public inbox for passt-dev@passt.top
 help / color / mirror / code / Atom feed
* [PATCH] tcp: Limit fast retransmit to one per loss event
@ 2026-11-19 12:57 Anshu Kumari
  2026-11-19 19:01 ` Stefano Brivio
  0 siblings, 1 reply; 2+ messages in thread
From: Anshu Kumari @ 2026-11-19 12:57 UTC (permalink / raw)
  To: sbrivio, passt-dev, ada.urbaniec+bugzilla; +Cc: anskuma

Every duplicate ACK unconditionally triggered a fast retransmit,
even when the ACK sequence hadn't advanced.

Add a retransmitting flag to track whether we are counting duplicate
ACKs or actively retransmitting. we trigger fast retransmit after
three duplicates ACKs (per RFC 5681, section 3.2). Clear both
retransmitting and retries when the ACK sequence advances.

Link: https://bugs.passt.top/show_bug.cgi?id=224
Signed-off-by: Anshu Kumari <anskuma@redhat.com>
---
 tcp.c      | 33 +++++++++++++++++++++++++++------
 tcp_conn.h |  8 ++++++--
 2 files changed, 33 insertions(+), 8 deletions(-)

diff --git a/tcp.c b/tcp.c
index 95f4f3e1..ff5d34fa 100644
--- a/tcp.c
+++ b/tcp.c
@@ -238,11 +238,12 @@
  *     - on zero read, send FIN to tap/guest, set TAP_FIN_SENT
  *   - on ACK from tap/guest:
  *     - set @ts_ack_from_tap
- *     - check if it's the second duplicated ACK
+ *     - count duplicate ACKs (same ack_seq and window, no data/FIN)
  *     - consume buffer by difference between new ack_seq and @seq_ack_from_tap
  *     - update @seq_ack_from_tap from ack_seq in header
- *     - on two duplicated ACKs, reset @seq_to_tap to @seq_ack_from_tap, and
- *       resend with steps listed above
+ *     - after three duplicate ACKs (RFC 5681), reset @seq_to_tap to
+ *       @seq_ack_from_tap, and resend with steps listed above, once per
+ *       loss event
  *
  * - from tap/guest to socket:
  *   - on packet from tap/guest:
@@ -1264,6 +1265,7 @@ static void tcp_update_seqack_from_tap(const struct ctx *c,
 			conn_flag(c, conn, ACK_FROM_TAP_DUE, now);
 
 		conn->retries = 0;
+		conn->retransmitting = 0;
 		conn->seq_ack_from_tap = seq;
 	}
 }
@@ -2082,7 +2084,13 @@ static int tcp_data_from_tap(const struct ctx *c, struct tcp_tap_conn *conn,
 
 	tcp_tap_window_update(c, conn, max_ack_seq_wnd, now);
 
-	if (retr) {
+	if (retr && !conn->retransmitting)
+		conn->retries++;
+
+	if (retr && !conn->retransmitting && conn->retries > 3) {
+		conn->retransmitting = 1;
+		conn->retries = 0;
+
 		flow_trace(conn,
 			   "fast re-transmit, ACK: %u, previous sequence: %u",
 			   conn->seq_ack_from_tap, conn->seq_to_tap);
@@ -2443,7 +2451,13 @@ int tcp_tap_handler(const struct ctx *c, uint8_t pif, sa_family_t af,
 			tcp_update_seqack_from_tap(c, conn, ntohl(th->ack_seq),
 						   now);
 
-		if (retr) {
+		if (retr && !conn->retransmitting)
+			conn->retries++;
+
+		if (retr && !conn->retransmitting && conn->retries > 3) {
+			conn->retransmitting = 1;
+			conn->retries = 0;
+
 			flow_trace(conn,
 				   "fast re-transmit, ACK: %u, previous sequence: %u",
 				   ntohl(th->ack_seq), conn->seq_to_tap);
@@ -2453,7 +2467,7 @@ int tcp_tap_handler(const struct ctx *c, uint8_t pif, sa_family_t af,
 		}
 
 		if (tcp_tap_window_update(c, conn, ntohs(th->window), now) ||
-		    retr)
+		    (retr && conn->retransmitting))
 			tcp_data_from_sock(c, conn, now);
 
 		if (conn->seq_ack_from_tap == conn->seq_to_tap) {
@@ -2737,6 +2751,11 @@ void tcp_timer_handler(const struct ctx *c, union epoll_ref ref,
 			if (!conn->wnd_from_tap)
 				conn->wnd_from_tap = 1; /* Zero-window probe */
 
