[PATCH v2 0/2] Fix reconnection latency caused by FIN/ACK handling race

From: SeongJae Park <redacted>

When closing a connection, the two acks that required to change closing
socket's status to FIN_WAIT_2 and then TIME_WAIT could be processed in
reverse order.  This is possible in RSS disabled environments such as a
connection inside a host.

For example, expected state transitions and required packets for the
disconnection will be similar to below flow.

	 00 (Process A)				(Process B)
	 01 ESTABLISHED				ESTABLISHED
	 02 close()
	 03 FIN_WAIT_1
	 04 		---FIN-->
	 05 					CLOSE_WAIT
	 06 		<--ACK---
	 07 FIN_WAIT_2
	 08 		<--FIN/ACK---
	 09 TIME_WAIT
	 10 		---ACK-->
	 11 					LAST_ACK
	 12 CLOSED				CLOSED

In some cases such as LINGER option applied socket, the FIN and FIN/ACK will be
substituted to RST and RST/ACK, but there is no difference in the main logic.

The acks in lines 6 and 8 are the acks.  If the line 8 packet is
processed before the line 6 packet, it will be just ignored as it is not
a expected packet, and the later process of the line 6 packet will
change the status of Process A to FIN_WAIT_2, but as it has already
handled line 8 packet, it will not go to TIME_WAIT and thus will not
send the line 10 packet to Process B.  Thus, Process B will left in
CLOSE_WAIT status, as below.

	 00 (Process A)				(Process B)
	 01 ESTABLISHED				ESTABLISHED
	 02 close()
	 03 FIN_WAIT_1
	 04 		---FIN-->
	 05 					CLOSE_WAIT
	 06 				(<--ACK---)
	 07	  			(<--FIN/ACK---)
	 08 				(fired in right order)
	 09 		<--FIN/ACK---
	 10 		<--ACK---
	 11 		(processed in reverse order)
	 12 FIN_WAIT_2

Later, if the Process B sends SYN to Process A for reconnection using
the same port, Process A will responds with an ACK for the last flow,
which has no increased sequence number.  Thus, Process A will send RST,
wait for TIMEOUT_INIT (one second in default), and then try
reconnection.  If reconnections are frequent, the one second latency
spikes can be a big problem.  Below is a tcpdump results of the problem:

    14.436259 IP 127.0.0.1.45150 > 127.0.0.1.4242: Flags [S], seq 2560603644
    14.436266 IP 127.0.0.1.4242 > 127.0.0.1.45150: Flags [.], ack 5, win 512
    14.436271 IP 127.0.0.1.45150 > 127.0.0.1.4242: Flags [R], seq 2541101298
    /* ONE SECOND DELAY */
    15.464613 IP 127.0.0.1.45150 > 127.0.0.1.4242: Flags [S], seq 2560603644

Patchset Organization
---------------------

The first patch fixes the problem by adjusting the first resend delay of
the SYN in the case.  The second one adds a user space test to reproduce
this problem.

The patches are based on the v5.5.  You can also clone the complete git
tree:

    $ git clone git://github.com/sjp38/linux -b patches/finack_lat/v2

The web is also available:
https://github.com/sjp38/linux/tree/patches/finack_lat/v2


Patchset History
----------------

From v1
(https://lore.kernel.org/linux-kselftest/20200131122421.23286-1-sjpark@amazon.com/)
 - Drop the trivial comment fix patch (Eric Dumazet)
 - Limit the delay adjustment to only the first SYN resend (Eric Dumazet)
 - selftest: Avoid use of hard-coded port number (Eric Dumazet)
 - Explain RST/ACK and FIN/ACK has no big difference (Neal Cardwell)

