TCPSocket::Connect() calls SocketAddress::GetAddressInfo() and tries to
connect any of them (in a for loop).
This used to work before commit 4f6d3a376c9f("[LLDB] Fix setting of
success in Socket::Close()") https://reviews.llvm.org/D116768.
As a side effect of that commit, TCPSocket can only connect to the first
address returned by SocketAddress::GetAddressInfo().
- If the attempt to connect to the first address fails, TCPSocket::Connect(), calls CLOSE_SOCKET(GetNativeSocket()), which closes the fd, but DOES NOT set m_socket to kInvalidSocketValue.
- On the second attempt, TCPSocket::CreateSocket() calls Socket::Close().
- Socket::Close() proceeds, because IsValid() is true (m_socket was not reset on step 1).
- Socket::Close() calls ::close(m_socket), which fails
- Since commit 4f6d3a376c9f("[LLDB] Fix setting of success in Socket::Close()"), this is error is detected. Socket::Close() returns an error.
- TCPSocket::CreateSocket() therefore returns an error.
- TCPSocket::Connect() detects the error and continues, skipping the second (and the third, fourth...) address.
This commit fixes the problem by changing step 1: by calling
Socket::Close, instead of directly calling close(m_socket), m_socket is
also se to kInvalidSocketValue. On step 3, Socket::Close() is going to
return immediately and, on step 6, TCPSocket::CreateSocket() does not
fail.
How to reproduce this problem:
On my system, getaddrinfo() resolves "localhost" to "::1" (first) and to
"127.0.0.1" (second).
Start a gdbserver that only listens on 127.0.0.1:
gdbserver 127.0.0.1:2159 /bin/cat Process /bin/cat created; pid = 2146709 Listening on port 2159
Start lldb and make it connect to "localhost:2159"
./bin/lldb (lldb) gdb-remote localhost:2159
Before 4f6d3a376c9f("[LLDB] Fix setting of success in Socket::Close()"),
this used to work. After that commit, it stopped working. This commit
fixes the problem.