fix(TCPServer): use poll() instead of select() to remove FD_SETSIZE limit

select() cannot watch file descriptors whose number is >= FD_SETSIZE (1024
on glibc). When the hosting process holds many descriptors (e.g. a JVM such
as MATLAB's), accepted socket FDs exceed that limit; the previous code either
rejected the connection (the set_size_exceeded guard added for select()) or
risked the "bit out of range" fd_set crash.

Replace the select()/fd_set machinery in TCPServer with poll() (WSAPoll() on
Windows), which has no FD_SETSIZE limitation. The pollfd set is rebuilt each
spin() from the listen socket plus client_fds_ (already tracked), so the
masterfds_/tempfds_/maxfd_ members and the FD_SET/FD_CLR/FD_ZERO/FD_ISSET and
set_size_exceeded code are removed entirely.

tests: add two TCPServer regression tests
- services_client_with_high_fd_number (POSIX): consumes low FDs so the
  accepted client FD exceeds FD_SETSIZE, then asserts connect, bidirectional
  data and disconnect all work. Fails on select(), passes on poll().
- receives_from_many_concurrent_clients: many clients send simultaneously;
  asserts the server observes activity on every client FD (guards the poll()
  revents loop).

Author: urrsk

include/ur_client_library/comm/tcp_server.h

@@ -223,10 +223,6 @@ class TCPServer
   std::atomic<socket_t> listen_fd_;
   int port_;
 
-  socket_t maxfd_;
-  fd_set masterfds_;
-  fd_set tempfds_;
-
   uint32_t max_clients_allowed_;
   std::vector<socket_t> client_fds_;
   std::mutex clients_mutex_;

src/comm/tcp_server.cpp

@@ -34,15 +34,19 @@
 
 #include <sstream>
 #include <cstring>
+#include <vector>
 #include "ur_client_library/comm/socket_t.h"
 #include <fcntl.h>
+#ifndef _WIN32
+#  include <poll.h>
+#endif
 
 namespace urcl
 {
 namespace comm
 {
 TCPServer::TCPServer(const int port, const size_t max_num_tries, const std::chrono::milliseconds reconnection_time)
-  : port_(port), maxfd_(0), max_clients_allowed_(0)
+  : port_(port), max_clients_allowed_(0)
 {
 #ifdef _WIN32
   WSAData data;
@@ -74,9 +78,6 @@ void TCPServer::init()
   ur_setsockopt(listen_fd_, SOL_SOCKET, SO_KEEPALIVE, &flag, sizeof(int));
 
   URCL_LOG_DEBUG("Created socket with FD %d", (int)listen_fd_);
-
-  FD_ZERO(&masterfds_);
-  FD_ZERO(&tempfds_);
 }
 
 void TCPServer::shutdown()
@@ -179,9 +180,6 @@ void TCPServer::bind(const size_t max_num_tries, const std::chrono::milliseconds
   } while (err == -1 && (connection_counter <= max_num_tries || max_num_tries == 0));
 
   URCL_LOG_DEBUG("Bound %d:%d to FD %d", server_addr.sin_addr.s_addr, port_, (int)listen_fd_);
-
-  FD_SET(listen_fd_, &masterfds_);
-  maxfd_ = listen_fd_;
 }
 
 void TCPServer::startListen()
@@ -220,33 +218,13 @@ void TCPServer::handleConnect()
     return;
   }
 
-#ifdef _WIN32
-  bool set_size_exceeded = client_fds_.size() >= FD_SETSIZE - 1;  // -1 because listen_fd_ also occupies one
-                                                                  // slot in masterfds_
-#else
-  bool set_size_exceeded = client_fd >= FD_SETSIZE;  // On Unix-like systems, the client FD itself must be less than
-                                                     // FD_SETSIZE, otherwise it cannot be added to the fd_set.
-#endif
-
-  if (set_size_exceeded)
-  {
-    URCL_LOG_ERROR("Accepted client FD %d exceeds FD_SETSIZE (%d). Closing connection.", (int)client_fd, FD_SETSIZE);
-    ur_close(client_fd);
-    return;
-  }
-
   bool accepted = false;
 
