ME: Autonomous&Lifecycle-bounded Pool Convergence

Co-Authored-By: brekotis <93345790+brekotis@users.noreply.github.com>

src/transport/middle_proxy/pool_writer/overlap.rs

@@ -0,0 +1,87 @@
+use super::*;
+use std::task::{Context, Poll};
+
+/// Writer-local physical lifetime, independent of registry visibility.
+pub(in crate::transport::middle_proxy) struct ReplacementHandoff {
+    state: parking_lot::Mutex<HandoffState>,
+}
+
+enum HandoffState {
+    Open(Option<WriterOpenReservation>),
+    Closed,
+}
+
+impl ReplacementHandoff {
+    /// Creates a handoff that may accept one physical-overlap token.
+    pub(in crate::transport::middle_proxy) fn new() -> Arc<Self> {
+        Arc::new(Self {
+            state: parking_lot::Mutex::new(HandoffState::Open(None)),
+        })
+    }
+
+    fn close(&self) {
+        *self.state.lock() = HandoffState::Closed;
+    }
+
+    /// Keeps attachment reversible until the final registry validation has succeeded.
+    pub(super) fn commit(
+        &self,
+        permit: &mut Option<WriterOpenReservation>,
+        validate: impl FnOnce() -> bool,
+    ) -> bool {
+        let mut state = self.state.lock();
+        match &mut *state {
+            HandoffState::Open(slot) if slot.is_none() && permit.is_some() && validate() => {
+                *slot = permit.take();
+                true
+            }
+            _ => false,
+        }
+    }
+}
+
+/// Destroys transport futures before releasing either opening or retiring capacity.
+pub(super) struct WriterTransport {
+    future: Option<Pin<Box<dyn Future<Output = ()> + Send + 'static>>>,
+    /// Shared with the registry but closed only by transport destruction.
+    pub(super) lifetime: Arc<ReplacementHandoff>,
+    /// Moves to the victim on replacement or is released after normal publication.
+    pub(super) opening: Option<WriterOpenReservation>,
+}
+
+impl WriterTransport {
+    /// Owns prepared transports even if their task is never polled or published.
+    pub(super) fn new(
+        future: Pin<Box<dyn Future<Output = ()> + Send + 'static>>,
+        opening: WriterOpenReservation,
+    ) -> Self {
+        Self {
+            future: Some(future),
+            lifetime: ReplacementHandoff::new(),
+            opening: Some(opening),
+        }
+    }
+}
+
+impl Future for WriterTransport {
+    type Output = ();
+
+    fn poll(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<()> {
+        match self.future.as_mut() {
+            Some(future) => future.as_mut().poll(cx),
+            None => Poll::Ready(()),
+        }
+    }
+}
+
+impl Drop for WriterTransport {
+    fn drop(&mut self) {
+        // Explicit order also applies to unpolled futures, panic unwinding and abort.
+        drop(self.future.take());
+        self.lifetime.close();
+        drop(self.opening.take());
+    }
+}
+
+#[cfg(test)]
+mod tests;

