len();
        let base_sessions_count = server.base_sessions_count;

        for cycle in 0..3 {
            let address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
            let listen_token = add_udp_listener(&mut server, address);
            activate(&mut server, address, ListenerType::Udp);
            if cycle % 2 == 0 {
                deactivate(&mut server, address, ListenerType::Udp);
            }

            server.notify_proxys(WorkerRequest {
                id: "test-remove".to_owned(),
                content: RequestType::RemoveListener(RemoveListener {
                    interface: None,
                    address,
                    proxy: ListenerType::Udp as i32,
                })
                .into(),
            });

            assert!(
                !server.sessions.borrow().slab.contains(listen_token.0),
                "cycle {cycle}: removing a listener must release its reserved slot"
            );
            assert_eq!(
                server.sessions.borrow().slab.len(),
                baseline,
                "cycle {cycle}: the slab must return to its baseline occupancy"
            );
            assert_eq!(
                server.base_sessions_count, base_sessions_count,
                "cycle {cycle}: base_sessions_count must track the reserved slots"
            );
        }
    }

    /// The UDP per-source flow limit is opt-in per cluster: a global
    /// `max_connections_per_ip` / `max_connections_per_subnet` — here set at
    /// runtime, as an operator would for HTTP/TCP — never reaches a UDP
    /// listener, while the cluster's own `max_connections_per_ip` does, and
    /// sheds a new port of an address that holds its limit.
    ///
    /// To SEE THIS RED: in `UdpProxy::cluster_config_for`, replace
    /// `per_ip.unwrap_or(0)` with
    /// `per_ip.unwrap_or(self.sessions.borrow().max_connections_per_ip)`, so
    /// UDP inherits the global default; the first assertion then reads `1`.
    #[test]
    fn udp_flow_limits_come_from_the_cluster_not_the_global_defaults() {
        use std::{net::SocketAddr, time::Instant};

        use crate::protocol::udp::{
            BackendId, BackendSource, ManagerInput, MetricEvent, Output, UdpManager,
        };

        struct OneBackend;
        impl BackendSource for OneBackend {
            fn select(
                &mut self,
                _cluster: &str,
                _key: Option<u64>,
                _now: Instant,
            ) -> Option<(BackendId, SocketAddr)> {
                Some(("b".to_owned(), SocketAddr::from(([127, 0, 0, 1], 5353))))
            }
        }
        // Feed one datagram from `port` of one address; `true` when it opened
        // a flow, `false` when it was shed at the per-source limit.
        fn admit(manager: &RefCell<UdpManager>, port: u16) -> bool {
            let mut manager = manager.borrow_mut();
            manager.handle_input(
                ManagerInput::ClientDatagram {
                    src: SocketAddr::from(([192, 0, 2, 1], port)),
                    payload: b"q",
                    backends: &mut OneBackend,
                },
                Instant::now(),
            );
            let mut opened = false;
            let mut shed = false;
            while let Some(out) = manager.poll_output() {
                opened |= matches!(out, Output::OpenUpstream { .. });
                shed |= matches!(out, Output::Metric(MetricEvent::FlowShedSourceLimit));
            }
            assert_ne!(opened, shed, "a datagram is either admitted or shed");
            opened
        }
        fn add_cluster(server: &mut Server, max_connections_per_ip: Option<u64>) {
            server.notify_proxys(WorkerRequest {
                id: "test-add-cluster".to_owned(),
                content: RequestType::AddCluster(Cluster {
                    cluster_id: "udp-limits".to_owned(),
                    max_connections_per_ip,
                    ..Default::default()
                })
                .into(),
            });
        }

        let mut server = bare_server();
        let address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let token = add_udp_listener(&mut server, address);
        server.notify(WorkerRequest {
            id: "test-set-per-ip".to_owned(),
            content: RequestType::SetMaxConnectionsPerIp(1).into(),
        });
        server.notify(WorkerRequest {
            id: "test-set-per-subnet".to_owned(),
            content: RequestType::SetMaxConnectionsPerSubnet(1).into(),
        });
        add_cluster(&mut server, None);
        server.notify_proxys(WorkerRequest {
            id: "test-add-udp-front".to_owned(),
            content: RequestType::AddUdpFrontend(RequestUdpFrontend {
                cluster_id: "udp-limits".to_owned(),
                address,
                tags: Default::default(),
            })
            .into(),
        });
        let manager = server
            .udp
            .borrow()
            .manager(token)
            .expect("the UDP listener must own a flow manager");

        {
            let manager = manager.borrow();
            let cfg = manager.cluster_config();
            assert_eq!(
                cfg.max_flows_per_ip, 0,
                "the global per-IP default must not apply to UDP"
            );
            assert_eq!(
                cfg.max_flows_per_subnet, 0,
                "the global per-subnet default must not apply to UDP"
            );
        }
        assert!(admit(&manager, 1000));
        assert!(
            admit(&manager, 1001),
            "a global limit of 1 must not shed a second UDP flow"
        );

        add_cluster(&mut server, Some(2));
        assert_eq!(manager.borrow().cluster_config().max_flows_per_ip, 2);
        assert!(
            !admit(&manager, 1002),
            "the cluster's own limit of 2 counts the two open flows and sheds a third"
        );
    }
}

#[cfg(test)]
mod add_cluster_validation_tests {
    use sozu_command::proto::command::{LoadBalancingParams, ShardMode};

    use super::listener_lifecycle_tests::bare_server;
    use super::*;
    use crate::{backends::ShardOutcome, load_balancing::affinity_key_from_ip};

    const CLUSTER: &str = "sharded";
    const ADD_CLUSTER_ID: &str = "test-add-invalid-shard";

    /// The worker mirrors `validate_shuffle_sharding` on `AddCluster`
    /// (#524): an `AddCluster` that reaches the worker without the master's
    /// check, as a TOML reload or a `LoadState` can, with `shard_percent`
    /// outside `1..=100` is answered with an error and leaves the cluster
    /// unsharded.
    ///
    /// TO SEE THIS RED: delete the `validate_shuffle_sharding` check from the
    /// `AddCluster` arm of `Server::notify_proxys`; the worker then answers
    /// `Ok` and arms a shard of two over the four backends.
    #[test]
    fn an_add_cluster_with_an_invalid_shard_percent_is_refused_and_unsharded() {
        let mut server = bare_server();
        // The cluster exists unsharded before the refused update: a backend
        // belongs to its cluster, and one of a cluster the worker does not
        // hold is refused (`ConfigState::add_backend`).
        server.notify_proxys(WorkerRequest {
            id: "test-add-unsharded-cluster".to_owned(),
            content: RequestType::AddCluster(Cluster {
                cluster_id: CLUSTER.to_owned(),
                load_balancing: LoadBalancingAlgorithms::RoundRobin as i32,
                ..Default::default()
            })
            .into(),
        });
        server.notify_proxys(WorkerRequest {
            id: ADD_CLUSTER_ID.to_owned(),
            content: RequestType::AddCluster(Cluster {
                cluster_id: CLUSTER.to_owned(),
                load_balancing: LoadBalancingAlgorithms::RoundRobin as i32,
                // The only invalid knob: `shard_min_backends` and
                // `shard_mode` are valid on their own.
                shard_percent: Some(0),
                shard_min_backends: Some(2),
                shard_mode: Some(ShardMode::Strict as i32),
                ..Default::default()
            })
            .into(),
        });
        let response = QUEUE
            .with(|queue| {
                queue
                    .borrow()
                    .iter()
                    .find(|response| response.id == ADD_CLUSTER_ID)
                    .cloned()
            })
            .expect("the worker must answer the AddCluster");
        assert_eq!(
            response.status,
            ResponseStatus::Failure as i32,
            "an invalid shard_percent must be refused: {response:?}"
        );
        assert_eq!(response.message, "shard_percent must lie in 1..=100");

        for index in 0..4u16 {
            server.notify_proxys(WorkerRequest {
                id: format!("test-add-backend-{index}"),
                content: RequestType::AddBackend(AddBackend {
                    cluster_id: CLUSTER.to_owned(),
                    backend_id: format!("{CLUSTER}-{index}"),
                    address: SocketAddr::from(([127, 0, 0, 1], 23_000 + index)).into(),
                    load_balancing_parameters: Some(LoadBalancingParams::default()),
                    sticky_id: None,
                    backup: None,
                    unix_path: None,
                })
                .into(),
            });
        }
        let key = affinity_key_from_ip(IpAddr::V4(Ipv4Addr::LOCALHOST));
        let mut backends = server.backends.borrow_mut();
        let list = backends
            .backends
            .get_mut(CLUSTER)
            .expect("the backends must have created the cluster's list");
        assert_eq!(
            list.select_with_key(Some(key), Instant::now()).1,
            ShardOutcome::Unsharded,
            "a refused AddCluster must not arm a shard"
        );
    }
}

/// A health check `port` targets a backend IP and a `unix_path` backend has
/// none: the worker refuses the pair in either order before its state, its
/// `BackendMap` or its `HealthChecker` change.
#[cfg(test)]
mod health_check_port_unix_tests {
    use sozu_command::{
        config::HEALTH_CHECK_PORT_UNIX_REFUSED,
        proto::command::{HealthCheckConfig, HealthCheckMode, SetHealthCheck, SocketAddress},
    };

    use super::listener_lifecycle_tests::bare_server;
    use super::*;

    const CLUSTER: &str = "port-unix-cluster";

    fn probe(port: Option<u32>) -> HealthCheckConfig {
        HealthCheckConfig {
            interval: 10,
            timeout: 5,
            healthy_threshold: 3,
            unhealthy_threshold: 3,
            mode: HealthCheckMode::Tcp as i32,
            port,
            ..Default::default()
        }
    }

    fn cluster(health_check: Option<HealthCheckConfig>) -> RequestType {
        RequestType::AddCluster(Cluster {
            cluster_id: CLUSTER.to_owned(),
            health_check,
            ..Default::default()
        })
    }

    fn unix_backend() -> RequestType {
        RequestType::AddUnixBackend(AddBackend {
            cluster_id: CLUSTER.to_owned(),
            backend_id: format!("{CLUSTER}-0"),
            address: SocketAddress::new_v4(127, 0, 0, 1, 1),
            unix_path: Some("/run/app.sock".to_owned()),
            ..Default::default()
        })
    }

    fn send(server: &mut Server, request: RequestType) -> Vec<WorkerResponse> {
        server.notify_proxys(WorkerRequest {
            id: "test-request".to_owned(),
            content: request.into(),
        });
        QUEUE.with(|queue| {
            queue
                .borrow_mut()
                .drain(..)
                .filter(|response| response.id != "EVENT")
                .collect()
        })
    }

    /// What the refusal must leave alone: the persisted state, the
    /// `BackendMap` (backend incarnations by allocation, probe configs) and
    /// the probes in flight.
    fn snapshot(server: &Server) -> (String, Vec<*const RefCell<Backend>>, Vec<String>, usize) {
        let backends = server.backends.borrow();
        (
            format!(
                "{:?}{:?}",
                server.config_state.clusters, server.config_state.backends
            ),
            backends
                .backends
                .get(CLUSTER)
                .map(|list| list.backends.iter().map(Rc::as_ptr).collect())
                .unwrap_or_default(),
            backends
                .health_check_configs
                .values()
                .map(|config| format!("{config:?}"))
                .collect(),
            server.health_checker.in_flight_len(),
        )
    }

    fn assert_refused(server: &mut Server, request: RequestType) {
        let before = snapshot(server);
        let responses = send(server, request);
        assert_eq!(responses.len(), 1, "one answer expected: {responses:?}");
        assert_eq!(responses[0].status, ResponseStatus::Failure as i32);
        assert!(
            responses[0]
                .message
                .contains(HEALTH_CHECK_PORT_UNIX_REFUSED),
            "the answer names the refusal: {responses:?}"
        );
        assert_eq!(before, snapshot(server), "a refusal changes nothing");
    }

    #[test]
    fn a_probe_port_already_set_refuses_a_unix_backend() {
        let mut server = bare_server();
        send(&mut server, cluster(Some(probe(Some(8081)))));
        assert_refused(&mut server, unix_backend());
        assert!(snapshot(&server).1.is_empty(), "no backend was added");
    }

    #[test]
    fn a_unix_backend_already_there_refuses_a_probe_port() {
        let mut server = bare_server();
        send(&mut server, cluster(Some(probe(None))));
        send(&mut server, unix_backend());
        assert_eq!(snapshot(&server).1.len(), 1, "test setup: the unix backend");

        assert_refused(
            &mut server,
            RequestType::SetHealthCheck(SetHealthCheck {
                cluster_id: CLUSTER.to_owned(),
                config: probe(Some(8081)),
            }),
        );
        assert_refused(&mut server, cluster(Some(probe(Some(8081)))));
        assert_eq!(
            server.backends.borrow().health_check_configs[CLUSTER].port,
            None,
            "the previous probe config stays"
        );
    }
}

