Skip to contentSkip to conversation

tcp: avoid too many retransmit packets

PATCH Applied Eric Dumazet 5 participants 5 messages Jul 10, 2024 – Jul 11, 2024

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 tcp: avoid too many retransmit packets

If a TCP socket is using TCP_USER_TIMEOUT, and the other peer retracted its window to zero, tcp_retransmit_timer() can retransmit a packet every two jiffies (2 ms for HZ=1000), for about 4 minutes after TCP_USER_TIMEOUT has 'expired'.

The fix is to make sure tcp_rtx_probe0_timed_out() takes icsk->icsk_user_timeout into account.

Before blamed commit, the socket would not timeout after icsk->icsk_user_timeout, but would use standard exponential backoff for the retransmits.

Also worth noting that before commit e89688e3e978 ("net: tcp:
fix unexcepted socket die when snd_wnd is 0"), the issue would last 2 minutes instead of 4.

Fixes b701a99e431d ("tcp: Add tcp_clamp_rto_to_user_timeout() helper to improve accuracy")
Cc Jon Maxwell <[email protected]>
Cc Neal Cardwell <[email protected]>
---
 net/ipv4/tcp_timer.c | 15 +++++++++++++--
 1 file changed, 13 insertions(+), 2 deletions(-)
net/ipv4/tcp_timer.c +13 -2
index db9d826560e57caf8274d1b7253c7af4dd7821a0..892c86657fbc243ce53a939157b77f1fe0410097 100644
@@ -483,15 +483,26 @@ static bool tcp_rtx_probe0_timed_out(const struct sock *sk,
 				     const struct sk_buff *skb,
 				     u32 rtx_delta)
 {
+	const struct inet_connection_sock *icsk = inet_csk(sk);
+	u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
 	const struct tcp_sock *tp = tcp_sk(sk);
-	const int timeout = TCP_RTO_MAX * 2;
+	int timeout = TCP_RTO_MAX * 2;
 	s32 rcv_delta;
 
+	if (user_timeout) {
+		/* If user application specified a TCP_USER_TIMEOUT,
+		 * it does not want win 0 packets to 'reset the timer'
+		 * while retransmits are not making progress.
+		 */
+		if (rtx_delta > user_timeout)
+			return true;
+		timeout = min_t(u32, timeout, msecs_to_jiffies(user_timeout));
+	}
 	/* Note: timer interrupt might have been delayed by at least one jiffy,
 	 * and tp->rcv_tstamp might very well have been written recently.
 	 * rcv_delta can thus be negative.
 	 */
-	rcv_delta = inet_csk(sk)->icsk_timeout - tp->rcv_tstamp;
+	rcv_delta = icsk->icsk_timeout - tp->rcv_tstamp;
 	if (rcv_delta <= timeout)
 		return false;
 
Signature · 1 line
2.45.2.993.g49e7a77208-goog
Reviewed-by on v1 1/1
#1 lore raw
Quoted 13 lines · Eric Dumazet

If a TCP socket is using TCP_USER_TIMEOUT, and the other peer retracted its window to zero, tcp_retransmit_timer() can retransmit a packet every two jiffies (2 ms for HZ=1000), for about 4 minutes after TCP_USER_TIMEOUT has 'expired'.

The fix is to make sure tcp_rtx_probe0_timed_out() takes icsk->icsk_user_timeout into account.

Before blamed commit, the socket would not timeout after icsk->icsk_user_timeout, but would use standard exponential backoff for the retransmits.

Quoted 8 lines

Also worth noting that before commit e89688e3e978 ("net: tcp:
fix unexcepted socket die when snd_wnd is 0"), the issue would last 2 minutes instead of 4.

Fixes b701a99e431d ("tcp: Add tcp_clamp_rto_to_user_timeout() helper to improve accuracy")
Cc Jon Maxwell <[email protected]>
Cc Neal Cardwell <[email protected]>

Interesting case. Thanks!

Quoted 41 lines
---
 net/ipv4/tcp_timer.c | 15 +++++++++++++--
 1 file changed, 13 insertions(+), 2 deletions(-)
net/ipv4/tcp_timer.c +13 -3
index db9d826560e57caf8274d1b7253c7af4dd7821a0..892c86657fbc243ce53a939157b77f1fe0410097 100644
@@ -483,15 +483,26 @@ static bool tcp_rtx_probe0_timed_out(const struct sock *sk,
                                     const struct sk_buff *skb,
                                     u32 rtx_delta)
 {
+       const struct inet_connection_sock *icsk = inet_csk(sk);
+       u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
        const struct tcp_sock *tp = tcp_sk(sk);
-       const int timeout = TCP_RTO_MAX * 2;
+       int timeout = TCP_RTO_MAX * 2;
        s32 rcv_delta;
 
+       if (user_timeout) {
+               /* If user application specified a TCP_USER_TIMEOUT,
+                * it does not want win 0 packets to 'reset the timer'
+                * while retransmits are not making progress.
+                */
+               if (rtx_delta > user_timeout)
+                       return true;
+               timeout = min_t(u32, timeout, msecs_to_jiffies(user_timeout));
+       }
        /* Note: timer interrupt might have been delayed by at least one jiffy,
         * and tp->rcv_tstamp might very well have been written recently.
         * rcv_delta can thus be negative.
         */
-       rcv_delta = inet_csk(sk)->icsk_timeout - tp->rcv_tstamp;
+       rcv_delta = icsk->icsk_timeout - tp->rcv_tstamp;
        if (rcv_delta <= timeout)
                return false;
 
--

2.45.2.993.g49e7a77208-goog

Reviewed-by on v1 1/1
#2 lore raw
In reply to Eric Dumazet
Quoted 13 lines · Eric Dumazet

If a TCP socket is using TCP_USER_TIMEOUT, and the other peer retracted its window to zero, tcp_retransmit_timer() can retransmit a packet every two jiffies (2 ms for HZ=1000), for about 4 minutes after TCP_USER_TIMEOUT has 'expired'.