SeongJae Park (2):
  tcp: Reduce SYN resend delay if a suspicous ACK is received
  selftests: net: Add FIN_ACK processing order related latency spike
    test

 net/ipv4/tcp_input.c                       |   8 +-
 tools/testing/selftests/net/.gitignore     |   2 +
 tools/testing/selftests/net/Makefile       |   2 +
 tools/testing/selftests/net/fin_ack_lat.c  | 151 +++++++++++++++++++++
 tools/testing/selftests/net/fin_ack_lat.sh |  35 +++++
 5 files changed, 197 insertions(+), 1 deletion(-)
 create mode 100644 tools/testing/selftests/net/fin_ack_lat.c
 create mode 100755 tools/testing/selftests/net/fin_ack_lat.sh

-- 
2.17.1

From: SeongJae Park <redacted>

When closing a connection, the two acks that required to change closing
socket's status to FIN_WAIT_2 and then TIME_WAIT could be processed in
reverse order.  This is possible in RSS disabled environments such as a
connection inside a host.

For example, expected state transitions and required packets for the
disconnection will be similar to below flow.

	 00 (Process A)				(Process B)
	 01 ESTABLISHED				ESTABLISHED
	 02 close()
	 03 FIN_WAIT_1
	 04 		---FIN-->
	 05 					CLOSE_WAIT
	 06 		<--ACK---
	 07 FIN_WAIT_2
	 08 		<--FIN/ACK---
	 09 TIME_WAIT
	 10 		---ACK-->
	 11 					LAST_ACK
	 12 CLOSED				CLOSED

In some cases such as LINGER option applied socket, the FIN and FIN/ACK
will be substituted to RST and RST/ACK, but there is no difference in
the main logic.

The acks in lines 6 and 8 are the acks.  If the line 8 packet is
processed before the line 6 packet, it will be just ignored as it is not
a expected packet, and the later process of the line 6 packet will
change the status of Process A to FIN_WAIT_2, but as it has already
handled line 8 packet, it will not go to TIME_WAIT and thus will not
send the line 10 packet to Process B.  Thus, Process B will left in
CLOSE_WAIT status, as below.

	 00 (Process A)				(Process B)
	 01 ESTABLISHED				ESTABLISHED
	 02 close()
	 03 FIN_WAIT_1
	 04 		---FIN-->
	 05 					CLOSE_WAIT
	 06 				(<--ACK---)
	 07	  			(<--FIN/ACK---)
	 08 				(fired in right order)
	 09 		<--FIN/ACK---
	 10 		<--ACK---
	 11 		(processed in reverse order)
	 12 FIN_WAIT_2

Later, if the Process B sends SYN to Process A for reconnection using
the same port, Process A will responds with an ACK for the last flow,
which has no increased sequence number.  Thus, Process A will send RST,
wait for TIMEOUT_INIT (one second in default), and then try
reconnection.  If reconnections are frequent, the one second latency
spikes can be a big problem.  Below is a tcpdump results of the problem:

    14.436259 IP 127.0.0.1.45150 > 127.0.0.1.4242: Flags [S], seq 2560603644
    14.436266 IP 127.0.0.1.4242 > 127.0.0.1.45150: Flags [.], ack 5, win 512
    14.436271 IP 127.0.0.1.45150 > 127.0.0.1.4242: Flags [R], seq 2541101298
    /* ONE SECOND DELAY */
    15.464613 IP 127.0.0.1.45150 > 127.0.0.1.4242: Flags [S], seq 2560603644

This commit mitigates the problem by reducing the delay for the next SYN
if the suspicous ACK is received while in SYN_SENT state.

Following commit will add a selftest, which can be also helpful for
understanding of this issue.

Signed-off-by: SeongJae Park <redacted>
---
 net/ipv4/tcp_input.c | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c
index 2a976f57f7e7..980bd04b9d95 100644
--- a/net/ipv4/tcp_input.c
+++ b/net/ipv4/tcp_input.c

@@ -5893,8 +5893,14 @@ static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb,
 		 *        the segment and return)"
 		 */
 		if (!after(TCP_SKB_CB(skb)->ack_seq, tp->snd_una) ||
-		    after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt))
+		    after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) {
+			/* Previous FIN/ACK or RST/ACK might be ignored. */
+			if (icsk->icsk_retransmits == 0)
+				inet_csk_reset_xmit_timer(sk,
+						ICSK_TIME_RETRANS, TCP_ATO_MIN,
+						TCP_RTO_MAX);
 			goto reset_and_undo;
+		}
 
 		if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr &&
 		    !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp,

-- 
2.17.1

From: SeongJae Park <redacted>

This commit adds a test for FIN_ACK process races related reconnection
latency spike issues.  The issue has described and solved by the
previous commit ("tcp: Reduce SYN resend delay if a suspicous ACK is
received").