/// Clearing a cluster's health check resets the backends a probe marked DOWN
/// (#1811): no probe will ever mark them UP again.
#[cfg(test)]
mod remove_health_check_tests {
    use sozu_command::proto::command::{HealthCheckConfig, LoadBalancingParams};

    use super::listener_lifecycle_tests::bare_server;
    use super::*;

    const CLUSTER: &str = "health-checked";

    fn health_check() -> HealthCheckConfig {
        HealthCheckConfig {
            uri: "/livez".to_owned(),
            interval: 1,
            timeout: 1,
            healthy_threshold: 1,
            unhealthy_threshold: 1,
            ..Default::default()
        }
    }

    fn add_cluster(server: &mut Server, health_check: Option<HealthCheckConfig>) {
        server.notify_proxys(WorkerRequest {
            id: "test-add-cluster".to_owned(),
            content: RequestType::AddCluster(Cluster {
                cluster_id: CLUSTER.to_owned(),
                health_check,
                ..Default::default()
            })
            .into(),
        });
    }

    /// A server whose cluster carries a health check and one backend a probe
    /// marked DOWN.
    fn server_with_down_backend() -> Server {
        let mut server = bare_server();
        add_cluster(&mut server, Some(health_check()));
        server.notify_proxys(WorkerRequest {
            id: "test-add-backend".to_owned(),
            content: RequestType::AddBackend(AddBackend {
                cluster_id: CLUSTER.to_owned(),
                backend_id: format!("{CLUSTER}-0"),
                address: SocketAddr::from(([127, 0, 0, 1], 23_100)).into(),
                load_balancing_parameters: Some(LoadBalancingParams::default()),
                sticky_id: None,
                backup: None,
                unix_path: None,
            })
            .into(),
        });
        server.backends.borrow().backends[CLUSTER].backends[0]
            .borrow_mut()
            .health
            .record_failure(1);
        assert!(!backend_is_healthy(&server), "test setup: backend DOWN");
        server
    }

    fn backend_is_healthy(server: &Server) -> bool {
        server.backends.borrow().backends[CLUSTER].backends[0]
            .borrow()
            .health
            .is_healthy()
    }

    /// TO SEE THIS RED: make `Server::remove_health_check_state` remove the
    /// cluster from `health_check_configs` directly instead of calling
    /// `set_health_check_config(cluster_id, None)`; the backend stays DOWN.
    #[test]
    fn remove_health_check_resets_a_backend_marked_down() {
        let mut server = server_with_down_backend();
        server.notify_proxys(WorkerRequest {
            id: "test-remove-health-check".to_owned(),
            content: RequestType::RemoveHealthCheck(CLUSTER.to_owned()).into(),
        });
        assert!(
            !server
                .backends
                .borrow()
                .health_check_configs
                .contains_key(CLUSTER),
            "RemoveHealthCheck must forget the cluster's configuration"
        );
        assert!(
            backend_is_healthy(&server),
            "RemoveHealthCheck must reset a backend a probe marked DOWN"
        );
    }

    /// An `AddCluster` upsert without a health check clears it the same way.
    ///
    /// TO SEE THIS RED: delete the `set_health_check_config` call from
    /// `Server::add_cluster`; the backend stays DOWN. The in-flight probes the
    /// `AddCluster` arm drops are not covered: that needs a live probe socket.
    #[test]
    fn add_cluster_without_health_check_resets_a_backend_marked_down() {
        let mut server = server_with_down_backend();
        add_cluster(&mut server, None);
        assert!(
            !server
                .backends
                .borrow()
                .health_check_configs
                .contains_key(CLUSTER),
            "an AddCluster without a health check must forget the configuration"
        );
        assert!(
            backend_is_healthy(&server),
            "an AddCluster without a health check must reset a backend marked DOWN"
        );
    }

    /// An idempotent reload re-sends the same `AddCluster` (health check
    /// `port` included) and the same `AddBackend`: the DOWN backend and its
    /// failure counters must survive, only a `None` health check resets them.
    ///
    /// TO SEE THIS RED: make `Server::add_cluster` pass `None` to
    /// `set_health_check_config`; the counters are cleared.
    #[test]
    fn an_idempotent_reload_keeps_a_backend_marked_down_and_its_counters() {
        let ported = HealthCheckConfig {
            port: Some(8081),
            ..health_check()
        };
        let mut server = bare_server();
        add_cluster(&mut server, Some(ported.clone()));
        let add_backend = || WorkerRequest {
            id: "test-add-backend".to_owned(),
            content: RequestType::AddBackend(AddBackend {
                cluster_id: CLUSTER.to_owned(),
                backend_id: format!("{CLUSTER}-0"),
                address: SocketAddr::from(([127, 0, 0, 1], 23_100)).into(),
                load_balancing_parameters: Some(LoadBalancingParams::default()),
                sticky_id: None,
                backup: None,
                unix_path: None,
            })
            .into(),
        };
        server.notify_proxys(add_backend());
        {
            let backends = server.backends.borrow();
            let mut backend = backends.backends[CLUSTER].backends[0].borrow_mut();
            // thresholds above the failure count keep the status stable; two
            // failures leave a nonzero streak, the third marks it DOWN
            for _ in 0..3 {
                backend.health.record_failure(3);
            }
        }
        let before = server.backends.borrow().backends[CLUSTER].backends[0]
            .borrow()
            .health
            .clone();
        assert!(!before.is_healthy(), "test setup: backend DOWN");
        assert_eq!(before.consecutive_failures, 3, "test setup: counters");

        // the order a reload sends: the cluster first, then its backend
        add_cluster(&mut server, Some(ported.clone()));
        server.notify_proxys(add_backend());

        let after = server.backends.borrow().backends[CLUSTER].backends[0]
            .borrow()
            .health
            .clone();
        assert_eq!(after, before, "a reload must not reset the runtime health");
        assert_eq!(
            server.backends.borrow().health_check_configs[CLUSTER],
            ported,
            "the health check configuration stays"
        );
    }
}

/// The SCM hand-off a worker upgrade performs: the retiring worker's listening
/// sockets travel to the new worker over the SCM socket, and the new worker
/// must ADOPT them instead of binding fresh ones.
/// `RemoveCluster` removes the cluster's frontends and backends from the
/// worker's proxies, the way `ConfigState::remove_cluster`
/// (`command/src/state.rs`) removes them from the state.
#[cfg(test)]
mod remove_cluster_cascade_tests {
    use sozu_command::{
        config::ListenerBuilder,
        proto::command::{
            LoadBalancingParams, PathRule, RequestHttpFrontend, RequestTcpFrontend,
            RequestUdpFrontend, RulePosition, SocketAddress,
        },
    };

    use super::listener_lifecycle_tests::bare_server;
    use super::*;
    use crate::{L7ListenerHandler, protocol::http::parser::Method, testing::provide_port};

    const REMOVED: &str = "cascade_removed";
    const KEPT: &str = "cascade_kept";

    /// The answers the worker queues for `request`. Events it queues along
    /// the way (a removed backend's, for instance) are not answers and are
    /// left out.
    fn send(server: &mut Server, id: &str, request: RequestType) -> Vec<WorkerResponse> {
        server.notify_proxys(WorkerRequest {
            id: id.to_owned(),
            content: request.into(),
        });
        QUEUE.with(|queue| {
            queue
                .borrow_mut()
                .drain(..)
                .filter(|response| response.id != "EVENT")
                .collect()
        })
    }

    fn http_front(cluster_id: Option<&str>, hostname: &str, path: &str) -> RequestHttpFrontend {
        RequestHttpFrontend {
            cluster_id: cluster_id.map(str::to_owned),
            hostname: hostname.to_owned(),
            path: PathRule::prefix(path.to_owned()),
            address: SocketAddress::new_v4(127, 0, 0, 1, 0),
            ..Default::default()
        }
    }

    /// A proxy can reject an order after the worker's `ConfigState` accepted
    /// it. The worker must then restore its state, or `QueryClusterById`
    /// reports an object the runtime never installed and disagrees with the
    /// main process after its own rollback.
    #[test]
    fn a_proxy_rejection_rolls_back_reversible_worker_state() {
        let cluster_id = "worker-rollback";
        let address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let http = http_front(Some(cluster_id), "rollback.test", "/");

        let requests = [
            RequestType::AddHttpFrontend(RequestHttpFrontend {
                address,
                ..http.clone()
            }),
            RequestType::AddHttpsFrontend(RequestHttpFrontend { address, ..http }),
            RequestType::AddTcpFrontend(RequestTcpFrontend {
                cluster_id: cluster_id.to_owned(),
                address,
                ..Default::default()
            }),
            RequestType::AddUdpFrontend(RequestUdpFrontend {
                cluster_id: cluster_id.to_owned(),
                address,
                ..Default::default()
            }),
        ];

        for (index, request) in requests.into_iter().enumerate() {
            let mut server = bare_server();
            for response in send(
                &mut server,
                &format!("cluster-{index}"),
                RequestType::AddCluster(Cluster {
                    cluster_id: cluster_id.to_owned(),
                    ..Default::default()
                }),
            ) {
                assert_eq!(response.status, ResponseStatus::Ok as i32);
            }

            let responses = send(&mut server, &format!("frontend-{index}"), request);
            assert_eq!(
                responses
                    .iter()
                    .map(|response| response.status)
                    .collect::<Vec<_>>(),
                vec![ResponseStatus::Failure as i32],
                "a frontend with no listener must be refused: {responses:?}"
            );

            let cluster = server
                .config_state
                .cluster_state(cluster_id)
                .expect("the setup cluster must stay in worker state");
            assert!(
                cluster.http_frontends.is_empty()
                    && cluster.https_frontends.is_empty()
                    && cluster.tcp_frontends.is_empty()
                    && cluster.udp_frontends.is_empty(),
                "a refused frontend must not remain in worker ConfigState: {cluster:?}"
            );
        }
    }

    /// Removing a frontend is keyed by only its routing identity. If the
    /// runtime refuses the removal after `ConfigState` accepted it, rollback
    /// must restore the complete stored frontend rather than manufacture an
    /// `Add` from the removal selector and lose metadata.
    #[test]
    fn a_refused_frontend_removal_restores_the_exact_stored_frontend() {
        let cluster_id = "worker-remove-rollback";
        let address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let tags = std::collections::BTreeMap::from([("owner".to_owned(), "preserved".to_owned())]);
        let mut http = http_front(Some(cluster_id), "remove.test", "/api");
        http.address = address;
        http.method = Some("POST".to_owned());
        http.position = RulePosition::Post.into();
        http.tags = tags.clone();
        http.rewrite_host = Some("internal.test".to_owned());

        let mut http_selector = http.clone();
        http_selector.position = RulePosition::Tree.into();
        http_selector.tags.clear();
        http_selector.rewrite_host = None;

        let mut https = http.clone();
        https.hostname = "secure-remove.test".to_owned();
        let mut https_selector = https.clone();
        https_selector.position = RulePosition::Tree.into();
        https_selector.tags.clear();
        https_selector.rewrite_host = None;

        let tcp = RequestTcpFrontend {
            cluster_id: cluster_id.to_owned(),
            address,
            tags: tags.clone(),
            sni: Some("REMOVE.TEST".to_owned()),
            alpn: vec!["http/1.1".to_owned(), "h2".to_owned()],
        };
        let tcp_selector = RequestTcpFrontend {
            tags: Default::default(),
            sni: Some("remove.test".to_owned()),
            alpn: vec!["h2".to_owned(), "http/1.1".to_owned()],
            ..tcp.clone()
        };

        let udp = RequestUdpFrontend {
            cluster_id: cluster_id.to_owned(),
            address,
            tags,
        };

        let cases = [
            (
                "http",
                RequestType::AddHttpFrontend(http),
                RequestType::RemoveHttpFrontend(http_selector),
            ),
            (
                "https",
                RequestType::AddHttpsFrontend(https),
                RequestType::RemoveHttpsFrontend(https_selector),
            ),
            (
                "tcp",
                RequestType::AddTcpFrontend(tcp),
                RequestType::RemoveTcpFrontend(tcp_selector),
            ),
            (
                "udp",
                RequestType::AddUdpFrontend(udp.clone()),
                RequestType::RemoveUdpFrontend(udp),
            ),
        ];

        for (protocol, add, remove) in cases {
            let mut server = bare_server();
            server
                .config_state
                .dispatch(
                    &RequestType::AddCluster(Cluster {
                        cluster_id: cluster_id.to_owned(),
                        ..Default::default()
                    })
                    .into(),
                )
                .expect("the setup cluster must enter worker state");
            server
                .config_state
                .dispatch(&add.into())
                .unwrap_or_else(|error| {
                    panic!("the {protocol} setup frontend must be stored: {error}")
                });
            let before = server
                .config_state
                .cluster_state(cluster_id)
                .expect("the setup cluster must expose its frontend");

            let responses = send(&mut server, &format!("remove-{protocol}"), remove);
            assert_eq!(
                responses
                    .iter()
                    .map(|response| response.status)
                    .collect::<Vec<_>>(),
                vec![ResponseStatus::Failure as i32],
                "a frontend removal with no runtime listener must be refused: {responses:?}"
            );
            assert_eq!(
                server.config_state.cluster_state(cluster_id),
                Some(before),
                "a refused {protocol} removal must restore the complete stored frontend"
            );
        }
    }

