diff options
Diffstat (limited to 'test/syscalls/linux/socket_inet_loopback.cc')
| -rw-r--r-- | test/syscalls/linux/socket_inet_loopback.cc | 443 |
1 files changed, 443 insertions, 0 deletions
diff --git a/test/syscalls/linux/socket_inet_loopback.cc b/test/syscalls/linux/socket_inet_loopback.cc index 322ee07ad..2eeee352e 100644 --- a/test/syscalls/linux/socket_inet_loopback.cc +++ b/test/syscalls/linux/socket_inet_loopback.cc @@ -13,9 +13,11 @@ // limitations under the License. #include <arpa/inet.h> +#include <linux/tcp.h> #include <netinet/in.h> #include <poll.h> #include <string.h> +#include <sys/epoll.h> #include <sys/socket.h> #include <atomic> @@ -30,6 +32,7 @@ #include "gtest/gtest.h" #include "absl/memory/memory.h" #include "absl/strings/str_cat.h" +#include "absl/time/clock.h" #include "absl/time/time.h" #include "test/syscalls/linux/socket_test_util.h" #include "test/util/file_descriptor.h" @@ -266,6 +269,340 @@ TEST_P(SocketInetLoopbackTest, TCPbacklog) { } } +// TCPFinWait2Test creates a pair of connected sockets then closes one end to +// trigger FIN_WAIT2 state for the closed endpoint. Then it binds the same local +// IP/port on a new socket and tries to connect. The connect should fail w/ +// an EADDRINUSE. Then we wait till the FIN_WAIT2 timeout is over and try the +// connect again with a new socket and this time it should succeed. +// +// TCP timers are not S/R today, this can cause this test to be flaky when run +// under random S/R due to timer being reset on a restore. +TEST_P(SocketInetLoopbackTest, TCPFinWait2Test_NoRandomSave) { + auto const& param = GetParam(); + TestAddress const& listener = param.listener; + TestAddress const& connector = param.connector; + + // Create the listening socket. + const FileDescriptor listen_fd = ASSERT_NO_ERRNO_AND_VALUE( + Socket(listener.family(), SOCK_STREAM, IPPROTO_TCP)); + sockaddr_storage listen_addr = listener.addr; + ASSERT_THAT(bind(listen_fd.get(), reinterpret_cast<sockaddr*>(&listen_addr), + listener.addr_len), + SyscallSucceeds()); + ASSERT_THAT(listen(listen_fd.get(), SOMAXCONN), SyscallSucceeds()); + + // Get the port bound by the listening socket. + socklen_t addrlen = listener.addr_len; + ASSERT_THAT(getsockname(listen_fd.get(), + reinterpret_cast<sockaddr*>(&listen_addr), &addrlen), + SyscallSucceeds()); + + uint16_t const port = + ASSERT_NO_ERRNO_AND_VALUE(AddrPort(listener.family(), listen_addr)); + + // Connect to the listening socket. + FileDescriptor conn_fd = ASSERT_NO_ERRNO_AND_VALUE( + Socket(connector.family(), SOCK_STREAM, IPPROTO_TCP)); + + // Lower FIN_WAIT2 state to 5 seconds for test. + constexpr int kTCPLingerTimeout = 5; + EXPECT_THAT(setsockopt(conn_fd.get(), IPPROTO_TCP, TCP_LINGER2, + &kTCPLingerTimeout, sizeof(kTCPLingerTimeout)), + SyscallSucceedsWithValue(0)); + + sockaddr_storage conn_addr = connector.addr; + ASSERT_NO_ERRNO(SetAddrPort(connector.family(), &conn_addr, port)); + ASSERT_THAT(RetryEINTR(connect)(conn_fd.get(), + reinterpret_cast<sockaddr*>(&conn_addr), + connector.addr_len), + SyscallSucceeds()); + + // Accept the connection. + auto accepted = + ASSERT_NO_ERRNO_AND_VALUE(Accept(listen_fd.get(), nullptr, nullptr)); + + // Get the address/port bound by the connecting socket. + sockaddr_storage conn_bound_addr; + socklen_t conn_addrlen = connector.addr_len; + ASSERT_THAT( + getsockname(conn_fd.get(), reinterpret_cast<sockaddr*>(&conn_bound_addr), + &conn_addrlen), + SyscallSucceeds()); + + // close the connecting FD to trigger FIN_WAIT2 on the