The test program is configured with a server and a client process.  The
server creates and binds a socket to a port that dynamically allocated,
listen on it, and start a infinite loop.  Inside the loop, it accepts
connection, reads 4 bytes from the socket, and closes the connection.
The client is constructed as an infinite loop.  Inside the loop, it
creates a socket with LINGER and NODELAY option, connect to the server,
send 4 bytes data, try read some data from server.  After the read()
returns, it measure the latency from the beginning of this loop to this
point and if the latency is larger than 1 second (spike), print a
message.

Signed-off-by: SeongJae Park <redacted>
---
 tools/testing/selftests/net/.gitignore     |   2 +
 tools/testing/selftests/net/Makefile       |   2 +
 tools/testing/selftests/net/fin_ack_lat.c  | 151 +++++++++++++++++++++
 tools/testing/selftests/net/fin_ack_lat.sh |  35 +++++
 4 files changed, 190 insertions(+)
 create mode 100644 tools/testing/selftests/net/fin_ack_lat.c
 create mode 100755 tools/testing/selftests/net/fin_ack_lat.sh

diff --git a/tools/testing/selftests/net/.gitignore b/tools/testing/selftests/net/.gitignore
index 8aefd81fbc86..1bcf7b5498dd 100644
--- a/tools/testing/selftests/net/.gitignore
+++ b/tools/testing/selftests/net/.gitignore

@@ -22,3 +22,5 @@ ipv6_flowlabel_mgr
 so_txtime
 tcp_fastopen_backup_key
 nettest
+fin_ack_lat_accept
+fin_ack_lat_connect

diff --git a/tools/testing/selftests/net/Makefile b/tools/testing/selftests/net/Makefile
index a8e04d665b69..b5694196430a 100644
--- a/tools/testing/selftests/net/Makefile
+++ b/tools/testing/selftests/net/Makefile

@@ -11,6 +11,7 @@ TEST_PROGS += udpgso_bench.sh fib_rule_tests.sh msg_zerocopy.sh psock_snd.sh
 TEST_PROGS += udpgro_bench.sh udpgro.sh test_vxlan_under_vrf.sh reuseport_addr_any.sh
 TEST_PROGS += test_vxlan_fdb_changelink.sh so_txtime.sh ipv6_flowlabel.sh
 TEST_PROGS += tcp_fastopen_backup_key.sh fcnal-test.sh l2tp.sh traceroute.sh
+TEST_PROGS += fin_ack_lat.sh
 TEST_PROGS_EXTENDED := in_netns.sh
 TEST_GEN_FILES =  socket nettest
 TEST_GEN_FILES += psock_fanout psock_tpacket msg_zerocopy reuseport_addr_any

@@ -18,6 +19,7 @@ TEST_GEN_FILES += tcp_mmap tcp_inq psock_snd txring_overwrite
 TEST_GEN_FILES += udpgso udpgso_bench_tx udpgso_bench_rx ip_defrag
 TEST_GEN_FILES += so_txtime ipv6_flowlabel ipv6_flowlabel_mgr
 TEST_GEN_FILES += tcp_fastopen_backup_key
+TEST_GEN_FILES += fin_ack_lat
 TEST_GEN_PROGS = reuseport_bpf reuseport_bpf_cpu reuseport_bpf_numa
 TEST_GEN_PROGS += reuseport_dualstack reuseaddr_conflict tls
 

diff --git a/tools/testing/selftests/net/fin_ack_lat.c b/tools/testing/selftests/net/fin_ack_lat.c
new file mode 100644
index 000000000000..70187494b57a
--- /dev/null
+++ b/tools/testing/selftests/net/fin_ack_lat.c