    /// A request-only inverse is safe only after the corresponding `Add`
    /// entered `ConfigState`. A duplicate is rejected before mutation; if the
    /// runtime then refuses it too, applying the inverse would delete the
    /// frontend that predated the request.
    #[test]
    fn a_refused_duplicate_frontend_add_preserves_the_preexisting_frontend() {
        let cluster_id = "worker-add-rollback";
        let address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let tags =
            std::collections::BTreeMap::from([("owner".to_owned(), "preexisting".to_owned())]);
        let mut http = http_front(Some(cluster_id), "duplicate.test", "/api");
        http.address = address;
        http.method = Some("PATCH".to_owned());
        http.position = RulePosition::Pre.into();
        http.tags = tags.clone();

        let mut https = http.clone();
        https.hostname = "secure-duplicate.test".to_owned();
        let tcp = RequestTcpFrontend {
            cluster_id: cluster_id.to_owned(),
            address,
            tags: tags.clone(),
            sni: Some("duplicate.test".to_owned()),
            alpn: vec!["h2".to_owned()],
        };
        let udp = RequestUdpFrontend {
            cluster_id: cluster_id.to_owned(),
            address,
            tags,
        };

        let cases = [
            ("http", RequestType::AddHttpFrontend(http)),
            ("https", RequestType::AddHttpsFrontend(https)),
            ("tcp", RequestType::AddTcpFrontend(tcp)),
            ("udp", RequestType::AddUdpFrontend(udp)),
        ];

        for (protocol, add) in cases {
            let mut server = bare_server();
            server
                .config_state
                .dispatch(
                    &RequestType::AddCluster(Cluster {
                        cluster_id: cluster_id.to_owned(),
                        ..Default::default()
                    })
                    .into(),
                )
                .expect("the setup cluster must enter worker state");
            server
                .config_state
                .dispatch(&add.clone().into())
                .unwrap_or_else(|error| {
                    panic!("the {protocol} setup frontend must be stored: {error}")
                });
            let before = server
                .config_state
                .cluster_state(cluster_id)
                .expect("the setup cluster must expose its frontend");

            let responses = send(&mut server, &format!("duplicate-{protocol}"), add);
            assert_eq!(
                responses
                    .iter()
                    .map(|response| response.status)
                    .collect::<Vec<_>>(),
                vec![ResponseStatus::Failure as i32],
                "a duplicate frontend with no runtime listener must be refused: {responses:?}"
            );
            assert_eq!(
                server.config_state.cluster_state(cluster_id),
                Some(before),
                "a refused duplicate {protocol} add must preserve the preexisting frontend"
            );
        }
    }

    /// Listener adds are request-invertible only on the vacant state path. A
    /// duplicate state error must leave rollback unarmed when the runtime also
    /// refuses the order.
    #[test]
    fn a_refused_duplicate_listener_add_preserves_the_preexisting_listener() {
        let address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let listener = ListenerBuilder::new_http(address)
            .to_http(None)
            .expect("the test HTTP listener configuration must build");
        let mut server = bare_server();
        server
            .config_state
            .dispatch(&RequestType::AddHttpListener(listener.clone()).into())
            .expect("the setup listener must enter worker state");
        let before = server.config_state.http_listeners.clone();

        // Fill the reserved slab headroom so the worker takes the explicit
        // refusal path before attempting to install a runtime listener.
        let threshold = server.sessions.borrow().accept_slab_threshold();
        while server.sessions.borrow().slab.len() < threshold {
            server
                .sessions
                .borrow_mut()
                .slab
                .insert(Rc::new(RefCell::new(ListenSession {
                    protocol: Protocol::HTTPListen,
                })));
        }
        assert!(server.sessions.borrow().at_capacity());

        let responses = send(
            &mut server,
            "duplicate-listener",
            RequestType::AddHttpListener(listener),
        );
        assert_eq!(
            responses
                .iter()
                .map(|response| response.status)
                .collect::<Vec<_>>(),
            vec![ResponseStatus::Failure as i32],
            "the full session slab must refuse the duplicate listener: {responses:?}"
        );
        assert_eq!(
            server.config_state.http_listeners, before,
            "a refused duplicate listener add must preserve the preexisting listener"
        );
    }

    fn backend(cluster_id: &str, port: u16) -> AddBackend {
        AddBackend {
            cluster_id: cluster_id.to_owned(),
            backend_id: format!("{cluster_id}-0"),
            address: SocketAddress::new_v4(127, 0, 0, 1, port),
            load_balancing_parameters: Some(LoadBalancingParams::default()),
            ..Default::default()
        }
    }

    /// Admission leases belong to live frontend-session tokens, not to the
    /// current cluster configuration object. Removing a cluster therefore
    /// leaves the old session's contribution in both counters; re-adding the
    /// same id lets a new session contribute alongside it; closing either
    /// session releases exactly that token's contribution.
    ///
    /// To SEE THIS RED: change `untrack_all_cluster_ip`'s per-IP decrement from
    /// `saturating_sub(1)` to `saturating_sub(0)`. The old lease then remains at
    /// two and the post-close assertion fails, proving this observes release
    /// rather than only reconfiguration.
    #[test]
    fn remove_then_readd_same_cluster_id_preserves_and_releases_live_admission_leases() {
        let mut server = bare_server();
        let cluster_id = ClusterId::from(REMOVED);
        let source: IpAddr = "203.0.113.9".parse().expect("test address must parse");
        let subnet = server.sessions.borrow().subnet_key(&source);
        let old_token = Token(60_001);
        let new_token = Token(60_002);

        for response in send(
            &mut server,
            "add-old-incarnation",
            RequestType::AddCluster(Cluster {
                cluster_id: REMOVED.to_owned(),
                max_connections_per_ip: Some(2),
                max_connections_per_subnet: Some(2),
                ..Default::default()
            }),
        ) {
            assert_eq!(response.status, ResponseStatus::Ok as i32);
        }
        server.sessions.borrow_mut().track_cluster_connection(
            old_token,
            cluster_id.clone(),
            source,
            Some(2),
        );

        for response in send(
            &mut server,
            "remove-old-incarnation",
            RequestType::RemoveCluster(REMOVED.to_owned()),
        ) {
            assert_eq!(response.status, ResponseStatus::Ok as i32);
        }
        for response in send(
            &mut server,
            "add-new-incarnation",
            RequestType::AddCluster(Cluster {
                cluster_id: REMOVED.to_owned(),
                max_connections_per_ip: Some(2),
                max_connections_per_subnet: Some(2),
                ..Default::default()
            }),
        ) {
            assert_eq!(response.status, ResponseStatus::Ok as i32);
        }

        {
            let mut sessions = server.sessions.borrow_mut();
            assert_eq!(sessions.connections_per_cluster_ip[REMOVED][&source], 1);
            assert_eq!(sessions.connections_per_cluster_subnet[REMOVED][&subnet], 1);
            sessions.track_cluster_connection(new_token, cluster_id, source, Some(2));
            assert_eq!(sessions.connections_per_cluster_ip[REMOVED][&source], 2);
            assert_eq!(sessions.connections_per_cluster_subnet[REMOVED][&subnet], 2);

            sessions.untrack_all_cluster_ip(old_token);
            assert_eq!(sessions.connections_per_cluster_ip[REMOVED][&source], 1);
            assert_eq!(sessions.connections_per_cluster_subnet[REMOVED][&subnet], 1);
            assert!(sessions.cluster_ip_tracks.contains_key(&new_token));
            assert!(sessions.cluster_subnet_tracks.contains_key(&new_token));

            sessions.untrack_all_cluster_ip(new_token);
            assert!(!sessions.connections_per_cluster_ip.contains_key(REMOVED));
            assert!(
                !sessions
                    .connections_per_cluster_subnet
                    .contains_key(REMOVED)
            );
        }
    }

    /// The route an HTTP request to `hostname` + `path` resolves to on the
    /// worker's only HTTP listener, as a cluster id (`Some(None)` for a route
    /// that names no cluster), or `None` when nothing routes it.
    fn http_route(
        server: &Server,
        address: SocketAddress,
        hostname: &str,
        path: &str,
    ) -> Option<Option<String>> {
        let http = server.http.borrow();
        let token = http
            .listener_token(&SocketAddr::from(address).into())
            .expect("the test HTTP listener must own a token");
        let listener = http
            .get_listener(&token)
            .expect("the test HTTP listener must exist");
        listener
            .borrow()
            .frontend_from_request(hostname, path, &Method::Get, &crate::router::NoHeaders)
            .ok()
            .map(|route| route.cluster_id.map(|id| id.to_string()))
    }

    /// A removed cluster's HTTP routes and backends are gone from the worker,
    /// while another cluster's and a route that names no cluster stay. The
    /// worker answers the `RemoveCluster` once, and a removal of one of the
    /// cluster's objects sent afterwards (what a main process that predates
    /// the cascade still sends) answers ok instead of failing on a route the
    /// cascade already removed.
    #[test]
    fn remove_cluster_removes_its_routes_and_backends_from_the_worker() {
        let mut server = bare_server();
        let http_address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let tcp_address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());

        let mut removed_front = http_front(Some(REMOVED), "removed.test", "/");
        removed_front.address = http_address;
        // A pre-rule: unlike a tree rule, removing it twice is an error in
        // `Router::remove_pre_rule` (`lib/src/router/mod.rs`).
        removed_front.position = RulePosition::Pre.into();
        let mut deny_front = http_front(None, "removed.test", "/deny");
        deny_front.address = http_address;
        let mut kept_front = http_front(Some(KEPT), "kept.test", "/");
        kept_front.address = http_address;
        let removed_tcp_front = RequestTcpFrontend {
            cluster_id: REMOVED.to_owned(),
            address: tcp_address,
            ..Default::default()
        };

        let setup = vec![
            RequestType::AddHttpListener(
                ListenerBuilder::new_http(http_address)
                    .to_http(None)
                    .expect("could not build the test HTTP listener config"),
            ),
            RequestType::AddTcpListener(
                ListenerBuilder::new_tcp(tcp_address)
                    .to_tcp(None)
                    .expect("could not build the test TCP listener config"),
            ),
            RequestType::AddCluster(Cluster {
                cluster_id: REMOVED.to_owned(),
                ..Default::default()
            }),
            RequestType::AddCluster(Cluster {
                cluster_id: KEPT.to_owned(),
                ..Default::default()
            }),
            RequestType::AddHttpFrontend(removed_front.clone()),
            RequestType::AddHttpFrontend(deny_front),
            RequestType::AddHttpFrontend(kept_front),
            RequestType::AddTcpFrontend(removed_tcp_front.clone()),
            RequestType::AddBackend(backend(REMOVED, 1026)),
            RequestType::AddBackend(backend(KEPT, 1027)),
        ];
        for (index, request) in setup.into_iter().enumerate() {
            for response in send(&mut server, &format!("setup-{index}"), request) {
                assert_eq!(
                    response.status,
                    ResponseStatus::Ok as i32,
                    "setup must succeed: {response:?}"
                );
            }
        }
        assert_eq!(
            http_route(&server, http_address, "removed.test", "/"),
            Some(Some(REMOVED.to_owned())),
            "premise: the removed cluster's hostname routes to it"
        );

        let responses = send(
            &mut server,
            "remove-cluster",
            RequestType::RemoveCluster(REMOVED.to_owned()),
        );
        assert_eq!(
            responses.iter().map(|r| r.id.as_str()).collect::<Vec<_>>(),
            vec!["remove-cluster"],
            "the worker must answer the RemoveCluster exactly once: {responses:?}"
        );
        assert_eq!(
            responses[0].status,
            ResponseStatus::Ok as i32,
            "RemoveCluster must succeed: {responses:?}"
        );

        assert_eq!(
            http_route(&server, http_address, "removed.test", "/"),
            None,
            "the removed cluster's hostname must no longer route"
        );
        assert_eq!(
            http_route(&server, http_address, "removed.test", "/deny"),
            Some(None),
            "a route that names no cluster must survive the removal of a cluster on its hostname"
        );
        assert_eq!(
            http_route(&server, http_address, "kept.test", "/"),
            Some(Some(KEPT.to_owned())),
            "another cluster's route must stay"
        );
        let backends = server.backends.borrow();
        assert!(
            !backends.backends.contains_key(REMOVED),
            "the removed cluster's backends must be gone from the BackendMap"
        );
        assert!(
            backends.backends.contains_key(KEPT),
            "another cluster's backends must stay"
        );
        drop(backends);

        for (index, request) in [
            RequestType::RemoveHttpFrontend(removed_front),
            RequestType::RemoveTcpFrontend(removed_tcp_front),
            RequestType::RemoveBackend(RemoveBackend {
                cluster_id: REMOVED.to_owned(),
                backend_id: format!("{REMOVED}-0"),
                address: SocketAddress::new_v4(127, 0, 0, 1, 1026),
            }),
        ]
        .into_iter()
        .enumerate()
        {
            let responses = send(&mut server, &format!("follow-up-{index}"), request);
            assert!(
                !responses.is_empty()
                    && responses
                        .iter()
                        .all(|r| r.status == ResponseStatus::Ok as i32),
                "a removal the cascade already performed must answer ok: {responses:?}"
            );
        }
    }

