kernel/linux.git/net/ipv4/tcp_timer.c, branch v6.1.168

tcp: fix race in tcp_write_err()

2024-08-03T06:49:04+00:00

[ Upstream commit 853c3bd7b7917670224c9fe5245bd045cac411dd ] I noticed flakes in a packetdrill test, expecting an epoll_wait() to return EPOLLERR | EPOLLHUP on a failed connect() attempt, after multiple SYN retransmits. It sometimes return EPOLLERR only. The issue is that tcp_write_err(): 1) writes an error in sk->sk_err, 2) calls sk_error_report(), 3) then calls tcp_done(). tcp_done() is writing SHUTDOWN_MASK into sk->sk_shutdown, among other things. Problem is that the awaken user thread (from 2) sk_error_report()) might call tcp_poll() before tcp_done() has written sk->sk_shutdown. tcp_poll() only sees a non zero sk->sk_err and returns EPOLLERR. This patch fixes the issue by making sure to call sk_error_report() after tcp_done(). tcp_write_err() also lacks an smp_wmb(). We can reuse tcp_done_with_error() to factor out the details, as Neal suggested. Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2") Signed-off-by: Eric Dumazet Acked-by: Neal Cardwell Link: https://lore.kernel.org/r/20240528125253.1966136-3-edumazet@google.com Signed-off-by: Jakub Kicinski Signed-off-by: Sasha Levin

tcp: annotate lockless access to sk->sk_err

2024-08-03T06:49:04+00:00

[ Upstream commit e13ec3da05d130f0d10da8e1fbe1be26dcdb0e27 ] tcp_poll() reads sk->sk_err without socket lock held/owned. We should used READ_ONCE() here, and update writers to use WRITE_ONCE(). Signed-off-by: Eric Dumazet Signed-off-by: David S. Miller Stable-dep-of: 853c3bd7b791 ("tcp: fix race in tcp_write_err()") Signed-off-by: Sasha Levin

tcp: annotate lockless accesses to sk->sk_err_soft

2024-08-03T06:49:04+00:00

[ Upstream commit cee1af825d65b8122627fc2efbc36c1bd51ee103 ] This field can be read/written without lock synchronization. Signed-off-by: Eric Dumazet Signed-off-by: David S. Miller Stable-dep-of: 853c3bd7b791 ("tcp: fix race in tcp_write_err()") Signed-off-by: Sasha Levin

tcp: avoid too many retransmit packets

2024-07-18T11:18:37+00:00

commit 97a9063518f198ec0adb2ecb89789de342bb8283 upstream. 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") Signed-off-by: Eric Dumazet Cc: Neal Cardwell Reviewed-by: Jason Xing Reviewed-by: Jon Maxwell Reviewed-by: Kuniyuki Iwashima Link: https://patch.msgid.link/20240710001402.2758273-1-edumazet@google.com Signed-off-by: Jakub Kicinski Signed-off-by: Greg Kroah-Hartman

tcp: use signed arithmetic in tcp_rtx_probe0_timed_out()

2024-07-18T11:18:37+00:00

commit 36534d3c54537bf098224a32dc31397793d4594d upstream. Due to timer wheel implementation, a timer will usually fire after its schedule. For instance, for HZ=1000, a timeout between 512ms and 4s has a granularity of 64ms. For this range of values, the extra delay could be up to 63ms. For TCP, this means that tp->rcv_tstamp may be after inet_csk(sk)->icsk_timeout whenever the timer interrupt finally triggers, if one packet came during the extra delay. We need to make sure tcp_rtx_probe0_timed_out() handles this case. Fixes: e89688e3e978 ("net: tcp: fix unexcepted socket die when snd_wnd is 0") Signed-off-by: Eric Dumazet Cc: Menglong Dong Acked-by: Neal Cardwell Reviewed-by: Jason Xing Link: https://lore.kernel.org/r/20240607125652.1472540-1-edumazet@google.com Signed-off-by: Jakub Kicinski Signed-off-by: Greg Kroah-Hartman

