On Thu, Feb 05, 2026 at 01:17:45AM +0100, Stefano Brivio wrote:

On Wed, 4 Feb 2026 21:41:37 +1000 David Gibson wrote:

...

There are several circumstances in which a live, but idle TCP connection can be forgotten by a guest, with no "on the wire" indication that this has happened. The most obvious is if the guest abruptly reboots. A more subtle case can happen with a half-closed connection, specifically one in FIN_WAIT_2 state on the guest. A connection can, legitimately, remain in this state indefinitely. If however, a socket in this state is closed by userspace, Linux at least will remove the kernel socket after 60s (or as configured in the net.ipv4.tcp_fin_timeout sysctl).

Because there's no on the wire indication in these cases, passt will pointlessly retain the connection in its flow table, at least until it is removed by the inactivity timeout after several hours.

To avoid keeping connections around for so long in this state, add functionality to periodically send TCP keepalive segments to the guest if we've seen no activity on the tap interface. If the guest is no longer aware of the connection, it should respond with an RST which will let passt remove the stale entry.

To do this we use a method similar to the inactivity timeout - a 1-bit page replacement / clock algorithm, but with a shorter interval, and only checking for tap side activity. Currently we use a 300s interval, meaning we'll send a keepalive after 5-10 minutes of (tap side) inactivity.

Link: https://bugs.passt.top/show_bug.cgi?id=179

Signed-off-by: David Gibson --- tcp.c | 39 +++++++++++++++++++++++++++++++++++++++ tcp.h | 2 ++ tcp_conn.h | 2 ++ 3 files changed, 43 insertions(+)

diff --git a/tcp.c b/tcp.c index acdac7df..bf57be23 100644 --- a/tcp.c +++ b/tcp.c @@ -206,6 +206,12 @@ * keepalives) will be removed between INACTIVITY_INTERVAL s and * 2*INACTIVITY_INTERVAL s after the last activity. * + * - KEEPALIVE_INTERVAL: if a connection has had no tap-side activity for an + * entire interval, send a tap-side keepalive. If the endpoint is no longer + * aware of the connection (due to a reboot, or a kernel timeout in FIN_WAIT_2 + * state) that should trigger an RST, so we won't keep track of connections + * that the guest endpoint no longer cares about. + * * Summary of data flows (with ESTABLISHED event) * ---------------------------------------------- * @@ -342,6 +348,7 @@ enum { #define RTO_INIT_AFTER_SYN_RETRIES 3 /* s, RFC 6298 */

#define INACTIVITY_INTERVAL 7200 /* s */ +#define KEEPALIVE_INTERVAL 30 /* s */

By the way, once we're done testing this, I'm not sure which value to use. MUST-28 in RFC 9293 (3.8.4. TCP Keep-Alives) says we should "default to no less than two hours". It looks a bit too long to be useful, to me.

I agree, and as we've discussed a bit, the purpose of these keepalives differs a bit from more common reasons for sending TCP keepalives, meaning we have different considerations for the interval. Also note that while RFC 9293 is pretty recent, the "no less than two hours" is unchanged since RFC 1122, from 1989. In 1989, keepalives every few minutes might have been a significant bandwidth concern, now, not so much. Especially since these are tap side only, which usually be a local high-speed connections (it only won't be if the guest routes to a VPN or something). I was thinking of going with a 300s interval, so timeout would happen after 5-10 minutes. That has the worse case timeout being just one (base 10) order of magnitude away from the default 60s value of net.ipv4.tcp_fin_timeout. I could be persuaded to increase the interval to 900s (15-30m timeout). Much longer than that and I don't think it would really accomplish the purpose we're looking for in bug179 - we'd still have stale connections hanging around for an hour or more.

...

#define LOW_RTT_TABLE_SIZE 8 #define LOW_RTT_THRESHOLD 10 /* us */ @@ -2263,6 +2270,7 @@ int tcp_tap_handler(const struct ctx *c, uint8_t pif, sa_family_t af, }

conn->inactive = false; + conn->tap_inactive = false;

if (th->ack && !(conn->events & ESTABLISHED)) tcp_update_seqack_from_tap(c, conn, ntohl(th->ack_seq)); @@ -2884,6 +2892,36 @@ int tcp_init(struct ctx *c) return 0; }

+/** + * tcp_keepalive() - Send keepalives for connections which need it + * @: Execution context

* @c: Execution context * @now: Current timestamp

Fixed.

