tcp: avoid too many retransmit packets
Subsystems (3)
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.
--- net/ipv4/tcp_timer.c | 15 +++++++++++++-- 1 file changed, 13 insertions(+), 2 deletions(-)
net/ipv4/tcp_timer.c +13 -2
@@ -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
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 afterTCP_USER_TIMEOUThas 'expired'.The fix is to make sure
tcp_rtx_probe0_timed_out()takesicsk->icsk_user_timeoutinto 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 whensnd_wndis 0"), the issue would last 2 minutes instead of 4.Fixes b701a99e431d ("tcp: Add tcp_clamp_rto_to_user_timeout() helper to improve accuracy")Signed-off-by Eric Dumazet <[email protected]>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@@ -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
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 afterTCP_USER_TIMEOUThas 'expired'.The fix is to make sure
tcp_rtx_probe0_timed_out()takesicsk->icsk_user_timeoutinto 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 whensnd_wndis 0"), the issue would last 2 minutes instead of 4.Fixes b701a99e431d ("tcp: Add tcp_clamp_rto_to_user_timeout() helper to improve accuracy")Signed-off-by Eric Dumazet <[email protected]>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@@ -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
PATCH 1/1 tcp: avoid too many retransmit packets
In reply to Eric DumazetFrom: 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 afterTCP_USER_TIMEOUThas 'expired'.The fix is to make sure
tcp_rtx_probe0_timed_out()takesicsk->icsk_user_timeoutinto 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 whensnd_wndis 0"), the issue would last 2 minutes instead of 4.Fixes b701a99e431d ("tcp: Add tcp_clamp_rto_to_user_timeout() helper to improve accuracy")Signed-off-by Eric Dumazet <[email protected]>
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 afterTCP_USER_TIMEOUThas 'expired'.The fix is to make sure
tcp_rtx_probe0_timed_out()takesicsk->icsk_user_timeoutinto 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