    /// A backend belongs to its cluster: the worker refuses an `AddBackend`
    /// for a cluster it does not hold, never added or already removed, and
    /// never adds it to its `BackendMap`, where it would outlive its cluster.
    #[test]
    fn a_backend_of_an_unknown_cluster_is_refused_by_the_worker() {
        const ORPHAN: &str = "orphan";
        let mut server = bare_server();
        let assert_refused = |server: &mut Server, id: &str, port: u16| {
            let responses = send(server, id, RequestType::AddBackend(backend(ORPHAN, port)));
            assert!(
                !responses.is_empty()
                    && responses
                        .iter()
                        .all(|r| r.status == ResponseStatus::Failure as i32),
                "a backend of an unknown cluster must be refused: {responses:?}"
            );
            assert!(
                !server.backends.borrow().backends.contains_key(ORPHAN),
                "a refused backend must not reach the BackendMap"
            );
        };

        assert_refused(&mut server, "never-added", 1026);

        for (index, request) in [
            RequestType::AddCluster(Cluster {
                cluster_id: ORPHAN.to_owned(),
                ..Default::default()
            }),
            RequestType::AddBackend(backend(ORPHAN, 1026)),
            RequestType::RemoveCluster(ORPHAN.to_owned()),
        ]
        .into_iter()
        .enumerate()
        {
            for response in send(&mut server, &format!("step-{index}"), request) {
                assert_eq!(
                    response.status,
                    ResponseStatus::Ok as i32,
                    "step must succeed: {response:?}"
                );
            }
        }

        assert_refused(&mut server, "removed", 1027);
    }

    /// The route key names no cluster, so cluster B can take the key cluster
    /// A's frontend used. A removal of A's frontend arriving afterwards must
    /// leave B's route in the worker's router, whether A is gone (the removal
    /// answers ok: A's frontend went with A) or still there (refused).
    #[test]
    fn a_stale_frontend_removal_never_removes_another_clusters_route() {
        let http_address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let route = |cluster_id: &str| {
            let mut front = http_front(Some(cluster_id), "reused.test", "/");
            front.address = http_address;
            front
        };
        let cluster = |cluster_id: &str| {
            RequestType::AddCluster(Cluster {
                cluster_id: cluster_id.to_owned(),
                ..Default::default()
            })
        };
        let listener = || {
            RequestType::AddHttpListener(
                ListenerBuilder::new_http(http_address)
                    .to_http(None)
                    .expect("could not build the test HTTP listener config"),
            )
        };

        for (a_is_gone, setup, expected_status) in [
            (
                true,
                vec![
                    listener(),
                    cluster("stale_a"),
                    RequestType::AddHttpFrontend(route("stale_a")),
                    RequestType::RemoveCluster("stale_a".to_owned()),
                    cluster("stale_b"),
                    RequestType::AddHttpFrontend(route("stale_b")),
                ],
                ResponseStatus::Ok,
            ),
            (
                false,
                vec![
                    listener(),
                    cluster("stale_a"),
                    cluster("stale_b"),
                    RequestType::AddHttpFrontend(route("stale_b")),
                ],
                ResponseStatus::Failure,
            ),
        ] {
            let mut server = bare_server();
            for (index, request) in setup.into_iter().enumerate() {
                for response in send(&mut server, &format!("setup-{index}"), request) {
                    assert_eq!(
                        response.status,
                        ResponseStatus::Ok as i32,
                        "setup must succeed: {response:?}"
                    );
                }
            }
            assert_eq!(
                http_route(&server, http_address, "reused.test", "/"),
                Some(Some("stale_b".to_owned())),
                "premise: cluster B owns the route"
            );

            let responses = send(
                &mut server,
                "stale-removal",
                RequestType::RemoveHttpFrontend(route("stale_a")),
            );
            assert_eq!(
                responses.iter().map(|r| r.status).collect::<Vec<_>>(),
                vec![expected_status as i32],
                "A gone: {a_is_gone}, answers: {responses:?}"
            );
            assert_eq!(
                http_route(&server, http_address, "reused.test", "/"),
                Some(Some("stale_b".to_owned())),
                "cluster B's route must survive a stale removal naming cluster A (A gone: {a_is_gone})"
            );
        }
    }

    /// TCP and UDP frontends are removed from their listener by address,
    /// whatever cluster owns it. Cluster A frees address X, B takes X, and A
    /// comes back with a frontend on Y. A stale removal of A's frontend on X
    /// then finds A's frontend list but no match in it (`NoChange`, not
    /// `NotFound`), and must still leave B's listener on X alone.
    #[test]
    fn a_stale_tcp_or_udp_frontend_removal_never_unmaps_another_clusters_listener() {
        let cluster = |cluster_id: &str| {
            RequestType::AddCluster(Cluster {
                cluster_id: cluster_id.to_owned(),
                ..Default::default()
            })
        };
        for udp in [false, true] {
            let x = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
            let y = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
            let listener = |address: SocketAddress| {
                if udp {
                    RequestType::AddUdpListener(
                        ListenerBuilder::new_udp(address)
                            .to_udp(None)
                            .expect("could not build the test UDP listener config"),
                    )
                } else {
                    RequestType::AddTcpListener(
                        ListenerBuilder::new_tcp(address)
                            .to_tcp(None)
                            .expect("could not build the test TCP listener config"),
                    )
                }
            };
            let front = |cluster_id: &str, address: SocketAddress| {
                if udp {
                    (
                        RequestType::AddUdpFrontend(RequestUdpFrontend {
                            cluster_id: cluster_id.to_owned(),
                            address,
                            ..Default::default()
                        }),
                        RequestType::RemoveUdpFrontend(RequestUdpFrontend {
                            cluster_id: cluster_id.to_owned(),
                            address,
                            ..Default::default()
                        }),
                    )
                } else {
                    (
                        RequestType::AddTcpFrontend(RequestTcpFrontend {
                            cluster_id: cluster_id.to_owned(),
                            address,
                            ..Default::default()
                        }),
                        RequestType::RemoveTcpFrontend(RequestTcpFrontend {
                            cluster_id: cluster_id.to_owned(),
                            address,
                            ..Default::default()
                        }),
                    )
                }
            };
            let owner = |server: &Server| {
                let key = SocketAddr::from(x).into();
                if udp {
                    server.udp.borrow().listener_cluster(&key)
                } else {
                    server.tcp.borrow().listener_cluster(&key)
                }
                .map(|id| id.to_string())
            };

            let mut server = bare_server();
            for (index, request) in [
                listener(x),
                listener(y),
                cluster("stale_a"),
                front("stale_a", x).0,
                RequestType::RemoveCluster("stale_a".to_owned()),
                cluster("stale_b"),
                front("stale_b", x).0,
                cluster("stale_a"),
                front("stale_a", y).0,
            ]
            .into_iter()
            .enumerate()
            {
                for response in send(&mut server, &format!("setup-{index}"), request) {
                    assert_eq!(
                        response.status,
                        ResponseStatus::Ok as i32,
                        "setup must succeed (udp: {udp}): {response:?}"
                    );
                }
            }
            assert_eq!(
                owner(&server).as_deref(),
                Some("stale_b"),
                "premise: cluster B owns the listener on X (udp: {udp})"
            );

            let responses = send(&mut server, "stale-removal", front("stale_a", x).1);
            assert_eq!(
                responses.iter().map(|r| r.status).collect::<Vec<_>>(),
                vec![ResponseStatus::Failure as i32],
                "the stale removal must be refused (udp: {udp}): {responses:?}"
            );
            assert_eq!(
                owner(&server).as_deref(),
                Some("stale_b"),
                "cluster B must keep the listener on X after a stale removal naming A (udp: {udp})"
            );
        }
    }
}

#[cfg(test)]
mod scm_listener_handoff_tests {
    use std::{
        net::{TcpStream as StdTcpStream, UdpSocket as StdUdpSocket},
        os::fd::{IntoRawFd, RawFd},
    };

    use sozu_command::{config::ListenerBuilder, proto::command::SocketAddress};

    use super::*;
    use crate::{
        ProxyError,
        socket::{server_bind, udp_bind},
        testing::{ServerParts, prebuild_server, provide_port},
    };

    /// A `ConfigState` holding one ACTIVE listener per protocol, i.e. what the
    /// main process carries into `produce_initial_state` for a worker upgrade.
    fn state_with_four_active_listeners(
        http_address: SocketAddress,
        https_address: SocketAddress,
        tcp_address: SocketAddress,
        udp_address: SocketAddress,
    ) -> ConfigState {
        let mut state = ConfigState::new();
        let adds: Vec<(Request, SocketAddress, ListenerType)> = vec![
            (
                RequestType::AddHttpListener(
                    ListenerBuilder::new_http(http_address)
                        .to_http(None)
                        .expect("could not build the test HTTP listener config"),
                )
                .into(),
                http_address,
                ListenerType::Http,
            ),
            (
                RequestType::AddHttpsListener(
                    ListenerBuilder::new_https(https_address)
                        .to_tls(None)
                        .expect("could not build the test HTTPS listener config"),
                )
                .into(),
                https_address,
                ListenerType::Https,
            ),
            (
                RequestType::AddTcpListener(
                    ListenerBuilder::new_tcp(tcp_address)
                        .to_tcp(None)
                        .expect("could not build the test TCP listener config"),
                )
                .into(),
                tcp_address,
                ListenerType::Tcp,
            ),
            (
                RequestType::AddUdpListener(
                    ListenerBuilder::new_udp(udp_address)
                        .to_udp(None)
                        .expect("could not build the test UDP listener config"),
                )
                .into(),
                udp_address,
                ListenerType::Udp,
            ),
        ];
        for (add, address, proxy) in adds {
            state.dispatch(&add).expect("could not add the listener");
            state
                .dispatch(
                    &RequestType::ActivateListener(ActivateListener {
                        interface: None,
                        address,
                        proxy: proxy.into(),
                        from_scm: false,
                    })
                    .into(),
                )
                .expect("could not mark the listener active");
        }
        state
    }

    /// The SCM table for `label`, so a test can assert on retention per protocol.
    fn inherited_table<'a>(server: &'a Server, label: &str) -> &'a Vec<(ListenerKey, RawFd)> {
        let scm_listeners = server
            .scm_listeners
            .as_ref()
            .expect("scm_listeners must be present");
        match label {
            "HTTP" => &scm_listeners.http,
            "HTTPS" => &scm_listeners.tls,
            "TCP" => &scm_listeners.tcp,
            "UDP" => &scm_listeners.udp,
            other => panic!("unknown protocol label {other}"),
        }
    }

    /// Whether `fd` is still the listening socket bound to `port`.
    ///
    /// Deliberately identity-aware rather than a bare "is it open" check: the
    /// test binary runs its tests in parallel, so a descriptor closed here can
    /// be handed straight back out to an unrelated file opened by another test
    /// and a liveness-only probe would call that "still open". Every port in
    /// these tests comes from `provide_port()` and is therefore unique to one
    /// test, so `getsockname` answering with that exact port is proof the
    /// descriptor is still the same socket.
    fn descriptor_is_socket_on_port(fd: RawFd, port: u16) -> bool {
        let mut storage: libc::sockaddr_storage = unsafe { std::mem::zeroed() };
        let mut length = std::mem::size_of::<libc::sockaddr_storage>() as libc::socklen_t;
        // SAFETY: `getsockname` writes at most `length` bytes into `storage`
        // and updates `length`; both outlive the call. A closed, reused or
        // non-socket descriptor answers -1, which is the answer under test.
        let result = unsafe {
            libc::getsockname(
                fd,
                (&raw mut storage).cast::<libc::sockaddr>(),
                &raw mut length,
            )
        };
        if result != 0 || storage.ss_family != libc::AF_INET as libc::sa_family_t {
            return false;
        }
        // SAFETY: the family check above proves the kernel filled a
        // `sockaddr_in`, which `sockaddr_storage` is aligned and sized for.
        let inet = unsafe { &*(&raw const storage).cast::<libc::sockaddr_in>() };
        u16::from_be(inet.sin_port) == port
    }

    /// The datagram queued on the inherited UDP socket before the hand-off.
    const QUEUED_DATAGRAM: &[u8] = b"queued-before-the-handoff";

