On Fri, Mar 13, 2020 at 06:39:30PM -0700, Andrii Nakryiko wrote:
> Previous attempt to make tcp_rtt more robust introduced a new race, in which
> server_done might be set to true before server can actually accept any
> connection. Fix this by unconditionally waiting for accept(). Given socket is
> non-blocking, if there are any problems with client side, it should eventually
> close listening FD and let server thread exit with failure.
>
> Fixes: 4cd729fa022c ("selftests/bpf: Make tcp_rtt test more robust to failures")
> Signed-off-by: Andrii Nakryiko <andriin@xxxxxx>
> ---
> tools/testing/selftests/bpf/prog_tests/tcp_rtt.c | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/tools/testing/selftests/bpf/prog_tests/tcp_rtt.c b/tools/testing/selftests/bpf/prog_tests/tcp_rtt.c
> index e08f6bb17700..e56b52ab41da 100644
> --- a/tools/testing/selftests/bpf/prog_tests/tcp_rtt.c
> +++ b/tools/testing/selftests/bpf/prog_tests/tcp_rtt.c
> @@ -226,7 +226,7 @@ static void *server_thread(void *arg)
> return ERR_PTR(err);
> }
>
> - while (!server_done) {
> + while (true) {
> client_fd = accept(fd, (struct sockaddr *)&addr, &len);
> if (client_fd == -1 && errno == EAGAIN) {
> usleep(50);
> @@ -272,7 +272,7 @@ void test_tcp_rtt(void)
> CHECK_FAIL(run_test(cgroup_fd, server_fd));
>
> server_done = true;
> - pthread_join(tid, &server_res);
> + CHECK_FAIL(pthread_join(tid, &server_res));
>From looking at run_test(),
I suspect without accept and server_thread, this may also work:
listen(server_fd, 1);
run_test(cgroup_fd, server_fd);
close(server_fd);
This change lgtm since it is NONBLOCK. Ideally, may be it should
also be time limited by epoll or setsockopt(SO_RCVTIMEO) in the future.
Acked-by: Martin KaFai Lau <kafai@xxxxxx>