connected fd. + conn_fd.reset(); + + // Now bind and connect a new socket. + const FileDescriptor conn_fd2 = ASSERT_NO_ERRNO_AND_VALUE( + Socket(connector.family(), SOCK_STREAM, IPPROTO_TCP)); + + // Disable cooperative saves after this point. As a save between the first + // bind/connect and the second one can cause the linger timeout timer to + // be restarted causing the final bind/connect to fail. + DisableSave ds; + + // TODO(gvisor.dev/issue/1030): Portmanager does not track all 5 tuple + // reservations which causes the bind() to succeed on gVisor but connect + // correctly fails. + if (IsRunningOnGvisor()) { + ASSERT_THAT( + bind(conn_fd2.get(), reinterpret_cast<sockaddr*>(&conn_bound_addr), + conn_addrlen), + SyscallSucceeds()); + ASSERT_THAT(RetryEINTR(connect)(conn_fd2.get(), + reinterpret_cast<sockaddr*>(&conn_addr), + conn_addrlen), + SyscallFailsWithErrno(EADDRINUSE)); + } else { + ASSERT_THAT( + bind(conn_fd2.get(), reinterpret_cast<sockaddr*>(&conn_bound_addr), + conn_addrlen), + SyscallFailsWithErrno(EADDRINUSE)); + } + + // Sleep for a little over the linger timeout to reduce flakiness in + // save/restore tests. + absl::SleepFor(absl::Seconds(kTCPLingerTimeout + 1)); + + ds.reset(); + + if (!IsRunningOnGvisor()) { + ASSERT_THAT( + bind(conn_fd2.get(), reinterpret_cast<sockaddr*>(&conn_bound_addr), + conn_addrlen), + SyscallSucceeds()); + } + ASSERT_THAT(RetryEINTR(connect)(conn_fd2.get(), + reinterpret_cast<sockaddr*>(&conn_addr), + conn_addrlen), + SyscallSucceeds()); +} + +// TCPLinger2TimeoutAfterClose creates a pair of connected sockets +// then closes one end to trigger FIN_WAIT2 state for the closed endpont. +// It then sleeps for the TCP_LINGER2 timeout and verifies that bind/ +// connecting the same address succeeds. +// +// TCP timers are not S/R today, this can cause this test to be flaky when run +// under random S/R due to timer being reset on a restore. +TEST_P(SocketInetLoopbackTest, TCPLinger2TimeoutAfterClose_NoRandomSave) { + auto const& param = GetParam(); + TestAddress const& listener = param.listener; + TestAddress const& connector = param.connector; + + // Create the listening socket. + const FileDescriptor listen_fd = ASSERT_NO_ERRNO_AND_VALUE( + Socket(listener.family(), SOCK_STREAM, IPPROTO_TCP)); + sockaddr_storage listen_addr = listener.addr; + ASSERT_THAT(bind(listen_fd.get(), reinterpret_cast<sockaddr*>(&listen_addr), + listener.addr_len), + SyscallSucceeds()); + ASSERT_THAT(listen(listen_fd.get(), SOMAXCONN), SyscallSucceeds()); + + // Get the port bound by the listening socket. + socklen_t addrlen = listener.addr_len; + ASSERT_THAT(getsockname(listen_fd.get(), + reinterpret_cast<sockaddr*>(&listen_addr), &addrlen), + SyscallSucceeds()); + + uint16_t const port = + ASSERT_NO_ERRNO_AND_VALUE(AddrPort(listener.family(), listen_addr)); + + // Connect to the listening socket. + FileDescriptor conn_fd = ASSERT_NO_ERRNO_AND_VALUE( + Socket(connector.family(), SOCK_STREAM, IPPROTO_TCP)); + + sockaddr_storage conn_addr = connector.addr; + ASSERT_NO_ERRNO(SetAddrPort(connector.family(), &conn_addr, port)); + ASSERT_THAT(RetryEINTR(connect)(conn_fd.get(), + reinterpret_cast<sockaddr*>(&conn_addr), + connector.addr_len), + SyscallSucceeds()); + + // Accept the connection. + auto accepted = + ASSERT_NO_ERRNO_AND_VALUE(Accept(listen_fd.get(), nullptr, nullptr)); + + // Get the address/port bound by the connecting socket. + sockaddr_storage conn_bound_addr; + socklen_t conn_addrlen = connector.addr_len; + ASSERT_THAT( + getsockname(conn_fd.get(), reinterpret_cast<sockaddr*>(&conn_bound_addr), + &conn_addrlen), + SyscallSucceeds()); + + constexpr int kTCPLingerTimeout = 5; + EXPECT_THAT(setsockopt(conn_fd.get(), IPPROTO_TCP, TCP_LINGER2, + &kTCPLingerTimeout, sizeof(kTCPLingerTimeout)), + SyscallSucceedsWithValue(0)); + + // close the connecting FD to trigger FIN_WAIT2 on the connected fd. + conn_fd.reset(); + + absl::SleepFor(absl::Seconds(kTCPLingerTimeout + 1)); + + // Now bind and connect a new socket and verify that we can immediately + // rebind the address bound by the conn_fd as it never entered TIME_WAIT. + const FileDescriptor conn_fd2 = ASSERT_NO_ERRNO_AND_VALUE( + Socket(connector.family(), SOCK_STREAM, IPPROTO_TCP)); + + ASSERT_THAT(bind(conn_fd2.get(), + reinterpret_cast<sockaddr*>(&conn_bound_addr), conn_addrlen), + SyscallSucceeds()); + ASSERT_THAT(RetryEINTR(connect)(conn_fd2.get(), + reinterpret_cast<sockaddr*>(&conn_addr), + conn_addrlen), + SyscallSucceeds()); +} + +// TCPResetAfterClose creates a pair of connected sockets then closes +// one end to trigger FIN_WAIT2 state for the closed endpoint verifies +// that we generate RSTs for any new data after the socket is fully +// closed. +TEST_P(SocketInetLoopbackTest, TCPResetAfterClose) { + auto const& param = GetParam(); + TestAddress const& listener = param.listener; + TestAddress const& connector = param.connector; + + // Create the listening socket. + const FileDescriptor listen_fd = ASSERT_NO_ERRNO_AND_VALUE( + Socket(listener.family(), SOCK_STREAM, IPPROTO_TCP)); + sockaddr_storage listen_addr = listener.addr; + ASSERT_THAT(bind(listen_fd.get(), reinterpret_cast<sockaddr*>(&listen_addr), + listener.addr_len), + SyscallSucceeds()); + ASSERT_THAT(listen(listen_fd.get(), SOMAXCONN), SyscallSucceeds()); + + // Get the port bound by the listening socket. + socklen_t addrlen = listener.addr_len; + ASSERT_THAT(getsockname(listen_fd.get(), + reinterpret_cast<sockaddr*>(&listen_addr), &addrlen), + SyscallSucceeds()); + + uint16_t const port = + ASSERT_NO_ERRNO_AND_VALUE(AddrPort(listener.family(), listen_addr)); + + // Connect to the listening socket. + FileDescriptor conn_fd = ASSERT_NO_ERRNO_AND_VALUE( + Socket(connector.family(), SOCK_STREAM, IPPROTO_TCP)); + + sockaddr_storage conn_addr = connector.addr; + ASSERT_NO_ERRNO(SetAddrPort(connector.family(), &conn_addr, port)); + ASSERT_THAT(RetryEINTR(connect)(conn_fd.get(), + reinterpret_cast<sockaddr*>(&conn_addr), + connector.addr_len), + SyscallSucceeds()); + + // Accept the connection. + auto accepted = + ASSERT_NO_ERRNO_AND_VALUE(Accept(listen_fd.get(), nullptr, nullptr)); + + // close the connecting FD to trigger FIN_WAIT2 on the connected fd. + conn_fd.reset(); + + int data = 1234; + + // Now send data which should trigger a RST as the other end should + // have timed out and closed the socket. + EXPECT_THAT(RetryEINTR(send)(accepted.get(), &data, sizeof(data), 0), + SyscallSucceeds()); + // Sleep for a shortwhile to get a RST back. + absl::SleepFor(absl::Seconds(1)); + + // Try writing again and we should get an EPIPE back. + EXPECT_THAT(RetryEINTR(send)(accepted.get(), &data, sizeof(data), 0), + SyscallFailsWithErrno(EPIPE)); + + // Trying to read should return zero as the other end did send + // us a FIN. We do it twice to verify that the RST does not cause an + // ECONNRESET on the read after EOF has been read by applicaiton. + EXPECT_THAT(RetryEINTR(recv)(accepted.get(), &data, sizeof(data), 