public inbox for passt-dev@passt.top
 help / color / mirror / code / Atom feed
* [PATCH v3] tcp: Limit fast retransmit to one per loss event
@ 2026-11-20 17:37 Anshu Kumari
  2026-11-20 19:48 ` Jon Maloy
  2026-11-20 20:18 ` Stefano Brivio
  0 siblings, 2 replies; 4+ messages in thread
From: Anshu Kumari @ 2026-11-20 17:37 UTC (permalink / raw)
  To: passt-dev, sbrivio, ada.urbaniec+bugzilla; +Cc: anskuma

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

Guard fast retransmit on !conn->retries, and increment retries
after the first retransmit. This is a quick fix to stop the
amplification storm. Full compliance with RFC 5681 Section 3.2
(three duplicate ACKs) is left to a later patch.

Link: https://bugs.passt.top/show_bug.cgi?id=224
Signed-off-by: Anshu Kumari <anskuma@redhat.com>
---
 tcp.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/tcp.c b/tcp.c
index 95f4f3e1..d3d65f76 100644
--- a/tcp.c
+++ b/tcp.c
@@ -2082,7 +2082,9 @@ 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->retries) {
+		conn->retries++;
+
 		flow_trace(conn,
 			   "fast re-transmit, ACK: %u, previous sequence: %u",
 			   conn->seq_ack_from_tap, conn->seq_to_tap);
@@ -2443,7 +2445,9 @@ 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->retries) {
+			conn->retries++;
+
 			flow_trace(conn,
 				   "fast re-transmit, ACK: %u, previous sequence: %u",
 				   ntohl(th->ack_seq), conn->seq_to_tap);
-- 
2.55.0


^ permalink raw reply	[flat|nested] 4+ messages in thread

* Re: [PATCH v3] tcp: Limit fast retransmit to one per loss event
  2026-11-20 17:37 [PATCH v3] tcp: Limit fast retransmit to one per loss event Anshu Kumari
@ 2026-11-20 19:48 ` Jon Maloy
  2026-11-20 20:18 ` Stefano Brivio
  1 sibling, 0 replies; 4+ messages in thread
From: Jon Maloy @ 2026-11-20 19:48 UTC (permalink / raw)
  To: passt-dev



On 2026-11-20 12:37, Anshu Kumari wrote:
> Every duplicate ACK unconditionally triggered a fast retransmit,
> even when the ACK sequence hadn't advanced.
> 
> Guard fast retransmit on !conn->retries, and increment retries
> after the first retransmit. This is a quick fix to stop the
> amplification storm. Full compliance with RFC 5681 Section 3.2
> (three duplicate ACKs) is left to a later patch.
> 
> Link: https://bugs.passt.top/show_bug.cgi?id=224
> Signed-off-by: Anshu Kumari <anskuma@redhat.com>
> ---
>   tcp.c | 8 ++++++--
>   1 file changed, 6 insertions(+), 2 deletions(-)
> 
> diff --git a/tcp.c b/tcp.c
> index 95f4f3e1..d3d65f76 100644
> --- a/tcp.c
> +++ b/tcp.c
> @@ -2082,7 +2082,9 @@ 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->retries) {
> +		conn->retries++;
> +
>   		flow_trace(conn,
>   			   "fast re-transmit, ACK: %u, previous sequence: %u",
>   			   conn->seq_ack_from_tap, conn->seq_to_tap);
> @@ -2443,7 +2445,9 @@ 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->retries) {
> +			conn->retries++;
> +
>   			flow_trace(conn,
>   				   "fast re-transmit, ACK: %u, previous sequence: %u",
>   				   ntohl(th->ack_seq), conn->seq_to_tap);

This looks correct to me, and matches what Stefano suggested for v2.
Two minor things:

- Missing changelog for v3 below the '---' line.

- In the TAP_FIN_RCVD path, tcp_data_from_sock() is still called if
   'retr' is set but we skip the retransmission because conn->retries
   is non-zero. It's probably harmless (no rewind, so we just send
   what the window allows), but it might be cleaner to fold the
   !conn->retries check into 'retr' there, so that both uses agree.

It might also be worth mentioning in the commit message that, as we
now account a fast retransmit in conn->retries, the next RTO is
doubled (tcp_timer_ctl() uses it as backoff exponent), and the
retransmission counts towards TCP_MAX_RETRIES.

With or without those changes:

Reviewed-by: Jon Maloy <jmaloy@redhat.com>


^ permalink raw reply	[flat|nested] 4+ messages in thread

* Re: [PATCH v3] tcp: Limit fast retransmit to one per loss event
  2026-11-20 17:37 [PATCH v3] tcp: Limit fast retransmit to one per loss event Anshu Kumari
  2026-11-20 19:48 ` Jon Maloy
@ 2026-11-20 20:18 ` Stefano Brivio
       [not found]   ` <CAA7t2yTP5oGNbyv4B3jt36buwTo2+_61OQoSRkE5m97xxdckzQ@mail.gmail.com>
  1 sibling, 1 reply; 4+ messages in thread
From: Stefano Brivio @ 2026-11-20 20:18 UTC (permalink / raw)
  To: ada.urbaniec+bugzilla; +Cc: Anshu Kumari, passt-dev

On Fri, 20 Nov 2026 23:07:56 +0530
Anshu Kumari <anskuma@redhat.com> wrote:

> Every duplicate ACK unconditionally triggered a fast retransmit,
> even when the ACK sequence hadn't advanced.
> 
> Guard fast retransmit on !conn->retries, and increment retries
> after the first retransmit. This is a quick fix to stop the
> amplification storm. Full compliance with RFC 5681 Section 3.2
> (three duplicate ACKs) is left to a later patch.

Ada, would you have a chance to try this patch out?  Would you like to
be credited with a Reported-by: tag?

-- 
Stefano


^ permalink raw reply	[flat|nested] 4+ messages in thread

* Fwd: [PATCH v3] tcp: Limit fast retransmit to one per loss event
       [not found]   ` <CAA7t2yTP5oGNbyv4B3jt36buwTo2+_61OQoSRkE5m97xxdckzQ@mail.gmail.com>
@ 2026-11-21 11:00     ` Ada Urbaniec
  0 siblings, 0 replies; 4+ messages in thread
From: Ada Urbaniec @ 2026-11-21 11:00 UTC (permalink / raw)
  To: Anshu Kumari, passt-dev

[-- Attachment #1: Type: text/plain, Size: 1918 bytes --]

Hi Stefano, Anshu,

Thanks! Yes, I would be happy to have a Reported-by tag. Happy to help!

I tested v3 as sent on my laptop, using source revision 6aa4f88 and Linux
7.0.0-43. I used an extended version of the reproducer I uploaded to bug
224 (attachment 73), with added periodic ACKs, delayed ACKs, zero-window
cases and retry timing measurements.

With both directions open, dropping the first data segment stops causing a
retransmission storm. If the segment remains missing, v3 retries after
about two seconds.

With a half-closed connection (the client sends FIN, then waits for the
response), v3 also stops the storm. Two cases still fail when the client
sends the same ACK every 10 ms:

* If the missing segment is never accepted, there is no further retry
during the four-second test.
* If the client starts accepting that segment after 100 ms, there is still
no retry, and the transfer does not complete within four seconds.

Without those repeated ACKs, the half-close recovery case completes after
about two seconds.

V3 stops the storm, but repeated ACKs can still prevent a half-closed
connection from retrying. The earlier patch with the timer fix does retry
under the same conditions.

On Fri, Nov 20, 2026 at 9:18 PM Stefano Brivio <sbrivio@redhat.com> wrote:

> On Fri, 20 Nov 2026 23:07:56 +0530
> Anshu Kumari <anskuma@redhat.com> wrote:
>
> > Every duplicate ACK unconditionally triggered a fast retransmit,
> > even when the ACK sequence hadn't advanced.
> >
> > Guard fast retransmit on !conn->retries, and increment retries
> > after the first retransmit. This is a quick fix to stop the
> > amplification storm. Full compliance with RFC 5681 Section 3.2
> > (three duplicate ACKs) is left to a later patch.
>
> Ada, would you have a chance to try this patch out?  Would you like to
> be credited with a Reported-by: tag?
>
> --
> Stefano
>
>

[-- Attachment #2: Type: text/html, Size: 2451 bytes --]

^ permalink raw reply	[flat|nested] 4+ messages in thread

end of thread, other threads:[~2026-11-21 11:00 UTC | newest]

Thread overview: 4+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2026-11-20 17:37 [PATCH v3] tcp: Limit fast retransmit to one per loss event Anshu Kumari
2026-11-20 19:48 ` Jon Maloy
2026-11-20 20:18 ` Stefano Brivio
     [not found]   ` <CAA7t2yTP5oGNbyv4B3jt36buwTo2+_61OQoSRkE5m97xxdckzQ@mail.gmail.com>
2026-11-21 11:00     ` Fwd: " Ada Urbaniec

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).