diff --git a/tensorpipe/common/deferred_executor.h b/tensorpipe/common/deferred_executor.h
index 95afc4041..653a03720 100644
--- a/tensorpipe/common/deferred_executor.h
+++ b/tensorpipe/common/deferred_executor.h
@@ -160,6 +160,12 @@ class EventLoopDeferredExecutor : public virtual DeferredExecutor {
   // subclasses and called inside the thread owned by this parent class.
   virtual void eventLoop() = 0;
 
+  // This is called after the event loop terminated, still within the thread
+  // that used to run that event loop. It will be called after this class has
+  // transitioned control to the on-demand deferred executor. It thus allows to
+  // clean up any resources without worrying about new work coming in.
+  virtual void cleanUpLoop() {}
+
   // This function is called by the parent class when a function is deferred to
   // it, and must be implemented by subclasses, which are required to have their
   // event loop call runDeferredFunctionsFromEventLoop as soon as possible. This
@@ -230,6 +236,8 @@ class EventLoopDeferredExecutor : public virtual DeferredExecutor {
         fn();
       }
     }
+
+    cleanUpLoop();
   }
 
   std::thread thread_;
diff --git a/tensorpipe/test/transport/connection_test.cc b/tensorpipe/test/transport/connection_test.cc
index af60258cc..a82725a62 100644
--- a/tensorpipe/test/transport/connection_test.cc
+++ b/tensorpipe/test/transport/connection_test.cc
@@ -245,3 +245,39 @@ TEST_P(TransportTest, DISABLED_Connection_EmptyBuffer) {
         peers_->join(PeerGroup::kClient);
       });
 }
+
+TEST_P(TransportTest, Connection_SpamAtClosing) {
+  using namespace std::chrono_literals;
+
+  std::shared_ptr<Context> ctx = GetParam()->getContext();
+  ctx->setId("loopback");
+
+  std::string addr = GetParam()->defaultAddr();
+  std::shared_ptr<Listener> listener = ctx->listen(addr);
+
+  std::atomic<bool> stopSpamming{false};
+  std::function<void()> spam = [&]() {
+    if (stopSpamming) {
+      return;
+    }
+    std::shared_ptr<Connection> conn = ctx->connect(addr);
+    conn->read(
+        [&](const Error& error, const void* /* unused */, size_t /* unused */) {
+          EXPECT_TRUE(error);
+          spam();
+        });
+    conn->close();
+  };
+
+  spam();
+
+  std::this_thread::sleep_for(10ms);
+
+  ctx->close();
+
+  std::this_thread::sleep_for(10ms);
+
+  stopSpamming = true;
+
+  ctx->join();
+}
diff --git a/tensorpipe/transport/connection_boilerplate.h b/tensorpipe/transport/connection_boilerplate.h
index df5d3e9a2..a1044a814 100644
--- a/tensorpipe/transport/connection_boilerplate.h
+++ b/tensorpipe/transport/connection_boilerplate.h
@@ -31,6 +31,8 @@ class ConnectionBoilerplate : public Connection {
       std::string id,
       Args... args);
 
+  explicit ConnectionBoilerplate(std::shared_ptr<TConn> connection);
+
   ConnectionBoilerplate(const ConnectionBoilerplate&) = delete;
   ConnectionBoilerplate(ConnectionBoilerplate&&) = delete;
   ConnectionBoilerplate& operator=(const ConnectionBoilerplate&) = delete;
@@ -79,6 +81,16 @@ ConnectionBoilerplate<TCtx, TList, TConn>::ConnectionBoilerplate(
   impl_->init();
 }
 
+template <typename TCtx, typename TList, typename TConn>
+ConnectionBoilerplate<TCtx, TList, TConn>::ConnectionBoilerplate(
+    std::shared_ptr<TConn> connection)
+    : impl_(std::move(connection)) {
+  static_assert(
+      std::is_base_of<ConnectionImplBoilerplate<TCtx, TList, TConn>, TConn>::
+          value,
+      "");
+}
+
 template <typename TCtx, typename TList, typename TConn>
 void ConnectionBoilerplate<TCtx, TList, TConn>::read(read_callback_fn fn) {
   impl_->read(std::move(fn));
diff --git a/tensorpipe/transport/connection_impl_boilerplate.h b/tensorpipe/transport/connection_impl_boilerplate.h
index b38962f3c..ec21b6075 100644
--- a/tensorpipe/transport/connection_impl_boilerplate.h
+++ b/tensorpipe/transport/connection_impl_boilerplate.h
@@ -126,11 +126,16 @@ class ConnectionImplBoilerplate : public std::enable_shared_from_this<TConn> {
   // Deal with an error.
   void handleError();
 