    /// A worker that inherits its predecessor's listening sockets must adopt
    /// them, not bind its own and leave the inherited descriptors unused.
    ///
    /// `fork_main_into_worker` (`bin/src/worker.rs`) sends the retiring
    /// worker's descriptors over the SCM socket immediately after the fork,
    /// before the main process sends anything else, and
    /// `ConfigState::produce_initial_state` carries one `ActivateListener` per
    /// active listener. `Server::new` must therefore have received those
    /// descriptors BEFORE it applies the initial state: otherwise every
    /// `ActivateListener` takes the `server_bind` / `udp_bind` branch — which
    /// succeeds only because of `SO_REUSEPORT` — and the inherited descriptor
    /// is dropped unused, discarding the accept backlog (TCP/HTTP/HTTPS) and
    /// the receive buffer (UDP) the hand-off exists to preserve.
    ///
    /// The three TCP-family assertions read `accept_queue`, which
    /// `notify_activate_listener` fills through `Server::accept` right after a
    /// successful activation; the UDP assertion reads the listener's own
    /// socket, since UDP has no accept path.
    ///
    /// To SEE THIS RED: in `Server::new`, move the `receive_listeners()` block
    /// back below the `if let Some(state) = initial_state` block. All four
    /// assertions fail — each listener binds its own socket, and the queued
    /// connections and datagram stay on the abandoned descriptors.
    ///
    /// The `TCP_NODELAY` assertions guard the backlog drain of
    /// sozu-proxy/sozu#1586. To SEE THEM RED: delete the
    /// `self.nodelay_backlog = true;` line from `HttpListener::activate`. The
    /// listener still gets the flag, but the connection queued before it did
    /// is accepted without it.
    #[test]
    fn inherited_listener_sockets_are_adopted_by_the_initial_activation() {
        let ServerParts {
            event_loop,
            sessions,
            pool,
            backends,
            client_scm_socket,
            server_scm_socket,
            server_config,
            ..
        } = prebuild_server(16, 16384, false).expect("could not prebuild a test server");
        let (_command_channel, proxy_channel) =
            Channel::generate(1000, 10000).expect("could not generate a test channel");

        let http_address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let https_address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let tcp_address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let udp_address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());

        // The retiring worker's sockets, bound exactly the way a live worker
        // binds them.
        let http_listener = server_bind(http_address.into(), None)
            .expect("could not bind the inherited HTTP socket");
        let https_listener = server_bind(https_address.into(), None)
            .expect("could not bind the inherited HTTPS socket");
        let tcp_listener =
            server_bind(tcp_address.into(), None).expect("could not bind the inherited TCP socket");
        let udp_socket =
            udp_bind(udp_address.into(), None).expect("could not bind the inherited UDP socket");

        // Queue on each inherited socket exactly what the hand-off exists to
        // preserve: a completed connection in the accept backlog, and a
        // datagram in the receive buffer. The clients stay alive for the whole
        // test so the connections are not closed from under the backlog.
        let http_client = StdTcpStream::connect::<SocketAddr>(http_address.into())
            .expect("could not queue a connection on the inherited HTTP socket");
        let https_client = StdTcpStream::connect::<SocketAddr>(https_address.into())
            .expect("could not queue a connection on the inherited HTTPS socket");
        let tcp_client = StdTcpStream::connect::<SocketAddr>(tcp_address.into())
            .expect("could not queue a connection on the inherited TCP socket");
        let datagram_sender =
            StdUdpSocket::bind("127.0.0.1:0").expect("could not bind the test datagram sender");
        datagram_sender
            .send_to(QUEUED_DATAGRAM, SocketAddr::from(udp_address))
            .expect("could not queue a datagram on the inherited UDP socket");

        // The main process's side of the hand-off: send the descriptors, then
        // close its own copies, exactly as `fork_main_into_worker` does.
        let listeners = Listeners {
            http: vec![(http_address.into(), http_listener.into_raw_fd())],
            tls: vec![(https_address.into(), https_listener.into_raw_fd())],
            tcp: vec![(tcp_address.into(), tcp_listener.into_raw_fd())],
            udp: vec![(udp_address.into(), udp_socket.into_raw_fd())],
        };
        client_scm_socket
            .send_listeners(&listeners)
            .expect("could not send the inherited listeners");
        listeners.close();

        // The state the main process writes for the new worker. Built through
        // `ConfigState` and `produce_initial_state` rather than hand-rolled, so
        // the emission order under test is the real generator's: this breaks if
        // `generate_requests` ever stops pairing each `AddXxxListener` with the
        // `ActivateListener` that follows it.
        let initial_state =
            state_with_four_active_listeners(http_address, https_address, tcp_address, udp_address)
                .produce_initial_state();
        assert_eq!(
            initial_state.requests.len(),
            8,
            "the generator must emit one Add and one Activate per active listener"
        );

        let server = Server::new(
            event_loop,
            proxy_channel,
            server_scm_socket,
            sessions,
            pool,
            backends,
            None,
            None,
            None,
            server_config,
            Some(initial_state),
            false,
        )
        .expect("could not build the test server");

        let accepted: Vec<Protocol> = server
            .accept_queue
            .iter()
            .map(|(_, _, protocol, _, _)| *protocol)
            .collect();
        assert!(
            accepted.contains(&Protocol::HTTPListen),
            "the connection queued on the inherited HTTP socket must survive the hand-off, got {accepted:?}"
        );
        assert!(
            accepted.contains(&Protocol::HTTPSListen),
            "the connection queued on the inherited HTTPS socket must survive the hand-off, got {accepted:?}"
        );
        assert!(
            accepted.contains(&Protocol::TCPListen),
            "the connection queued on the inherited TCP socket must survive the hand-off, got {accepted:?}"
        );

        // What those queued connections are handed to the sessions with. They
        // completed their handshake BEFORE the new worker set `TCP_NODELAY` on
        // the inherited socket, and a socket inherits the flag when its
        // handshake completes, not when it is accepted, so the listener alone
        // cannot have given it to them: the first accept drain after an
        // activation must. And the peer address queued with each one is the
        // one `accept(2)` returned, which is the client's own address.
        for (protocol, client) in [
            (Protocol::HTTPListen, &http_client),
            (Protocol::HTTPSListen, &https_client),
            (Protocol::TCPListen, &tcp_client),
        ] {
            let (sock, _, _, _, peer) = server
                .accept_queue
                .iter()
                .find(|(_, _, queued, _, _)| *queued == protocol)
                .expect("the queued connection was asserted above");
            assert_eq!(
                socket2::SockRef::from(sock).tcp_nodelay().ok(),
                Some(true),
                "a {protocol:?} connection queued before the hand-off must carry TCP_NODELAY"
            );
            assert_eq!(
                *peer,
                client.local_addr().expect("the test client has an address"),
                "the {protocol:?} peer address must be the client's"
            );
        }

        let (_, adopted_udp_socket) = server
            .udp
            .borrow_mut()
            .give_back_listeners()
            .pop()
            .expect("the UDP listener must hold a socket after activation");
        let mut buffer = [0u8; 64];
        let received = adopted_udp_socket.recv_from(&mut buffer);
        match received {
            Ok((length, _)) => assert_eq!(
                &buffer[..length],
                QUEUED_DATAGRAM,
                "the datagram queued on the inherited UDP socket must survive the hand-off"
            ),
            Err(error) => panic!(
                "the datagram queued on the inherited UDP socket must survive the hand-off, got {error:?}"
            ),
        }
    }

    /// An `ActivateListener` for a listener that is ALREADY active must dispose
    /// of the inherited descriptor deliberately, not leave it in the table.
    ///
    /// `activate()` short-circuits on `if self.active`, so nothing adopts that
    /// descriptor while the listener holds the address — a deactivate plus
    /// reactivate would, which is why `InheritedSocketFate::Refused` documents
    /// the close as a trade rather than an impossibility. Leaving it in the SCM
    /// table is not neutral either: it is the retiring worker's listening
    /// socket, still bound with `SO_REUSEPORT` and registered with no event
    /// loop, which is expected to keep taking a share of new connections that
    /// nothing accepts. Nothing here measures that expectation — the assertion
    /// below checks only that the descriptor is gone from the table and closed.
    ///
    /// Receiving the SCM listeners before the initial state means the ordinary
    /// upgrade no longer reaches this state, so the test constructs it directly.
    ///
    /// To SEE THIS RED: make any arm's `InheritedSocketFate::Refused` branch
    /// return `None` without discarding — that protocol's descriptor stays in the
    /// table and open, and both assertions below fail. This guards the
    /// retain-everything regression; the ORIGINAL defect closed the descriptor
    /// here too, by dropping the wrapper, and is guarded by
    /// `inherited_listener_sockets_are_adopted_by_the_initial_activation`
    /// instead.
    #[test]
    fn a_repeated_activation_closes_the_descriptor_it_cannot_adopt_now() {
        let ServerParts {
            event_loop,
            sessions,
            pool,
            backends,
            server_scm_socket,
            server_config,
            ..
        } = prebuild_server(16, 16384, true).expect("could not prebuild a test server");
        let (_command_channel, proxy_channel) =
            Channel::generate(1000, 10000).expect("could not generate a test channel");

        let http_address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let https_address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let tcp_address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let udp_address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());

        // A worker whose four listeners are up on their OWN sockets: the initial
        // state activates them and `scm_listeners` is empty, every ordinary start.
        let initial_state =
            state_with_four_active_listeners(http_address, https_address, tcp_address, udp_address)
                .produce_initial_state();
        let mut server = Server::new(
            event_loop,
            proxy_channel,
            server_scm_socket,
            sessions,
            pool,
            backends,
            None,
            None,
            None,
            server_config,
            Some(initial_state),
            false,
        )
        .expect("could not build the test server");

        // Now hand that already-active worker a descriptor per address.
        let http_fd = server_bind(http_address.into(), None)
            .expect("could not bind the spare HTTP socket")
            .into_raw_fd();
        let https_fd = server_bind(https_address.into(), None)
            .expect("could not bind the spare HTTPS socket")
            .into_raw_fd();
        let tcp_fd = server_bind(tcp_address.into(), None)
            .expect("could not bind the spare TCP socket")
            .into_raw_fd();
        let udp_fd = udp_bind(udp_address.into(), None)
            .expect("could not bind the spare UDP socket")
            .into_raw_fd();
        {
            let scm_listeners = server
                .scm_listeners
                .as_mut()
                .expect("Server::new must have populated scm_listeners");
            scm_listeners.http.push((http_address.into(), http_fd));
            scm_listeners.tls.push((https_address.into(), https_fd));
            scm_listeners.tcp.push((tcp_address.into(), tcp_fd));
            scm_listeners.udp.push((udp_address.into(), udp_fd));
        }

        for (address, proxy) in [
            (http_address, ListenerType::Http),
            (https_address, ListenerType::Https),
            (tcp_address, ListenerType::Tcp),
            (udp_address, ListenerType::Udp),
        ] {
            let response = server.notify_activate_listener(
                "test-repeat",
                &ActivateListener {
                    interface: None,
                    address,
                    proxy: proxy.into(),
                    from_scm: false,
                },
            );
            assert_eq!(
                response.status,
                ResponseStatus::Ok as i32,
                "a repeated {proxy:?} activation must still answer ok: {response:?}"
            );
        }

        let scm_listeners = server
            .scm_listeners
            .as_ref()
            .expect("scm_listeners must still be present");
        for (label, table, fd, port) in [
            (
                "HTTP",
                &scm_listeners.http,
                http_fd,
                http_address.port as u16,
            ),
            (
                "HTTPS",
                &scm_listeners.tls,
                https_fd,
                https_address.port as u16,
            ),
            ("TCP", &scm_listeners.tcp, tcp_fd, tcp_address.port as u16),
            ("UDP", &scm_listeners.udp, udp_fd, udp_address.port as u16),
        ] {
            assert!(
                !table.iter().any(|(_, entry)| *entry == fd),
                "the inherited {label} descriptor this activation cannot adopt must leave the SCM table"
            );
            assert!(
                !descriptor_is_socket_on_port(fd, port),
                "the inherited {label} descriptor this activation cannot adopt must be closed, \
                 not left bound in the address's SO_REUSEPORT group, where it is expected — \
                 inferred, not measured here — to go on taking a share of new connections"
            );
        }
    }

    /// An inherited descriptor whose address has NO listener yet must stay in
    /// the SCM table, and a later `AddListener` + `ActivateListener` must adopt
    /// it — backlog and all.
    ///
    /// This is a real upgrade shape, not a hypothetical: `Worker::upgrade`
    /// (`e2e/src/sozu/worker.rs`) hands the new worker a state whose listeners
    /// are inactive and sends `generate_activate_requests()` afterwards, and
    /// `UpgradeWorkerTask` (`bin/src/command/upgrade.rs`) scatters those
    /// requests only once `Server::new` has returned. Closing leftovers at the
    /// end of the initial state would destroy exactly this case, which is why
    /// `InheritedSocketFate::Unclaimed` keeps them.
    ///
    /// To SEE THIS RED: make any arm's `InheritedSocketFate::Unclaimed` branch
    /// take the descriptor out of the table like the other two — the first
    /// activation consumes and closes it, so the listener added afterwards binds
    /// its own socket and that protocol's assertions fail.
    #[test]
    fn an_inherited_descriptor_with_no_listener_yet_is_adopted_by_a_later_activation() {
        let ServerParts {
            event_loop,
            sessions,
            pool,
            backends,
            server_scm_socket,
            server_config,
            ..
        } = prebuild_server(16, 16384, true).expect("could not prebuild a test server");
        let (_command_channel, proxy_channel) =
            Channel::generate(1000, 10000).expect("could not generate a test channel");

        let http_address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let https_address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let tcp_address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let udp_address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());

        let mut server = Server::new(
            event_loop,
            proxy_channel,
            server_scm_socket,
            sessions,
            pool,
            backends,
            None,
            None,
            None,
            server_config,
            None,
            false,
        )
        .expect("could not build the test server");

