[<prev] [next>] [thread-next>] [day] [month] [year] [list]
Message-ID: <20240326133412.47cf6d99@kernel.org>
Date: Tue, 26 Mar 2024 13:34:12 -0700
From: Jakub Kicinski <kuba@...nel.org>
To: Eric Dumazet <edumazet@...gle.com>, Paolo Abeni <pabeni@...hat.com>,
ncardwell@...gle.com
Cc: netdev@...r.kernel.org
Subject: ICMP_PARAMETERPROB and ICMP_TIME_EXCEEDED during connect
Hi!
I got a report from a user surprised/displeased that ICMP_TIME_EXCEEDED
breaks connect(), while TCP RFCs say it shouldn't. Even pointing a
finger at Linux, RFC5461:
A number of TCP implementations have modified their reaction to all
ICMP soft errors and treat them as hard errors when they are received
for connections in the SYN-SENT or SYN-RECEIVED states. For example,
this workaround has been implemented in the Linux kernel since
version 2.0.0 (released in 1996) [Linux]. However, it should be
noted that this change violates section 4.2.3.9 of [RFC1122], which
states that these ICMP error messages indicate soft error conditions
and that, therefore, TCP MUST NOT abort the corresponding connection.
Is there any reason we continue with this behavior or is it just that
nobody ever sent a patch?
Somewhat related in tcp_v4_err() we do:
switch (sk->sk_state) {
case TCP_SYN_SENT:
case TCP_SYN_RECV:
[...]
if (!sock_owned_by_user(sk)) {
WRITE_ONCE(sk->sk_err, err);
sk_error_report(sk);
tcp_done(sk);
} else {
WRITE_ONCE(sk->sk_err_soft, err);
}
goto out;
}
So the error is soft if socket is locked, and I can't find anything
in backlog processing which would pay attention. So it seems that
under certain conditions we already ignore it.
Can we ignore it always, or perhaps conditionally based on IP_RECVERR?
Powered by blists - more mailing lists