[BUG] srt_accept() in blocking mode always prints an error when the socket is closed.

[fchasse-rd]

Describe the bug
I have a thread that continuously accepts connections from a listener socket. The socket is configured in blocking mode, hence the only way to exit an srt_accept call is to close the socket. This results in this log always appearing when I shutdown my app:

17:38:38.926767/eveSrtTests*E:SRT.cn: srt_accept: listener socket @899642338 is already closed

To Reproduce

#include <chrono>
#include <future>
#include <iostream>
#include <thread>

#include <catch2/catch_test_macros.hpp>
#include <srt/srt.h>

TEST_CASE("srt_accept error")
{
    REQUIRE(srt_startup() >= 0);

    auto listener = srt_create_socket();
    REQUIRE(listener != SRT_INVALID_SOCK);

    sockaddr_in bindTarget;
    bindTarget.sin_family      = AF_INET;
    bindTarget.sin_addr.s_addr = inet_addr("127.0.0.1");
    bindTarget.sin_port        = htons(0);

    REQUIRE(srt_bind(listener, (sockaddr*)&bindTarget, sizeof(bindTarget)) == 0);
    REQUIRE(srt_listen(listener, 1) == 0);

    std::future<int> acceptFuture = std::async([listener] {
        sockaddr_in sa;
        int         saLen = sizeof(sa);
        std::ignore       = srt_accept(listener, (sockaddr*)&sa, &saLen);
        return srt_getlasterror(nullptr);
    });

    // Wait a bit to ensure srt_accept() is waiting on the accept.
    std::this_thread::sleep_for(std::chrono::milliseconds(1'000));

    REQUIRE(srt_close(listener) == 0);
    int acceptError = acceptFuture.get();
    CHECK((acceptError == SRT_EINVSOCK || acceptError == SRT_ESCLOSED));

    std::ignore = srt_cleanup();
}

Expected behavior
As this is the intended use, I would not expect an error log to appear.

Suggestion
Could the log level be reduced to info (or removed since we already have the throw) if we're in blocking mode?

[ethouris]

The error log is displayed because this evidently is an erroneous situation. If you don't like the error log, you can turn off logging, or lift the level to fatal. And it's not a normal way to stop listening, hence the error message. By something like historical reasons (including "similarity" requirements in the old API to the system accept function - although this one is even worse than srt_accept because it won't exit even if you close the listener socket; killing the thread or sending a signal to the app is the only way) there's no timeout for accept, as well as there's no mechanism for interrupting a waiting function in SRT.

You can use srt_accept_bond if that is of any help for you - although it's actually a wrapper around epoll, that is, for the internals of this function it sets temporarily non-blocking mode on the listener sockets, then waits for any reporting-in sockets, and restores the original blocking settings before exiting. This is intended for groups, but only as something to make things easier because you can specify multiple listener sockets there, as well as the timeout. This function returns with the first reporting-in connection, or on timeout.

[fchasse-rd]

srt_accept_bond does seem to work nicely, thank you!

Slight note: when running the same unit test as I listed initially, the return code I get is SRT_EPOLLEMPTY, which is not documented as a return code of srt_accept_bond. But the documentation of SRT_EPOLLEMPTY itself does make it seem expected.

Thanks again,

Fred

[ethouris]

Well, this has likely slipped through my hands; that's not aligned with the logics of this function. We'll likely need a fix here. Thanks for reporting it.