        // Four inherited descriptors and no listener anywhere. The TCP-family
        // ones each carry a completed connection so the adoption can be shown to
        // preserve the backlog, not merely the descriptor.
        let http_listener = server_bind(http_address.into(), None)
            .expect("could not bind the inherited HTTP socket");
        let https_listener = server_bind(https_address.into(), None)
            .expect("could not bind the inherited HTTPS socket");
        let tcp_listener =
            server_bind(tcp_address.into(), None).expect("could not bind the inherited TCP socket");
        let udp_socket =
            udp_bind(udp_address.into(), None).expect("could not bind the inherited UDP socket");
        let _http_client = StdTcpStream::connect::<SocketAddr>(http_address.into())
            .expect("could not queue a connection on the inherited HTTP socket");
        let _https_client = StdTcpStream::connect::<SocketAddr>(https_address.into())
            .expect("could not queue a connection on the inherited HTTPS socket");
        let _tcp_client = StdTcpStream::connect::<SocketAddr>(tcp_address.into())
            .expect("could not queue a connection on the inherited TCP socket");
        let (http_fd, https_fd, tcp_fd, udp_fd) = (
            http_listener.into_raw_fd(),
            https_listener.into_raw_fd(),
            tcp_listener.into_raw_fd(),
            udp_socket.into_raw_fd(),
        );
        {
            let scm_listeners = server
                .scm_listeners
                .as_mut()
                .expect("Server::new must have populated scm_listeners");
            scm_listeners.http.push((http_address.into(), http_fd));
            scm_listeners.tls.push((https_address.into(), https_fd));
            scm_listeners.tcp.push((tcp_address.into(), tcp_fd));
            scm_listeners.udp.push((udp_address.into(), udp_fd));
        }

        let add_requests =
            state_with_four_active_listeners(http_address, https_address, tcp_address, udp_address);
        let cases = [
            ("HTTP", http_address, ListenerType::Http, http_fd),
            ("HTTPS", https_address, ListenerType::Https, https_fd),
            ("TCP", tcp_address, ListenerType::Tcp, tcp_fd),
            ("UDP", udp_address, ListenerType::Udp, udp_fd),
        ];

        // No listener at any of the addresses: every activation fails, and every
        // descriptor must survive it untouched.
        for (label, address, proxy, fd) in cases {
            let response = server.notify_activate_listener(
                "test-early",
                &ActivateListener {
                    interface: None,
                    address,
                    proxy: proxy.into(),
                    from_scm: false,
                },
            );
            assert_ne!(
                response.status,
                ResponseStatus::Ok as i32,
                "activating {label} with no listener must fail: {response:?}"
            );
            assert!(
                inherited_table(&server, label)
                    .iter()
                    .any(|(_, e)| *e == fd),
                "the {label} descriptor must stay in the SCM table while no listener owns \
                 its address"
            );
            assert!(
                descriptor_is_socket_on_port(fd, address.port as u16),
                "the {label} descriptor must stay open while no listener owns its address"
            );
        }

        // The listeners arrive, then their activations: every descriptor is
        // adopted, so nothing is left in the table and nothing was closed unused.
        for request in add_requests.produce_initial_state().requests {
            server.notify_proxys(request);
        }
        for (label, _address, _proxy, fd) in cases {
            assert!(
                !inherited_table(&server, label)
                    .iter()
                    .any(|(_, e)| *e == fd),
                "the retained {label} descriptor must be adopted by the later activation"
            );
        }
        let accepted: Vec<Protocol> = server
            .accept_queue
            .iter()
            .map(|(_, _, protocol, _, _)| *protocol)
            .collect();
        for expected in [
            Protocol::HTTPListen,
            Protocol::HTTPSListen,
            Protocol::TCPListen,
        ] {
            assert!(
                accepted.contains(&expected),
                "the connection queued on the retained {expected:?} descriptor must be served \
                 by the later activation, got {accepted:?}"
            );
        }
    }

    /// A socket parked by a failed registration must not be mistaken for a
    /// live, registered one.
    ///
    /// Every consumer of the listener's socket field reads its `Some` as
    /// "registered and live": `give_back_listener` hands it back as an
    /// activated socket and answers `Ok`, `soft_stop` / `hard_stop` deregister
    /// it and fold a failure into `ProxyError::SoftStop` / `HardStop`, `accept`
    /// accepts on it. Parking there made a listener that had simply never come
    /// up look activated — and, for a registration failure that leaves the
    /// socket unregistered, made `deregister` answer `ENOENT` and report a
    /// failed shutdown. The park therefore lives in its own `parked_listener` /
    /// `parked_socket` field, which nothing but `activate()` reads.
    ///
    /// `give_back_listener` is the assertion that falsifies the mutation:
    /// `ProxyError::UnactivatedListener` keys on the live field being `None`,
    /// so parking into it turns the answer into `Ok`. The stop assertions ride
    /// along to pin the shutdown behaviour, but they cannot redden on this
    /// repro — the EEXIST it uses means the descriptor IS registered (under
    /// another token), so `deregister` succeeds either way.
    ///
    /// To SEE THIS RED: in any `activate()`, replace
    /// `self.parked_listener = Some(listener)` (or `self.parked_socket = Some(socket)`)
    /// with `self.listener = Some(listener)` / `self.socket = Some(socket)` —
    /// that protocol's `give_back_listener` assertion fails.
    #[test]
    fn a_socket_parked_by_a_failed_registration_is_not_mistaken_for_a_live_one() {
        let ServerParts {
            registry,
            sessions,
            pool,
            backends,
            ..
        } = prebuild_server(16, 16384, true).expect("could not prebuild a test server");

        let token_for = |protocol: Protocol| {
            let mut manager = sessions.borrow_mut();
            let entry = manager.slab.vacant_entry();
            let token = Token(entry.key());
            entry.insert(Rc::new(RefCell::new(ListenSession { protocol })));
            token
        };
        // A socket whose descriptor is already registered on the same epoll
        // instance, so `activate()`'s own `register` answers EEXIST and parks.
        let pre_registered = |address: SocketAddress| {
            let mut socket = server_bind(address.into(), None)
                .expect("could not bind the test listening socket");
            registry
                .register(&mut socket, Token(usize::MAX - 2), Interest::READABLE)
                .expect("could not pre-register the test socket");
            socket
        };

        let http_address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let mut http_proxy = http::HttpProxy::new(
            registry.try_clone().expect("could not clone the registry"),
            sessions.clone(),
            pool.clone(),
            backends.clone(),
        );
        http_proxy
            .add_listener(
                ListenerBuilder::new_http(http_address)
                    .to_http(None)
                    .expect("could not build the test HTTP listener config"),
                token_for(Protocol::HTTPListen),
            )
            .expect("could not add the test HTTP listener");
        assert!(
            http_proxy
                .activate_listener(&http_address.into(), Some(pre_registered(http_address)))
                .is_err(),
            "the pre-registered descriptor must make the HTTP registration fail"
        );
        assert!(
            matches!(
                http_proxy.give_back_listener(&http_address.into()),
                Err(ProxyError::UnactivatedListener)
            ),
            "an HTTP listener holding only a parked socket must not answer as activated"
        );
        assert!(
            http_proxy.soft_stop().is_ok(),
            "a parked socket must not fail an HTTP soft stop"
        );

        let https_address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let mut https_proxy = https::HttpsProxy::new(
            registry.try_clone().expect("could not clone the registry"),
            sessions.clone(),
            pool.clone(),
            backends.clone(),
        );
        https_proxy
            .add_listener(
                ListenerBuilder::new_https(https_address)
                    .to_tls(None)
                    .expect("could not build the test HTTPS listener config"),
                token_for(Protocol::HTTPSListen),
            )
            .expect("could not add the test HTTPS listener");
        assert!(
            https_proxy
                .activate_listener(&https_address.into(), Some(pre_registered(https_address)))
                .is_err(),
            "the pre-registered descriptor must make the HTTPS registration fail"
        );
        assert!(
            matches!(
                https_proxy.give_back_listener(&https_address.into()),
                Err(ProxyError::UnactivatedListener)
            ),
            "an HTTPS listener holding only a parked socket must not answer as activated"
        );
        assert!(
            https_proxy.hard_stop().is_ok(),
            "a parked socket must not fail an HTTPS hard stop"
        );

        let tcp_address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let mut tcp_proxy = tcp::TcpProxy::new(
            registry.try_clone().expect("could not clone the registry"),
            sessions.clone(),
            pool.clone(),
            backends.clone(),
        );
        tcp_proxy
            .add_listener(
                ListenerBuilder::new_tcp(tcp_address)
                    .to_tcp(None)
                    .expect("could not build the test TCP listener config"),
                token_for(Protocol::TCPListen),
            )
            .expect("could not add the test TCP listener");
        assert!(
            tcp_proxy
                .activate_listener(&tcp_address.into(), Some(pre_registered(tcp_address)))
                .is_err(),
            "the pre-registered descriptor must make the TCP registration fail"
        );
        assert!(
            matches!(
                tcp_proxy.give_back_listener(&tcp_address.into()),
                Err(ProxyError::UnactivatedListener)
            ),
            "a TCP listener holding only a parked socket must not answer as activated"
        );

        let udp_address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let mut udp_proxy = udp::UdpProxy::new(
            registry.try_clone().expect("could not clone the registry"),
            sessions.clone(),
            pool.clone(),
            backends.clone(),
            16,
            16384,
        );
        udp_proxy
            .add_listener(
                ListenerBuilder::new_udp(udp_address)
                    .to_udp(None)
                    .expect("could not build the test UDP listener config"),
                token_for(Protocol::UDP),
            )
            .expect("could not add the test UDP listener");
        let mut udp_socket =
            udp_bind(udp_address.into(), None).expect("could not bind the test UDP socket");
        registry
            .register(&mut udp_socket, Token(usize::MAX - 2), Interest::READABLE)
            .expect("could not pre-register the test UDP socket");
        assert!(
            udp_proxy
                .activate_listener(&udp_address.into(), Some(udp_socket))
                .is_err(),
            "the pre-registered descriptor must make the UDP registration fail"
        );
        assert!(
            matches!(
                udp_proxy.give_back_listener(&udp_address.into()),
                Err(ProxyError::UnactivatedListener)
            ),
            "a UDP listener holding only a parked socket must not answer as activated"
        );
    }

    /// A failed mio registration must not close the socket it was handed.
    ///
    /// `activate()` moves the socket into a local to register it. Propagating a
    /// registration error straight from that local drops it — and closes a
    /// descriptor `Listeners::get_*` has already removed from the SCM table,
    /// which is the sozu#1342 close in a second place. `Registry::register`
    /// answers `EEXIST` for a descriptor already registered on the same epoll
    /// instance, which is what the test arranges and what a `deactivate` whose
    /// `deregister` only logged leaves behind.
    ///
    /// To SEE THIS RED: in any `activate()`, go back to
    /// `registry.register(..).map_err(..)?` followed by `self.listener = Some(listener)`
    /// — the `?` drops the local and that protocol's assertion fails.
    #[test]
    fn a_failed_registration_does_not_close_the_socket_it_was_handed() {
        let ServerParts {
            registry,
            sessions,
            pool,
            backends,
            ..
        } = prebuild_server(16, 16384, true).expect("could not prebuild a test server");

        let token_for = |protocol: Protocol| {
            let mut manager = sessions.borrow_mut();
            let entry = manager.slab.vacant_entry();
            let token = Token(entry.key());
            entry.insert(Rc::new(RefCell::new(ListenSession { protocol })));
            token
        };

