robertleahy / asio-curl Goto Github PK

Running the code in the attached file (uploaded as ZIP, because direct cpp is not supported, for whatever reason), which is based mostly on the example from the main project page, with include files and proper termination added, leads to a crash.

System: Ubuntu 20.04 LTS
Boost Version: 1.73.0 (self compiled)
libcurl-gnutls as packaged with Ubuntu 20.04 LTS

The crash happens in the main thread in the call to curl.add(easy), which in turn calls curl_multi_add_handle(), which calls back to asiocurl::io_service::timer(), which calls asiocurl::io_service::do_action(), which calls curl_multi_socket_action(), which calls asiocurl::io_service::socket().

When breaking with GDB into asiocurl::io_service::socket() and analyzing the passed in userp, the good news is, that it seems to point to a valid asiocurl::io_service. The bad news is, that the sockets_ field of that io_service is an empty std::unordered_map . Somehow, curl is calling back asiocurl::io_service::socket(), when the io_service is not even aware of any socket, yet.

No need to say, that the following line of asiocurl::io_service::socket()

auto & ss=self.sockets_.find(socket)->second;

doesn't find anything and thus continues with an invalid pointer, which soon leads to a SEGFAULT.

With all the changes of both boost::asio and libcurl during the last years, I am not going to debug this problem any further.
test.zip

Two problems:

1. Depends on Boost
2. Depends specifically on Boost.ASIO

It would be nice if this library could be used without having to include all of Boost. Also, it would be nice if it was possible to use the version of ASIO that isn't part of Boost.

I understand this code is old and may not be maintained, but I noticed a issue while using it as a example reference.

If you request something that can't fit in a single buffer eg 16KB the code will stall indefinitely.

This is because the code is only calling socket read/write once in response to ::socket calls.
Instead it should enter a async loop calling read/write over and over unless otherwise instructed by another call to ::socket.

In order to fix it I suggest moving the socket read/write logic into another function and calling that function again instead of do_action after a read/write completes.

Since I only used this as a reference I don't have changes for a pull request, sorry.