src/transport/middle_proxy/pool_writer/overlap/tests.rs

@@ -0,0 +1,238 @@
+use super::*;
+use crate::transport::middle_proxy::pool_writer_security_tests::make_pool;
+use std::sync::atomic::AtomicUsize;
+
+struct DestructionProbe {
+    counter: Arc<AtomicUsize>,
+    destroyed: Arc<AtomicBool>,
+}
+
+impl Drop for DestructionProbe {
+    fn drop(&mut self) {
+        assert!(
+            self.counter.load(Ordering::Acquire) > 0,
+            "transport must be destroyed before its overlap token"
+        );
+        self.destroyed.store(true, Ordering::Release);
+    }
+}
+
+async fn transport(pool: &Arc<MePool>, destroyed: Arc<AtomicBool>) -> WriterTransport {
+    let permit = pool
+        .reserve_writer_open(
+            WriterContour::Active,
+            WriterOpenIntent::Replacement,
+            2,
+            "127.0.0.1:443".parse().unwrap(),
+        )
+        .await
+        .unwrap();
+    let probe = DestructionProbe {
+        counter: pool.writer_replacement_open_reserved.clone(),
+        destroyed,
+    };
+    WriterTransport::new(
+        Box::pin(async move {
+            let _probe = probe;
+            std::future::pending::<()>().await;
+        }),
+        permit,
+    )
+}
+
+#[tokio::test]
+async fn unpolled_transport_and_aborted_task_release_after_destruction() {
+    let pool = make_pool().await;
+    for abort in [false, true] {
+        let destroyed = Arc::new(AtomicBool::new(false));
+        let task = transport(&pool, destroyed.clone()).await;
+        if abort {
+            let task = tokio::spawn(task);
+            task.abort();
+            let _ = task.await;
+        } else {
+            drop(task);
+        }
+        assert!(destroyed.load(Ordering::Acquire));
+        assert_eq!(
+            pool.writer_replacement_open_reserved
+                .load(Ordering::Acquire),
+            0
+        );
+    }
+}
+
+#[tokio::test]
+async fn retiring_transport_owns_capacity_after_logical_removal() {
+    let pool = make_pool().await;
+    let destroyed = Arc::new(AtomicBool::new(false));
+    let mut victim = transport(&pool, destroyed.clone()).await;
+    drop(victim.opening.take());
+    let lifetime = victim.lifetime.clone();
+    let mut successor = transport(&pool, Arc::new(AtomicBool::new(false))).await;
+    successor.future = Some(Box::pin(std::future::pending()));
+    assert!(lifetime.commit(&mut successor.opening, || true));
+    assert!(successor.opening.is_none());
+    // Registry removal only drops its handle; the real transport retains the token.
+    drop(lifetime);
+    assert_eq!(
+        pool.writer_replacement_open_reserved
+            .load(Ordering::Acquire),
+        1
+    );
+    drop(victim);
+    assert!(destroyed.load(Ordering::Acquire));
+    assert_eq!(
+        pool.writer_replacement_open_reserved
+            .load(Ordering::Acquire),
+        0
+    );
+    // Successor is no longer overlap after its victim has physically gone.
+    drop(successor);
+}
+
+#[tokio::test]
+async fn failed_or_closed_handoff_keeps_the_prepared_token() {
+    let pool = make_pool().await;
+    let mut prepared = transport(&pool, Arc::new(AtomicBool::new(false))).await;
+    let lifetime = ReplacementHandoff::new();
+    assert!(!lifetime.commit(&mut prepared.opening, || false));
+    assert!(prepared.opening.is_some());
+    lifetime.close();
+    assert!(!lifetime.commit(&mut prepared.opening, || true));
+    assert!(prepared.opening.is_some());
+    drop(prepared);
+    assert_eq!(
+        pool.writer_replacement_open_reserved
+            .load(Ordering::Acquire),
+        0
+    );
+}
+
+#[tokio::test]
+async fn overlap_limit_survives_commit_and_cap_shrink_until_teardown() {
+    let pool = make_pool().await;
+    pool.floor_runtime
+        .me_adaptive_floor_cpu_cores_override