   {
     std::lock_guard<std::mutex> lk(clients_mutex_);
     if (client_fds_.size() < max_clients_allowed_ || max_clients_allowed_ == 0)
     {
       client_fds_.push_back(client_fd);
-      FD_SET(client_fd, &masterfds_);
-      if (client_fd > maxfd_)
-      {
-        maxfd_ = client_fd;
-      }
       accepted = true;
     }
     else
@@ -268,27 +246,40 @@ void TCPServer::handleConnect()
 
 void TCPServer::spin()
 {
-  tempfds_ = masterfds_;
-
-  timeval timeout;
-  timeout.tv_sec = 1;
-  timeout.tv_usec = 0;
+  // Build the poll set fresh each iteration from the listen socket plus all currently connected
+  // clients. poll() is used on both platforms (WSAPoll() on Windows) because it has no
+  // FD_SETSIZE limit on file descriptor numbers, unlike select(). This matters when the hosting
+  // process holds many file descriptors (e.g. a JVM), pushing socket FDs past FD_SETSIZE (1024).
+  std::vector<struct pollfd> pollfds;
+  pollfds.push_back({ static_cast<socket_t>(listen_fd_), POLLIN, 0 });
+  {
+    std::lock_guard<std::mutex> lk(clients_mutex_);
+    for (const auto& client_fd : client_fds_)
+    {
+      pollfds.push_back({ client_fd, POLLIN, 0 });
+    }
+  }
 
-  // blocks until activity on any socket from tempfds
-  int sel = select(static_cast<int>(maxfd_ + 1), &tempfds_, NULL, NULL, &timeout);
-  if (sel < 0)
+  // Block for up to 1 s waiting for activity on any socket. A shutdown wakes this immediately by
+  // connecting to the listen socket (see shutdown()).
+#ifdef _WIN32
+  int ready = ::WSAPoll(pollfds.data(), static_cast<ULONG>(pollfds.size()), 1000);
+#else
+  int ready = ::poll(pollfds.data(), pollfds.size(), 1000);
+#endif
+  if (ready < 0)
   {
-    URCL_LOG_ERROR("select() failed. Shutting down socket event handler.");
+    URCL_LOG_ERROR("poll() failed. Shutting down socket event handler.");
     keep_running_ = false;
     return;
   }
 
-  if (!keep_running_ || sel == 0)
+  if (!keep_running_ || ready == 0)
   {
     return;
   }
 
-  if (FD_ISSET(listen_fd_, &tempfds_))
+  if (pollfds[0].revents & POLLIN)
   {
     URCL_LOG_DEBUG("Activity on listen FD %d", (int)listen_fd_);
     handleConnect();
@@ -297,15 +288,13 @@ void TCPServer::spin()
   std::vector<socket_t> disconnected_clients;
   std::vector<socket_t> client_fds_with_activity;
 
+  // pollfds[0] is the listen socket; client entries start at index 1.
+  for (size_t i = 1; i < pollfds.size(); ++i)
   {
-    std::lock_guard<std::mutex> lk(clients_mutex_);
-    for (const auto& client_fd : client_fds_)
+    if (pollfds[i].revents & (POLLIN | POLLHUP | POLLERR))
     {
-      if (FD_ISSET(client_fd, &tempfds_))
-      {
-        URCL_LOG_DEBUG("Activity on client FD %d", (int)client_fd);
-        client_fds_with_activity.push_back(client_fd);
-      }
+      URCL_LOG_DEBUG("Activity on client FD %d", (int)pollfds[i].fd);
+      client_fds_with_activity.push_back(static_cast<socket_t>(pollfds[i].fd));
     }
   }
   // We handle client activity outside the clients_mutex_ lock to avoid holding it during potentially slow I/O and
@@ -331,7 +320,6 @@ void TCPServer::handleDisconnect(const socket_t fd)
   {
     std::lock_guard<std::mutex> lk(clients_mutex_);
     ur_close(fd);
-    FD_CLR(fd, &masterfds_);
 
     for (size_t i = 0; i < client_fds_.size(); ++i)
     {
@@ -341,15 +329,6 @@ void TCPServer::handleDisconnect(const socket_t fd)
         break;
       }
     }
-
-    maxfd_ = listen_fd_;
-    for (const auto& client_fd : client_fds_)
-    {
-      if (client_fd > maxfd_)
-      {
-        maxfd_ = client_fd;
-      }
-    }
   }
 
   {

tests/test_tcp_server.cpp

@@ -34,8 +34,14 @@
 #include <condition_variable>
 #include <chrono>
 #include <memory>
+#include <mutex>
 #include <thread>
 #include <vector>
+#ifndef _WIN32
+#  include <fcntl.h>
+#  include <unistd.h>
+#  include <sys/resource.h>
+#endif
 #include "test_utils.h"
 
 #include <ur_client_library/comm/tcp_server.h>
@@ -497,6 +503,134 @@ TEST_F(TCPServerTest, shutdown_during_active_writes)
   writer.join();
 }
 