+			if (!conn->retransmitting) {
+				conn->retries = 0;
+				conn->retransmitting = 1;
+			}
+
 			conn->retries++;
 			if (tcp_rewind_seq(c, conn, now))
 				return;
@@ -3550,6 +3569,7 @@ int tcp_flow_migrate_source(int fd, struct tcp_tap_conn *conn)
 {
 	struct tcp_tap_transfer t = {
 		.retries		= conn->retries,
+		.retransmitting		= conn->retransmitting,
 		.ws_from_tap		= conn->ws_from_tap,
 		.ws_to_tap		= conn->ws_to_tap,
 		.events			= conn->events,
@@ -3826,6 +3846,7 @@ int tcp_flow_migrate_target(struct ctx *c, int fd)
 	conn = FLOW_SET_TYPE(flow, FLOW_TCP, tcp);
 
 	conn->retries			= t.retries;
+	conn->retransmitting		= t.retransmitting;
 	conn->ws_from_tap		= t.ws_from_tap;
 	conn->ws_to_tap			= t.ws_to_tap;
 	conn->events			= t.events;
diff --git a/tcp_conn.h b/tcp_conn.h
index 3173ddb9..e8e2edf0 100644
--- a/tcp_conn.h
+++ b/tcp_conn.h
@@ -16,12 +16,13 @@
 /**
  * struct tcp_tap_conn - Descriptor for a TCP connection (not spliced)
  * @f:			Generic flow information
- * @retries:		Number of retries occurred due to timeouts
+ * @retries:		Number of retries or duplicate ACKs, see @retransmitting
  * @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
  * @inactive:		No activity within the current INACTIVITY_INTERVAL
+ * @retransmitting:	0 if counting duplicate ACKs, 1 if retransmitting
  * @sock:		Socket descriptor number
  * @events:		Connection events, implying connection states
  * @timer:		timerfd descriptor for timeout events
@@ -72,6 +73,7 @@ struct tcp_tap_conn {
 
 	bool		tap_inactive	:1;
 	bool		inactive	:1;
+	bool		retransmitting	:1;
 
 	int		sock		:FD_REF_BITS;
 
@@ -122,7 +124,8 @@ struct tcp_tap_conn {
  * struct tcp_tap_transfer - Migrated TCP data, flow table part, network order
  * @pif:		Interfaces for each side of the flow
  * @side:		Addresses and ports for each side of the flow
- * @retries:		Number of retries occurred due to timeouts
+ * @retries:		Number of retries / duplicate ACKs
+ * @retransmitting:	0 if counting duplicate ACKs, 1 if retransmitting
  * @ws_from_tap:	Window scaling factor advertised from tap/guest
  * @ws_to_tap:		Window scaling factor advertised to tap/guest
  * @events:		Connection events, implying connection states
@@ -143,6 +146,7 @@ struct tcp_tap_transfer {
 	struct flowside	side[SIDES];
 
 	uint8_t		retries;
+	uint8_t		retransmitting;
 	uint8_t		ws_from_tap;
 	uint8_t		ws_to_tap;
 	uint8_t		events;
-- 
2.55.0


^ permalink raw reply	[flat|nested] 2+ messages in thread

* Re: [PATCH] tcp: Limit fast retransmit to one per loss event
  2026-11-19 12:57 [PATCH] tcp: Limit fast retransmit to one per loss event Anshu Kumari
@ 2026-11-19 19:01 ` Stefano Brivio
  0 siblings, 0 replies; 2+ messages in thread
From: Stefano Brivio @ 2026-11-19 19:01 UTC (permalink / raw)
  To: Anshu Kumari; +Cc: passt-dev, ada.urbaniec+bugzilla

On Thu, 19 Nov 2026 18:27:52 +0530
Anshu Kumari <anskuma@redhat.com> wrote:

> Every duplicate ACK unconditionally triggered a fast retransmit,
> even when the ACK sequence hadn't advanced.
> 
> Add a retransmitting flag to track whether we are counting duplicate
> ACKs or actively retransmitting. we trigger fast retransmit after
> three duplicates ACKs (per RFC 5681, section 3.2). Clear both
> retransmitting and retries when the ACK sequence advances.

