Posix remote_address() may throw, but it is marked noexcept #2379
Comments
|
But the class name suggests that the socket is connected, while kernel thinks otherwise. Can you shed more light on this -- how did you end up having unconnected socket wrapped with connected_socket_impl object? |
|
This happened when I used an external tcpkill process to rapidly terminate connections, for testing purposes. What is supposed to happen to a |
Ah, I see. In that case this can be a nice unit test too -- a local listening posix socket, a local established connection, then server-side closes and client-side tries to call
It won't transform into some other class, that's for sure. No I understand where the issue could be. |
|
Can you show me your full test? The ENOTCON is returned for SYN_SENT and CLOSE states. The latter doesn't happen until client closes its side, and if it does it's no longer valid to call remote_address() on connected_socket(). So I assume we're in SYN_SENT state, but don't see how it happens until connected_socket is instantiated. |
|
hi Pavel, Here is an (ugly) reproducer (no attempts to cleanly shutdown services, just a fairly minimal reproducer to trigger the bug): I first run, in another terminal, tcpkill, and leave it running: Then I run the following program: (I compiled my test in dev mode) #include "stop_signal.hh"
#include <seastar/core/abort_source.hh>
#include <seastar/core/app-template.hh>
#include <seastar/core/future.hh>
#include <seastar/core/sleep.hh>
#include <seastar/core/thread.hh>
#include <seastar/core/when_all.hh>
#include <seastar/util/log.hh>
static seastar::logger mylog("main");
using namespace seastar;
using namespace std::chrono_literals;
abort_source as;
future<> client()
{
ipv4_addr addr(8000);
connected_socket s = co_await connect(addr);
auto out = s.output();
auto in = s.input();
std::string hello{"hello"};
co_await out.write(hello);
co_await out.flush();
auto payload = co_await in.read();
}
future<> server()
{
connected_socket srv_sock;
listen_options opts;
opts.reuse_address = true;
opts.set_fixed_cpu(0);
server_socket socket = listen(make_ipv4_address(8000), opts);
accept_result ar = co_await socket.accept();
srv_sock = std::move(ar.connection);
auto in = srv_sock.input();
auto out = srv_sock.output();
while (!as.abort_requested()) {
if (as.abort_requested()) {
break;
}
uint16_t acc{};
for (int i = 0; i < 10000; ++i) {
auto ra = srv_sock.remote_address();
acc += ra.port();
}
mylog.info("port sum: {}", acc);
}
}
int main(int argc, char** argv)
{
app_template app;
return app.run(argc, argv, [] {
return seastar::async([] {
seastar_apps_lib::stop_signal s;
auto server_fut = server();
auto client_fut = client();
s.wait().get();
});
});
}When the program is run, tcpkill has output like this: The seastar program then crashes with: |
|
@xemul Is this helpful? |
Polite "ping" @xemul . See reproducer above. |
I ran into the following error:
Partial backtrace:
I believe the cause is
posix_connected_socket_impl::remote_address()is marked noexcept, but it callsposix_connected_socket_operations::remote_address(), which is not noexcept, and in fact the latter callsfile_desc::get_remote_address()which explicitly throws.See:
https://github.com/scylladb/seastar/blob/master/src/net/posix-stack.cc#L288
https://github.com/scylladb/seastar/blob/master/src/net/posix-stack.cc#L110
https://github.com/scylladb/seastar/blob/master/include/seastar/core/posix.hh#L298
The text was updated successfully, but these errors were encountered: