On Mon, Aug 7, 2023 at 10:20 PM Eric Dumazet [off-list ref] wrote:

On Mon, Aug 7, 2023 at 3:47 PM [off-list ref] wrote:

quoted

From: Menglong Dong <redacted>

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 is not precise enough, as it base on timer wheel.

quoted

Sorry that I am not good at timer, but I know the concept of time-wheel.

Can you remove this line, can we keep focused on the actual patch instead ?

Regardless of timer-wheel, a timer can be delayed under load.

Okay!

quoted

The longer of the timer, the rougher it will be. So the timeout is not
triggered after TCP_RTO_MAX, but 122877ms 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. Meanwhile, using
"max(tp->retrans_stamp, tp->rcv_tstamp)" as the last updated timestamp in
the receiving path, as "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.

Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
Link: https://lore.kernel.org/netdev/CADxym3YyMiO+zMD4zj03YPM3FBi-1LHi6gSD2XT8pyAMM096pg@mail.gmail.com/ (local)
Signed-off-by: Menglong Dong <redacted>
---
v2:
- consider the case of the connection restart from idle, as Neal comment
---
 net/ipv4/tcp_timer.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/net/ipv4/tcp_timer.c b/net/ipv4/tcp_timer.c
index d45c96c7f5a4..e4b2d8706cae 100644
--- a/net/ipv4/tcp_timer.c
+++ b/net/ipv4/tcp_timer.c

@@ -454,6 +454,14 @@ static void tcp_fastopen_synack_timer(struct sock *sk, struct request_sock *req)
                          req->timeout << req->num_timeout, TCP_RTO_MAX);
 }

+static bool tcp_rtx_probe0_timed_out(struct sock *sk)
+{
+       struct tcp_sock *tp = tcp_sk(sk);
+       u32 last_ts;
+
+       last_ts = max(tp->retrans_stamp, tp->rcv_tstamp);

u32     retrans_stamp;  /* Timestamp of the last retransmit,
u32     rcv_tstamp;     /* timestamp of last received ACK (for keepalives) */

Both fields receive tcp_jiffies32 values, which wrap every 2^32 ticks.

So max(A, B) won't work as you expect...

You need to use before(), after() or something like that.

Now I know how the before()/after() works. I thought they
simply compared the arguments.

I'll use brefore/after here instead.

Thanks!
Menglong Dong

https://en.wikipedia.org/wiki/Serial_number_arithmetic#General_Solution

quoted

+       return inet_csk(sk)->icsk_timeout - last_ts > TCP_RTO_MAX;
+}

 /**
  *  tcp_retransmit_timer() - The TCP retransmit timeout handler

@@ -519,7 +527,7 @@ void tcp_retransmit_timer(struct sock *sk)
                                            tp->snd_una, tp->snd_nxt);
                }
 #endif
-               if (tcp_jiffies32 - tp->rcv_tstamp > TCP_RTO_MAX) {
+               if (tcp_rtx_probe0_timed_out(sk)) {
                        tcp_write_err(sk);
                        goto out;
                }
--

2.40.1