Is this change really doing what the title says?  To me it looks like a
mix of two things, but the second one is not complete, so it might
actually break fast retransmit.

That is, limiting the retransmissions to one per retransmit request is
one part, that I was indicating as "quick fix" for #224 (I thought it
was two lines, it turns out to be four), and that looks correct, and
the most critical part of the fix, that we want to have soon.

But the part counting three duplicate ACKs to implement better
compliance to RFC 5681 is just a partial implementation (which I would
leave as a second step, as I was suggesting, because it's definitely
more involved).  The concrete issue with that is:

> Link: https://bugs.passt.top/show_bug.cgi?id=224
> Signed-off-by: Anshu Kumari <anskuma@redhat.com>
> ---
>  tcp.c      | 33 +++++++++++++++++++++++++++------
>  tcp_conn.h |  8 ++++++--
>  2 files changed, 33 insertions(+), 8 deletions(-)
> 
> diff --git a/tcp.c b/tcp.c
> index 95f4f3e1..ff5d34fa 100644
> --- a/tcp.c
> +++ b/tcp.c
> @@ -238,11 +238,12 @@
>   *     - on zero read, send FIN to tap/guest, set TAP_FIN_SENT
>   *   - on ACK from tap/guest:
>   *     - set @ts_ack_from_tap
> - *     - check if it's the second duplicated ACK
> + *     - count duplicate ACKs (same ack_seq and window, no data/FIN)
>   *     - consume buffer by difference between new ack_seq and @seq_ack_from_tap
>   *     - update @seq_ack_from_tap from ack_seq in header
> - *     - on two duplicated ACKs, reset @seq_to_tap to @seq_ack_from_tap, and
> - *       resend with steps listed above
> + *     - after three duplicate ACKs (RFC 5681), reset @seq_to_tap to
> + *       @seq_ack_from_tap, and resend with steps listed above, once per
> + *       loss event
>   *
>   * - from tap/guest to socket:
>   *   - on packet from tap/guest:
> @@ -1264,6 +1265,7 @@ static void tcp_update_seqack_from_tap(const struct ctx *c,
>  			conn_flag(c, conn, ACK_FROM_TAP_DUE, now);
>  
>  		conn->retries = 0;
> +		conn->retransmitting = 0;
>  		conn->seq_ack_from_tap = seq;
>  	}
>  }
> @@ -2082,7 +2084,13 @@ static int tcp_data_from_tap(const struct ctx *c, struct tcp_tap_conn *conn,
>  
>  	tcp_tap_window_update(c, conn, max_ack_seq_wnd, now);
>  
> -	if (retr) {
> +	if (retr && !conn->retransmitting)
> +		conn->retries++;
> +
> +	if (retr && !conn->retransmitting && conn->retries > 3) {
> +		conn->retransmitting = 1;
> +		conn->retries = 0;
> +
>  		flow_trace(conn,
>  			   "fast re-transmit, ACK: %u, previous sequence: %u",
>  			   conn->seq_ack_from_tap, conn->seq_to_tap);
> @@ -2443,7 +2451,13 @@ int tcp_tap_handler(const struct ctx *c, uint8_t pif, sa_family_t af,
>  			tcp_update_seqack_from_tap(c, conn, ntohl(th->ack_seq),
>  						   now);
>  
> -		if (retr) {
> +		if (retr && !conn->retransmitting)
> +			conn->retries++;

...that when we set 'retr' in the loop above, we're typically in the
middle of a batch.  We might have four segments there, with the same
ACK sequence, but 'retr' is a boolean, so it can only convey the fact
that there was at least one duplicate ACK.

A 'true' value won't distinguish between one, two, or three.

If we reach this point, we'll increment conn->retries once, and that's
it.  We might have several duplicate ACKs but never trigger the
retransmission, below.

Tracking 'retr' in the loop is a bit more complicated than this and I'd
leave it to a later stage.