        // Hand each proxy a socket whose descriptor is ALREADY registered on the
        // same epoll instance, so `activate()`'s own `register` answers EEXIST.
        let handed_out = |address: SocketAddress| {
            let mut socket = server_bind(address.into(), None)
                .expect("could not bind the test listening socket");
            registry
                .register(&mut socket, Token(usize::MAX - 1), Interest::READABLE)
                .expect("could not pre-register the test socket");
            let fd = socket.as_raw_fd();
            (socket, fd)
        };

        let http_address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let token = token_for(Protocol::HTTPListen);
        let mut http_proxy = http::HttpProxy::new(
            registry.try_clone().expect("could not clone the registry"),
            sessions.clone(),
            pool.clone(),
            backends.clone(),
        );
        http_proxy
            .add_listener(
                ListenerBuilder::new_http(http_address)
                    .to_http(None)
                    .expect("could not build the test HTTP listener config"),
                token,
            )
            .expect("could not add the test HTTP listener");
        let (socket, http_fd) = handed_out(http_address);
        let result = http_proxy.activate_listener(&http_address.into(), Some(socket));
        assert!(
            result.is_err(),
            "the pre-registered descriptor must make the HTTP registration fail"
        );
        assert!(
            descriptor_is_socket_on_port(http_fd, http_address.port as u16),
            "a failed HTTP registration must not close the socket it was handed"
        );

        let tcp_address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let token = token_for(Protocol::TCPListen);
        let mut tcp_proxy = tcp::TcpProxy::new(
            registry.try_clone().expect("could not clone the registry"),
            sessions.clone(),
            pool.clone(),
            backends.clone(),
        );
        tcp_proxy
            .add_listener(
                ListenerBuilder::new_tcp(tcp_address)
                    .to_tcp(None)
                    .expect("could not build the test TCP listener config"),
                token,
            )
            .expect("could not add the test TCP listener");
        let (socket, tcp_fd) = handed_out(tcp_address);
        let result = tcp_proxy.activate_listener(&tcp_address.into(), Some(socket));
        assert!(
            result.is_err(),
            "the pre-registered descriptor must make the TCP registration fail"
        );
        assert!(
            descriptor_is_socket_on_port(tcp_fd, tcp_address.port as u16),
            "a failed TCP registration must not close the socket it was handed"
        );

        let https_address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let token = token_for(Protocol::HTTPSListen);
        let mut https_proxy = https::HttpsProxy::new(
            registry.try_clone().expect("could not clone the registry"),
            sessions.clone(),
            pool.clone(),
            backends.clone(),
        );
        https_proxy
            .add_listener(
                ListenerBuilder::new_https(https_address)
                    .to_tls(None)
                    .expect("could not build the test HTTPS listener config"),
                token,
            )
            .expect("could not add the test HTTPS listener");
        let (socket, https_fd) = handed_out(https_address);
        let result = https_proxy.activate_listener(&https_address.into(), Some(socket));
        assert!(
            result.is_err(),
            "the pre-registered descriptor must make the HTTPS registration fail"
        );
        assert!(
            descriptor_is_socket_on_port(https_fd, https_address.port as u16),
            "a failed HTTPS registration must not close the socket it was handed"
        );

        let udp_address = SocketAddress::new_v4(127, 0, 0, 1, provide_port());
        let token = token_for(Protocol::UDP);
        let mut udp_proxy = udp::UdpProxy::new(
            registry.try_clone().expect("could not clone the registry"),
            sessions.clone(),
            pool.clone(),
            backends.clone(),
            16,
            16384,
        );
        udp_proxy
            .add_listener(
                ListenerBuilder::new_udp(udp_address)
                    .to_udp(None)
                    .expect("could not build the test UDP listener config"),
                token,
            )
            .expect("could not add the test UDP listener");
        let mut socket =
            udp_bind(udp_address.into(), None).expect("could not bind the test UDP socket");
        registry
            .register(&mut socket, Token(usize::MAX - 1), Interest::READABLE)
            .expect("could not pre-register the test UDP socket");
        let udp_fd = socket.as_raw_fd();
        let result = udp_proxy.activate_listener(&udp_address.into(), Some(socket));
        assert!(
            result.is_err(),
            "the pre-registered descriptor must make the UDP registration fail"
        );
        assert!(
            descriptor_is_socket_on_port(udp_fd, udp_address.port as u16),
            "a failed UDP registration must not close the socket it was handed"
        );
    }
}

#[cfg(test)]
mod tests {
    use super::{ClusterId, SessionManager, mask_ipv4, mask_ipv6, subnet_key};
    use mio::Token;
    use slab::Slab;
    use std::net::{IpAddr, Ipv4Addr, Ipv6Addr};

    /// A cluster override can drop to `Some(0)` AFTER a session was
    /// tracked under a positive limit: `AddCluster` is an upsert and
    /// clears no counters, and only the GLOBAL `SetMaxConnectionsPerSubnet(0)`
    /// drains the bookkeeping. The next track for that still-open session
    /// then takes the disabled-limiter early return with its slot already
    /// recorded, which must be an ordinary no-op rather than a panic.
    #[test]
    fn a_cluster_override_dropping_to_zero_after_a_track_is_a_no_op() {
        let manager = SessionManager::new(Slab::with_capacity(8), 8, 0, 0, 0, 24, 56);
        let token = Token(1);
        let cluster = ClusterId::from("api");
        let ip: IpAddr = "203.0.113.9".parse().expect("test address must parse");

        // Tracked under a positive per-cluster override.
        manager
            .borrow_mut()
            .track_cluster_subnet(token, cluster.clone(), ip, Some(5));
        assert!(
            !manager.borrow().subnet_tracking_is_empty(),
            "premise: the positive override must have recorded a slot"
        );

        // The operator upserts the cluster with the cap disabled. The same
        // session opens another stream: the early return must not trip on
        // the slot it legitimately left behind.
        manager
            .borrow_mut()
            .track_cluster_subnet(token, cluster.clone(), ip, Some(0));

        // And the stale slot is still released normally on close.
        manager.borrow_mut().untrack_all_cluster_ip(token);
        assert!(
            manager.borrow().subnet_tracking_is_empty(),
            "teardown must drain the slot even though the cap was disabled in between"
        );
    }

    /// The two ends of the prefix range are the ones that break: a
    /// full-width prefix would shift by the whole word, and a zero prefix
    /// would shift by it from the other side. Both are handled by an
    /// explicit branch, and this pins that neither reaches the shift.
    #[test]
    fn masking_handles_both_shift_overflowing_prefix_bounds() {
        let v4: Ipv4Addr = "203.0.113.45".parse().expect("test address must parse");
        assert_eq!(
            mask_ipv4(v4, 32),
            v4,
            "a /32 must be the address itself, not a wrapped-shift artefact"
        );
        assert_eq!(
            mask_ipv4(v4, 0),
            Ipv4Addr::UNSPECIFIED,
            "a /0 must mask every bit away"
        );

        let v6: Ipv6Addr = "2001:db8:dead:beef::1"
            .parse()
            .expect("test address must parse");
        assert_eq!(
            mask_ipv6(v6, 128),
            v6,
            "a /128 must be the address itself, not a wrapped-shift artefact"
        );
        assert_eq!(
            mask_ipv6(v6, 0),
            Ipv6Addr::UNSPECIFIED,
            "a /0 must mask every bit away"
        );

        // Out-of-range prefixes cannot reach the shift either. Config load
        // rejects them, but the hot path does not rely on that having run.
        assert_eq!(
            mask_ipv4(v4, 33),
            v4,
            "an over-wide IPv4 prefix must clamp to the address, not panic or wrap"
        );
        assert_eq!(
            mask_ipv6(v6, 129),
            v6,
            "an over-wide IPv6 prefix must clamp to the address, not panic or wrap"
        );
    }

    /// The recommended prefixes from sozu-proxy/sozu#1270 — /24 and /56 —
    /// must group the addresses an operator expects them to.
    #[test]
    fn masking_groups_addresses_at_the_recommended_prefixes() {
        let parse_v4 = |s: &str| -> Ipv4Addr { s.parse().expect("test address must parse") };
        assert_eq!(
            mask_ipv4(parse_v4("203.0.113.1"), 24),
            mask_ipv4(parse_v4("203.0.113.254"), 24),
            "two hosts in one /24 must share a subnet key"
        );
        assert_ne!(
            mask_ipv4(parse_v4("203.0.113.1"), 24),
            mask_ipv4(parse_v4("203.0.114.1"), 24),
            "hosts in different /24s must NOT share a subnet key"
        );

        let parse_v6 = |s: &str| -> Ipv6Addr { s.parse().expect("test address must parse") };
        assert_eq!(
            mask_ipv6(parse_v6("2001:db8:abcd:0000::1"), 56),
            mask_ipv6(parse_v6("2001:db8:abcd:00ff::9999"), 56),
            "two addresses in one /56 must share a subnet key"
        );
        assert_ne!(
            mask_ipv6(parse_v6("2001:db8:abcd:0000::1"), 56),
            mask_ipv6(parse_v6("2001:db8:abcd:0100::1"), 56),
            "addresses in different /56s must NOT share a subnet key"
        );

        // The issue's own scenario: a /64 holder taking a fresh address
        // per connection lands in one bucket at /56.
        assert_eq!(
            mask_ipv6(parse_v6("2001:db8:abcd:1234::1"), 56),
            mask_ipv6(parse_v6("2001:db8:abcd:1234:ffff:ffff:ffff:ffff"), 56),
            "every address a /64 holder can mint must collapse to one /56 key — that \
             collapse IS the mitigation for sozu-proxy/sozu#1270"
        );
    }

    /// A dual-stack `[::]` listener hands IPv4 clients back as
    /// `::ffff:a.b.c.d`. Masking those with an IPv6 prefix would collapse
    /// the entire IPv4 internet into one `::/56` bucket, so the mapped
    /// form is canonicalised to IPv4 and charged to the IPv4 prefix.
    #[test]
    fn an_ipv4_mapped_source_is_charged_to_the_ipv4_prefix() {
        let mapped: IpAddr = "::ffff:203.0.113.7".parse().expect("must parse");
        let native: IpAddr = "203.0.113.7".parse().expect("must parse");
        assert_eq!(
            subnet_key(&mapped, 24, 56),
            subnet_key(&native, 24, 56),
            "a mapped IPv4 source must key identically to the same address arriving natively"
        );
        assert!(
            subnet_key(&mapped, 24, 56).is_ipv4(),
            "the key for a mapped source must be an IPv4 key"
        );

        // Negative space: two mapped addresses from DIFFERENT /24s must
        // not collide. Under the bug this guards against, both would mask
        // to `::/56` and share one counter.
        let other: IpAddr = "::ffff:198.51.100.7".parse().expect("must parse");
        assert_ne!(
            subnet_key(&mapped, 24, 56),
            subnet_key(&other, 24, 56),
            "mapped sources from different IPv4 /24s must NOT share a subnet key — if they \
             do, every IPv4 client on a dual-stack listener shares one bucket"
        );

        // A genuine IPv6 source is still keyed as IPv6.
        let genuine: IpAddr = "2001:db8::1".parse().expect("must parse");
        assert!(
            subnet_key(&genuine, 24, 56).is_ipv6(),
            "a genuine IPv6 source must keep an IPv6 subnet key"
        );
    }
}

#[cfg(test)]
mod initial_protected_bootstrap_tests {
    use std::thread;

    use sozu_command::{
        channel::Channel,
        proto::command::{
            CountRequests, HealthCheckConfig, HealthCheckHeader, HealthCheckHeaderValue,
            InitialState, ProtectedHealthCheckOrder, ProtectedHealthCheckPhase,
            QueryProtectedHealthCheckSupport, SetProtectedHealthCheck, Status,
        },
        secret::{FingerprintKey, ProtectedHeader, SecretValue, Zeroizing},
    };

    use super::*;
    use crate::testing::{ServerParts, prebuild_server};

    const CLUSTER: &str = "bootstrap-cluster";
    const HEADER: &str = "X-Synthetic-Bootstrap";
    const SECRET: &[u8] = b"synthetic-bootstrap-secret";
    const DEADLINE: Duration = Duration::from_secs(20);

    type Phase = ProtectedHealthCheckPhase;

    fn request(id: &str, request_type: RequestType) -> WorkerRequest {
        WorkerRequest {
            id: id.to_owned(),
            content: request_type.into(),
        }
    }

    fn capability(id: &str) -> WorkerRequest {
        request(
            id,
            RequestType::QueryProtectedHealthCheckSupport(QueryProtectedHealthCheckSupport {}),
        )
    }

    fn status(id: &str) -> WorkerRequest {
        request(id, RequestType::Status(Status {}))
    }

    fn foreign(id: &str) -> WorkerRequest {
        request(id, RequestType::CountRequests(CountRequests {}))
    }

    fn payload(cluster: &str) -> SetProtectedHealthCheck {
        let key = FingerprintKey::from_parts(Zeroizing::new([7u8; 32]), [9u8; 16]);
        let mut carrier = HealthCheckHeaderValue {
            name: HEADER.to_owned(),
            value: SECRET.to_vec(),
        };
        let secret = SecretValue::take_from_ingress(&mut carrier).expect("a valid value");
        let header = ProtectedHeader::new(&key, HEADER.to_owned(), secret);
        SetProtectedHealthCheck {
            cluster_id: cluster.to_owned(),
            config: HealthCheckConfig {
                uri: "/".to_owned(),
                interval: 10,
                timeout: 5,
                healthy_threshold: 3,
                unhealthy_threshold: 3,
                headers: vec![HealthCheckHeader {
                    name: HEADER.to_owned(),
                    fingerprint: header.fingerprint().to_string(),
                }],
                ..Default::default()
            },
            values: vec![HealthCheckHeaderValue {
                name: HEADER.to_owned(),
                value: SECRET.to_vec(),
            }],
        }
    }

    fn order(id: &str, phase: Phase, cluster: &str, version: u64) -> WorkerRequest {
        let prepare = (phase == Phase::ProtectedHealthCheckPrepare).then(|| payload(cluster));
        request(
            id,
            RequestType::ProtectedHealthCheckOrder(ProtectedHealthCheckOrder {
                phase: phase as i32,
                cluster_id: cluster.to_owned(),
                version,
                prepare,
            }),
        )
    }

    fn prepare(id: &str, version: u64) -> WorkerRequest {
        order(id, Phase::ProtectedHealthCheckPrepare, CLUSTER, version)
    }

    fn commit(id: &str, version: u64) -> WorkerRequest {
        order(id, Phase::ProtectedHealthCheckCommit, CLUSTER, version)
    }

    fn release(id: &str, version: u64) -> WorkerRequest {
        order(id, Phase::ProtectedHealthCheckRelease, CLUSTER, version)
    }

    fn abort(id: &str, version: u64) -> WorkerRequest {
        order(id, Phase::ProtectedHealthCheckAbort, CLUSTER, version)
    }

    fn assert_reply(response: &WorkerResponse, phase: Phase, cluster: &str, version: u64) {
        assert_eq!(response.status, ResponseStatus::Ok as i32, "{response:?}");
        let Some(ContentType::ProtectedHealthCheckReply(reply)) = response
            .content
            .as_ref()
            .and_then(|content| content.content_type.as_ref())
        else {
            panic!("the answer must be a typed protected reply: {response:?}");
        };
        assert_eq!(reply.phase, phase as i32);
        assert_eq!(reply.cluster_id, cluster);
        assert_eq!(reply.version, version);
    }

    fn assert_refused(response: &WorkerResponse, id: &str) {
        assert_eq!(response.id, id);
        assert_eq!(
            response.status,
            ResponseStatus::Failure as i32,
            "{response:?}"
        );
        assert!(response.content.is_none(), "{response:?}");
    }

    fn assert_no_secret(responses: &[WorkerResponse]) {
        let shown = format!("{responses:?}");
        assert!(!shown.contains(HEADER), "{shown}");
        assert!(!shown.contains("synthetic-bootstrap-secret"), "{shown}");
    }