-  ClosingReceiver closingReceiver_;
-
   // A sequence number for the calls to read and write.
   uint64_t nextBufferBeingRead_{0};
   uint64_t nextBufferBeingWritten_{0};
+
+  // Contexts and listeners do sometimes need to call directly into initFromLoop
+  // and closeForLoop, in order to make sure that some of their operations can
+  // happen "atomically" on the connection, without possibly other operations
+  // occurring in between (e.g., an error).
+  friend ContextImplBoilerplate<TCtx, TList, TConn>;
+  friend ListenerImplBoilerplate<TCtx, TList, TConn>;
 };
 
 template <typename TCtx, typename TList, typename TConn>
@@ -138,9 +143,7 @@ ConnectionImplBoilerplate<TCtx, TList, TConn>::ConnectionImplBoilerplate(
     ConstructorToken /* unused */,
     std::shared_ptr<TCtx> context,
     std::string id)
-    : context_(std::move(context)),
-      id_(std::move(id)),
-      closingReceiver_(context_, context_->getClosingEmitter()) {}
+    : context_(std::move(context)), id_(std::move(id)) {}
 
 template <typename TCtx, typename TList, typename TConn>
 void ConnectionImplBoilerplate<TCtx, TList, TConn>::init() {
@@ -150,7 +153,14 @@ void ConnectionImplBoilerplate<TCtx, TList, TConn>::init() {
 
 template <typename TCtx, typename TList, typename TConn>
 void ConnectionImplBoilerplate<TCtx, TList, TConn>::initFromLoop() {
-  closingReceiver_.activate(*this);
+  if (context_->closed()) {
+    // Set the error without calling setError because we do not want to invoke
+    // the subclass's handleErrorImpl as it would find itself in a weird state
+    // (since initFromLoop wouldn't have been called).
+    error_ = TP_CREATE_ERROR(ConnectionClosedError);
+    TP_VLOG(7) << "Connection " << id_ << " is closing (without initing)";
+    return;
+  }
 
   initImplFromLoop();
 }
diff --git a/tensorpipe/transport/context_impl_boilerplate.h b/tensorpipe/transport/context_impl_boilerplate.h
index 7d42cdcec..16dc5025a 100644
--- a/tensorpipe/transport/context_impl_boilerplate.h
+++ b/tensorpipe/transport/context_impl_boilerplate.h
@@ -9,12 +9,12 @@
 #pragma once
 
 #include <atomic>
+#include <future>
 #include <memory>
 #include <string>
 #include <unordered_map>
 #include <utility>
 
-#include <tensorpipe/common/callback.h>
 #include <tensorpipe/common/defs.h>
 #include <tensorpipe/transport/connection_boilerplate.h>
 #include <tensorpipe/transport/listener_boilerplate.h>
@@ -51,7 +51,9 @@ class ContextImplBoilerplate : public virtual DeferredExecutor,
   void unenroll(TList& listener);
   void unenroll(TConn& connection);
 
-  ClosingEmitter& getClosingEmitter();
+  // Return whether the context is in a closed state. To avoid race conditions,
+  // this must be called from within the loop.
+  bool closed();
 
   void setId(std::string id);
 
@@ -73,7 +75,6 @@ class ContextImplBoilerplate : public virtual DeferredExecutor,
  private:
   std::atomic<bool> closed_{false};
   std::atomic<bool> joined_{false};
-  ClosingEmitter closingEmitter_;
 
   const std::string domainDescriptor_;
 