> +
> +		if (retr && !conn->retransmitting && conn->retries > 3) {
> +			conn->retransmitting = 1;
> +			conn->retries = 0;
> +
>  			flow_trace(conn,
>  				   "fast re-transmit, ACK: %u, previous sequence: %u",
>  				   ntohl(th->ack_seq), conn->seq_to_tap);
> @@ -2453,7 +2467,7 @@ int tcp_tap_handler(const struct ctx *c, uint8_t pif, sa_family_t af,
>  		}
>  
>  		if (tcp_tap_window_update(c, conn, ntohs(th->window), now) ||
> -		    retr)
> +		    (retr && conn->retransmitting))
>  			tcp_data_from_sock(c, conn, now);
>  
>  		if (conn->seq_ack_from_tap == conn->seq_to_tap) {
> @@ -2737,6 +2751,11 @@ void tcp_timer_handler(const struct ctx *c, union epoll_ref ref,
>  			if (!conn->wnd_from_tap)
>  				conn->wnd_from_tap = 1; /* Zero-window probe */
>  
> +			if (!conn->retransmitting) {
> +				conn->retries = 0;
> +				conn->retransmitting = 1;
> +			}
> +
>  			conn->retries++;
>  			if (tcp_rewind_seq(c, conn, now))
>  				return;
> @@ -3550,6 +3569,7 @@ int tcp_flow_migrate_source(int fd, struct tcp_tap_conn *conn)
>  {
>  	struct tcp_tap_transfer t = {
>  		.retries		= conn->retries,
> +		.retransmitting		= conn->retransmitting,
>  		.ws_from_tap		= conn->ws_from_tap,
>  		.ws_to_tap		= conn->ws_to_tap,
>  		.events			= conn->events,
> @@ -3826,6 +3846,7 @@ int tcp_flow_migrate_target(struct ctx *c, int fd)
>  	conn = FLOW_SET_TYPE(flow, FLOW_TCP, tcp);
>  
>  	conn->retries			= t.retries;
> +	conn->retransmitting		= t.retransmitting;
>  	conn->ws_from_tap		= t.ws_from_tap;
>  	conn->ws_to_tap			= t.ws_to_tap;
>  	conn->events			= t.events;
> diff --git a/tcp_conn.h b/tcp_conn.h
> index 3173ddb9..e8e2edf0 100644
> --- a/tcp_conn.h
> +++ b/tcp_conn.h
> @@ -16,12 +16,13 @@
>  /**
>   * struct tcp_tap_conn - Descriptor for a TCP connection (not spliced)
>   * @f:			Generic flow information
> - * @retries:		Number of retries occurred due to timeouts
> + * @retries:		Number of retries or duplicate ACKs, see @retransmitting
>   * @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
>   * @inactive:		No activity within the current INACTIVITY_INTERVAL
> + * @retransmitting:	0 if counting duplicate ACKs, 1 if retransmitting
>   * @sock:		Socket descriptor number
>   * @events:		Connection events, implying connection states
>   * @timer:		timerfd descriptor for timeout events
> @@ -72,6 +73,7 @@ struct tcp_tap_conn {
>  
>  	bool		tap_inactive	:1;
>  	bool		inactive	:1;
> +	bool		retransmitting	:1;
>  
>  	int		sock		:FD_REF_BITS;
>  
> @@ -122,7 +124,8 @@ struct tcp_tap_conn {
>   * struct tcp_tap_transfer - Migrated TCP data, flow table part, network order
>   * @pif:		Interfaces for each side of the flow
>   * @side:		Addresses and ports for each side of the flow
> - * @retries:		Number of retries occurred due to timeouts
> + * @retries:		Number of retries / duplicate ACKs
> + * @retransmitting:	0 if counting duplicate ACKs, 1 if retransmitting
>   * @ws_from_tap:	Window scaling factor advertised from tap/guest
>   * @ws_to_tap:		Window scaling factor advertised to tap/guest
>   * @events:		Connection events, implying connection states
> @@ -143,6 +146,7 @@ struct tcp_tap_transfer {
>  	struct flowside	side[SIDES];
>  
>  	uint8_t		retries;

I would suggest to use a union here, so that this can be called
something like 'dup_acks' if it's counting duplicate ACKs, and
'retries' if it's counting retries.

> +	uint8_t		retransmitting;
>  	uint8_t		ws_from_tap;
>  	uint8_t		ws_to_tap;
>  	uint8_t		events;

-- 
Stefano


^ permalink raw reply	[flat|nested] 2+ messages in thread

end of thread, other threads:[~2026-11-19 19:01 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2026-11-19 12:57 [PATCH] tcp: Limit fast retransmit to one per loss event Anshu Kumari
2026-11-19 19:01 ` Stefano Brivio

Code repositories for project(s) associated with this public inbox

	https://passt.top/passt

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for IMAP folder(s).