I'm using channel, created with libssh2_channel_direct_tcpip2(). Sometimes when the program writes a lot of data, writing process stops and all calls of libssh2_channel_write() returns LIBSSH2_ERROR_EAGAIN.

After some debug I found the reason and fix the problem:
There is key exchange procedure, that starts by _libssh2_transport_send().
Also the _libssh2_transport_send tryes to continue sending existing packet if any by calling send_existing().
send_existing() checks that we are trying to send the same packet that allready exists (if any). If packets are not the same, send_existing() exits with LIBSSH2_ERROR_EAGAIN.
But what happens if we need key exchange and send existing at the same time?

1. _libssh2_transport_send() calls _libssh2_kex_exchange()
2. _libssh2_kex_exchange() makes it's own packet and tryies to send it with _libssh2_transport_send()
3. _libssh2_transport_send() calls send_existing()
4. send_existing() compares new packet address and existing packet address. Test fails and we get LIBSSH2_ERROR_EAGAIN
5. All subsequent write calls exits with LIBSSH2_ERROR_EAGAIN by this scheme

My suggestion is simply call send_existing() before _libssh2_kex_exchange() in _libssh2_transport_send(). That solves the problem.

To Reproduce
The bug is rather random. It's hard to reproduce. You need to create channel with libssh2_channel_direct_tcpip2() and write data as many as possible.

Expected behavior
Succesful write attempts sooner or latter.

Version:

• OS: Ubuntu 22.04 jammy
• Kernel: x86_64 Linux 6.8.0-83-generic
• libssh2 version: 1.11.1
• crypto backend and version: OpenSSL from repository

Additional context
Also I'm using read/write callbacks to run SSH session through custom IO layer, but it doesnt matter I gues.