...

+ */ +static void tcp_keepalive(struct ctx *c, const struct timespec *now) +{ + union flow *flow; + + if (now->tv_sec - c->tcp.keepalive_run < KEEPALIVE_INTERVAL) + return; + + c->tcp.keepalive_run = now->tv_sec; + + flow_foreach(flow) { + struct tcp_tap_conn *conn = &flow->tcp; + + if (flow->f.type != FLOW_TCP) + continue; + + if (conn->tap_inactive) { + flow_dbg(conn, "No tap activity for least %us, send keepalive",

s/least// (or "for at least")

Oops. "for at least" was what I intended, since that's more accurate.

...

+ KEEPALIVE_INTERVAL); + tcp_send_flag(c, conn, KEEPALIVE);

Do we have to make sure we don't interpret replies to keepalives as duplicate ACKs (by setting some internal flow flag)? I haven't really checked that path.

That's... an interesting question. I think the answer is "not in a way that matters". If we get here, we haven't heard anything from the guest for at least 5 minutes (or whatever). That implies either: a) it's already acked everything we sent. So, even if we see the keepalive reply as a dup ack, there's nothing to retransmit, so it's harmless. or b) it's not listening at all, in which case it's unlikely to respond to the keepalive either or c) we're already retransmitting, and have backed off to pretty long timeouts (greater than the keepalive interval), in which case a dup-ack is just a bit more data for us

...

+ } + + /* Ready to check fot next interval */

for

Copypasta. Fixed.

...

+ conn->tap_inactive = true; + } +} + /** * tcp_inactivity() - Scan for and close long-inactive connections * @: Execution context @@ -2927,6 +2965,7 @@ void tcp_timer(struct ctx *c, const struct timespec *now) if (c->mode == MODE_PASTA) tcp_splice_refill(c);

+ tcp_keepalive(c, now); tcp_inactivity(c, now); }

diff --git a/tcp.h b/tcp.h index e104d453..2739f309 100644 --- a/tcp.h +++ b/tcp.h @@ -38,6 +38,7 @@ extern bool peek_offset_cap; * @rto_max: Maximum retry timeout (in s) * @syn_retries: SYN retries using exponential backoff timeout * @syn_linear_timeouts: SYN retries before using exponential backoff timeout + * @keepalive_run: Time we last issued tap-side keepalives * @inactivity_run: Time we last scanned for inactive connections */ struct tcp_ctx { @@ -48,6 +49,7 @@ struct tcp_ctx { int rto_max; uint8_t syn_retries; uint8_t syn_linear_timeouts; + time_t keepalive_run; time_t inactivity_run; };

diff --git a/tcp_conn.h b/tcp_conn.h index 7197ff63..c82e1441 100644 --- a/tcp_conn.h +++ b/tcp_conn.h @@ -16,6 +16,7 @@ * @ws_from_tap: Window scaling factor advertised from tap/guest * @ws_to_tap: Window scaling factor advertised to tap/guest * @tap_mss: MSS advertised by tap/guest, rounded to 2 ^ TCP_MSS_BITS + * @tapinactive: No tao activity within the current KEEPALIVE_INTERVAL

tap_inactive, tap activity

The typos keep coming. I'm going to blame the laptop keyboard and an uncomfortable public library chair :).

...

* @inactive: No activity within the current INACTIVITY_INTERVAL * @sock: Socket descriptor number * @events: Connection events, implying connection states @@ -58,6 +59,7 @@ struct tcp_tap_conn { (conn->rtt_exp = MIN(RTT_EXP_MAX, ilog2(MAX(1, rtt / RTT_STORE_MIN)))) #define RTT_GET(conn) (RTT_STORE_MIN << conn->rtt_exp)

+ bool tap_inactive :1; bool inactive :1;

int sock :FD_REF_BITS;

The rest looks all good to me, but I didn't test it at all.

-- Stefano

-- David Gibson (he or they) | I'll have my music baroque, and my code david AT gibson.dropbear.id.au | minimalist, thank you, not the other way | around. http://www.ozlabs.org/~dgibson