+        .store(1, Ordering::Relaxed);
+    pool.floor_runtime
+        .me_adaptive_floor_max_active_writers_per_core
+        .store(1, Ordering::Relaxed);
+    pool.floor_runtime
+        .me_adaptive_floor_max_active_writers_global
+        .store(1, Ordering::Relaxed);
+    pool.set_adaptive_floor_runtime_caps(pool.floor_authority(), 1, 2, 1, 1, 1, 0, 0);
+    let mut victim = transport(&pool, Arc::new(AtomicBool::new(false))).await;
+    drop(victim.opening.take());
+    let mut successor = transport(&pool, Arc::new(AtomicBool::new(false))).await;
+    successor.future = Some(Box::pin(std::future::pending()));
+    assert!(victim.lifetime.commit(&mut successor.opening, || true));
+    let second = transport(&pool, Arc::new(AtomicBool::new(false))).await;
+    pool.set_adaptive_floor_runtime_caps(pool.floor_authority(), 1, 1, 1, 1, 1, 1, 0);
+    let addr = "127.0.0.1:443".parse().unwrap();
+    assert!(
+        pool.reserve_writer_open(
+            WriterContour::Active,
+            WriterOpenIntent::Replacement,
+            2,
+            addr
+        )
+        .await
+        .is_none()
+    );
+    drop(second);
+    assert!(
+        pool.reserve_writer_open(
+            WriterContour::Active,
+            WriterOpenIntent::Replacement,
+            2,
+            addr
+        )
+        .await
+        .is_none()
+    );
+    drop(victim);
+    assert!(
+        pool.reserve_writer_open(
+            WriterContour::Active,
+            WriterOpenIntent::Replacement,
+            2,
+            addr
+        )
+        .await
+        .is_some()
+    );
+    drop(successor);
+    assert_eq!(
+        pool.writer_replacement_open_reserved
+            .load(Ordering::Acquire),
+        0
+    );
+}
+
+#[tokio::test]
+async fn panicking_transport_releases_capacity_after_unwinding_its_resources() {
+    let pool = make_pool().await;
+    let permit = pool
+        .reserve_writer_open(
+            WriterContour::Active,
+            WriterOpenIntent::Replacement,
+            2,
+            "127.0.0.1:443".parse().unwrap(),
+        )
+        .await
+        .unwrap();
+    let destroyed = Arc::new(AtomicBool::new(false));
+    let probe = DestructionProbe {
+        counter: pool.writer_replacement_open_reserved.clone(),
+        destroyed: destroyed.clone(),
+    };
+    let task = WriterTransport::new(
+        Box::pin(async move {
+            let _probe = probe;
+            panic!("injected transport panic");
+        }),
+        permit,
+    );
+    assert!(tokio::spawn(task).await.unwrap_err().is_panic());
+    assert!(destroyed.load(Ordering::Acquire));
+    assert_eq!(
+        pool.writer_replacement_open_reserved
+            .load(Ordering::Acquire),
+        0
+    );
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+async fn racing_commit_and_close_release_each_token_exactly_once() {
+    let pool = make_pool().await;
+    for _ in 0..128 {
+        let permit = pool
+            .reserve_writer_open(
+                WriterContour::Active,
+                WriterOpenIntent::Replacement,
+                2,
+                "127.0.0.1:443".parse().unwrap(),
+            )
+            .await
+            .unwrap();
+        let lifetime = ReplacementHandoff::new();
+        let closing = lifetime.clone();
+        let close = tokio::spawn(async move {
+            tokio::task::yield_now().await;
+            closing.close();
+        });
+        let commit = tokio::spawn(async move {
+            let mut permit = Some(permit);
+            lifetime.commit(&mut permit, || true);
+            drop(permit);
+        });
+        close.await.unwrap();
+        commit.await.unwrap();
+        assert_eq!(
+            pool.writer_replacement_open_reserved
+                .load(Ordering::Acquire),
+            0
+        );
+    }
+}