0), + SyscallSucceedsWithValue(0)); + EXPECT_THAT(RetryEINTR(recv)(accepted.get(), &data, sizeof(data), 0), + SyscallSucceedsWithValue(0)); +} + +// This test is disabled under random save as the the restore run +// results in the stack.Seed() being different which can cause +// sequence number of final connect to be one that is considered +// old and can cause the test to be flaky. +TEST_P(SocketInetLoopbackTest, TCPTimeWaitTest_NoRandomSave) { + auto const& param = GetParam(); + TestAddress const& listener = param.listener; + TestAddress const& connector = param.connector; + + // Create the listening socket. + const FileDescriptor listen_fd = ASSERT_NO_ERRNO_AND_VALUE( + Socket(listener.family(), SOCK_STREAM, IPPROTO_TCP)); + sockaddr_storage listen_addr = listener.addr; + ASSERT_THAT(bind(listen_fd.get(), reinterpret_cast<sockaddr*>(&listen_addr), + listener.addr_len), + SyscallSucceeds()); + ASSERT_THAT(listen(listen_fd.get(), SOMAXCONN), SyscallSucceeds()); + + // Get the port bound by the listening socket. + socklen_t addrlen = listener.addr_len; + ASSERT_THAT(getsockname(listen_fd.get(), + reinterpret_cast<sockaddr*>(&listen_addr), &addrlen), + SyscallSucceeds()); + + uint16_t const port = + ASSERT_NO_ERRNO_AND_VALUE(AddrPort(listener.family(), listen_addr)); + + // Connect to the listening socket. + FileDescriptor conn_fd = ASSERT_NO_ERRNO_AND_VALUE( + Socket(connector.family(), SOCK_STREAM, IPPROTO_TCP)); + + // We disable saves after this point as a S/R causes the netstack seed + // to be regenerated which changes what ports/ISN is picked for a given + // tuple (src ip,src port, dst ip, dst port). This can cause the final + // SYN to use a sequence number that looks like one from the current + // connection in TIME_WAIT and will not be accepted causing the test + // to timeout. + // + // TODO(gvisor.dev/issue/940): S/R portSeed/portHint + DisableSave ds; + sockaddr_storage conn_addr = connector.addr; + ASSERT_NO_ERRNO(SetAddrPort(connector.family(), &conn_addr, port)); + ASSERT_THAT(RetryEINTR(connect)(conn_fd.get(), + reinterpret_cast<sockaddr*>(&conn_addr), + connector.addr_len), + SyscallSucceeds()); + + // Accept the connection. + auto accepted = + ASSERT_NO_ERRNO_AND_VALUE(Accept(listen_fd.get(), nullptr, nullptr)); + + // Get the address/port bound by the connecting socket. + sockaddr_storage conn_bound_addr; + socklen_t conn_addrlen = connector.addr_len; + ASSERT_THAT( + getsockname(conn_fd.get(), reinterpret_cast<sockaddr*>(&conn_bound_addr), + &conn_addrlen), + SyscallSucceeds()); + + // close the accept FD to trigger TIME_WAIT on the accepted socket which + // should cause the conn_fd to follow CLOSE_WAIT->LAST_ACK->CLOSED instead of + // TIME_WAIT. + accepted.reset(); + absl::SleepFor(absl::Seconds(1)); + conn_fd.reset(); + absl::SleepFor(absl::Seconds(1)); + + // Now bind and connect a new socket and verify that we can immediately + // rebind the address bound by the conn_fd as it never entered TIME_WAIT. + const FileDescriptor conn_fd2 = ASSERT_NO_ERRNO_AND_VALUE( + Socket(connector.family(), SOCK_STREAM, IPPROTO_TCP)); + + ASSERT_THAT(bind(conn_fd2.get(), + reinterpret_cast<sockaddr*>(&conn_bound_addr), conn_addrlen), + SyscallSucceeds()); + ASSERT_THAT(RetryEINTR(connect)(conn_fd2.get(), + reinterpret_cast<sockaddr*>(&conn_addr), + conn_addrlen), + SyscallSucceeds()); +} + INSTANTIATE_TEST_SUITE_P( All, SocketInetLoopbackTest, ::testing::Values( @@ -516,6 +853,112 @@ TEST_P(SocketInetReusePortTest, UdpPortReuseMultiThread) { EquivalentWithin((kConnectAttempts / kThreadCount), 0.10)); } +TEST_P(SocketInetReusePortTest, UdpPortReuseMultiThreadShort) { + auto const& param = GetParam(); + + TestAddress const& listener = param.listener; + TestAddress const& connector = param.connector; + sockaddr_storage listen_addr = listener.addr; + sockaddr_storage conn_addr = connector.addr; + constexpr int kThreadCount = 3; + + // TODO(b/141211329): endpointsByNic.seed has to be saved/restored. + const DisableSave ds141211329; + + // Create listening sockets. + FileDescriptor listener_fds[kThreadCount]; + for (int i = 0; i < kThreadCount; i++) { + listener_fds[i] = + ASSERT_NO_ERRNO_AND_VALUE(Socket(listener.family(), SOCK_DGRAM, 0)); + int fd = listener_fds[i].get(); + + ASSERT_THAT(setsockopt(fd, SOL_SOCKET, SO_REUSEPORT, &kSockOptOn, + sizeof(kSockOptOn)), + SyscallSucceeds()); + ASSERT_THAT( + bind(fd, reinterpret_cast<sockaddr*>(&listen_addr), listener.addr_len), + SyscallSucceeds()); + + // On the first bind we need to determine which port was bound. + if (i != 0) { + continue; + } + + // Get the port bound by the listening socket. + socklen_t addrlen = listener.addr_len; + ASSERT_THAT( + getsockname(listener_fds[0].get(), + reinterpret_cast<sockaddr*>(&listen_addr), &addrlen), + SyscallSucceeds()); + uint16_t const port = + ASSERT_NO_ERRNO_AND_VALUE(AddrPort(listener.family(), listen_addr)); + ASSERT_NO_ERRNO(SetAddrPort(listener.family(), &listen_addr, port)); + ASSERT_NO_ERRNO(SetAddrPort(connector.family(), &conn_addr, port)); + } + + constexpr int kConnectAttempts = 10; + FileDescriptor client_fds[kConnectAttempts]; + + // Do the first run without save/restore. + DisableSave ds; + for (int i = 0; i < kConnectAttempts; i++) { + client_fds[i] = + ASSERT_NO_ERRNO_AND_VALUE(Socket(connector.family(), SOCK_DGRAM, 0)); + EXPECT_THAT(RetryEINTR(sendto)(client_fds[i].get(), &i, sizeof(i), 0, + reinterpret_cast<sockaddr*>(&conn_addr), + connector.addr_len), + SyscallSucceedsWithValue(sizeof(i))); + } + ds.reset(); + + // Check that a mapping of client and server sockets has + // not been change after save/restore. + for (int i = 0; i < kConnectAttempts; i++) { + EXPECT_THAT(RetryEINTR(sendto)(client_fds[i].get(), &i, sizeof(i), 0, + reinterpret_cast<sockaddr*>(&conn_addr), + connector.addr_len), + SyscallSucceedsWithValue(sizeof(i))); + } + + int epollfd; + ASSERT_THAT(epollfd = epoll_create1(0), SyscallSucceeds()); + + for (int i = 0; i < kThreadCount; i++) { + int fd = listener_fds[i].get(); + struct epoll_event ev; + ev.data.fd = fd; + ev.events = EPOLLIN; + ASSERT_THAT(epoll_ctl(epollfd, EPOLL_CTL_ADD, fd, &ev), SyscallSucceeds()); + } + + std::map<uint16_t, int> portToFD; + + for (int i = 0; i < kConnectAttempts * 2; i++) { + struct sockaddr_storage addr = {}; + socklen_t addrlen = sizeof(addr); + struct epoll_event ev; + int data, fd; + + ASSERT_THAT(epoll_wait(epollfd, &ev, 1, -1), SyscallSucceedsWithValue(1)); + + fd = ev.data.fd; + EXPECT_THAT(RetryEINTR(recvfrom)(fd, &data, sizeof(data), 0, + reinterpret_cast<struct sockaddr*>(&addr), + &addrlen), + SyscallSucceedsWithValue(sizeof(data))); + uint16_t const port = + ASSERT_NO_ERRNO_AND_VALUE(AddrPort(connector.family(), addr)); + auto prev_port = portToFD.find(port); + // Check that all packets from one client have been delivered to the same + // server socket. + if (prev_port == portToFD.end()) { + portToFD[port] = ev.data.fd; + } else { + EXPECT_EQ(portToFD[port], ev.data.fd); + } + } +} + INSTANTIATE_TEST_SUITE_P( All, SocketInetReusePortTest, ::testing::Values( |