@@ -162,9 +163,9 @@ void ContextImplBoilerplate<TCtx, TList, TConn>::unenroll(TConn& connection) {
 }
 
 template <typename TCtx, typename TList, typename TConn>
-ClosingEmitter& ContextImplBoilerplate<TCtx, TList, TConn>::
-    getClosingEmitter() {
-  return closingEmitter_;
+bool ContextImplBoilerplate<TCtx, TList, TConn>::closed() {
+  TP_DCHECK(inLoop());
+  return closed_;
 };
 
 template <typename TCtx, typename TList, typename TConn>
@@ -175,14 +176,32 @@ void ContextImplBoilerplate<TCtx, TList, TConn>::setId(std::string id) {
 
 template <typename TCtx, typename TList, typename TConn>
 void ContextImplBoilerplate<TCtx, TList, TConn>::close() {
-  if (!closed_.exchange(true)) {
-    TP_VLOG(7) << "Transport context " << id_ << " is closing";
-
-    closingEmitter_.close();
-    closeImpl();
-
-    TP_VLOG(7) << "Transport context " << id_ << " done closing";
-  }
+  // Defer this to the loop so that it won't race with other code accessing it
+  // (in other words: any code in the loop can assume that this won't change).
+  deferToLoop([this]() {
+    if (!closed_.exchange(true)) {
+      TP_VLOG(7) << "Transport context " << id_ << " is closing";
+
+      // Make a copy as they could unenroll themselves inline.
+      decltype(listeners_) listenersCopy = listeners_;
+      decltype(connections_) connectionsCopy = connections_;
+      // We call closeFromLoop, rather than just close, because we need these
+      // objects to transition _immediately_ to error, "atomically". If we just
+      // deferred closing to later, this could come after some already-enqueued
+      // operations that could try to access the context, which would be closed,
+      // and this could fail.
+      for (auto& iter : listenersCopy) {
+        iter.second->closeFromLoop();
+      }
+      for (auto& iter : connectionsCopy) {
+        iter.second->closeFromLoop();
+      }
+
+      closeImpl();
+
+      TP_VLOG(7) << "Transport context " << id_ << " done closing";
+    }
+  });
 }
 
 template <typename TCtx, typename TList, typename TConn>
@@ -192,6 +211,14 @@ void ContextImplBoilerplate<TCtx, TList, TConn>::join() {
   if (!joined_.exchange(true)) {
     TP_VLOG(7) << "Transport context " << id_ << " is joining";
 
+    // As closing is deferred to the loop, we must wait for closeImpl to be
+    // actually called before we call joinImpl, to avoid race conditions. For
+    // this, we defer another task to the loop, which we know will run after the
+    // closing, and then we wait for that task to be run.
+    std::promise<void> hasClosed;
+    deferToLoop([&]() { hasClosed.set_value(); });
+    hasClosed.get_future().wait();
+
     joinImpl();
 
     TP_VLOG(7) << "Transport context " << id_ << " done joining";
diff --git a/tensorpipe/transport/ibv/listener_impl.cc b/tensorpipe/transport/ibv/listener_impl.cc
index 50aa92a7c..4558d7321 100644
--- a/tensorpipe/transport/ibv/listener_impl.cc
+++ b/tensorpipe/transport/ibv/listener_impl.cc
@@ -146,7 +146,7 @@ void ListenerImpl::handleEventsFromLoop(int events) {
   if (fns_.empty()) {
     context_->unregisterDescriptor(socket_.fd());
   }
-  fn(Error::kSuccess, createConnection(std::move(socket)));
+  fn(Error::kSuccess, createAndInitConnection(std::move(socket)));
 }
 
 } // namespace ibv
diff --git a/tensorpipe/transport/listener_impl_boilerplate.h b/tensorpipe/transport/listener_impl_boilerplate.h
index a169df956..efb13a27b 100644
--- a/tensorpipe/transport/listener_impl_boilerplate.h
+++ b/tensorpipe/transport/listener_impl_boilerplate.h
@@ -78,7 +78,7 @@ class ListenerImplBoilerplate : public std::enable_shared_from_this<TList> {
   Error error_{Error::kSuccess};
 
   template <typename... Args>
-  std::shared_ptr<Connection> createConnection(Args&&... args);
+  std::shared_ptr<Connection> createAndInitConnection(Args&&... args);
 
   // An identifier for the listener, composed of the identifier for the context,
   // combined with an increasing sequence number. It will be used as a prefix
@@ -104,8 +104,6 @@ class ListenerImplBoilerplate : public std::enable_shared_from_this<TList> {
   // Deal with an error.
   void handleError();
 