src/transport/middle_proxy/pool_writer/publication.rs

@@ -11,9 +11,8 @@ impl MePool {
             tx,
             byte_budget,
             task_registration,
-            writer_task,
+            mut writer_task,
             intent,
-            _open_reservation,
         } = prepared;
         let writer_id = writer.id;
         let mut writers = self.writers.write().await;
@@ -22,9 +21,15 @@ impl MePool {
         let contour = self.authorize_writer_publication(&writer, &coordinator)?;
         self.authorize_writer_publication_capacity(&writer, contour, intent, writers.as_slice())?;
         writer.contour.store(contour.as_u8(), Ordering::Release);
-        registry_registration.install(writer_id, tx, byte_budget);
+        registry_registration.install_with_lifetime(
+            writer_id,
+            tx,
+            byte_budget,
+            writer_task.lifetime.clone(),
+        );
         writers.push(writer);
         self.conn_count.fetch_add(1, Ordering::Relaxed);
+        drop(writer_task.opening.take());
         self.lifecycle
             .spawn_registered_writer(task_registration, writer_task);
         drop(coordinator);

src/transport/middle_proxy/pool_writer/recovery_tests.rs

@@ -36,7 +36,7 @@ fn prepared<'a>(
     pool: &'a Arc<MePool>,
     writer: MeWriter,
     intent: WriterOpenIntent,
-    reservation: WriterOpenReservation<'a>,
+    reservation: WriterOpenReservation,
     task_started: Arc<AtomicBool>,
 ) -> PreparedWriter<'a> {
     let cancel = writer.cancel.clone();
@@ -45,12 +45,14 @@ fn prepared<'a>(
         byte_budget: writer.byte_budget.clone(),
         writer,
         task_registration: pool.lifecycle.try_register().unwrap(),
-        writer_task: Box::pin(async move {
-            task_started.store(true, Ordering::Release);
-            cancel.cancelled().await;
-        }),
+        writer_task: WriterTransport::new(
+            Box::pin(async move {
+                task_started.store(true, Ordering::Release);
+                cancel.cancelled().await;
+            }),
+            reservation,
+        ),
         intent,
-        _open_reservation: reservation,
     }
 }
 

src/transport/middle_proxy/pool_writer/replacement.rs

@@ -44,9 +44,8 @@ impl MePool {
             tx,
             byte_budget,
             task_registration,
-            writer_task,
-            intent,
-            _open_reservation,
+            mut writer_task,
+            intent: _,
         } = prepared;
         let replacement_writer_id = writer.id;
         let victim_writer_id = reservation.writer_id();
@@ -57,7 +56,25 @@ impl MePool {
         let mut registry_registration = self.registry.prepare_writer_registration().await;
         let coordinator = self.reinit.coordinator.lock();
         let contour = self.authorize_writer_publication(&writer, &coordinator)?;
-        self.authorize_writer_publication_capacity(&writer, contour, intent, writers.as_slice())?;
+        // Only this victim-backed transaction may replace Active coverage during suppression.
+        let family = if writer.addr.is_ipv4() {
+            crate::network::IpFamily::V4
+        } else {
+            crate::network::IpFamily::V6
+        };
+        let configured = if writer.addr.is_ipv4() {
+            self.decision.ipv4_me
+        } else {
+            self.decision.ipv6_me
+        };
+        if !configured
+            || (contour != WriterContour::Active
+                && self.is_family_temporarily_suppressed(family, Self::now_epoch_secs()))
+        {
+            return Err(ProxyError::Proxy(
+                "ME replacement family lost publication authority".into(),
+            ));
+        }
         writer.contour.store(contour.as_u8(), Ordering::Release);
         let Some(victim_pos) = writers
             .iter()
@@ -73,6 +90,17 @@ impl MePool {
                 "ME replacement victim changed role before commit".into(),
             ));
         }
+        if !matches!(
+            purpose,
+            WriterReplacementPurpose::FloorRebalance { .. }
+                | WriterReplacementPurpose::CoverageTransfer { .. }
+        ) && (expected_victim_role.dc != writer.writer_dc
+            || expected_victim_role.family != family)
+        {
+            return Err(ProxyError::Proxy(
+                "ME replacement purpose cannot move coverage between groups".into(),
+            ));
+        }
