Websocket lifecycle fixes

.gitignore

@@ -11,6 +11,9 @@ CMakeUserPresets.json
 **/CMakeCache.txt
 **/Testing/
 
+# Claude Code
+.claude/
+
 # IDE files
 .idea/
 .vs/

cmake/asio-iocp-fix.cmake

@@ -0,0 +1,84 @@
+# Patch for ASIO issue #312: IOCP out-of-resources error reporting.
+# https://github.com/chriskohlhoff/asio/issues/312
+#
+# When PostQueuedCompletionStatus fails due to resource exhaustion, the
+# completion key (overlapped_contains_result) is lost. The operation falls
+# back to an internal queue but do_one() dispatches it with the wrong
+# error code and bytes_transferred, causing undefined behavior / crashes.
+#
+# Fix: store the completion key on the operation object before posting,
+# so it survives the fallback path.
+#
+# Based on MongoDB's patch:
+# https://github.com/mongodb-forks/asio/commit/d03c2e7002131305645374e735a8ece4191f2fc5
+
+function(apply_asio_iocp_fix asio_include_dir)
+    set(OP_HEADER "${asio_include_dir}/asio/detail/win_iocp_operation.hpp")
+    set(CTX_IMPL "${asio_include_dir}/asio/detail/impl/win_iocp_io_context.ipp")
+
+    if(NOT EXISTS "${OP_HEADER}" OR NOT EXISTS "${CTX_IMPL}")
+        message(WARNING "ASIO IOCP fix: headers not found at ${asio_include_dir}, skipping patch")
+        return()
+    endif()
+
+    # Check if already patched (look for our added member)
+    file(READ "${OP_HEADER}" op_content)
+    if(op_content MATCHES "completionKey_")
+        message(STATUS "ASIO IOCP fix: already applied")
+        return()
+    endif()
+
+    message(STATUS "Applying ASIO IOCP fix (issue #312)")
+
+    # --- Patch win_iocp_operation.hpp ---
+    # Add completionKey_ member and accessor
+
+    # Add member after "long ready_;"
+    string(REPLACE
+        "long ready_;"
+        "long ready_;\n  ULONG_PTR completionKey_;"
+        op_content "${op_content}")
+
+    # Initialize in constructor: find "ready_(0)" and add completionKey_(0)
+    string(REPLACE
+        "ready_(0)"
+        "ready_(0), completionKey_(0)"
+        op_content "${op_content}")
+
+    # Add accessor method before the "private:" or "protected:" section
+    # Find "func_type func_;" and add the accessor before the member block
+    string(REPLACE
+        "win_iocp_operation* next_;"
+        "ULONG_PTR& completionKey() { return completionKey_; }\n\n  win_iocp_operation* next_;"
+        op_content "${op_content}")
+
+    file(WRITE "${OP_HEADER}" "${op_content}")
+
+    # --- Patch win_iocp_io_context.ipp ---
+    file(READ "${CTX_IMPL}" ctx_content)
+
+    # Fix post_deferred_completion: pass op->completionKey() instead of 0 for the key
+    # Original: PostQueuedCompletionStatus(iocp_.handle, 0, 0, op)
+    # Fixed:    PostQueuedCompletionStatus(iocp_.handle, 0, op->completionKey(), op)
+    string(REPLACE
+        "::PostQueuedCompletionStatus(iocp_.handle, 0, 0, op)"
+        "::PostQueuedCompletionStatus(iocp_.handle, 0, op->completionKey(), op)"
+        ctx_content "${ctx_content}")
+
+    # Fix on_pending/on_completion: store key on op before posting
+    # Original pattern:
+    #   if (!::PostQueuedCompletionStatus(iocp_.handle,
+    #         0, overlapped_contains_result, op))
+    # Fixed pattern:
+    #   op->completionKey() = overlapped_contains_result;
+    #   if (!::PostQueuedCompletionStatus(iocp_.handle,
+    #         0, op->completionKey(), op))
+    string(REPLACE
+        "if (!::PostQueuedCompletionStatus(iocp_.handle,\n          0, overlapped_contains_result, op))"
+        "op->completionKey() = overlapped_contains_result;\n      if (!::PostQueuedCompletionStatus(iocp_.handle,\n          0, op->completionKey(), op))"
+        ctx_content "${ctx_content}")
+
+    file(WRITE "${CTX_IMPL}" "${ctx_content}")
+
+    message(STATUS "ASIO IOCP fix applied successfully")
+endfunction()

include/glaze/net/websocket_client.hpp

@@ -223,7 +223,32 @@ namespace glz
                resolver_.reset();
             }
 