tcp: fix incorrect undo caused by DSACK of TLP retransmit

2024-07-18T11:18:32+00:00

[ Upstream commit 0ec986ed7bab6801faed1440e8839dcc710331ff ] Loss recovery undo_retrans bookkeeping had a long-standing bug where a DSACK from a spurious TLP retransmit packet could cause an erroneous undo of a fast recovery or RTO recovery that repaired a single really-lost packet (in a sequence range outside that of the TLP retransmit). Basically, because the loss recovery state machine didn't account for the fact that it sent a TLP retransmit, the DSACK for the TLP retransmit could erroneously be implicitly be interpreted as corresponding to the normal fast recovery or RTO recovery retransmit that plugged a real hole, thus resulting in an improper undo. For example, consider the following buggy scenario where there is a real packet loss but the congestion control response is improperly undone because of this bug: + send packets P1, P2, P3, P4 + P1 is really lost + send TLP retransmit of P4 + receive SACK for original P2, P3, P4 + enter fast recovery, fast-retransmit P1, increment undo_retrans to 1 + receive DSACK for TLP P4, decrement undo_retrans to 0, undo (bug!) + receive cumulative ACK for P1-P4 (fast retransmit plugged real hole) The fix: when we initialize undo machinery in tcp_init_undo(), if there is a TLP retransmit in flight, then increment tp->undo_retrans so that we make sure that we receive a DSACK corresponding to the TLP retransmit, as well as DSACKs for all later normal retransmits, before triggering a loss recovery undo. Note that we also have to move the line that clears tp->tlp_high_seq for RTO recovery, so that upon RTO we remember the tp->tlp_high_seq value until tcp_init_undo() and clear it only afterward. Also note that the bug dates back to the original 2013 TLP implementation, commit 6ba8a3b19e76 ("tcp: Tail loss probe (TLP)"). However, this patch will only compile and work correctly with kernels that have tp->tlp_retrans, which was added only in v5.8 in 2020 in commit 76be93fc0702 ("tcp: allow at most one TLP probe per flight"). So we associate this fix with that later commit. Fixes: 76be93fc0702 ("tcp: allow at most one TLP probe per flight") Signed-off-by: Neal Cardwell Reviewed-by: Eric Dumazet Cc: Yuchung Cheng Cc: Kevin Yang Link: https://patch.msgid.link/20240703171246.1739561-1-ncardwell.sw@gmail.com Signed-off-by: Jakub Kicinski Signed-off-by: Sasha Levin

net: tcp: fix unexcepted socket die when snd_wnd is 0

2023-09-13T07:42:32+00:00

[ Upstream commit e89688e3e97868451a5d05b38a9d2633d6785cd4 ] In tcp_retransmit_timer(), a window shrunk connection will be regarded as timeout if 'tcp_jiffies32 - tp->rcv_tstamp > TCP_RTO_MAX'. This is not right all the time. The retransmits will become zero-window probes in tcp_retransmit_timer() if the 'snd_wnd==0'. Therefore, the icsk->icsk_rto will come up to TCP_RTO_MAX sooner or later. However, the timer can be delayed and be triggered after 122877ms, not TCP_RTO_MAX, as I tested. Therefore, 'tcp_jiffies32 - tp->rcv_tstamp > TCP_RTO_MAX' is always true once the RTO come up to TCP_RTO_MAX, and the socket will die. Fix this by replacing the 'tcp_jiffies32' with '(u32)icsk->icsk_timeout', which is exact the timestamp of the timeout. However, "tp->rcv_tstamp" can restart from idle, then tp->rcv_tstamp could already be a long time (minutes or hours) in the past even on the first RTO. So we double check the timeout with the duration of the retransmission. Meanwhile, making "2 * TCP_RTO_MAX" as the timeout to avoid the socket dying too soon. Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2") Link: https://lore.kernel.org/netdev/CADxym3YyMiO+zMD4zj03YPM3FBi-1LHi6gSD2XT8pyAMM096pg@mail.gmail.com/ Signed-off-by: Menglong Dong Reviewed-by: Eric Dumazet Signed-off-by: David S. Miller Signed-off-by: Sasha Levin

