[BUG] srt_rendezvous hangs after calling srt_close in different thread

[alexpokotilo]

Describe the bug
Problem reproduced only for SRT socket working in blocking mode. When srt_rendezvous call is blocked trying to connect to the peer and srt_close is called for the socket in parallel thread then both srt_rendezvous and srt_close calls hung till SRTO_CONNTIMEO.

This problem is introduced by this commit 03dafd8 and particularly by removing this line:
03dafd8#diff-70625c5a9e97bc3670abd5bfa05dd3cb0f847b9c6a81618cde939d6f104b0a86L1901

To Reproduce

#include <srt.h>

#include <thread>
#include <chrono>

TEST(SRT, SyncRendezvousHangs) {
    ASSERT_EQ(srt_startup(), 0);

    int yes = 1;

    SRTSOCKET m_bindsock = srt_socket(AF_INET, SOCK_DGRAM, 0);
    ASSERT_NE(m_bindsock, SRT_ERROR);

    ASSERT_NE(srt_setsockopt(m_bindsock, 0, SRTO_TSBPDMODE, &yes, sizeof yes), SRT_ERROR);
    ASSERT_NE(srt_setsockflag(m_bindsock, SRTO_SENDER, &yes, sizeof yes), SRT_ERROR);
    ASSERT_EQ(srt_setsockopt(m_bindsock, 0, SRTO_RENDEZVOUS, &yes, sizeof yes), 0);

    const int connection_timeout_ms = 500; // rendezvous timeout is x10 hence 5seconds
    ASSERT_EQ(srt_setsockopt(m_bindsock, 0, SRTO_CONNTIMEO, &connection_timeout_ms, sizeof connection_timeout_ms), 0);

    sockaddr_in local_sa={};
    local_sa.sin_family = AF_INET;
    local_sa.sin_port = htons(9999);
    local_sa.sin_addr.s_addr = INADDR_ANY;

    sockaddr_in peer_sa= {};
    peer_sa.sin_family = AF_INET;
    peer_sa.sin_port = htons(9998);
    ASSERT_EQ(inet_pton(AF_INET, "127.0.0.1", &peer_sa.sin_addr), 1);

    uint64_t duration = 0;

    std::thread close_thread([&m_bindsock, &duration] {
        std::this_thread::sleep_for(std::chrono::seconds(1)); // wait till srt_rendezvous is called
        auto start = std::chrono::steady_clock::now();
        srt_close(m_bindsock);
        auto end = std::chrono::steady_clock::now();

        duration = std::chrono::duration_cast<std::chrono::seconds>(end - start).count();
    });

    ASSERT_EQ(srt_rendezvous(m_bindsock, (sockaddr*)&local_sa, sizeof local_sa,
              (sockaddr*)&peer_sa, sizeof peer_sa), SRT_ERROR);

    close_thread.join();
    ASSERT_LE(duration, 1);
    srt_close(m_bindsock);
    srt_cleanup();
}

Expected behavior
srt_close should return right after call.
srt_rendezvous should return immediately after srt_close is called for the socket in parallel thread.

Desktop (please provide the following information):

• OS: all OSes including Android/IOS where it was found
• SRT Version / commit ID: 1.4.3/03dafd8981e680fec8ddb6437144227ca39284ff

[ethouris]

Could you please check with the fix in #2032? I tested it in the environment with other fixes related to "thread insanity" problems, but I made the fix as little invasive as possible.

[alexpokotilo]

@ethouris,
we've checked on several OSes and fix works fine. Thank you!