@@ -0,0 +1,151 @@
+// SPDX-License-Identifier: GPL-2.0
+
+#include <arpa/inet.h>
+#include <errno.h>
+#include <error.h>
+#include <netinet/in.h>
+#include <netinet/tcp.h>
+#include <signal.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <sys/socket.h>
+#include <sys/time.h>
+#include <unistd.h>
+
+static int child_pid;
+
+static unsigned long timediff(struct timeval s, struct timeval e)
+{
+	unsigned long s_us, e_us;
+
+	s_us = s.tv_sec * 1000000 + s.tv_usec;
+	e_us = e.tv_sec * 1000000 + e.tv_usec;
+	if (s_us > e_us)
+		return 0;
+	return e_us - s_us;
+}
+
+static void client(int port)
+{
+	int sock = 0;
+	struct sockaddr_in addr, laddr;
+	socklen_t len = sizeof(laddr);
+	struct linger sl;
+	int flag = 1;
+	int buffer;
+	struct timeval start, end;
+	unsigned long lat, sum_lat = 0, nr_lat = 0;
+
+	while (1) {
+		gettimeofday(&start, NULL);
+
+		sock = socket(AF_INET, SOCK_STREAM, 0);
+		if (sock < 0)
+			error(-1, errno, "socket creation");
+
+		sl.l_onoff = 1;
+		sl.l_linger = 0;
+		if (setsockopt(sock, SOL_SOCKET, SO_LINGER, &sl, sizeof(sl)))
+			error(-1, errno, "setsockopt(linger)");
+
+		if (setsockopt(sock, IPPROTO_TCP, TCP_NODELAY,
+					&flag, sizeof(flag)))
+			error(-1, errno, "setsockopt(nodelay)");
+
+		addr.sin_family = AF_INET;
+		addr.sin_port = htons(port);
+
+		if (inet_pton(AF_INET, "127.0.0.1", &addr.sin_addr) <= 0)
+			error(-1, errno, "inet_pton");
+
+		if (connect(sock, (struct sockaddr *)&addr, sizeof(addr)) < 0)
+			error(-1, errno, "connect");
+
+		send(sock, &buffer, sizeof(buffer), 0);
+		if (read(sock, &buffer, sizeof(buffer)) == -1)
+			error(-1, errno, "waiting read");
+
+		gettimeofday(&end, NULL);
+		lat = timediff(start, end);
+		sum_lat += lat;
+		nr_lat++;
+		if (lat < 100000)
+			goto close;
+
+		if (getsockname(sock, (struct sockaddr *)&laddr, &len) == -1)
+			error(-1, errno, "getsockname");
+		printf("port: %d, lat: %lu, avg: %lu, nr: %lu\n",
+				ntohs(laddr.sin_port), lat,
+				sum_lat / nr_lat, nr_lat);
+close:
+		fflush(stdout);
+		close(sock);
+	}
+}
+
+static void server(int sock, struct sockaddr_in address)
+{
+	int accepted;
+	int addrlen = sizeof(address);
+	int buffer;
+
+	while (1) {
+		accepted = accept(sock, (struct sockaddr *)&address,
+				(socklen_t *)&addrlen);
+		if (accepted < 0)
+			error(-1, errno, "accept");
+
+		if (read(accepted, &buffer, sizeof(buffer)) == -1)
+			error(-1, errno, "read");
+		close(accepted);
+	}
+}
+
+static void sig_handler(int signum)
+{
+	kill(SIGTERM, child_pid);
+	exit(0);
+}
+
+int main(int argc, char const *argv[])
+{
+	int sock;
+	int opt = 1;
+	struct sockaddr_in address;
+	struct sockaddr_in laddr;
+	socklen_t len = sizeof(laddr);
+
+	if (signal(SIGTERM, sig_handler) == SIG_ERR)
+		error(-1, errno, "signal");
+
+	sock = socket(AF_INET, SOCK_STREAM, 0);
+	if (sock < 0)
+		error(-1, errno, "socket");
+
+	if (setsockopt(sock, SOL_SOCKET, SO_REUSEADDR | SO_REUSEPORT,
+				&opt, sizeof(opt)) == -1)
+		error(-1, errno, "setsockopt");
+
+	address.sin_family = AF_INET;
+	address.sin_addr.s_addr = INADDR_ANY;
+	/* dynamically allocate unused port */
+	address.sin_port = 0;
+
+	if (bind(sock, (struct sockaddr *)&address, sizeof(address)) < 0)
+		error(-1, errno, "bind");
+
+	if (listen(sock, 3) < 0)
+		error(-1, errno, "listen");
+
+	if (getsockname(sock, (struct sockaddr *)&laddr, &len) == -1)
+		error(-1, errno, "getsockname");
+
+	fprintf(stderr, "server port: %d\n", ntohs(laddr.sin_port));
+	child_pid = fork();
+	if (!child_pid)
+		client(ntohs(laddr.sin_port));
+	else
+		server(sock, laddr);
+
+	return 0;
+}

