Programming

Oh yeah, or set MSG_NOSIGNAL as last parameter to send()

Yes! That's it. The problem was that there appeared SIGPIPE during second communicate() invocation. It was killing the program (SIGPIPE default action is to terminate process).
But there is still one strange thing. On the client side, during server breakdown, first send_request does not return any error, only the second one (I have checked, it get SIGPIPE, so I can handle it properly). Why? What is about this first execution that it does not fail? Neither send nor recv get SIGPIPE.

TCP buffers, you know. You might not get all the data on first read() so there could be data left over from last time if you weren't careful to get it all. And it also buffers on the way out. It's not psychic either, so may not know the instant when the connection does.

Because send() just copies it to a local buffer to be sent out, it is not until the other end fails to recv() that the OS can detect a broken pipe.

I understand. So how to discover server breakdown on the first request sending?

You should make sure you're reading all the data in the first place. Just because you do one successful read() doesn't mean you got everything the server was sending. If you have no way of knowing how much the server was sending, you'll have to add that data to your protocol.

That way, next time you try and communicate with the server, there won't be old data lying in the buffer that it will mistake for a reply; it may succeed in sending data, not knowing the connection's broken yet, but won't get any reply, and time out instead.

Maybe you would like to put a timeout on your select(), "No response from server." and you did change your recv() to check for 0?

Your code is beautiful, btw.