-  ClosingReceiver closingReceiver_;
-
   // A sequence number for the calls to accept.
   uint64_t nextConnectionBeingAccepted_{0};
 
@@ -113,6 +111,12 @@ class ListenerImplBoilerplate : public std::enable_shared_from_this<TList> {
   // create their identifiers based off this listener's identifier. They will
   // only be used for logging and debugging.
   std::atomic<uint64_t> connectionCounter_{0};
+
+  // Contexts do sometimes need to call directly into closeForLoop, in order to
+  // make sure that some of their operations can happen "atomically" on the
+  // connection, without possibly other operations occurring in between (e.g.,
+  // an error).
+  friend ContextImplBoilerplate<TCtx, TList, TConn>;
 };
 
 template <typename TCtx, typename TList, typename TConn>
@@ -120,9 +124,7 @@ ListenerImplBoilerplate<TCtx, TList, TConn>::ListenerImplBoilerplate(
     ConstructorToken /* unused */,
     std::shared_ptr<TCtx> context,
     std::string id)
-    : context_(std::move(context)),
-      id_(std::move(id)),
-      closingReceiver_(context_, context_->getClosingEmitter()) {}
+    : context_(std::move(context)), id_(std::move(id)) {}
 
 template <typename TCtx, typename TList, typename TConn>
 void ListenerImplBoilerplate<TCtx, TList, TConn>::init() {
@@ -132,7 +134,14 @@ void ListenerImplBoilerplate<TCtx, TList, TConn>::init() {
 
 template <typename TCtx, typename TList, typename TConn>
 void ListenerImplBoilerplate<TCtx, TList, TConn>::initFromLoop() {
-  closingReceiver_.activate(*this);
+  if (context_->closed()) {
+    // Set the error without calling setError because we do not want to invoke
+    // the subclass's handleErrorImpl as it would find itself in a weird state
+    // (since initFromLoop wouldn't have been called).
+    error_ = TP_CREATE_ERROR(ListenerClosedError);
+    TP_VLOG(7) << "Listener " << id_ << " is closing (without initing)";
+    return;
+  }
 
   initImplFromLoop();
 }
@@ -189,15 +198,24 @@ std::string ListenerImplBoilerplate<TCtx, TList, TConn>::addrFromLoop() const {
 template <typename TCtx, typename TList, typename TConn>
 template <typename... Args>
 std::shared_ptr<Connection> ListenerImplBoilerplate<TCtx, TList, TConn>::
-    createConnection(Args&&... args) {
+    createAndInitConnection(Args&&... args) {
+  TP_DCHECK(context_->inLoop());
   std::string connectionId = id_ + ".c" + std::to_string(connectionCounter_++);
   TP_VLOG(7) << "Listener " << id_ << " is opening connection " << connectionId;
-  return std::make_shared<ConnectionBoilerplate<TCtx, TList, TConn>>(
+  auto connection = std::make_shared<TConn>(
       typename ConnectionImplBoilerplate<TCtx, TList, TConn>::
           ConstructorToken(),
       context_,
       std::move(connectionId),
       std::forward<Args>(args)...);
+  // We initialize the connection from the loop immediately, inline, because the
+  // initialization of a connection accepted by a listener typically happens
+  // partly in the listener (e.g., opening and accepting the socket) and partly
+  // in the connection's initFromLoop, and we need these two steps to happen
+  // "atomicically" to make it impossible for an error to occur in between.
+  connection->initFromLoop();
+  return std::make_shared<ConnectionBoilerplate<TCtx, TList, TConn>>(
+      std::move(connection));
 }
 
 template <typename TCtx, typename TList, typename TConn>
diff --git a/tensorpipe/transport/shm/listener_impl.cc b/tensorpipe/transport/shm/listener_impl.cc
index 953d8b217..126d9b316 100644
--- a/tensorpipe/transport/shm/listener_impl.cc
+++ b/tensorpipe/transport/shm/listener_impl.cc
@@ -136,7 +136,7 @@ void ListenerImpl::handleEventsFromLoop(int events) {
   if (fns_.empty()) {
     context_->unregisterDescriptor(socket_.fd());
   }
-  fn(Error::kSuccess, createConnection(std::move(socket)));
+  fn(Error::kSuccess, createAndInitConnection(std::move(socket)));
 }
 