         if let WriterReplacementPurpose::FloorRebalance {
             donor_floor,
             receiver_floor,
@@ -122,19 +150,103 @@ impl MePool {
                         }) == receiver_family
                 })
                 .count();
-            if donor_count <= donor_floor || receiver_count >= receiver_floor {
+            let protected_floor = self.required_writers_for_dc_with_floor_mode(
+                endpoint_snapshot
+                    .endpoints_for_dc_family(expected_victim_role.dc, expected_victim_role.family)
+                    .len(),
+                false,
+            );
+            let current_receiver_floor = self.required_writers_for_dc_with_floor_mode(
+                endpoint_snapshot
+                    .endpoints_for_dc_family(writer.writer_dc, receiver_family)
+                    .len(),
+                false,
+            );
+            if donor_count <= donor_floor.max(protected_floor)
+                || receiver_count >= receiver_floor.min(current_receiver_floor)
+            {
                 return Err(ProxyError::Proxy(
                     "ME floor rebalance became unnecessary before commit".into(),
                 ));
             }
         }
-        if !registry_registration.prepare_replacement_commit(reservation) {
+        if let WriterReplacementPurpose::CoverageTransfer {
+            authority,
+            receiver_floor,
+        } = purpose
+        {
+            if authority
+                != (
+                    coordinator.endpoint_revision,
+                    coordinator.floor_policy_revision,
+                )
+                || contour != WriterContour::Active
+                || expected_victim_role.contour != WriterContour::Active
+            {
+                return Err(ProxyError::Proxy(
+                    "ME coverage transfer lost policy authority".into(),
+                ));
+            }
+            let endpoints = self.endpoint_snapshot.load();
+            let count = |dc, family| {
+                writers
+                    .iter()
+                    .filter(|candidate| {
+                        !candidate.draining.load(Ordering::Acquire)
+                            && candidate.generation == coordinator.active_generation
+                            && candidate.writer_dc == dc
+                            && WriterContour::from_u8(candidate.contour.load(Ordering::Acquire))
+                                == WriterContour::Active
+                            && candidate.addr.is_ipv4()
+                                == matches!(family, crate::network::IpFamily::V4)
+                            && endpoints.contains_dc_endpoint(dc, candidate.addr)
+                    })
+                    .count()
+            };
+            let donor_is_current = victim.generation == coordinator.active_generation
+                && endpoints.contains_dc_endpoint(victim.writer_dc, victim.addr);
+            let donor_floor = self.required_writers_for_dc_with_floor_mode(
+                endpoints
+                    .endpoints_for_dc_family(victim.writer_dc, expected_victim_role.family)
+                    .len(),
+                false,
+            );
+            if (donor_is_current
+                && count(victim.writer_dc, expected_victim_role.family) <= donor_floor)
+                || count(writer.writer_dc, family)
+                    >= receiver_floor.min(
+                        self.required_writers_for_dc_with_floor_mode(
+                            endpoints
+                                .endpoints_for_dc_family(writer.writer_dc, family)
+                                .len(),
+                            false,
+                        ),
+                    )
+            {
+                return Err(ProxyError::Proxy(
+                    "ME coverage transfer no longer preserves floors".into(),
+                ));
+            }
+        }
+        let lifetime = registry_registration
+            .writer_lifetime(victim_writer_id)
+            .ok_or_else(|| {
+                ProxyError::Proxy("ME replacement victim transport disappeared".into())
+            })?;
+        if !lifetime.commit(&mut writer_task.opening, || {
+            registry_registration.prepare_replacement_commit(reservation)
+        }) {
             return Err(ProxyError::Proxy(
                 "ME replacement victim became active before commit".into(),
             ));
         }
 