diff --git a/tools/testing/selftests/net/fin_ack_lat.sh b/tools/testing/selftests/net/fin_ack_lat.sh
new file mode 100755
index 000000000000..a3ff6e0b2c7a
--- /dev/null
+++ b/tools/testing/selftests/net/fin_ack_lat.sh

@@ -0,0 +1,35 @@
+#!/bin/bash
+# SPDX-License-Identifier: GPL-2.0
+#
+# Test latency spikes caused by FIN/ACK handling race.
+
+set +x
+set -e
+
+tmpfile=$(mktemp /tmp/fin_ack_latency.XXXX.log)
+
+cleanup() {
+	kill $(pidof fin_ack_lat)
+	rm -f $tmpfile
+}
+
+trap cleanup EXIT
+
+do_test() {
+	RUNTIME=$1
+
+	./fin_ack_lat | tee $tmpfile &
+	PID=$!
+
+	sleep $RUNTIME
+	NR_SPIKES=$(wc -l $tmpfile | awk '{print $1}')
+	if [ $NR_SPIKES -gt 0 ]
+	then
+		echo "FAIL: $NR_SPIKES spikes detected"
+		return 1
+	fi
+	return 0
+}
+
+do_test "30"
+echo "test done"

-- 
2.17.1

On Sat, Feb 1, 2020 at 2:19 AM [off-list ref] wrote:

Scheduling a timer for TCP_ATO_MIN, typically 40ms, sounds like it
might be a bit on the slow side. How about TCP_TIMEOUT_MIN, which is
typically 2ms on a HZ=1000 kernel?

I think this would be closer to what Eric mentioned: "sending the SYN
a few ms after the RST seems way better than waiting 1 second as if we
received no packet at all."

neal

On 1/31/20 11:18 PM, sj38.park@gmail.com wrote:

From: SeongJae Park <redacted>

This commit adds a test for FIN_ACK process races related reconnection
latency spike issues.  The issue has described and solved by the
previous commit ("tcp: Reduce SYN resend delay if a suspicous ACK is
received").

The test program is configured with a server and a client process.  The
server creates and binds a socket to a port that dynamically allocated,
listen on it, and start a infinite loop.  Inside the loop, it accepts
connection, reads 4 bytes from the socket, and closes the connection.
The client is constructed as an infinite loop.  Inside the loop, it
creates a socket with LINGER and NODELAY option, connect to the server,
send 4 bytes data, try read some data from server.  After the read()
returns, it measure the latency from the beginning of this loop to this
point and if the latency is larger than 1 second (spike), print a
message.

Signed-off-by: SeongJae Park <redacted>

Reviewed-by: Eric Dumazet <edumazet@google.com>