    /// Run `Server::new` on this thread against a blocking peer that plays
    /// `script`, one request at a time, and returns every answer. The script
    /// ends with a valid `Status`, the only message that may release the
    /// constructor; a constructor that returned earlier leaves the peer
    /// unanswered and the read deadline reports it.
    fn bootstrap(
        script: Vec<WorkerRequest>,
    ) -> (Server, Rc<RefCell<BackendMap>>, Vec<WorkerResponse>) {
        let ServerParts {
            event_loop,
            sessions,
            pool,
            backends,
            server_scm_socket,
            server_config,
            ..
        } = prebuild_server(16, 16384, true).expect("could not prebuild a test server");
        let (mut peer, proxy_channel) =
            Channel::generate(1000, 10000).expect("could not generate a test channel");
        let initial_state = InitialState {
            requests: vec![request(
                "initial",
                RequestType::AddCluster(Cluster {
                    cluster_id: CLUSTER.to_owned(),
                    ..Default::default()
                }),
            )],
        };

        let driver = thread::spawn(move || {
            let mut answers = Vec::with_capacity(script.len());
            for message in script {
                peer.write_message(&message)
                    .map_err(|error| format!("peer write: {error}"))?;
                answers.push(
                    peer.read_message_blocking_timeout(Some(DEADLINE))
                        .map_err(|error| format!("peer read: {error}"))?,
                );
            }
            Ok::<_, String>(answers)
        });

        let held = backends.to_owned();
        let server = Server::new(
            event_loop,
            proxy_channel,
            server_scm_socket,
            sessions,
            pool,
            backends,
            None,
            None,
            None,
            server_config,
            Some(initial_state),
            true,
        )
        .expect("the constructor must return once the Status arrives");
        let answers = driver
            .join()
            .expect("the peer must not panic")
            .expect("every request must be answered");
        (server, held, answers)
    }

    fn assert_holds_the_values(backends: &Rc<RefCell<BackendMap>>) {
        let backends = backends.borrow();
        let config = backends
            .health_check_configs
            .get(CLUSTER)
            .expect("the commit must have installed the metadata");
        assert_eq!(config.headers.len(), 1);
        assert_eq!(config.headers[0].name, HEADER);
        let set = backends
            .protected_health_check_headers
            .get(CLUSTER)
            .expect("the commit must have installed the values");
        assert_eq!(set.len(), 1);
        assert_eq!(set.headers()[0].name(), HEADER);
        assert_eq!(set.headers()[0].value().expose_secret_bytes(), SECRET);
    }

    fn assert_holds_nothing(backends: &Rc<RefCell<BackendMap>>) {
        let backends = backends.borrow();
        assert!(
            !backends
                .protected_health_check_headers
                .contains_key(CLUSTER)
        );
        assert!(!backends.health_check_configs.contains_key(CLUSTER));
    }

    /// The valid cycle that closes every scenario below.
    fn finish(mut script: Vec<WorkerRequest>, version: u64) -> Vec<WorkerRequest> {
        script.extend([
            prepare("p", version),
            commit("c", version),
            release("r", version),
            status("s"),
        ]);
        script
    }

    #[test]
    fn initial_protected_bootstrap_installs_the_values_before_the_constructor_returns() {
        let (_server, backends, answers) = bootstrap(vec![
            capability("cap"),
            prepare("prep", 1),
            commit("com", 1),
            release("rel", 1),
            status("st"),
        ]);
        assert_reply(&answers[0], Phase::ProtectedHealthCheckCapability, "", 0);
        assert_reply(&answers[1], Phase::ProtectedHealthCheckPrepare, CLUSTER, 1);
        assert_reply(&answers[2], Phase::ProtectedHealthCheckCommit, CLUSTER, 1);
        assert_reply(&answers[3], Phase::ProtectedHealthCheckRelease, CLUSTER, 1);
        assert_eq!(answers[4], WorkerResponse::ok("st"));
        assert_eq!(answers.len(), 5);
        assert_holds_the_values(&backends);
        assert_no_secret(&answers);
    }

    #[test]
    fn initial_protected_bootstrap_still_accepts_a_direct_status() {
        let (_server, backends, answers) = bootstrap(vec![status("st")]);
        assert_eq!(answers, vec![WorkerResponse::ok("st")]);
        assert_holds_nothing(&backends);
    }

    #[test]
    fn initial_protected_bootstrap_refuses_every_out_of_order_request() {
        // each script holds exactly one request named `bad`, which must be
        // refused, and ends with a valid cycle and `Status`
        let cases: Vec<(&str, Vec<WorkerRequest>)> = vec![
            (
                "prepare before the capability",
                finish(vec![prepare("bad", 1), capability("cap")], 1),
            ),
            (
                "commit before the capability",
                finish(vec![commit("bad", 1), capability("cap")], 1),
            ),
            (
                "commit before prepare",
                finish(vec![capability("cap"), commit("bad", 1)], 1),
            ),
            (
                "release before commit",
                finish(vec![capability("cap"), release("bad", 1)], 1),
            ),
            (
                "foreign request after the capability",
                finish(vec![capability("cap"), foreign("bad")], 1),
            ),
            (
                "foreign request before the capability",
                finish(vec![foreign("bad"), capability("cap")], 1),
            ),
            (
                "repeated capability",
                finish(vec![capability("cap"), capability("bad")], 1),
            ),
            (
                "abort without a started cycle",
                finish(vec![capability("cap"), abort("bad", 1)], 1),
            ),
            ("commit of the wrong version", {
                let mut script = vec![capability("cap"), prepare("prep", 1)];
                script.push(commit("bad", 2));
                script.extend([commit("com", 1), release("rel", 1), status("st")]);
                script
            }),
            ("commit of the wrong cluster", {
                let mut script = vec![capability("cap"), prepare("prep", 1)];
                script.push(order(
                    "bad",
                    Phase::ProtectedHealthCheckCommit,
                    "other-cluster",
                    1,
                ));
                script.extend([commit("com", 1), release("rel", 1), status("st")]);
                script
            }),
            ("second prepare while one is staged", {
                let mut script = vec![capability("cap"), prepare("prep", 1)];
                script.push(prepare("bad", 2));
                script.extend([commit("com", 1), release("rel", 1), status("st")]);
                script
            }),
            ("repeated commit", {
                let mut script = vec![capability("cap"), prepare("prep", 1), commit("com", 1)];
                script.push(commit("bad", 1));
                script.extend([release("rel", 1), status("st")]);
                script
            }),
            ("release of the wrong version", {
                let mut script = vec![capability("cap"), prepare("prep", 1), commit("com", 1)];
                script.push(release("bad", 2));
                script.extend([release("rel", 1), status("st")]);
                script
            }),
            ("status while a prepare is outstanding", {
                let mut script = vec![capability("cap"), prepare("prep", 1)];
                script.push(status("bad"));
                script.extend([commit("com", 1), release("rel", 1), status("st")]);
                script
            }),
            ("status while a commit is outstanding", {
                let mut script = vec![capability("cap"), prepare("prep", 1), commit("com", 1)];
                script.push(status("bad"));
                script.extend([release("rel", 1), status("st")]);
                script
            }),
            ("stale prepare after a released cycle", {
                let mut script = vec![
                    capability("cap"),
                    prepare("prep", 2),
                    commit("com", 2),
                    release("rel", 2),
                ];
                script.push(prepare("bad", 1));
                script.push(status("st"));
                script
            }),
        ];
        for (name, script) in cases {
            let refused: Vec<usize> = script
                .iter()
                .enumerate()
                .filter(|(_, message)| message.id == "bad")
                .map(|(index, _)| index)
                .collect();
            assert_eq!(refused.len(), 1, "{name}: exactly one refused request");
            let (_server, backends, answers) = bootstrap(script);
            assert_refused(&answers[refused[0]], "bad");
            assert!(
                answers
                    .iter()
                    .enumerate()
                    .all(|(index, answer)| index == refused[0]
                        || answer.status == ResponseStatus::Ok as i32),
                "{name}: only the refused request fails: {answers:?}"
            );
            assert_holds_the_values(&backends);
            assert_no_secret(&answers);
        }
    }

    #[test]
    fn initial_protected_bootstrap_forwards_a_handler_refusal_and_stays_blocked() {
        let unknown = order(
            "bad",
            Phase::ProtectedHealthCheckPrepare,
            "unknown-cluster",
            1,
        );
        let (_server, backends, answers) = bootstrap(finish(vec![capability("cap"), unknown], 1));
        assert_refused(&answers[1], "bad");
        assert_eq!(
            answers[1].message,
            "the cluster is not known to this worker"
        );
        assert_holds_the_values(&backends);
        assert_no_secret(&answers);
    }

    #[test]
    fn initial_protected_bootstrap_abort_drops_the_staged_change() {
        let (_server, backends, answers) = bootstrap(vec![
            capability("cap"),
            prepare("prep", 1),
            abort("abort", 1),
            status("st"),
        ]);
        assert_reply(&answers[2], Phase::ProtectedHealthCheckAbort, CLUSTER, 1);
        assert_eq!(answers[3], WorkerResponse::ok("st"));
        assert_holds_nothing(&backends);
        assert_no_secret(&answers);
    }

    #[test]
    fn initial_protected_bootstrap_abort_restores_an_applied_change() {
        let (_server, backends, answers) = bootstrap(vec![
            capability("cap"),
            prepare("prep", 1),
            commit("com", 1),
            abort("abort", 1),
            status("st"),
        ]);
        assert_reply(&answers[3], Phase::ProtectedHealthCheckAbort, CLUSTER, 1);
        assert_holds_nothing(&backends);
        assert_no_secret(&answers);
    }

    #[test]
    fn initial_protected_bootstrap_refuses_a_mismatched_abort_then_completes() {
        let mut script = vec![capability("cap"), prepare("prep", 1)];
        script.push(abort("bad", 9));
        script.extend([abort("abort", 1), prepare("again", 2)]);
        script.extend([commit("com", 2), release("rel", 2), status("st")]);
        let (_server, backends, answers) = bootstrap(script);
        assert_refused(&answers[2], "bad");
        assert_reply(&answers[3], Phase::ProtectedHealthCheckAbort, CLUSTER, 1);
        assert_reply(&answers[6], Phase::ProtectedHealthCheckRelease, CLUSTER, 2);
        assert_holds_the_values(&backends);
        assert_no_secret(&answers);
    }

    #[test]
    fn initial_protected_bootstrap_returns_an_error_when_the_channel_closes() {
        let ServerParts {
            event_loop,
            sessions,
            pool,
            backends,
            server_scm_socket,
            server_config,
            ..
        } = prebuild_server(16, 16384, true).expect("could not prebuild a test server");
        let (peer, proxy_channel) =
            Channel::generate(1000, 10000).expect("could not generate a test channel");
        drop(peer);
        let result = Server::new(
            event_loop,
            proxy_channel,
            server_scm_socket,
            sessions,
            pool,
            backends,
            None,
            None,
            None,
            server_config,
            None,
            true,
        );
        let Err(error) = result else {
            panic!("a closed channel must not release the constructor");
        };
        assert!(!error.to_string().is_empty());
    }

    #[test]
    fn initial_status_ack_failure_fails_the_constructor() {
        let ServerParts {
            event_loop,
            sessions,
            pool,
            backends,
            server_scm_socket,
            server_config,
            ..
        } = prebuild_server(16, 16384, true).expect("could not prebuild a test server");
        let (mut peer, proxy_channel) =
            Channel::generate(1000, 10000).expect("could not generate a test channel");
        // the Status is queued, then the peer goes away: the worker reads it
        // but can no longer deliver the ACK
        peer.write_message(&status("st"))
            .expect("the peer must queue the Status");
        drop(peer);
        let result = Server::new(
            event_loop,
            proxy_channel,
            server_scm_socket,
            sessions,
            pool,
            backends,
            None,
            None,
            None,
            server_config,
            None,
            true,
        );
        let Err(error) = result else {
            panic!("a lost final ACK must not release the constructor");
        };
        assert!(
            matches!(error, ServerError::WriteInitialAnswer(_)),
            "{error:?}"
        );
    }
}