boost asio error category is null

Question

In my asio based application I have problems with the error handling. I use asio single-threaded/single io_service with the async_read / async_write functions, and when the error-code passed to me in the completion handler is not null, any call to its .message() method results in a segfault, because the error-category pointer points to null. However, the actual value is always correct.

Weirdly I am not able to reproduce this in a test application. This is essentially what I'm doing, but in case of an error, calling message() here works as expected:

void do_write(std::shared_ptr<std::vector<unsigned char>> , std::shared_ptr<boost::asio::ip::tcp::socket> );

void do_read(std::shared_ptr<std::vector<unsigned char>> buf, std::shared_ptr<boost::asio::ip::tcp::socket> sock)
{
   boost::asio::async_read(*sock, boost::asio::buffer(*buf), boost::asio::transfer_exactly(100000),
                           [sock, buf](const boost::system::error_code& ec, std::size_t bytes)
   {
      std::cout << "read\n";
      if (ec) {
         std::cout << ec.message() << std::flush;
      } else {
         do_write(buf, sock);
      }
   });
}

void do_write(std::shared_ptr<std::vector<unsigned char>> buf, std::shared_ptr<boost::asio::ip::tcp::socket> sock)
{
   boost::asio::async_write(*sock, boost::asio::buffer(*buf), [sock, buf](const boost::system::error_code& ec, std::size_t bytes)
   {
      std::cout << "write\n";
      if (ec) {
         std::cout << ec.message() << std::flush;
      } else {
         do_read(buf, sock);
      }
   });
}

int main()
{
   auto buf = std::make_shared<std::vector<unsigned char>>(100000);

   unsigned short port = 1113;
   boost::asio::io_service ios;
   boost::asio::ip::tcp::acceptor acptr {ios, boost::asio::ip::tcp::endpoint{boost::asio::ip::tcp::v4(), port}};

   auto socket = std::make_shared<boost::asio::ip::tcp::socket>(acptr.get_io_service());
   acptr.async_accept(*socket, [socket, buf](const boost::system::error_code& ec)
   {
      do_read(buf, socket);
   });
   ios.run();
}

I do not think it is a problem with lifetimes of my application, AddressSanitizer does not find any potential errors and I can shift hundreds of megabytes of data without any problems. Also if I just uncomment the call to message(), the server continues to run fine and correctly handles new connections. Anyway, here is the callstack when the crash occurs:

https://gist.github.com/mariusherzog/82f24caf9eea4d94946706aa8c025ef1 The category points to null from frame 11 upwards.

I use linux with boost 1.62 with clang 3.9.1 and gcc 5.4.0

Answer 1

The only thing I can think of is right now that a piece of code initializes as part of constructing global static data. If it grabs a category reference at that time and there are multiple translation units, the category instance might not have been constructed yet.

This situation is affectionately referred to as the Static Initialization Order Fiasco.

---

Other than that, the converse comes to mind (the global category has been destructed). This seems a little less likely because 1. you'd have realized it happens at shutdown 2. references don't usually become null in such a scenario - although an implementation might