+// Verifies that the server receives data from many clients that all send simultaneously. This
+// exercises the poll() revents loop across many client file descriptors and guards against
+// missed read events when several sockets are readable at once.
+TEST_F(TCPServerTest, receives_from_many_concurrent_clients)
+{
+  comm::TCPServer server(0);
+
+  std::mutex mtx;
+  std::condition_variable cv;
+  std::atomic<int> message_count{ 0 };
+
+  server.setMessageCallback([&](const socket_t, char*, int) {
+    message_count.fetch_add(1);
+    std::lock_guard<std::mutex> lk(mtx);
+    cv.notify_all();
+  });
+  server.start();
+
+#ifdef _WIN32
+  // Windows allows a maximum of 64 sockets per process by default.
+  constexpr int num_clients = 50;
+#else
+  constexpr int num_clients = 100;
+#endif
+
+  std::vector<std::unique_ptr<Client>> clients;
+  for (int i = 0; i < num_clients; ++i)
+  {
+    clients.push_back(std::make_unique<Client>(server.getPort()));
+  }
+
+  // Every client sends a single message concurrently.
+  std::vector<std::thread> senders;
+  for (auto& client : clients)
+  {
+    senders.emplace_back([&client]() { client->send("ping\n"); });
+  }
+  for (auto& t : senders)
+  {
+    t.join();
+  }
+
+  // The server's poll() loop must observe activity on every client FD and deliver all messages.
+  std::unique_lock<std::mutex> lk(mtx);
+  EXPECT_TRUE(cv.wait_for(lk, std::chrono::seconds(5), [&]() { return message_count.load() >= num_clients; }));
+  EXPECT_EQ(message_count.load(), num_clients);
+}
+
+#ifndef _WIN32
+// Regression test for the FD_SETSIZE limitation of select(): a client whose accepted socket file
+// descriptor number is >= FD_SETSIZE (1024) must still be serviced normally. This is the exact
+// scenario that occurs when the hosting process (e.g. a JVM) holds many file descriptors. The old
+// select()-based implementation rejected/crashed on such descriptors; poll() handles them.
+TEST_F(TCPServerTest, services_client_with_high_fd_number)
+{
+  // Make sure we are allowed to open more than FD_SETSIZE descriptors; raise the soft limit if
+  // needed and skip the test if the hard limit does not allow it.
+  struct rlimit rl;
+  ASSERT_EQ(getrlimit(RLIMIT_NOFILE, &rl), 0);
+  const rlim_t needed = static_cast<rlim_t>(FD_SETSIZE) + 64;
+  if (rl.rlim_cur < needed)
+  {
+    rl.rlim_cur = std::min<rlim_t>(needed, rl.rlim_max);
+    if (setrlimit(RLIMIT_NOFILE, &rl) != 0 || rl.rlim_cur < needed)
+    {
+      GTEST_SKIP() << "Cannot raise RLIMIT_NOFILE above FD_SETSIZE; skipping high-fd test.";
+    }
+  }
+
+  // Consume the low-numbered descriptors so that subsequently created sockets are assigned fd
+  // numbers beyond FD_SETSIZE.
+  std::vector<int> fd_hogs;
+  while (true)
+  {
+    int fd = ::open("/dev/null", O_RDONLY);
+    if (fd < 0)
+    {
+      break;
+    }
+    fd_hogs.push_back(fd);
+    if (fd > static_cast<int>(FD_SETSIZE) + 8)
+    {
+      break;
+    }
+  }
+  const bool pushed_past_limit = !fd_hogs.empty() && fd_hogs.back() > static_cast<int>(FD_SETSIZE);
+  if (!pushed_past_limit)
+  {
+    for (int fd : fd_hogs)
+    {
+      ::close(fd);
+    }
+    GTEST_SKIP() << "Could not allocate descriptors beyond FD_SETSIZE; skipping.";
+  }
+
+  TestableTcpServer server(port_);
+  server.start();
+
+  Client client(port_);
+  EXPECT_TRUE(server.waitForConnectionCallback(2000));
+
+  // The server-side accepted client FD should exceed FD_SETSIZE -- the case that breaks select().
+  auto client_fds = server.getClientFDs();
+  ASSERT_FALSE(client_fds.empty());
+  EXPECT_GT(client_fds.back(), static_cast<socket_t>(FD_SETSIZE));
+
+  // Data must flow both ways on the high-numbered descriptor.
+  const std::string message = "high fd message\n";
+  client.send(message);
+  EXPECT_TRUE(server.waitForMessageCallback(2000));
+  EXPECT_EQ(server.getReceivedMessage(), message);
+
+  size_t written;
+  const auto* data = reinterpret_cast<const uint8_t*>(message.c_str());
+  ASSERT_TRUE(server.write(data, message.size(), written));
+  EXPECT_EQ(client.recv(), message);
+
+  // Disconnect must also be detected on the high-numbered descriptor.
+  client.close();
+  EXPECT_TRUE(server.waitForDisconnectionCallback(2000));
+
+  for (int fd : fd_hogs)
+  {
+    ::close(fd);
+  }
+}
+#endif
+
 int main(int argc, char* argv[])
 {
   ::testing::InitGoogleTest(&argc, argv);