The fix is to make sure tcp_rtx_probe0_timed_out() takes icsk->icsk_user_timeout into account.

Before blamed commit, the socket would not timeout after icsk->icsk_user_timeout, but would use standard exponential backoff for the retransmits.

Quoted 8 lines

Also worth noting that before commit e89688e3e978 ("net: tcp:
fix unexcepted socket die when snd_wnd is 0"), the issue would last 2 minutes instead of 4.

Fixes b701a99e431d ("tcp: Add tcp_clamp_rto_to_user_timeout() helper to improve accuracy")
Cc Jon Maxwell <[email protected]>
Cc Neal Cardwell <[email protected]>

Nice catch Eric. Thanks for fixing LGTM.

Quoted 40 lines
---
 net/ipv4/tcp_timer.c | 15 +++++++++++++--
 1 file changed, 13 insertions(+), 2 deletions(-)
net/ipv4/tcp_timer.c +13 -3
index db9d826560e57caf8274d1b7253c7af4dd7821a0..892c86657fbc243ce53a939157b77f1fe0410097 100644
@@ -483,15 +483,26 @@ static bool tcp_rtx_probe0_timed_out(const struct sock *sk,
                                     const struct sk_buff *skb,
                                     u32 rtx_delta)
 {
+       const struct inet_connection_sock *icsk = inet_csk(sk);
+       u32 user_timeout = READ_ONCE(icsk->icsk_user_timeout);
        const struct tcp_sock *tp = tcp_sk(sk);
-       const int timeout = TCP_RTO_MAX * 2;
+       int timeout = TCP_RTO_MAX * 2;
        s32 rcv_delta;
 
+       if (user_timeout) {
+               /* If user application specified a TCP_USER_TIMEOUT,
+                * it does not want win 0 packets to 'reset the timer'
+                * while retransmits are not making progress.
+                */
+               if (rtx_delta > user_timeout)
+                       return true;
+               timeout = min_t(u32, timeout, msecs_to_jiffies(user_timeout));
+       }
        /* Note: timer interrupt might have been delayed by at least one jiffy,
         * and tp->rcv_tstamp might very well have been written recently.
         * rcv_delta can thus be negative.
         */
-       rcv_delta = inet_csk(sk)->icsk_timeout - tp->rcv_tstamp;
+       rcv_delta = icsk->icsk_timeout - tp->rcv_tstamp;
        if (rcv_delta <= timeout)
                return false;
 
--

2.45.2.993.g49e7a77208-goog

#3 lore raw

PATCH 1/1 tcp: avoid too many retransmit packets

In reply to Eric Dumazet

From: Eric Dumazet <[email protected]>
Date: Wed, 10 Jul 2024 00:14:01 +0000

Quoted 12 lines

If a TCP socket is using TCP_USER_TIMEOUT, and the other peer retracted its window to zero, tcp_retransmit_timer() can retransmit a packet every two jiffies (2 ms for HZ=1000), for about 4 minutes after TCP_USER_TIMEOUT has 'expired'.

The fix is to make sure tcp_rtx_probe0_timed_out() takes icsk->icsk_user_timeout into account.

Before blamed commit, the socket would not timeout after icsk->icsk_user_timeout, but would use standard exponential backoff for the retransmits.

Quoted 6 lines

Also worth noting that before commit e89688e3e978 ("net: tcp:
fix unexcepted socket die when snd_wnd is 0"), the issue would last 2 minutes instead of 4.

Fixes b701a99e431d ("tcp: Add tcp_clamp_rto_to_user_timeout() helper to improve accuracy")
#4 lore raw
In reply to Eric Dumazet
This patch was applied to netdev/net.git (main) by Jakub Kicinski <[email protected]>:

Hello:

This patch was applied to netdev/net.git (main)
by Jakub Kicinski <[email protected]>:

Quoted 9 lines · On Wed, 10 Jul 2024 00:14:01 +0000 you wrote:

If a TCP socket is using TCP_USER_TIMEOUT, and the other peer retracted its window to zero, tcp_retransmit_timer() can retransmit a packet every two jiffies (2 ms for HZ=1000), for about 4 minutes after TCP_USER_TIMEOUT has 'expired'.

The fix is to make sure tcp_rtx_probe0_timed_out() takes icsk->icsk_user_timeout into account.

[...]

Here is the summary with links:
- [net] tcp: avoid too many retransmit packets

    https://git.kernel.org/netdev/net/c/97a9063518f1

You are awesome, thank you!

Signature · 2 lines
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html