-        registry_registration.install(replacement_writer_id, tx, byte_budget);
+        registry_registration.install_with_lifetime(
+            replacement_writer_id,
+            tx,
+            byte_budget,
+            writer_task.lifetime.clone(),
+        );
         writers.push(writer);
         self.conn_count.fetch_add(1, Ordering::Relaxed);
         writers.publish_current();
@@ -158,6 +270,7 @@ impl MePool {
 
 #[cfg(test)]
 mod tests {
+    mod convergence;
     use std::collections::HashMap;
     use std::net::{IpAddr, Ipv4Addr};
 
@@ -248,14 +361,54 @@ mod tests {
             tx,
             byte_budget,
             task_registration,
-            writer_task: Box::pin(async move {
-                drop(rx);
-            }),
+            writer_task: WriterTransport::new(
+                Box::pin(async move {
+                    drop(rx);
+                }),
+                open_reservation,
+            ),
             intent: WriterOpenIntent::Replacement,
-            _open_reservation: open_reservation,
         }
     }
 
+    #[tokio::test]
+    async fn suppressed_victim_backed_replacement_keeps_coverage() {
+        let pool = make_pool().await;
+        let addr = endpoint(8);
+        pool.update_proxy_maps(HashMap::from([(2, vec![(addr.ip(), addr.port())])]), None)
+            .await;
+        let victim = install_writer(&pool, 8001, 2, addr).await;
+        pool.set_family_runtime_state(
+            crate::network::IpFamily::V4,
+            crate::transport::middle_proxy::pool::MeFamilyRuntimeState::Suppressed,
+            MePool::now_epoch_secs(),
+            MePool::now_epoch_secs() + 3600,
+            5,
+            0,
+        );
+        let mut reservation = pool
+            .registry
+            .try_reserve_writer_replacement(victim.id)
+            .await
+            .unwrap();
+        let prepared = prepared_writer(&pool, 8002, 2, addr).await;
+        pool.publish_prepared_replacement_writer(
+            prepared,
+            WriterRole::from_writer(&victim),
+            WriterReplacementPurpose::IdleRefresh,
+            &mut reservation,
+        )
+        .await
+        .expect("a reserved active victim must remain replaceable during suppression");
+        assert!(victim.draining.load(Ordering::Acquire));
+        assert_eq!(
+            pool.writer_replacement_open_reserved
+                .load(Ordering::Acquire),
+            1,
+            "overlap is owned until the victim transport is destroyed"
+        );
+    }
+
     #[tokio::test]
     async fn replacement_commit_publishes_successor_before_draining_victim() {
         let pool = make_pool().await;