 } // namespace shm
diff --git a/tensorpipe/transport/uv/connection_impl.cc b/tensorpipe/transport/uv/connection_impl.cc
index 21366d181..12a25b161 100644
--- a/tensorpipe/transport/uv/connection_impl.cc
+++ b/tensorpipe/transport/uv/connection_impl.cc
@@ -52,6 +52,8 @@ ConnectionImpl::ConnectionImpl(
 void ConnectionImpl::initImplFromLoop() {
   context_->enroll(*this);
 
+  TP_VLOG(9) << "Connection " << id_ << " is initializing in loop";
+
   if (sockaddr_.has_value()) {
     handle_->initFromLoop();
     handle_->connectFromLoop(sockaddr_.value(), [this](int status) {
diff --git a/tensorpipe/transport/uv/listener_impl.cc b/tensorpipe/transport/uv/listener_impl.cc
index f052f1d33..b9ed1f448 100644
--- a/tensorpipe/transport/uv/listener_impl.cc
+++ b/tensorpipe/transport/uv/listener_impl.cc
@@ -36,6 +36,8 @@ ListenerImpl::ListenerImpl(
 void ListenerImpl::initImplFromLoop() {
   context_->enroll(*this);
 
+  TP_VLOG(9) << "Listener " << id_ << " is initializing in loop";
+
   handle_->initFromLoop();
   auto rv = handle_->bindFromLoop(sockaddr_);
   TP_THROW_UV_IF(rv < 0, rv);
@@ -67,7 +69,8 @@ void ListenerImpl::connectionCallbackFromLoop(int status) {
   auto connection = context_->createHandle();
   connection->initFromLoop();
   handle_->acceptFromLoop(*connection);
-  callback_.trigger(Error::kSuccess, createConnection(std::move(connection)));
+  callback_.trigger(
+      Error::kSuccess, createAndInitConnection(std::move(connection)));
 }
 
 void ListenerImpl::closeCallbackFromLoop() {
diff --git a/tensorpipe/transport/uv/loop.cc b/tensorpipe/transport/uv/loop.cc
index 52152ce99..2c59c0026 100644
--- a/tensorpipe/transport/uv/loop.cc
+++ b/tensorpipe/transport/uv/loop.cc
@@ -58,6 +58,10 @@ void Loop::eventLoop() {
   rv = uv_run(&loop_, UV_RUN_DEFAULT);
   TP_THROW_ASSERT_IF(rv > 0)
       << ": uv_run returned with active handles or requests";
+}
+
+void Loop::cleanUpLoop() {
+  int rv;
 
   uv_ref(reinterpret_cast<uv_handle_t*>(&async_));
   uv_close(reinterpret_cast<uv_handle_t*>(&async_), nullptr);
diff --git a/tensorpipe/transport/uv/loop.h b/tensorpipe/transport/uv/loop.h
index 2383f122d..a6e31f8bc 100644
--- a/tensorpipe/transport/uv/loop.h
+++ b/tensorpipe/transport/uv/loop.h
@@ -32,6 +32,10 @@ class Loop final : public EventLoopDeferredExecutor {
     return &loop_;
   }
 
+  bool closed() {
+    return closed_;
+  }
+
   void close();
 
   void join();
@@ -42,6 +46,9 @@ class Loop final : public EventLoopDeferredExecutor {
   // Event loop thread entry function.
   void eventLoop() override;
 
+  // Clean up after event loop transitioned to on-demand.
+  void cleanUpLoop() override;
+
   // Wake up the event loop.
   void wakeupEventLoopToDeferFunction() override;
 
diff --git a/tensorpipe/transport/uv/uv.h b/tensorpipe/transport/uv/uv.h
index 01771e0a2..41ba616f7 100644
--- a/tensorpipe/transport/uv/uv.h
+++ b/tensorpipe/transport/uv/uv.h
@@ -261,6 +261,7 @@ class TCPHandle : public StreamHandle<TCPHandle, uv_tcp_t> {
 
   void initFromLoop() {
     TP_DCHECK(this->loop_.inLoop());
+    TP_THROW_ASSERT_IF(loop_.closed());
     int rv;
     rv = uv_tcp_init(loop_.ptr(), this->ptr());
     TP_THROW_UV_IF(rv < 0, rv);