-            // Reset socket pointers (sockets already closed via force_close above)
+            // Cancel pending operations on raw sockets before destroying them.
+            // During the connection/handshake phase, the websocket_connection doesn't
+            // exist yet, so force_close() above is a no-op. We must cancel any pending
+            // operations on the raw sockets directly.
+            {
+               asio::error_code ec;
+               if (tcp_socket_) {
+                  tcp_socket_->cancel(ec);
+                  tcp_socket_->close(ec);
+               }
+#ifdef GLZ_ENABLE_SSL
+               if (ssl_socket_) {
+                  ssl_socket_->lowest_layer().cancel(ec);
+                  ssl_socket_->lowest_layer().close(ec);
+               }
+#endif
+            }
+
+            // Drain any pending completion handlers (e.g. IOCP cancellation completions)
+            // so they are processed before the io_context or sockets are destroyed.
+            if (ctx) {
+               ctx->restart();
+               ctx->poll();
+            }
+
+            // Now safe to reset socket pointers - all pending operations have completed
             tcp_socket_.reset();
 #ifdef GLZ_ENABLE_SSL
             ssl_socket_.reset();
@@ -452,17 +477,19 @@ namespace glz
                                       ws_conn->set_client_mode(true);
                                       ws_conn->set_max_message_size(self->max_message_size);
 
+                                      // Set handlers before processing initial data, so that any
+                                      // frames in the initial data are properly dispatched.
+                                      if (self->on_message && *self->on_message) ws_conn->on_message(*self->on_message);
+                                      if (self->on_close && *self->on_close) ws_conn->on_close(*self->on_close);
+                                      if (self->on_error && *self->on_error) ws_conn->on_error(*self->on_error);
+
                                       if (response_buf->size() > 0) {
                                          std::string_view initial_data{
                                             static_cast<const char*>(response_buf->data().data()),
                                             response_buf->size()};
                                          ws_conn->set_initial_data(initial_data);
                                       }
 
-                                      if (self->on_message && *self->on_message) ws_conn->on_message(*self->on_message);
-                                      if (self->on_close && *self->on_close) ws_conn->on_close(*self->on_close);
-                                      if (self->on_error && *self->on_error) ws_conn->on_error(*self->on_error);
-
                                       ws_conn->start_read();
 
                                       {

tests/CMakeLists.txt

@@ -62,6 +62,13 @@ else()
             GIT_SHALLOW TRUE
         )
         FetchContent_MakeAvailable(asio)
+
+        # Apply IOCP fix for issue #312 (Windows crash in op->complete)
+        if(WIN32)
+            include(${CMAKE_CURRENT_SOURCE_DIR}/../cmake/asio-iocp-fix.cmake)
+            apply_asio_iocp_fix("${asio_SOURCE_DIR}/asio/include")
+        endif()
+
         add_library(glz_asio INTERFACE)
         target_include_directories(glz_asio INTERFACE ${asio_SOURCE_DIR}/asio/include)
     else()

tests/networking_tests/websocket_test/CMakeLists.txt

@@ -34,6 +34,15 @@ target_link_libraries(${PROJECT_NAME} PRIVATE glz_test_exceptions)
 
 add_test(NAME ${PROJECT_NAME} COMMAND ${PROJECT_NAME})
 
+# WebSocket client lifetime tests (issue #2409)
+project(websocket_client_lifetime_test)
+
+add_executable(${PROJECT_NAME} ${PROJECT_NAME}.cpp)
+
+target_link_libraries(${PROJECT_NAME} PRIVATE glz_test_exceptions)
+
+add_test(NAME ${PROJECT_NAME} COMMAND ${PROJECT_NAME})
+
 # Shared Context Bug Test
 project(shared_context_bug_test)