src/transport/middle_proxy/pool_writer/replacement/tests/convergence.rs

@@ -0,0 +1,198 @@
+use super::*;
+use crate::transport::middle_proxy::registry::ConnMeta;
+
+#[tokio::test]
+async fn busy_obsolete_donor_preserves_clients_and_rechecks_policy() {
+    for stale in [false, true] {
+        let pool = make_pool().await;
+        let obsolete = endpoint(21);
+        let receiver = endpoint(22);
+        pool.update_proxy_maps(
+            HashMap::from([(2, vec![(receiver.ip(), receiver.port())])]),
+            None,
+        )
+        .await;
+        let victim = install_writer(&pool, 9001, 1, obsolete).await;
+        let (client, _messages) = pool.registry.register().await;
+        assert!(
+            pool.registry
+                .bind_writer(
+                    client,
+                    victim.id,
+                    ConnMeta {
+                        target_dc: 1,
+                        client_addr: obsolete,
+                        our_addr: obsolete,
+                        proto_flags: 0,
+                    }
+                )
+                .await
+        );
+        let authority = pool.floor_authority();
+        let mut reservation = pool
+            .registry
+            .try_reserve_writer_replacement_preserving_clients(victim.id)
+            .await
+            .unwrap();
+        let prepared = prepared_writer(&pool, 9002, 2, receiver).await;
+        if stale {
+            pool.reinit.coordinator.lock().floor_policy_revision += 1;
+        }
+        let result = pool
+            .publish_prepared_replacement_writer(
+                prepared,
+                WriterRole::from_writer(&victim),
+                WriterReplacementPurpose::CoverageTransfer {
+                    authority,
+                    receiver_floor: pool.required_writers_for_dc(1),
+                },
+                &mut reservation,
+            )
+            .await;
+        assert_eq!(result.is_ok(), !stale);
+        assert_eq!(victim.draining.load(Ordering::Acquire), !stale);
+        assert_eq!(
+            pool.registry.get_writer(client).await.unwrap().writer_id,
+            victim.id
+        );
+        assert_eq!(
+            pool.writer_replacement_open_reserved
+                .load(Ordering::Acquire),
+            usize::from(!stale)
+        );
+    }
+}
+
+#[tokio::test]
+async fn transfer_commit_matches_independent_donor_floor_model() {
+    // The model counts only current endpoints; obsolete sockets contribute no coverage.
+    for obsolete in [false, true] {
+        for donor_delta in [0usize, 1] {
+            for receiver_full in [false, true] {
+                let pool = make_pool().await;
+                let donor = endpoint(31);
+                let receiver = endpoint(32);
+                pool.update_proxy_maps(
+                    HashMap::from([
+                        (1, vec![(donor.ip(), donor.port())]),
+                        (2, vec![(receiver.ip(), receiver.port())]),
+                    ]),
+                    None,
+                )
+                .await;
+                let floor = pool.required_writers_for_dc(1);
+                let mut victim = None;
+                for id in 0..(floor + donor_delta) {
+                    let writer = install_writer(&pool, 10000 + id as u64, 1, donor).await;
+                    if id == 0 {
+                        victim = Some(writer);
+                    }
+                }
+                let mut victim = victim.unwrap();
+                if obsolete {
+                    let mut writers = pool.writers.write().await;
+                    writers
+                        .iter_mut()
+                        .find(|writer| writer.id == victim.id)
+                        .unwrap()
+                        .addr = endpoint(33);
+                    victim.addr = endpoint(33);
+                }
+                if receiver_full {
+                    for id in 0..floor {
+                        install_writer(&pool, 11000 + id as u64, 2, receiver).await;
+                    }
+                }
+                let mut reservation = pool
+                    .registry
+                    .try_reserve_writer_replacement(victim.id)
+                    .await
+                    .unwrap();
+                let prepared = prepared_writer(&pool, 12000, 2, receiver).await;
+                let result = pool
+                    .publish_prepared_replacement_writer(
+                        prepared,
+                        WriterRole::from_writer(&victim),
+                        WriterReplacementPurpose::CoverageTransfer {
+                            authority: pool.floor_authority(),
+                            receiver_floor: floor,
+                        },
+                        &mut reservation,
+                    )
+                    .await;
+                assert_eq!(
+                    result.is_ok(),
+                    !receiver_full && (obsolete || donor_delta > 0),
+                    "obsolete={obsolete}, surplus={donor_delta}, receiver_full={receiver_full}"
+                );
+            }
+        }
+    }
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+async fn concurrent_transfers_cannot_overfill_receiver_or_steal_donor_floor() {
+    let pool = make_pool().await;
+    let donor = endpoint(41);
+    let receiver = endpoint(42);
+    pool.update_proxy_maps(
+        HashMap::from([
+            (1, vec![(donor.ip(), donor.port())]),
+            (2, vec![(receiver.ip(), receiver.port())]),
+        ]),
+        None,
+    )
+    .await;
+    pool.set_adaptive_floor_runtime_caps(pool.floor_authority(), 128, 128, 128, 128, 6, 0, 0);
+    let floor = pool.required_writers_for_dc(1);
+    let count = floor + 8;
+    let mut victims = Vec::new();
+    for id in 0..count {
+        victims.push(install_writer(&pool, 20000 + id as u64, 1, donor).await);
+    }
+    let mut tasks = tokio::task::JoinSet::new();
+    for (index, victim) in victims.into_iter().enumerate() {
+        let pool = pool.clone();
+        tasks.spawn(async move {
+            let authority = pool.floor_authority();
+            let mut reservation = pool
+                .registry
+                .try_reserve_writer_replacement(victim.id)
+                .await
+                .unwrap();
+            let prepared = prepared_writer(&pool, 21000 + index as u64, 2, receiver).await;
+            tokio::task::yield_now().await;
+            pool.publish_prepared_replacement_writer(
+                prepared,
+                WriterRole::from_writer(&victim),
+                WriterReplacementPurpose::CoverageTransfer {
+                    authority,
+                    receiver_floor: floor,
+                },
+                &mut reservation,
+            )
+            .await
+            .is_ok()
+        });
+    }
+    let mut committed = 0;
+    while let Some(result) = tasks.join_next().await {
+        committed += usize::from(result.unwrap());
+    }
+    assert_eq!(committed, floor);
+    let writers = pool.writers.read().await;
+    let live = |dc| {
+        writers
+            .iter()
+            .filter(|writer| writer.writer_dc == dc && !writer.draining.load(Ordering::Acquire))
+            .count()
+    };
+    assert_eq!(live(2), floor);
+    assert_eq!(live(1), count - floor);
+    assert!(live(1) >= floor);
+    assert_eq!(
+        pool.writer_replacement_open_reserved
+            .load(Ordering::Acquire),
+        committed
+    );
+}

src/transport/middle_proxy/pool_writer/runtime.rs

@@ -291,9 +291,8 @@ impl MePool {
             tx,
             byte_budget,
             task_registration,
-            writer_task,
+            writer_task: WriterTransport::new(writer_task, writer_open_reservation),
             intent,
-            _open_reservation: writer_open_reservation,
         })
     }