net: fix the RTO timer retransmitting skb every 1ms if linear option is enabled

2023-08-23T15:52:42+00:00

commit e4dd0d3a2f64b8bd8029ec70f52bdbebd0644408 upstream. In the real workload, I encountered an issue which could cause the RTO timer to retransmit the skb per 1ms with linear option enabled. The amount of lost-retransmitted skbs can go up to 1000+ instantly. The root cause is that if the icsk_rto happens to be zero in the 6th round (which is the TCP_THIN_LINEAR_RETRIES value), then it will always be zero due to the changed calculation method in tcp_retransmit_timer() as follows: icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX); Above line could be converted to icsk->icsk_rto = min(0 << 1, TCP_RTO_MAX) = 0 Therefore, the timer expires so quickly without any doubt. I read through the RFC 6298 and found that the RTO value can be rounded up to a certain value, in Linux, say TCP_RTO_MIN as default, which is regarded as the lower bound in this patch as suggested by Eric. Fixes: 36e31b0af587 ("net: TCP thin linear timeouts") Suggested-by: Eric Dumazet Signed-off-by: Jason Xing Reviewed-by: Eric Dumazet Signed-off-by: David S. Miller Signed-off-by: Greg Kroah-Hartman

tcp: fix mishandling when the sack compression is deferred.

2023-06-09T08:34:05+00:00

[ Upstream commit 30c6f0bf9579debce27e45fac34fdc97e46acacc ] In this patch, we mainly try to handle sending a compressed ack correctly if it's deferred. Here are more details in the old logic: When sack compression is triggered in the tcp_compressed_ack_kick(), if the sock is owned by user, it will set TCP_DELACK_TIMER_DEFERRED and then defer to the release cb phrase. Later once user releases the sock, tcp_delack_timer_handler() should send a ack as expected, which, however, cannot happen due to lack of ICSK_ACK_TIMER flag. Therefore, the receiver would not sent an ack until the sender's retransmission timeout. It definitely increases unnecessary latency. Fixes: 5d9f4262b7ea ("tcp: add SACK compression") Suggested-by: Eric Dumazet Signed-off-by: fuyuanli Signed-off-by: Jason Xing Link: https://lore.kernel.org/netdev/20230529113804.GA20300@didi-ThinkCentre-M920t-N000/ Reviewed-by: Eric Dumazet Link: https://lore.kernel.org/r/20230531080150.GA20424@didi-ThinkCentre-M920t-N000 Signed-off-by: Paolo Abeni Signed-off-by: Sasha Levin

tcp: Make SYN ACK RTO tunable by BPF programs with TFO

2022-08-17T09:19:22+00:00

Instead of the hardcoded TCP_TIMEOUT_INIT, this diff calls tcp_timeout_init to initiate req->timeout like the non TFO SYN ACK case. Tested using the following packetdrill script, on a host with a BPF program that sets the initial connect timeout to 10ms. `../../common/defaults.sh` // Initialize connection 0 socket(..., SOCK_STREAM, IPPROTO_TCP) = 3 +0 setsockopt(3, SOL_TCP, TCP_FASTOPEN, [1], 4) = 0 +0 bind(3, ..., ...) = 0 +0 listen(3, 1) = 0 +0 < S 0:0(0) win 32792 +0 > S. 0:0(0) ack 1 +.01 > S. 0:0(0) ack 1 +.02 > S. 0:0(0) ack 1 +.04 > S. 0:0(0) ack 1 +.01 < . 1:1(0) ack 1 win 32792 +0 accept(3, ..., ...) = 4 Signed-off-by: Jie Meng Acked-by: Neal Cardwell Acked-by: Yuchung Cheng Acked-by: Martin KaFai Lau Signed-off-by: David S. Miller